1 /*
2 * Copyright (c) 1998, 2010, 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
46 // Optimization - Graph Style
47
48 //------------------------------implicit_null_check----------------------------
49 // Detect implicit-null-check opportunities. Basically, find NULL checks
50 // with suitable memory ops nearby. Use the memory op to do the NULL check.
51 // I can generate a memory op if there is not one nearby.
52 // The proj is the control projection for the not-null case.
53 // The val is the pointer being checked for nullness or
54 // decodeHeapOop_not_null node if it did not fold into address.
55 void Block::implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons) {
56 // Assume if null check need for 0 offset then always needed
57 // Intel solaris doesn't support any null checks yet and no
58 // mechanism exists (yet) to set the switches at an os_cpu level
59 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
60
61 // Make sure the ptr-is-null path appears to be uncommon!
62 float f = end()->as_MachIf()->_prob;
63 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
64 if( f > PROB_UNLIKELY_MAG(4) ) return;
65
66 uint bidx = 0; // Capture index of value into memop
67 bool was_store; // Memory op is a store op
68
69 // Get the successor block for if the test ptr is non-null
70 Block* not_null_block; // this one goes with the proj
71 Block* null_block;
72 if (_nodes[_nodes.size()-1] == proj) {
73 null_block = _succs[0];
74 not_null_block = _succs[1];
75 } else {
76 assert(_nodes[_nodes.size()-2] == proj, "proj is one or the other");
77 not_null_block = _succs[0];
78 null_block = _succs[1];
79 }
80 while (null_block->is_Empty() == Block::empty_with_goto) {
81 null_block = null_block->_succs[0];
82 }
83
84 // Search the exception block for an uncommon trap.
85 // (See Parse::do_if and Parse::do_ifnull for the reason
86 // we need an uncommon trap. Briefly, we need a way to
87 // detect failure of this optimization, as in 6366351.)
88 {
89 bool found_trap = false;
90 for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) {
91 Node* nn = null_block->_nodes[i1];
92 if (nn->is_MachCall() &&
93 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
94 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
95 if (trtype->isa_int() && trtype->is_int()->is_con()) {
96 jint tr_con = trtype->is_int()->get_con();
97 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
98 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
99 assert((int)reason < (int)BitsPerInt, "recode bit map");
100 if (is_set_nth_bit(allowed_reasons, (int) reason)
101 && action != Deoptimization::Action_none) {
102 // This uncommon trap is sure to recompile, eventually.
103 // When that happens, C->too_many_traps will prevent
104 // this transformation from happening again.
105 found_trap = true;
106 }
107 }
108 break;
109 }
110 }
111 if (!found_trap) {
112 // We did not find an uncommon trap.
113 return;
114 }
115 }
116
117 // Check for decodeHeapOop_not_null node which did not fold into address
118 bool is_decoden = ((intptr_t)val) & 1;
119 val = (Node*)(((intptr_t)val) & ~1);
120
121 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() &&
122 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity");
123
124 // Search the successor block for a load or store who's base value is also
125 // the tested value. There may be several.
126 Node_List *out = new Node_List(Thread::current()->resource_area());
127 MachNode *best = NULL; // Best found so far
128 for (DUIterator i = val->outs(); val->has_out(i); i++) {
129 Node *m = val->out(i);
130 if( !m->is_Mach() ) continue;
131 MachNode *mach = m->as_Mach();
132 was_store = false;
133 int iop = mach->ideal_Opcode();
134 switch( iop ) {
135 case Op_LoadB:
136 case Op_LoadUS:
137 case Op_LoadD:
138 case Op_LoadF:
139 case Op_LoadI:
140 case Op_LoadL:
141 case Op_LoadP:
142 case Op_LoadN:
143 case Op_LoadS:
144 case Op_LoadKlass:
145 case Op_LoadNKlass:
146 case Op_LoadRange:
147 case Op_LoadD_unaligned:
148 case Op_LoadL_unaligned:
149 assert(mach->in(2) == val, "should be address");
150 break;
151 case Op_StoreB:
152 case Op_StoreC:
153 case Op_StoreCM:
154 case Op_StoreD:
155 case Op_StoreF:
156 case Op_StoreI:
157 case Op_StoreL:
158 case Op_StoreP:
159 case Op_StoreN:
160 was_store = true; // Memory op is a store op
161 // Stores will have their address in slot 2 (memory in slot 1).
