Mon, 24 Nov 2014 07:29:03 -0800
8058148: MaxNodeLimit and LiveNodeCountInliningCutoff
Reviewed-by: kvn, roland
1 /*
2 * Copyright (c) 1998, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "opto/block.hpp"
28 #include "opto/c2compiler.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/cfgnode.hpp"
31 #include "opto/machnode.hpp"
32 #include "opto/runtime.hpp"
33 #ifdef TARGET_ARCH_MODEL_x86_32
34 # include "adfiles/ad_x86_32.hpp"
35 #endif
36 #ifdef TARGET_ARCH_MODEL_x86_64
37 # include "adfiles/ad_x86_64.hpp"
38 #endif
39 #ifdef TARGET_ARCH_MODEL_sparc
40 # include "adfiles/ad_sparc.hpp"
41 #endif
42 #ifdef TARGET_ARCH_MODEL_zero
43 # include "adfiles/ad_zero.hpp"
44 #endif
45 #ifdef TARGET_ARCH_MODEL_arm
46 # include "adfiles/ad_arm.hpp"
47 #endif
48 #ifdef TARGET_ARCH_MODEL_ppc_32
49 # include "adfiles/ad_ppc_32.hpp"
50 #endif
51 #ifdef TARGET_ARCH_MODEL_ppc_64
52 # include "adfiles/ad_ppc_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() > 0) { // 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 if the offset is not too high for implicit exception
258 {
259 intptr_t offset = 0;
260 const TypePtr *adr_type = NULL; // Do not need this return value here
261 const Node* base = mach->get_base_and_disp(offset, adr_type);
262 if (base == NULL || base == NodeSentinel) {
263 // Narrow oop address doesn't have base, only index
264 if( val->bottom_type()->isa_narrowoop() &&
265 MacroAssembler::needs_explicit_null_check(offset) )
266 continue; // Give up if offset is beyond page size
267 // cannot reason about it; is probably not implicit null exception
268 } else {
269 const TypePtr* tptr;
270 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 ||
271 Universe::narrow_klass_shift() == 0)) {
272 // 32-bits narrow oop can be the base of address expressions
273 tptr = base->get_ptr_type();
274 } else {
275 // only regular oops are expected here
276 tptr = base->bottom_type()->is_ptr();
277 }
278 // Give up if offset is not a compile-time constant
279 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
280 continue;
281 offset += tptr->_offset; // correct if base is offseted
282 if( MacroAssembler::needs_explicit_null_check(offset) )
283 continue; // Give up is reference is beyond 4K page size
284 }
285 }
287 // Check ctrl input to see if the null-check dominates the memory op
288 Block *cb = get_block_for_node(mach);
289 cb = cb->_idom; // Always hoist at least 1 block
290 if( !was_store ) { // Stores can be hoisted only one block
291 while( cb->_dom_depth > (block->_dom_depth + 1))
292 cb = cb->_idom; // Hoist loads as far as we want
293 // The non-null-block should dominate the memory op, too. Live
294 // range spilling will insert a spill in the non-null-block if it is
295 // needs to spill the memory op for an implicit null check.
296 if (cb->_dom_depth == (block->_dom_depth + 1)) {
297 if (cb != not_null_block) continue;
298 cb = cb->_idom;
299 }
300 }
301 if( cb != block ) continue;
303 // Found a memory user; see if it can be hoisted to check-block
304 uint vidx = 0; // Capture index of value into memop
305 uint j;
306 for( j = mach->req()-1; j > 0; j-- ) {
307 if( mach->in(j) == val ) {
308 vidx = j;
309 // Ignore DecodeN val which could be hoisted to where needed.
310 if( is_decoden ) continue;
311 }
312 // Block of memory-op input
313 Block *inb = get_block_for_node(mach->in(j));
314 Block *b = block; // Start from nul check
315 while( b != inb && b->_dom_depth > inb->_dom_depth )
316 b = b->_idom; // search upwards for input
317 // See if input dominates null check
318 if( b != inb )
319 break;
320 }
321 if( j > 0 )
322 continue;
323 Block *mb = get_block_for_node(mach);
324 // Hoisting stores requires more checks for the anti-dependence case.
