1 /*
2 * Copyright (c) 1998, 2011, 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/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
49 # include "adfiles/ad_ppc.hpp"
50 #endif
51
52 // Optimization - Graph Style
53
54 //------------------------------implicit_null_check----------------------------
55 // Detect implicit-null-check opportunities. Basically, find NULL checks
56 // with suitable memory ops nearby. Use the memory op to do the NULL check.
57 // I can generate a memory op if there is not one nearby.
58 // The proj is the control projection for the not-null case.
59 // The val is the pointer being checked for nullness or
60 // decodeHeapOop_not_null node if it did not fold into address.
61 void Block::implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons) {
62 // Assume if null check need for 0 offset then always needed
63 // Intel solaris doesn't support any null checks yet and no
64 // mechanism exists (yet) to set the switches at an os_cpu level
65 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
66
67 // Make sure the ptr-is-null path appears to be uncommon!
68 float f = end()->as_MachIf()->_prob;
69 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
70 if( f > PROB_UNLIKELY_MAG(4) ) return;
71
72 uint bidx = 0; // Capture index of value into memop
73 bool was_store; // Memory op is a store op
74
75 // Get the successor block for if the test ptr is non-null
76 Block* not_null_block; // this one goes with the proj
77 Block* null_block;
78 if (_nodes[_nodes.size()-1] == proj) {
79 null_block = _succs[0];
80 not_null_block = _succs[1];
81 } else {
82 assert(_nodes[_nodes.size()-2] == proj, "proj is one or the other");
83 not_null_block = _succs[0];
84 null_block = _succs[1];
85 }
86 while (null_block->is_Empty() == Block::empty_with_goto) {
87 null_block = null_block->_succs[0];
88 }
89
90 // Search the exception block for an uncommon trap.
91 // (See Parse::do_if and Parse::do_ifnull for the reason
92 // we need an uncommon trap. Briefly, we need a way to
93 // detect failure of this optimization, as in 6366351.)
94 {
95 bool found_trap = false;
96 for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) {
97 Node* nn = null_block->_nodes[i1];
98 if (nn->is_MachCall() &&
99 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
100 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
101 if (trtype->isa_int() && trtype->is_int()->is_con()) {
102 jint tr_con = trtype->is_int()->get_con();
103 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
104 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
105 assert((int)reason < (int)BitsPerInt, "recode bit map");
106 if (is_set_nth_bit(allowed_reasons, (int) reason)
107 && action != Deoptimization::Action_none) {
108 // This uncommon trap is sure to recompile, eventually.
109 // When that happens, C->too_many_traps will prevent
110 // this transformation from happening again.
111 found_trap = true;
112 }
113 }
114 break;
115 }
116 }
117 if (!found_trap) {
118 // We did not find an uncommon trap.
119 return;
120 }
121 }
122
123 // Check for decodeHeapOop_not_null node which did not fold into address
124 bool is_decoden = ((intptr_t)val) & 1;
125 val = (Node*)(((intptr_t)val) & ~1);
126
127 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() &&
128 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity");
129
130 // Search the successor block for a load or store who's base value is also
131 // the tested value. There may be several.
132 Node_List *out = new Node_List(Thread::current()->resource_area());
133 MachNode *best = NULL; // Best found so far
134 for (DUIterator i = val->outs(); val->has_out(i); i++) {
135 Node *m = val->out(i);
136 if( !m->is_Mach() ) continue;
137 MachNode *mach = m->as_Mach();
138 was_store = false;
139 int iop = mach->ideal_Opcode();
140 switch( iop ) {
141 case Op_LoadB:
142 case Op_LoadUS:
143 case Op_LoadD:
144 case Op_LoadF:
145 case Op_LoadI:
146 case Op_LoadL:
147 case Op_LoadP:
148 case Op_LoadN:
149 case Op_LoadS:
150 case Op_LoadKlass:
151 case Op_LoadNKlass:
152 case Op_LoadRange:
153 case Op_LoadD_unaligned:
154 case Op_LoadL_unaligned:
155 assert(mach->in(2) == val, "should be address");
156 break;
157 case Op_StoreB:
158 case Op_StoreC:
159 case Op_StoreCM:
160 case Op_StoreD:
161 case Op_StoreF:
162 case Op_StoreI:
163 case Op_StoreL:
164 case Op_StoreP:
165 case Op_StoreN:
166 was_store = true; // Memory op is a store op
167 // Stores will have their address in slot 2 (memory in slot 1).
168 // If the value being nul-checked is in another slot, it means we
169 // are storing the checked value, which does NOT check the value!
170 if( mach->in(2) != val ) continue;
171 break; // Found a memory op?