162 // If the value being nul-checked is in another slot, it means we
163 // are storing the checked value, which does NOT check the value!
164 if( mach->in(2) != val ) continue;
165 break; // Found a memory op?
166 case Op_StrComp:
167 case Op_StrEquals:
168 case Op_StrIndexOf:
169 case Op_AryEq:
170 // Not a legit memory op for implicit null check regardless of
171 // embedded loads
172 continue;
173 default: // Also check for embedded loads
174 if( !mach->needs_anti_dependence_check() )
175 continue; // Not an memory op; skip it
176 if( must_clone[iop] ) {
177 // Do not move nodes which produce flags because
178 // RA will try to clone it to place near branch and
179 // it will cause recompilation, see clone_node().
180 continue;
181 }
182 {
183 // Check that value is used in memory address in
184 // instructions with embedded load (CmpP val1,(val2+off)).
185 Node* base;
186 Node* index;
187 const MachOper* oper = mach->memory_inputs(base, index);
188 if (oper == NULL || oper == (MachOper*)-1) {
189 continue; // Not an memory op; skip it
190 }
191 if (val == base ||
192 val == index && val->bottom_type()->isa_narrowoop()) {
193 break; // Found it
194 } else {
195 continue; // Skip it
196 }
197 }
198 break;
199 }
200 // check if the offset is not too high for implicit exception
201 {
202 intptr_t offset = 0;
203 const TypePtr *adr_type = NULL; // Do not need this return value here
204 const Node* base = mach->get_base_and_disp(offset, adr_type);
205 if (base == NULL || base == NodeSentinel) {
206 // Narrow oop address doesn't have base, only index
207 if( val->bottom_type()->isa_narrowoop() &&
208 MacroAssembler::needs_explicit_null_check(offset) )
209 continue; // Give up if offset is beyond page size
210 // cannot reason about it; is probably not implicit null exception
211 } else {
212 const TypePtr* tptr;
213 if (UseCompressedOops && Universe::narrow_oop_shift() == 0) {
214 // 32-bits narrow oop can be the base of address expressions
215 tptr = base->bottom_type()->make_ptr();
216 } else {
217 // only regular oops are expected here
218 tptr = base->bottom_type()->is_ptr();
219 }
220 // Give up if offset is not a compile-time constant
221 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
222 continue;
223 offset += tptr->_offset; // correct if base is offseted
224 if( MacroAssembler::needs_explicit_null_check(offset) )
225 continue; // Give up is reference is beyond 4K page size
226 }
227 }
228
229 // Check ctrl input to see if the null-check dominates the memory op
230 Block *cb = cfg->_bbs[mach->_idx];
231 cb = cb->_idom; // Always hoist at least 1 block
232 if( !was_store ) { // Stores can be hoisted only one block
233 while( cb->_dom_depth > (_dom_depth + 1))
234 cb = cb->_idom; // Hoist loads as far as we want
235 // The non-null-block should dominate the memory op, too. Live
236 // range spilling will insert a spill in the non-null-block if it is
237 // needs to spill the memory op for an implicit null check.
238 if (cb->_dom_depth == (_dom_depth + 1)) {
239 if (cb != not_null_block) continue;
240 cb = cb->_idom;
241 }
242 }
243 if( cb != this ) continue;
244
245 // Found a memory user; see if it can be hoisted to check-block
246 uint vidx = 0; // Capture index of value into memop
247 uint j;
248 for( j = mach->req()-1; j > 0; j-- ) {
249 if( mach->in(j) == val ) {
250 vidx = j;
251 // Ignore DecodeN val which could be hoisted to where needed.