325 // Give up hoisting if we have to move the store past any load.
326 if( was_store ) {
327 Block *b = mb; // Start searching here for a local load
328 // mach use (faulting) trying to hoist
329 // n might be blocker to hoisting
330 while( b != block ) {
331 uint k;
332 for( k = 1; k < b->number_of_nodes(); k++ ) {
333 Node *n = b->get_node(k);
334 if( n->needs_anti_dependence_check() &&
335 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
336 break; // Found anti-dependent load
337 }
338 if( k < b->number_of_nodes() )
339 break; // Found anti-dependent load
340 // Make sure control does not do a merge (would have to check allpaths)
341 if( b->num_preds() != 2 ) break;
342 b = get_block_for_node(b->pred(1)); // Move up to predecessor block
343 }
344 if( b != block ) continue;
345 }
347 // Make sure this memory op is not already being used for a NullCheck
348 Node *e = mb->end();
349 if( e->is_MachNullCheck() && e->in(1) == mach )
350 continue; // Already being used as a NULL check
352 // Found a candidate! Pick one with least dom depth - the highest
353 // in the dom tree should be closest to the null check.
354 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) {
355 best = mach;
356 bidx = vidx;
357 }
358 }
359 // No candidate!
360 if (best == NULL) {
361 return;
362 }
364 // ---- Found an implicit null check
365 extern int implicit_null_checks;
366 implicit_null_checks++;
368 if( is_decoden ) {
369 // Check if we need to hoist decodeHeapOop_not_null first.
370 Block *valb = get_block_for_node(val);
371 if( block != valb && block->_dom_depth < valb->_dom_depth ) {
372 // Hoist it up to the end of the test block.
373 valb->find_remove(val);
374 block->add_inst(val);
375 map_node_to_block(val, block);
376 // DecodeN on x86 may kill flags. Check for flag-killing projections
377 // that also need to be hoisted.
378 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
379 Node* n = val->fast_out(j);
380 if( n->is_MachProj() ) {
381 get_block_for_node(n)->find_remove(n);
382 block->add_inst(n);
383 map_node_to_block(n, block);
384 }
385 }
386 }
387 }
388 // Hoist the memory candidate up to the end of the test block.
389 Block *old_block = get_block_for_node(best);
390 old_block->find_remove(best);
391 block->add_inst(best);
392 map_node_to_block(best, block);
394 // Move the control dependence
395 if (best->in(0) && best->in(0) == old_block->head())
396 best->set_req(0, block->head());
398 // Check for flag-killing projections that also need to be hoisted
399 // Should be DU safe because no edge updates.
400 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
401 Node* n = best->fast_out(j);
402 if( n->is_MachProj() ) {
403 get_block_for_node(n)->find_remove(n);
404 block->add_inst(n);
405 map_node_to_block(n, block);
406 }
407 }
409 // proj==Op_True --> ne test; proj==Op_False --> eq test.
410 // One of two graph shapes got matched:
411 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
412 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
413 // NULL checks are always branch-if-eq. If we see a IfTrue projection
414 // then we are replacing a 'ne' test with a 'eq' NULL check test.
415 // We need to flip the projections to keep the same semantics.
416 if( proj->Opcode() == Op_IfTrue ) {
417 // Swap order of projections in basic block to swap branch targets
418 Node *tmp1 = block->get_node(block->end_idx()+1);
419 Node *tmp2 = block->get_node(block->end_idx()+2);
420 block->map_node(tmp2, block->end_idx()+1);
421 block->map_node(tmp1, block->end_idx()+2);
422 Node *tmp = new (C) Node(C->top()); // Use not NULL input
423 tmp1->replace_by(tmp);
424 tmp2->replace_by(tmp1);
425 tmp->replace_by(tmp2);
426 tmp->destruct();
427 }
429 // Remove the existing null check; use a new implicit null check instead.
430 // Since schedule-local needs precise def-use info, we need to correct
431 // it as well.