172 case Op_StrComp:
173 case Op_StrEquals:
174 case Op_StrIndexOf:
175 case Op_AryEq:
176 // Not a legit memory op for implicit null check regardless of
177 // embedded loads
178 continue;
179 default: // Also check for embedded loads
180 if( !mach->needs_anti_dependence_check() )
181 continue; // Not an memory op; skip it
182 if( must_clone[iop] ) {
183 // Do not move nodes which produce flags because
184 // RA will try to clone it to place near branch and
185 // it will cause recompilation, see clone_node().
186 continue;
187 }
188 {
189 // Check that value is used in memory address in
190 // instructions with embedded load (CmpP val1,(val2+off)).
191 Node* base;
192 Node* index;
193 const MachOper* oper = mach->memory_inputs(base, index);
194 if (oper == NULL || oper == (MachOper*)-1) {
195 continue; // Not an memory op; skip it
196 }
197 if (val == base ||
198 val == index && val->bottom_type()->isa_narrowoop()) {
199 break; // Found it
200 } else {
201 continue; // Skip it
202 }
203 }
204 break;
205 }
206 // check if the offset is not too high for implicit exception
207 {
208 intptr_t offset = 0;
209 const TypePtr *adr_type = NULL; // Do not need this return value here
210 const Node* base = mach->get_base_and_disp(offset, adr_type);
211 if (base == NULL || base == NodeSentinel) {
212 // Narrow oop address doesn't have base, only index
213 if( val->bottom_type()->isa_narrowoop() &&
214 MacroAssembler::needs_explicit_null_check(offset) )
215 continue; // Give up if offset is beyond page size
216 // cannot reason about it; is probably not implicit null exception
217 } else {
218 const TypePtr* tptr;
219 if (UseCompressedOops && Universe::narrow_oop_shift() == 0) {
220 // 32-bits narrow oop can be the base of address expressions
221 tptr = base->bottom_type()->make_ptr();
222 } else {
223 // only regular oops are expected here
224 tptr = base->bottom_type()->is_ptr();
225 }
226 // Give up if offset is not a compile-time constant
227 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
228 continue;
229 offset += tptr->_offset; // correct if base is offseted
230 if( MacroAssembler::needs_explicit_null_check(offset) )
231 continue; // Give up is reference is beyond 4K page size
232 }
233 }
234
235 // Check ctrl input to see if the null-check dominates the memory op
236 Block *cb = cfg->_bbs[mach->_idx];
237 cb = cb->_idom; // Always hoist at least 1 block
238 if( !was_store ) { // Stores can be hoisted only one block
239 while( cb->_dom_depth > (_dom_depth + 1))
240 cb = cb->_idom; // Hoist loads as far as we want
241 // The non-null-block should dominate the memory op, too. Live
242 // range spilling will insert a spill in the non-null-block if it is
243 // needs to spill the memory op for an implicit null check.
244 if (cb->_dom_depth == (_dom_depth + 1)) {
245 if (cb != not_null_block) continue;
246 cb = cb->_idom;
247 }
248 }
249 if( cb != this ) continue;
250
251 // Found a memory user; see if it can be hoisted to check-block
252 uint vidx = 0; // Capture index of value into memop
253 uint j;
254 for( j = mach->req()-1; j > 0; j-- ) {
255 if( mach->in(j) == val ) {
256 vidx = j;
257 // Ignore DecodeN val which could be hoisted to where needed.
258 if( is_decoden ) continue;
259 }
260 // Block of memory-op input
261 Block *inb = cfg->_bbs[mach->in(j)->_idx];
262 Block *b = this; // Start from nul check
263 while( b != inb && b->_dom_depth > inb->_dom_depth )
264 b = b->_idom; // search upwards for input
265 // See if input dominates null check
266 if( b != inb )
267 break;
268 }
269 if( j > 0 )
270 continue;
271 Block *mb = cfg->_bbs[mach->_idx];
272 // Hoisting stores requires more checks for the anti-dependence case.
273 // Give up hoisting if we have to move the store past any load.
274 if( was_store ) {
275 Block *b = mb; // Start searching here for a local load
276 // mach use (faulting) trying to hoist
277 // n might be blocker to hoisting
278 while( b != this ) {
279 uint k;
280 for( k = 1; k < b->_nodes.size(); k++ ) {
281 Node *n = b->_nodes[k];
282 if( n->needs_anti_dependence_check() &&
283 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
284 break; // Found anti-dependent load
285 }
286 if( k < b->_nodes.size() )
287 break; // Found anti-dependent load
288 // Make sure control does not do a merge (would have to check allpaths)
289 if( b->num_preds() != 2 ) break;
290 b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block
291 }
292 if( b != this ) continue;
293 }
294
295 // Make sure this memory op is not already being used for a NullCheck
296 Node *e = mb->end();
297 if( e->is_MachNullCheck() && e->in(1) == mach )
298 continue; // Already being used as a NULL check
299
300 // Found a candidate! Pick one with least dom depth - the highest
301 // in the dom tree should be closest to the null check.