252 if( is_decoden ) continue;
253 }
254 // Block of memory-op input
255 Block *inb = cfg->_bbs[mach->in(j)->_idx];
256 Block *b = this; // Start from nul check
257 while( b != inb && b->_dom_depth > inb->_dom_depth )
258 b = b->_idom; // search upwards for input
259 // See if input dominates null check
260 if( b != inb )
261 break;
262 }
263 if( j > 0 )
264 continue;
265 Block *mb = cfg->_bbs[mach->_idx];
266 // Hoisting stores requires more checks for the anti-dependence case.
267 // Give up hoisting if we have to move the store past any load.
268 if( was_store ) {
269 Block *b = mb; // Start searching here for a local load
270 // mach use (faulting) trying to hoist
271 // n might be blocker to hoisting
272 while( b != this ) {
273 uint k;
274 for( k = 1; k < b->_nodes.size(); k++ ) {
275 Node *n = b->_nodes[k];
276 if( n->needs_anti_dependence_check() &&
277 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
278 break; // Found anti-dependent load
279 }
280 if( k < b->_nodes.size() )
281 break; // Found anti-dependent load
282 // Make sure control does not do a merge (would have to check allpaths)
283 if( b->num_preds() != 2 ) break;
284 b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block
285 }
286 if( b != this ) continue;
287 }
288
289 // Make sure this memory op is not already being used for a NullCheck
290 Node *e = mb->end();
291 if( e->is_MachNullCheck() && e->in(1) == mach )
292 continue; // Already being used as a NULL check
293
294 // Found a candidate! Pick one with least dom depth - the highest
295 // in the dom tree should be closest to the null check.
296 if( !best ||
297 cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) {
298 best = mach;
299 bidx = vidx;
300
301 }
302 }
303 // No candidate!
304 if( !best ) return;
305
306 // ---- Found an implicit null check
307 extern int implicit_null_checks;
308 implicit_null_checks++;
309
310 if( is_decoden ) {
311 // Check if we need to hoist decodeHeapOop_not_null first.
312 Block *valb = cfg->_bbs[val->_idx];
313 if( this != valb && this->_dom_depth < valb->_dom_depth ) {
314 // Hoist it up to the end of the test block.
315 valb->find_remove(val);
316 this->add_inst(val);
317 cfg->_bbs.map(val->_idx,this);
318 // DecodeN on x86 may kill flags. Check for flag-killing projections
319 // that also need to be hoisted.
320 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
321 Node* n = val->fast_out(j);
322 if( n->Opcode() == Op_MachProj ) {
323 cfg->_bbs[n->_idx]->find_remove(n);
324 this->add_inst(n);
325 cfg->_bbs.map(n->_idx,this);
326 }
327 }
328 }
329 }
330 // Hoist the memory candidate up to the end of the test block.
331 Block *old_block = cfg->_bbs[best->_idx];
332 old_block->find_remove(best);
333 add_inst(best);
334 cfg->_bbs.map(best->_idx,this);
335
336 // Move the control dependence
337 if (best->in(0) && best->in(0) == old_block->_nodes[0])
338 best->set_req(0, _nodes[0]);
339
340 // Check for flag-killing projections that also need to be hoisted
341 // Should be DU safe because no edge updates.
342 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
343 Node* n = best->fast_out(j);
344 if( n->Opcode() == Op_MachProj ) {
345 cfg->_bbs[n->_idx]->find_remove(n);
346 add_inst(n);
347 cfg->_bbs.map(n->_idx,this);
348 }
349 }
350
351 Compile *C = cfg->C;
352 // proj==Op_True --> ne test; proj==Op_False --> eq test.