432 Node *old_tst = proj->in(0);
433 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
434 block->map_node(nul_chk, block->end_idx());
435 map_node_to_block(nul_chk, block);
436 // Redirect users of old_test to nul_chk
437 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
438 old_tst->last_out(i2)->set_req(0, nul_chk);
439 // Clean-up any dead code
440 for (uint i3 = 0; i3 < old_tst->req(); i3++)
441 old_tst->set_req(i3, NULL);
443 latency_from_uses(nul_chk);
444 latency_from_uses(best);
445 }
448 //------------------------------select-----------------------------------------
449 // Select a nice fellow from the worklist to schedule next. If there is only
450 // one choice, then use it. Projections take top priority for correctness
451 // reasons - if I see a projection, then it is next. There are a number of
452 // other special cases, for instructions that consume condition codes, et al.
453 // These are chosen immediately. Some instructions are required to immediately
454 // precede the last instruction in the block, and these are taken last. Of the
455 // remaining cases (most), choose the instruction with the greatest latency
456 // (that is, the most number of pseudo-cycles required to the end of the
457 // routine). If there is a tie, choose the instruction with the most inputs.
458 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) {
460 // If only a single entry on the stack, use it
461 uint cnt = worklist.size();
462 if (cnt == 1) {
463 Node *n = worklist[0];
464 worklist.map(0,worklist.pop());
465 return n;
466 }
468 uint choice = 0; // Bigger is most important
469 uint latency = 0; // Bigger is scheduled first
470 uint score = 0; // Bigger is better
471 int idx = -1; // Index in worklist
472 int cand_cnt = 0; // Candidate count
474 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
475 // Order in worklist is used to break ties.
476 // See caller for how this is used to delay scheduling
477 // of induction variable increments to after the other
478 // uses of the phi are scheduled.
479 Node *n = worklist[i]; // Get Node on worklist
481 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
482 if( n->is_Proj() || // Projections always win
483 n->Opcode()== Op_Con || // So does constant 'Top'
484 iop == Op_CreateEx || // Create-exception must start block
485 iop == Op_CheckCastPP
486 ) {
487 worklist.map(i,worklist.pop());
488 return n;
489 }
491 // Final call in a block must be adjacent to 'catch'
492 Node *e = block->end();
493 if( e->is_Catch() && e->in(0)->in(0) == n )
494 continue;
496 // Memory op for an implicit null check has to be at the end of the block
497 if( e->is_MachNullCheck() && e->in(1) == n )
498 continue;
500 // Schedule IV increment last.
501 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd &&
502 e->in(1)->in(1) == n && n->is_iteratively_computed())
503 continue;
505 uint n_choice = 2;
507 // See if this instruction is consumed by a branch. If so, then (as the
508 // branch is the last instruction in the basic block) force it to the
509 // end of the basic block
510 if ( must_clone[iop] ) {
511 // See if any use is a branch
512 bool found_machif = false;
514 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
515 Node* use = n->fast_out(j);
517 // The use is a conditional branch, make them adjacent
518 if (use->is_MachIf() && get_block_for_node(use) == block) {
519 found_machif = true;
520 break;
521 }
523 // More than this instruction pending for successor to be ready,
524 // don't choose this if other opportunities are ready
525 if (ready_cnt.at(use->_idx) > 1)
526 n_choice = 1;
527 }
529 // loop terminated, prefer not to use this instruction
530 if (found_machif)
531 continue;
532 }
534 // See if this has a predecessor that is "must_clone", i.e. sets the
535 // condition code. If so, choose this first
536 for (uint j = 0; j < n->req() ; j++) {
537 Node *inn = n->in(j);
538 if (inn) {
539 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
540 n_choice = 3;
541 break;
542 }
543 }
544 }
546 // MachTemps should be scheduled last so they are near their uses
547 if (n->is_MachTemp()) {
548 n_choice = 1;
549 }
551 uint n_latency = get_latency_for_node(n);