302 if( !best ||
303 cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) {
304 best = mach;
305 bidx = vidx;
306
307 }
308 }
309 // No candidate!
310 if( !best ) return;
311
312 // ---- Found an implicit null check
313 extern int implicit_null_checks;
314 implicit_null_checks++;
315
316 if( is_decoden ) {
317 // Check if we need to hoist decodeHeapOop_not_null first.
318 Block *valb = cfg->_bbs[val->_idx];
319 if( this != valb && this->_dom_depth < valb->_dom_depth ) {
320 // Hoist it up to the end of the test block.
321 valb->find_remove(val);
322 this->add_inst(val);
323 cfg->_bbs.map(val->_idx,this);
324 // DecodeN on x86 may kill flags. Check for flag-killing projections
325 // that also need to be hoisted.
326 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
327 Node* n = val->fast_out(j);
328 if( n->is_MachProj() ) {
329 cfg->_bbs[n->_idx]->find_remove(n);
330 this->add_inst(n);
331 cfg->_bbs.map(n->_idx,this);
332 }
333 }
334 }
335 }
336 // Hoist the memory candidate up to the end of the test block.
337 Block *old_block = cfg->_bbs[best->_idx];
338 old_block->find_remove(best);
339 add_inst(best);
340 cfg->_bbs.map(best->_idx,this);
341
342 // Move the control dependence
343 if (best->in(0) && best->in(0) == old_block->_nodes[0])
344 best->set_req(0, _nodes[0]);
345
346 // Check for flag-killing projections that also need to be hoisted
347 // Should be DU safe because no edge updates.
348 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
349 Node* n = best->fast_out(j);
350 if( n->is_MachProj() ) {
351 cfg->_bbs[n->_idx]->find_remove(n);
352 add_inst(n);
353 cfg->_bbs.map(n->_idx,this);
354 }
355 }
356
357 Compile *C = cfg->C;
358 // proj==Op_True --> ne test; proj==Op_False --> eq test.
359 // One of two graph shapes got matched:
360 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
361 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
362 // NULL checks are always branch-if-eq. If we see a IfTrue projection
363 // then we are replacing a 'ne' test with a 'eq' NULL check test.
364 // We need to flip the projections to keep the same semantics.
365 if( proj->Opcode() == Op_IfTrue ) {
366 // Swap order of projections in basic block to swap branch targets
367 Node *tmp1 = _nodes[end_idx()+1];
368 Node *tmp2 = _nodes[end_idx()+2];
369 _nodes.map(end_idx()+1, tmp2);
370 _nodes.map(end_idx()+2, tmp1);
371 Node *tmp = new (C, 1) Node(C->top()); // Use not NULL input
372 tmp1->replace_by(tmp);
373 tmp2->replace_by(tmp1);
374 tmp->replace_by(tmp2);
375 tmp->destruct();
376 }
377
378 // Remove the existing null check; use a new implicit null check instead.
379 // Since schedule-local needs precise def-use info, we need to correct
380 // it as well.
381 Node *old_tst = proj->in(0);
382 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
383 _nodes.map(end_idx(),nul_chk);
384 cfg->_bbs.map(nul_chk->_idx,this);
385 // Redirect users of old_test to nul_chk
386 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
387 old_tst->last_out(i2)->set_req(0, nul_chk);
388 // Clean-up any dead code
389 for (uint i3 = 0; i3 < old_tst->req(); i3++)
390 old_tst->set_req(i3, NULL);
391
392 cfg->latency_from_uses(nul_chk);
393 cfg->latency_from_uses(best);
394 }
395
396
397 //------------------------------select-----------------------------------------
398 // Select a nice fellow from the worklist to schedule next. If there is only
399 // one choice, then use it. Projections take top priority for correctness
400 // reasons - if I see a projection, then it is next. There are a number of
401 // other special cases, for instructions that consume condition codes, et al.
402 // These are chosen immediately. Some instructions are required to immediately
403 // precede the last instruction in the block, and these are taken last. Of the
404 // remaining cases (most), choose the instruction with the greatest latency
405 // (that is, the most number of pseudo-cycles required to the end of the
406 // routine). If there is a tie, choose the instruction with the most inputs.