552 uint n_score = n->req(); // Many inputs get high score to break ties
554 // Keep best latency found
555 cand_cnt++;
556 if (choice < n_choice ||
557 (choice == n_choice &&
558 ((StressLCM && Compile::randomized_select(cand_cnt)) ||
559 (!StressLCM &&
560 (latency < n_latency ||
561 (latency == n_latency &&
562 (score < n_score))))))) {
563 choice = n_choice;
564 latency = n_latency;
565 score = n_score;
566 idx = i; // Also keep index in worklist
567 }
568 } // End of for all ready nodes in worklist
570 assert(idx >= 0, "index should be set");
571 Node *n = worklist[(uint)idx]; // Get the winner
573 worklist.map((uint)idx, worklist.pop()); // Compress worklist
574 return n;
575 }
578 //------------------------------set_next_call----------------------------------
579 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) {
580 if( next_call.test_set(n->_idx) ) return;
581 for( uint i=0; i<n->len(); i++ ) {
582 Node *m = n->in(i);
583 if( !m ) continue; // must see all nodes in block that precede call
584 if (get_block_for_node(m) == block) {
585 set_next_call(block, m, next_call);
586 }
587 }
588 }
590 //------------------------------needed_for_next_call---------------------------
591 // Set the flag 'next_call' for each Node that is needed for the next call to
592 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
593 // next subroutine call get priority - basically it moves things NOT needed
594 // for the next call till after the call. This prevents me from trying to
595 // carry lots of stuff live across a call.
596 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) {
597 // Find the next control-defining Node in this block
598 Node* call = NULL;
599 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
600 Node* m = this_call->fast_out(i);
601 if (get_block_for_node(m) == block && // Local-block user
602 m != this_call && // Not self-start node
603 m->is_MachCall()) {
604 call = m;
605 break;
606 }
607 }
608 if (call == NULL) return; // No next call (e.g., block end is near)
609 // Set next-call for all inputs to this call
610 set_next_call(block, call, next_call);
611 }
613 //------------------------------add_call_kills-------------------------------------
614 // helper function that adds caller save registers to MachProjNode
615 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
616 // Fill in the kill mask for the call
617 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
618 if( !regs.Member(r) ) { // Not already defined by the call
619 // Save-on-call register?
620 if ((save_policy[r] == 'C') ||
621 (save_policy[r] == 'A') ||
622 ((save_policy[r] == 'E') && exclude_soe)) {
623 proj->_rout.Insert(r);
624 }
625 }
626 }
627 }
630 //------------------------------sched_call-------------------------------------
631 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) {
632 RegMask regs;
634 // Schedule all the users of the call right now. All the users are
635 // projection Nodes, so they must be scheduled next to the call.
636 // Collect all the defined registers.
637 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
638 Node* n = mcall->fast_out(i);
639 assert( n->is_MachProj(), "" );
640 int n_cnt = ready_cnt.at(n->_idx)-1;
641 ready_cnt.at_put(n->_idx, n_cnt);
642 assert( n_cnt == 0, "" );
643 // Schedule next to call
644 block->map_node(n, node_cnt++);
645 // Collect defined registers
646 regs.OR(n->out_RegMask());
647 // Check for scheduling the next control-definer
648 if( n->bottom_type() == Type::CONTROL )
649 // Warm up next pile of heuristic bits
650 needed_for_next_call(block, n, next_call);
652 // Children of projections are now all ready
653 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
654 Node* m = n->fast_out(j); // Get user
655 if(get_block_for_node(m) != block) {
656 continue;
657 }
658 if( m->is_Phi() ) continue;
659 int m_cnt = ready_cnt.at(m->_idx)-1;
660 ready_cnt.at_put(m->_idx, m_cnt);
661 if( m_cnt == 0 )
662 worklist.push(m);
663 }
665 }
667 // Act as if the call defines the Frame Pointer.
668 // Certainly the FP is alive and well after the call.
669 regs.Insert(_matcher.c_frame_pointer());
671 // Set all registers killed and not already defined by the call.
672 uint r_cnt = mcall->tf()->range()->cnt();
673 int op = mcall->ideal_Opcode();
674 MachProjNode *proj = new (C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
675 map_node_to_block(proj, block);
676 block->insert_node(proj, node_cnt++);
678 // Select the right register save policy.
679 const char * save_policy;
680 switch (op) {
681 case Op_CallRuntime:
682 case Op_CallLeaf:
683 case Op_CallLeafNoFP:
684 // Calling C code so use C calling convention
685 save_policy = _matcher._c_reg_save_policy;
686 break;
688 case Op_CallStaticJava:
689 case Op_CallDynamicJava:
690 // Calling Java code so use Java calling convention
691 save_policy = _matcher._register_save_policy;
692 break;
694 default:
695 ShouldNotReachHere();
696 }
698 // When using CallRuntime mark SOE registers as killed by the call
699 // so values that could show up in the RegisterMap aren't live in a
700 // callee saved register since the register wouldn't know where to
701 // find them. CallLeaf and CallLeafNoFP are ok because they can't
702 // have debug info on them. Strictly speaking this only needs to be
703 // done for oops since idealreg2debugmask takes care of debug info
704 // references but there no way to handle oops differently than other
705 // pointers as far as the kill mask goes.
706 bool exclude_soe = op == Op_CallRuntime;
708 // If the call is a MethodHandle invoke, we need to exclude the
709 // register which is used to save the SP value over MH invokes from
710 // the mask. Otherwise this register could be used for
711 // deoptimization information.
712 if (op == Op_CallStaticJava) {
713 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
714 if (mcallstaticjava->_method_handle_invoke)
715 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
716 }
718 add_call_kills(proj, regs, save_policy, exclude_soe);
720 return node_cnt;
721 }
724 //------------------------------schedule_local---------------------------------
725 // Topological sort within a block. Someday become a real scheduler.
726 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) {
727 // Already "sorted" are the block start Node (as the first entry), and
728 // the block-ending Node and any trailing control projections. We leave
729 // these alone. PhiNodes and ParmNodes are made to follow the block start
730 // Node. Everything else gets topo-sorted.
732 #ifndef PRODUCT
733 if (trace_opto_pipelining()) {
734 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order);
735 for (uint i = 0;i < block->number_of_nodes(); i++) {
736 tty->print("# ");
737 block->get_node(i)->fast_dump();
738 }
739 tty->print_cr("#");
740 }
741 #endif
743 // RootNode is already sorted
744 if (block->number_of_nodes() == 1) {
745 return true;
746 }
748 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
749 uint node_cnt = block->end_idx();
750 uint phi_cnt = 1;
751 uint i;
752 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
753 Node *n = block->get_node(i);
754 if( n->is_Phi() || // Found a PhiNode or ParmNode
755 (n->is_Proj() && n->in(0) == block->head()) ) {
756 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
757 block->map_node(block->get_node(phi_cnt), i);
758 block->map_node(n, phi_cnt++); // swap Phi/Parm up front
759 } else { // All others
760 // Count block-local inputs to 'n'
761 uint cnt = n->len(); // Input count
762 uint local = 0;
763 for( uint j=0; j<cnt; j++ ) {
764 Node *m = n->in(j);
765 if( m && get_block_for_node(m) == block && !m->is_top() )
766 local++; // One more block-local input
767 }
768 ready_cnt.at_put(n->_idx, local); // Count em up
770 #ifdef ASSERT
771 if( UseConcMarkSweepGC || UseG1GC ) {
772 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
773 // Check the precedence edges
774 for (uint prec = n->req(); prec < n->len(); prec++) {
775 Node* oop_store = n->in(prec);
776 if (oop_store != NULL) {
777 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark");
778 }
779 }
780 }
781 }
782 #endif
784 // A few node types require changing a required edge to a precedence edge
785 // before allocation.
786 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
787 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
788 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
789 // MemBarAcquire could be created without Precedent edge.
790 // del_req() replaces the specified edge with the last input edge
791 // and then removes the last edge. If the specified edge > number of
792 // edges the last edge will be moved outside of the input edges array
793 // and the edge will be lost. This is why this code should be
794 // executed only when Precedent (== TypeFunc::Parms) edge is present.