407 Node *Block::select(PhaseCFG *cfg, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) {
408
409 // If only a single entry on the stack, use it
410 uint cnt = worklist.size();
411 if (cnt == 1) {
412 Node *n = worklist[0];
413 worklist.map(0,worklist.pop());
414 return n;
415 }
416
417 uint choice = 0; // Bigger is most important
418 uint latency = 0; // Bigger is scheduled first
419 uint score = 0; // Bigger is better
420 int idx = -1; // Index in worklist
421
422 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
423 // Order in worklist is used to break ties.
424 // See caller for how this is used to delay scheduling
425 // of induction variable increments to after the other
426 // uses of the phi are scheduled.
427 Node *n = worklist[i]; // Get Node on worklist
428
429 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
430 if( n->is_Proj() || // Projections always win
431 n->Opcode()== Op_Con || // So does constant 'Top'
432 iop == Op_CreateEx || // Create-exception must start block
433 iop == Op_CheckCastPP
434 ) {
435 worklist.map(i,worklist.pop());
436 return n;
437 }
438
439 // Final call in a block must be adjacent to 'catch'
440 Node *e = end();
441 if( e->is_Catch() && e->in(0)->in(0) == n )
442 continue;
443
444 // Memory op for an implicit null check has to be at the end of the block
445 if( e->is_MachNullCheck() && e->in(1) == n )
446 continue;
447
448 uint n_choice = 2;
449
450 // See if this instruction is consumed by a branch. If so, then (as the
451 // branch is the last instruction in the basic block) force it to the
452 // end of the basic block
453 if ( must_clone[iop] ) {
454 // See if any use is a branch
455 bool found_machif = false;
456
457 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
458 Node* use = n->fast_out(j);
459
460 // The use is a conditional branch, make them adjacent
461 if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) {
462 found_machif = true;
463 break;
464 }
465
466 // More than this instruction pending for successor to be ready,
467 // don't choose this if other opportunities are ready
468 if (ready_cnt.at(use->_idx) > 1)
469 n_choice = 1;
470 }
471
472 // loop terminated, prefer not to use this instruction
473 if (found_machif)
474 continue;
475 }
476
477 // See if this has a predecessor that is "must_clone", i.e. sets the
478 // condition code. If so, choose this first
479 for (uint j = 0; j < n->req() ; j++) {
480 Node *inn = n->in(j);
481 if (inn) {
482 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
483 n_choice = 3;
484 break;
485 }
486 }
487 }
488
489 // MachTemps should be scheduled last so they are near their uses
490 if (n->is_MachTemp()) {
491 n_choice = 1;
492 }
493
494 uint n_latency = cfg->_node_latency->at_grow(n->_idx);
495 uint n_score = n->req(); // Many inputs get high score to break ties
496
497 // Keep best latency found
498 if( choice < n_choice ||
499 ( choice == n_choice &&
500 ( latency < n_latency ||
501 ( latency == n_latency &&
502 ( score < n_score ))))) {
503 choice = n_choice;
504 latency = n_latency;
505 score = n_score;
506 idx = i; // Also keep index in worklist
507 }
508 } // End of for all ready nodes in worklist
509
510 assert(idx >= 0, "index should be set");
511 Node *n = worklist[(uint)idx]; // Get the winner
512
513 worklist.map((uint)idx, worklist.pop()); // Compress worklist
514 return n;
515 }
516
517
518 //------------------------------set_next_call----------------------------------
519 void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) {
520 if( next_call.test_set(n->_idx) ) return;
521 for( uint i=0; i<n->len(); i++ ) {
522 Node *m = n->in(i);
523 if( !m ) continue; // must see all nodes in block that precede call
524 if( bbs[m->_idx] == this )
525 set_next_call( m, next_call, bbs );
526 }
527 }
528
529 //------------------------------needed_for_next_call---------------------------
530 // Set the flag 'next_call' for each Node that is needed for the next call to
531 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
532 // next subroutine call get priority - basically it moves things NOT needed
533 // for the next call till after the call. This prevents me from trying to
534 // carry lots of stuff live across a call.
535 void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) {
536 // Find the next control-defining Node in this block
537 Node* call = NULL;
538 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
539 Node* m = this_call->fast_out(i);
540 if( bbs[m->_idx] == this && // Local-block user
541 m != this_call && // Not self-start node
542 m->is_MachCall() )
543 call = m;
544 break;
545 }
546 if (call == NULL) return; // No next call (e.g., block end is near)
547 // Set next-call for all inputs to this call
548 set_next_call(call, next_call, bbs);
549 }
550
551 //------------------------------add_call_kills-------------------------------------
552 void Block::add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
553 // Fill in the kill mask for the call
554 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
555 if( !regs.Member(r) ) { // Not already defined by the call
556 // Save-on-call register?
557 if ((save_policy[r] == 'C') ||
558 (save_policy[r] == 'A') ||
559 ((save_policy[r] == 'E') && exclude_soe)) {
560 proj->_rout.Insert(r);
561 }
562 }
563 }
564 }
565
566
567 //------------------------------sched_call-------------------------------------
568 uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, GrowableArray<int> &ready_cnt, MachCallNode *mcall, VectorSet &next_call ) {
569 RegMask regs;
570
571 // Schedule all the users of the call right now. All the users are
572 // projection Nodes, so they must be scheduled next to the call.
573 // Collect all the defined registers.
574 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
575 Node* n = mcall->fast_out(i);
576 assert( n->is_MachProj(), "" );
577 int n_cnt = ready_cnt.at(n->_idx)-1;
578 ready_cnt.at_put(n->_idx, n_cnt);
579 assert( n_cnt == 0, "" );
580 // Schedule next to call
581 _nodes.map(node_cnt++, n);
582 // Collect defined registers
583 regs.OR(n->out_RegMask());
584 // Check for scheduling the next control-definer
585 if( n->bottom_type() == Type::CONTROL )
586 // Warm up next pile of heuristic bits
587 needed_for_next_call(n, next_call, bbs);
588
589 // Children of projections are now all ready
590 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
591 Node* m = n->fast_out(j); // Get user
592 if( bbs[m->_idx] != this ) continue;
593 if( m->is_Phi() ) continue;
594 int m_cnt = ready_cnt.at(m->_idx)-1;
595 ready_cnt.at_put(m->_idx, m_cnt);
596 if( m_cnt == 0 )
597 worklist.push(m);
598 }
599
600 }
601
602 // Act as if the call defines the Frame Pointer.
603 // Certainly the FP is alive and well after the call.
604 regs.Insert(matcher.c_frame_pointer());
605
606 // Set all registers killed and not already defined by the call.
607 uint r_cnt = mcall->tf()->range()->cnt();
608 int op = mcall->ideal_Opcode();
609 MachProjNode *proj = new (matcher.C, 1) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
610 bbs.map(proj->_idx,this);
611 _nodes.insert(node_cnt++, proj);
612
613 // Select the right register save policy.
614 const char * save_policy;
615 switch (op) {
616 case Op_CallRuntime:
617 case Op_CallLeaf:
618 case Op_CallLeafNoFP:
619 // Calling C code so use C calling convention
620 save_policy = matcher._c_reg_save_policy;
621 break;
622
623 case Op_CallStaticJava:
624 case Op_CallDynamicJava:
625 // Calling Java code so use Java calling convention
626 save_policy = matcher._register_save_policy;
627 break;
628
629 default:
630 ShouldNotReachHere();
631 }
632
633 // When using CallRuntime mark SOE registers as killed by the call
634 // so values that could show up in the RegisterMap aren't live in a
635 // callee saved register since the register wouldn't know where to
636 // find them. CallLeaf and CallLeafNoFP are ok because they can't
637 // have debug info on them. Strictly speaking this only needs to be
638 // done for oops since idealreg2debugmask takes care of debug info
639 // references but there no way to handle oops differently than other
640 // pointers as far as the kill mask goes.
641 bool exclude_soe = op == Op_CallRuntime;
642
643 // If the call is a MethodHandle invoke, we need to exclude the
644 // register which is used to save the SP value over MH invokes from
645 // the mask. Otherwise this register could be used for
646 // deoptimization information.
647 if (op == Op_CallStaticJava) {
648 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
649 if (mcallstaticjava->_method_handle_invoke)
650 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
651 }
652
653 add_call_kills(proj, regs, save_policy, exclude_soe);
654
655 return node_cnt;
656 }
657
658
659 //------------------------------schedule_local---------------------------------
660 // Topological sort within a block. Someday become a real scheduler.
661 bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, GrowableArray<int> &ready_cnt, VectorSet &next_call) {
662 // Already "sorted" are the block start Node (as the first entry), and
663 // the block-ending Node and any trailing control projections. We leave
664 // these alone. PhiNodes and ParmNodes are made to follow the block start
665 // Node. Everything else gets topo-sorted.