795 Node *x = n->in(TypeFunc::Parms);
796 n->del_req(TypeFunc::Parms);
797 n->add_prec(x);
798 }
799 }
800 }
801 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count
802 ready_cnt.at_put(block->get_node(i2)->_idx, 0);
804 // All the prescheduled guys do not hold back internal nodes
805 uint i3;
806 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
807 Node *n = block->get_node(i3); // Get pre-scheduled
808 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
809 Node* m = n->fast_out(j);
810 if (get_block_for_node(m) == block) { // Local-block user
811 int m_cnt = ready_cnt.at(m->_idx)-1;
812 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
813 }
814 }
815 }
817 Node_List delay;
818 // Make a worklist
819 Node_List worklist;
820 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
821 Node *m = block->get_node(i4);
822 if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
823 if (m->is_iteratively_computed()) {
824 // Push induction variable increments last to allow other uses
825 // of the phi to be scheduled first. The select() method breaks
826 // ties in scheduling by worklist order.
827 delay.push(m);
828 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
829 // Force the CreateEx to the top of the list so it's processed
830 // first and ends up at the start of the block.
831 worklist.insert(0, m);
832 } else {
833 worklist.push(m); // Then on to worklist!
834 }
835 }
836 }
837 while (delay.size()) {
838 Node* d = delay.pop();
839 worklist.push(d);
840 }
842 // Warm up the 'next_call' heuristic bits
843 needed_for_next_call(block, block->head(), next_call);
845 #ifndef PRODUCT
846 if (trace_opto_pipelining()) {
847 for (uint j=0; j< block->number_of_nodes(); j++) {
848 Node *n = block->get_node(j);
849 int idx = n->_idx;
850 tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
851 tty->print("latency:%3d ", get_latency_for_node(n));
852 tty->print("%4d: %s\n", idx, n->Name());
853 }
854 }
855 #endif
857 uint max_idx = (uint)ready_cnt.length();
858 // Pull from worklist and schedule
859 while( worklist.size() ) { // Worklist is not ready
861 #ifndef PRODUCT
862 if (trace_opto_pipelining()) {
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 #endif
872 // Select and pop a ready guy from worklist
873 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt);
874 block->map_node(n, phi_cnt++); // Schedule him next
876 #ifndef PRODUCT
877 if (trace_opto_pipelining()) {
878 tty->print("# select %d: %s", n->_idx, n->Name());
879 tty->print(", latency:%d", get_latency_for_node(n));
880 n->dump();
881 if (Verbose) {
882 tty->print("# ready list:");
883 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
884 Node *n = worklist[i]; // Get Node on worklist
885 tty->print(" %d", n->_idx);
886 }
887 tty->cr();
888 }
889 }
891 #endif
892 if( n->is_MachCall() ) {
893 MachCallNode *mcall = n->as_MachCall();
894 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call);
895 continue;
896 }
898 if (n->is_Mach() && n->as_Mach()->has_call()) {
899 RegMask regs;
900 regs.Insert(_matcher.c_frame_pointer());
901 regs.OR(n->out_RegMask());
903 MachProjNode *proj = new (C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
904 map_node_to_block(proj, block);
905 block->insert_node(proj, phi_cnt++);
907 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false);
908 }
910 // Children are now all ready
911 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
912 Node* m = n->fast_out(i5); // Get user
913 if (get_block_for_node(m) != block) {
914 continue;
915 }
916 if( m->is_Phi() ) continue;
917 if (m->_idx >= max_idx) { // new node, skip it
918 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
919 continue;
920 }
921 int m_cnt = ready_cnt.at(m->_idx)-1;
922 ready_cnt.at_put(m->_idx, m_cnt);
923 if( m_cnt == 0 )
924 worklist.push(m);
925 }
926 }
928 if( phi_cnt != block->end_idx() ) {
929 // did not schedule all. Retry, Bailout, or Die
930 if (C->subsume_loads() == true && !C->failing()) {
931 // Retry with subsume_loads == false
932 // If this is the first failure, the sentinel string will "stick"
933 // to the Compile object, and the C2Compiler will see it and retry.