666
667 #ifndef PRODUCT
668 if (cfg->trace_opto_pipelining()) {
669 tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order);
670 for (uint i = 0;i < _nodes.size();i++) {
671 tty->print("# ");
672 _nodes[i]->fast_dump();
673 }
674 tty->print_cr("#");
675 }
676 #endif
677
678 // RootNode is already sorted
679 if( _nodes.size() == 1 ) return true;
680
681 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
682 uint node_cnt = end_idx();
683 uint phi_cnt = 1;
684 uint i;
685 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
686 Node *n = _nodes[i];
687 if( n->is_Phi() || // Found a PhiNode or ParmNode
688 (n->is_Proj() && n->in(0) == head()) ) {
689 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
690 _nodes.map(i,_nodes[phi_cnt]);
691 _nodes.map(phi_cnt++,n); // swap Phi/Parm up front
692 } else { // All others
693 // Count block-local inputs to 'n'
694 uint cnt = n->len(); // Input count
695 uint local = 0;
696 for( uint j=0; j<cnt; j++ ) {
697 Node *m = n->in(j);
698 if( m && cfg->_bbs[m->_idx] == this && !m->is_top() )
699 local++; // One more block-local input
700 }
701 ready_cnt.at_put(n->_idx, local); // Count em up
702
703 #ifdef ASSERT
704 if( UseConcMarkSweepGC || UseG1GC ) {
705 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
706 // Check the precedence edges
707 for (uint prec = n->req(); prec < n->len(); prec++) {
708 Node* oop_store = n->in(prec);
709 if (oop_store != NULL) {
710 assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark");
711 }
712 }
713 }
714 }
715 #endif
716
717 // A few node types require changing a required edge to a precedence edge
718 // before allocation.
719 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
720 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
721 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
722 // MemBarAcquire could be created without Precedent edge.
723 // del_req() replaces the specified edge with the last input edge
724 // and then removes the last edge. If the specified edge > number of
725 // edges the last edge will be moved outside of the input edges array
726 // and the edge will be lost. This is why this code should be
727 // executed only when Precedent (== TypeFunc::Parms) edge is present.
728 Node *x = n->in(TypeFunc::Parms);
729 n->del_req(TypeFunc::Parms);
730 n->add_prec(x);
731 }
732 }
733 }
734 for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count
735 ready_cnt.at_put(_nodes[i2]->_idx, 0);
736
737 // All the prescheduled guys do not hold back internal nodes
738 uint i3;
739 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
740 Node *n = _nodes[i3]; // Get pre-scheduled
741 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
742 Node* m = n->fast_out(j);
743 if( cfg->_bbs[m->_idx] ==this ) { // Local-block user
744 int m_cnt = ready_cnt.at(m->_idx)-1;
745 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
746 }
747 }
748 }
749
750 Node_List delay;
751 // Make a worklist
752 Node_List worklist;
753 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
754 Node *m = _nodes[i4];
755 if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
756 if (m->is_iteratively_computed()) {
757 // Push induction variable increments last to allow other uses
758 // of the phi to be scheduled first. The select() method breaks
759 // ties in scheduling by worklist order.
760 delay.push(m);
761 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
762 // Force the CreateEx to the top of the list so it's processed
763 // first and ends up at the start of the block.
764 worklist.insert(0, m);
765 } else {
766 worklist.push(m); // Then on to worklist!
767 }
768 }
769 }
770 while (delay.size()) {
771 Node* d = delay.pop();
772 worklist.push(d);
773 }
774
775 // Warm up the 'next_call' heuristic bits
776 needed_for_next_call(_nodes[0], next_call, cfg->_bbs);
777
778 #ifndef PRODUCT
779 if (cfg->trace_opto_pipelining()) {
780 for (uint j=0; j<_nodes.size(); j++) {
781 Node *n = _nodes[j];
782 int idx = n->_idx;
783 tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
784 tty->print("latency:%3d ", cfg->_node_latency->at_grow(idx));
785 tty->print("%4d: %s\n", idx, n->Name());
786 }
787 }
788 #endif
789
790 uint max_idx = (uint)ready_cnt.length();
791 // Pull from worklist and schedule
792 while( worklist.size() ) { // Worklist is not ready
793
794 #ifndef PRODUCT
795 if (cfg->trace_opto_pipelining()) {
796 tty->print("# ready list:");
797 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
798 Node *n = worklist[i]; // Get Node on worklist
799 tty->print(" %d", n->_idx);
800 }
801 tty->cr();
802 }
803 #endif
804
805 // Select and pop a ready guy from worklist
806 Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt);
807 _nodes.map(phi_cnt++,n); // Schedule him next
808
809 #ifndef PRODUCT
810 if (cfg->trace_opto_pipelining()) {
811 tty->print("# select %d: %s", n->_idx, n->Name());
812 tty->print(", latency:%d", cfg->_node_latency->at_grow(n->_idx));
813 n->dump();
814 if (Verbose) {
815 tty->print("# ready list:");
816 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
817 Node *n = worklist[i]; // Get Node on worklist
818 tty->print(" %d", n->_idx);
819 }
820 tty->cr();
821 }
822 }
823
824 #endif
825 if( n->is_MachCall() ) {
826 MachCallNode *mcall = n->as_MachCall();
827 phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call);
828 continue;
829 }
830
831 if (n->is_Mach() && n->as_Mach()->has_call()) {
832 RegMask regs;
833 regs.Insert(matcher.c_frame_pointer());
834 regs.OR(n->out_RegMask());
835
836 MachProjNode *proj = new (matcher.C, 1) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
837 cfg->_bbs.map(proj->_idx,this);
838 _nodes.insert(phi_cnt++, proj);
839
840 add_call_kills(proj, regs, matcher._c_reg_save_policy, false);
841 }
842
843 // Children are now all ready
844 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
845 Node* m = n->fast_out(i5); // Get user
846 if( cfg->_bbs[m->_idx] != this ) continue;
847 if( m->is_Phi() ) continue;
848 if (m->_idx >= max_idx) { // new node, skip it
849 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
850 continue;
851 }
852 int m_cnt = ready_cnt.at(m->_idx)-1;
853 ready_cnt.at_put(m->_idx, m_cnt);
854 if( m_cnt == 0 )
855 worklist.push(m);
856 }
857 }
858
859 if( phi_cnt != end_idx() ) {
860 // did not schedule all. Retry, Bailout, or Die
861 Compile* C = matcher.C;
862 if (C->subsume_loads() == true && !C->failing()) {
863 // Retry with subsume_loads == false
864 // If this is the first failure, the sentinel string will "stick"
865 // to the Compile object, and the C2Compiler will see it and retry.