934 C->record_failure(C2Compiler::retry_no_subsuming_loads());
935 }
936 // assert( phi_cnt == end_idx(), "did not schedule all" );
937 return false;
938 }
940 #ifndef PRODUCT
941 if (trace_opto_pipelining()) {
942 tty->print_cr("#");
943 tty->print_cr("# after schedule_local");
944 for (uint i = 0;i < block->number_of_nodes();i++) {
945 tty->print("# ");
946 block->get_node(i)->fast_dump();
947 }
948 tty->cr();
949 }
950 #endif
953 return true;
954 }
956 //--------------------------catch_cleanup_fix_all_inputs-----------------------
957 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
958 for (uint l = 0; l < use->len(); l++) {
959 if (use->in(l) == old_def) {
960 if (l < use->req()) {
961 use->set_req(l, new_def);
962 } else {
963 use->rm_prec(l);
964 use->add_prec(new_def);
965 l--;
966 }
967 }
968 }
969 }
971 //------------------------------catch_cleanup_find_cloned_def------------------
972 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
973 assert( use_blk != def_blk, "Inter-block cleanup only");
975 // The use is some block below the Catch. Find and return the clone of the def
976 // that dominates the use. If there is no clone in a dominating block, then
977 // create a phi for the def in a dominating block.
979 // Find which successor block dominates this use. The successor
980 // blocks must all be single-entry (from the Catch only; I will have
981 // split blocks to make this so), hence they all dominate.
982 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
983 use_blk = use_blk->_idom;
985 // Find the successor
986 Node *fixup = NULL;
988 uint j;
989 for( j = 0; j < def_blk->_num_succs; j++ )
990 if( use_blk == def_blk->_succs[j] )
991 break;
993 if( j == def_blk->_num_succs ) {
994 // Block at same level in dom-tree is not a successor. It needs a
995 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
996 Node_Array inputs = new Node_List(Thread::current()->resource_area());
997 for(uint k = 1; k < use_blk->num_preds(); k++) {
998 Block* block = get_block_for_node(use_blk->pred(k));
999 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx));
1000 }
1002 // Check to see if the use_blk already has an identical phi inserted.
1003 // If it exists, it will be at the first position since all uses of a
1004 // def are processed together.
1005 Node *phi = use_blk->get_node(1);
1006 if( phi->is_Phi() ) {
1007 fixup = phi;
1008 for (uint k = 1; k < use_blk->num_preds(); k++) {
1009 if (phi->in(k) != inputs[k]) {
1010 // Not a match
1011 fixup = NULL;
1012 break;
1013 }
1014 }
1015 }
1017 // If an existing PhiNode was not found, make a new one.
1018 if (fixup == NULL) {
1019 Node *new_phi = PhiNode::make(use_blk->head(), def);
1020 use_blk->insert_node(new_phi, 1);
1021 map_node_to_block(new_phi, use_blk);
1022 for (uint k = 1; k < use_blk->num_preds(); k++) {
1023 new_phi->set_req(k, inputs[k]);
1024 }
1025 fixup = new_phi;
1026 }
1028 } else {
1029 // Found the use just below the Catch. Make it use the clone.
1030 fixup = use_blk->get_node(n_clone_idx);
1031 }
1033 return fixup;
1034 }
1036 //--------------------------catch_cleanup_intra_block--------------------------
1037 // Fix all input edges in use that reference "def". The use is in the same
1038 // block as the def and both have been cloned in each successor block.
1039 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
1041 // Both the use and def have been cloned. For each successor block,
1042 // get the clone of the use, and make its input the clone of the def
1043 // found in that block.