866 C->record_failure(C2Compiler::retry_no_subsuming_loads());
867 }
868 // assert( phi_cnt == end_idx(), "did not schedule all" );
869 return false;
870 }
871
872 #ifndef PRODUCT
873 if (cfg->trace_opto_pipelining()) {
874 tty->print_cr("#");
875 tty->print_cr("# after schedule_local");
876 for (uint i = 0;i < _nodes.size();i++) {
877 tty->print("# ");
878 _nodes[i]->fast_dump();
879 }
880 tty->cr();
881 }
882 #endif
883
884
885 return true;
886 }
887
888 //--------------------------catch_cleanup_fix_all_inputs-----------------------
889 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
890 for (uint l = 0; l < use->len(); l++) {
891 if (use->in(l) == old_def) {
892 if (l < use->req()) {
893 use->set_req(l, new_def);
894 } else {
895 use->rm_prec(l);
896 use->add_prec(new_def);
897 l--;
898 }
899 }
900 }
901 }
902
903 //------------------------------catch_cleanup_find_cloned_def------------------
904 static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
905 assert( use_blk != def_blk, "Inter-block cleanup only");
906
907 // The use is some block below the Catch. Find and return the clone of the def
908 // that dominates the use. If there is no clone in a dominating block, then
909 // create a phi for the def in a dominating block.
910
911 // Find which successor block dominates this use. The successor
912 // blocks must all be single-entry (from the Catch only; I will have
913 // split blocks to make this so), hence they all dominate.
914 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
915 use_blk = use_blk->_idom;
916
917 // Find the successor
918 Node *fixup = NULL;
919
920 uint j;
921 for( j = 0; j < def_blk->_num_succs; j++ )
922 if( use_blk == def_blk->_succs[j] )
923 break;
924
925 if( j == def_blk->_num_succs ) {
926 // Block at same level in dom-tree is not a successor. It needs a
927 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
928 Node_Array inputs = new Node_List(Thread::current()->resource_area());
929 for(uint k = 1; k < use_blk->num_preds(); k++) {
930 inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx));
931 }
932
933 // Check to see if the use_blk already has an identical phi inserted.
934 // If it exists, it will be at the first position since all uses of a
935 // def are processed together.
936 Node *phi = use_blk->_nodes[1];
937 if( phi->is_Phi() ) {
938 fixup = phi;
939 for (uint k = 1; k < use_blk->num_preds(); k++) {
940 if (phi->in(k) != inputs[k]) {
941 // Not a match
942 fixup = NULL;
943 break;
944 }
945 }
946 }
947
948 // If an existing PhiNode was not found, make a new one.
949 if (fixup == NULL) {
950 Node *new_phi = PhiNode::make(use_blk->head(), def);
951 use_blk->_nodes.insert(1, new_phi);
952 bbs.map(new_phi->_idx, use_blk);
953 for (uint k = 1; k < use_blk->num_preds(); k++) {
954 new_phi->set_req(k, inputs[k]);
955 }
956 fixup = new_phi;
957 }
958
959 } else {
960 // Found the use just below the Catch. Make it use the clone.
961 fixup = use_blk->_nodes[n_clone_idx];
962 }
963
964 return fixup;
965 }
966
967 //--------------------------catch_cleanup_intra_block--------------------------
968 // Fix all input edges in use that reference "def". The use is in the same
969 // block as the def and both have been cloned in each successor block.