1045 uint use_idx = blk->find_node(use);
1046 uint offset_idx = use_idx - beg;
1047 for( uint k = 0; k < blk->_num_succs; k++ ) {
1048 // Get clone in each successor block
1049 Block *sb = blk->_succs[k];
1050 Node *clone = sb->get_node(offset_idx+1);
1051 assert( clone->Opcode() == use->Opcode(), "" );
1053 // Make use-clone reference the def-clone
1054 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx));
1055 }
1056 }
1058 //------------------------------catch_cleanup_inter_block---------------------
1059 // Fix all input edges in use that reference "def". The use is in a different
1060 // block than the def.
1061 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1062 if( !use_blk ) return; // Can happen if the use is a precedence edge
1064 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx);
1065 catch_cleanup_fix_all_inputs(use, def, new_def);
1066 }
1068 //------------------------------call_catch_cleanup-----------------------------
1069 // If we inserted any instructions between a Call and his CatchNode,
1070 // clone the instructions on all paths below the Catch.
1071 void PhaseCFG::call_catch_cleanup(Block* block) {
1073 // End of region to clone
1074 uint end = block->end_idx();
1075 if( !block->get_node(end)->is_Catch() ) return;
1076 // Start of region to clone
1077 uint beg = end;
1078 while(!block->get_node(beg-1)->is_MachProj() ||
1079 !block->get_node(beg-1)->in(0)->is_MachCall() ) {
1080 beg--;
1081 assert(beg > 0,"Catch cleanup walking beyond block boundary");
1082 }
1083 // Range of inserted instructions is [beg, end)
1084 if( beg == end ) return;
1086 // Clone along all Catch output paths. Clone area between the 'beg' and
1087 // 'end' indices.
1088 for( uint i = 0; i < block->_num_succs; i++ ) {
1089 Block *sb = block->_succs[i];
1090 // Clone the entire area; ignoring the edge fixup for now.
1091 for( uint j = end; j > beg; j-- ) {
1092 // It is safe here to clone a node with anti_dependence
1093 // since clones dominate on each path.
1094 Node *clone = block->get_node(j-1)->clone();
1095 sb->insert_node(clone, 1);
1096 map_node_to_block(clone, sb);
1097 }
1098 }
1101 // Fixup edges. Check the def-use info per cloned Node
1102 for(uint i2 = beg; i2 < end; i2++ ) {
1103 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1104 Node *n = block->get_node(i2); // Node that got cloned
1105 // Need DU safe iterator because of edge manipulation in calls.
1106 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1107 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1108 out->push(n->fast_out(j1));
1109 }
1110 uint max = out->size();
1111 for (uint j = 0; j < max; j++) {// For all users
1112 Node *use = out->pop();
1113 Block *buse = get_block_for_node(use);
1114 if( use->is_Phi() ) {
1115 for( uint k = 1; k < use->req(); k++ )
1116 if( use->in(k) == n ) {
1117 Block* b = get_block_for_node(buse->pred(k));
1118 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx);
1119 use->set_req(k, fixup);
1120 }
1121 } else {
1122 if (block == buse) {
1123 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx);
1124 } else {
1125 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx);
1126 }
1127 }
1128 } // End for all users
1130 } // End of for all Nodes in cloned area
1132 // Remove the now-dead cloned ops
1133 for(uint i3 = beg; i3 < end; i3++ ) {
1134 block->get_node(beg)->disconnect_inputs(NULL, C);
1135 block->remove_node(beg);
1136 }
1138 // If the successor blocks have a CreateEx node, move it back to the top
1139 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) {
1140 Block *sb = block->_succs[i4];
1141 uint new_cnt = end - beg;
1142 // Remove any newly created, but dead, nodes.
1143 for( uint j = new_cnt; j > 0; j-- ) {
1144 Node *n = sb->get_node(j);
1145 if (n->outcnt() == 0 &&
1146 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1147 n->disconnect_inputs(NULL, C);
1148 sb->remove_node(j);
1149 new_cnt--;
1150 }
1151 }
1152 // If any newly created nodes remain, move the CreateEx node to the top
1153 if (new_cnt > 0) {
1154 Node *cex = sb->get_node(1+new_cnt);
1155 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1156 sb->remove_node(1+new_cnt);
1157 sb->insert_node(cex, 1);
1158 }
1159 }
1160 }
1161 }