970 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
971
972 // Both the use and def have been cloned. For each successor block,
973 // get the clone of the use, and make its input the clone of the def
974 // found in that block.
975
976 uint use_idx = blk->find_node(use);
977 uint offset_idx = use_idx - beg;
978 for( uint k = 0; k < blk->_num_succs; k++ ) {
979 // Get clone in each successor block
980 Block *sb = blk->_succs[k];
981 Node *clone = sb->_nodes[offset_idx+1];
982 assert( clone->Opcode() == use->Opcode(), "" );
983
984 // Make use-clone reference the def-clone
985 catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]);
986 }
987 }
988
989 //------------------------------catch_cleanup_inter_block---------------------
990 // Fix all input edges in use that reference "def". The use is in a different
991 // block than the def.
992 static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
993 if( !use_blk ) return; // Can happen if the use is a precedence edge
994
995 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx);
996 catch_cleanup_fix_all_inputs(use, def, new_def);
997 }
998
999 //------------------------------call_catch_cleanup-----------------------------
1000 // If we inserted any instructions between a Call and his CatchNode,
1001 // clone the instructions on all paths below the Catch.
1002 void Block::call_catch_cleanup(Block_Array &bbs) {
1003
1004 // End of region to clone
1005 uint end = end_idx();
1006 if( !_nodes[end]->is_Catch() ) return;
1007 // Start of region to clone
1008 uint beg = end;
1009 while(!_nodes[beg-1]->is_MachProj() ||
1010 !_nodes[beg-1]->in(0)->is_MachCall() ) {
1011 beg--;
1012 assert(beg > 0,"Catch cleanup walking beyond block boundary");
1013 }
1014 // Range of inserted instructions is [beg, end)
1015 if( beg == end ) return;
1016
1017 // Clone along all Catch output paths. Clone area between the 'beg' and
1018 // 'end' indices.
1019 for( uint i = 0; i < _num_succs; i++ ) {
1020 Block *sb = _succs[i];
1021 // Clone the entire area; ignoring the edge fixup for now.
1022 for( uint j = end; j > beg; j-- ) {
1023 // It is safe here to clone a node with anti_dependence
1024 // since clones dominate on each path.
1025 Node *clone = _nodes[j-1]->clone();
1026 sb->_nodes.insert( 1, clone );
1027 bbs.map(clone->_idx,sb);
1028 }
1029 }
1030
1031
1032 // Fixup edges. Check the def-use info per cloned Node
1033 for(uint i2 = beg; i2 < end; i2++ ) {
1034 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1035 Node *n = _nodes[i2]; // Node that got cloned
1036 // Need DU safe iterator because of edge manipulation in calls.
1037 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1038 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1039 out->push(n->fast_out(j1));
1040 }
1041 uint max = out->size();
1042 for (uint j = 0; j < max; j++) {// For all users
1043 Node *use = out->pop();
1044 Block *buse = bbs[use->_idx];
1045 if( use->is_Phi() ) {
1046 for( uint k = 1; k < use->req(); k++ )
1047 if( use->in(k) == n ) {
1048 Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx);
1049 use->set_req(k, fixup);
1050 }
1051 } else {
1052 if (this == buse) {
1053 catch_cleanup_intra_block(use, n, this, beg, n_clone_idx);
1054 } else {
1055 catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx);
1056 }
1057 }
1058 } // End for all users
1059
1060 } // End of for all Nodes in cloned area
1061
1062 // Remove the now-dead cloned ops
1063 for(uint i3 = beg; i3 < end; i3++ ) {
1064 _nodes[beg]->disconnect_inputs(NULL);
1065 _nodes.remove(beg);
1066 }
1067
1068 // If the successor blocks have a CreateEx node, move it back to the top
1069 for(uint i4 = 0; i4 < _num_succs; i4++ ) {
1070 Block *sb = _succs[i4];
1071 uint new_cnt = end - beg;
1072 // Remove any newly created, but dead, nodes.
1073 for( uint j = new_cnt; j > 0; j-- ) {
1074 Node *n = sb->_nodes[j];
1075 if (n->outcnt() == 0 &&
1076 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1077 n->disconnect_inputs(NULL);
1078 sb->_nodes.remove(j);
1079 new_cnt--;
1080 }
1081 }
1082 // If any newly created nodes remain, move the CreateEx node to the top
1083 if (new_cnt > 0) {
1084 Node *cex = sb->_nodes[1+new_cnt];
1085 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1086 sb->_nodes.remove(1+new_cnt);
1087 sb->_nodes.insert(1,cex);
1088 }
1089 }
1090 }
1091 }