353 // One of two graph shapes got matched:
354 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
355 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
356 // NULL checks are always branch-if-eq. If we see a IfTrue projection
357 // then we are replacing a 'ne' test with a 'eq' NULL check test.
358 // We need to flip the projections to keep the same semantics.
359 if( proj->Opcode() == Op_IfTrue ) {
360 // Swap order of projections in basic block to swap branch targets
361 Node *tmp1 = _nodes[end_idx()+1];
362 Node *tmp2 = _nodes[end_idx()+2];
363 _nodes.map(end_idx()+1, tmp2);
364 _nodes.map(end_idx()+2, tmp1);
365 Node *tmp = new (C, 1) Node(C->top()); // Use not NULL input
366 tmp1->replace_by(tmp);
367 tmp2->replace_by(tmp1);
368 tmp->replace_by(tmp2);
369 tmp->destruct();
370 }
371
372 // Remove the existing null check; use a new implicit null check instead.
373 // Since schedule-local needs precise def-use info, we need to correct
374 // it as well.
375 Node *old_tst = proj->in(0);
376 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
377 _nodes.map(end_idx(),nul_chk);
378 cfg->_bbs.map(nul_chk->_idx,this);
379 // Redirect users of old_test to nul_chk
380 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
381 old_tst->last_out(i2)->set_req(0, nul_chk);
382 // Clean-up any dead code
383 for (uint i3 = 0; i3 < old_tst->req(); i3++)
384 old_tst->set_req(i3, NULL);
385
386 cfg->latency_from_uses(nul_chk);
387 cfg->latency_from_uses(best);
388 }
389
390
391 //------------------------------select-----------------------------------------
392 // Select a nice fellow from the worklist to schedule next. If there is only
393 // one choice, then use it. Projections take top priority for correctness
394 // reasons - if I see a projection, then it is next. There are a number of
395 // other special cases, for instructions that consume condition codes, et al.
396 // These are chosen immediately. Some instructions are required to immediately
397 // precede the last instruction in the block, and these are taken last. Of the
398 // remaining cases (most), choose the instruction with the greatest latency
399 // (that is, the most number of pseudo-cycles required to the end of the
400 // routine). If there is a tie, choose the instruction with the most inputs.
401 Node *Block::select(PhaseCFG *cfg, Node_List &worklist, int *ready_cnt, VectorSet &next_call, uint sched_slot) {
402
403 // If only a single entry on the stack, use it
404 uint cnt = worklist.size();
405 if (cnt == 1) {
406 Node *n = worklist[0];
407 worklist.map(0,worklist.pop());
408 return n;
409 }
410
411 uint choice = 0; // Bigger is most important
412 uint latency = 0; // Bigger is scheduled first
413 uint score = 0; // Bigger is better
414 int idx = -1; // Index in worklist
415
416 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
417 // Order in worklist is used to break ties.
418 // See caller for how this is used to delay scheduling
419 // of induction variable increments to after the other
420 // uses of the phi are scheduled.
421 Node *n = worklist[i]; // Get Node on worklist
422
423 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
424 if( n->is_Proj() || // Projections always win
425 n->Opcode()== Op_Con || // So does constant 'Top'
426 iop == Op_CreateEx || // Create-exception must start block
427 iop == Op_CheckCastPP
428 ) {
429 worklist.map(i,worklist.pop());
430 return n;
431 }
432
433 // Final call in a block must be adjacent to 'catch'
434 Node *e = end();
435 if( e->is_Catch() && e->in(0)->in(0) == n )
436 continue;
437
438 // Memory op for an implicit null check has to be at the end of the block
439 if( e->is_MachNullCheck() && e->in(1) == n )
440 continue;
441
442 uint n_choice = 2;
443
444 // See if this instruction is consumed by a branch. If so, then (as the
445 // branch is the last instruction in the basic block) force it to the
446 // end of the basic block
447 if ( must_clone[iop] ) {
448 // See if any use is a branch
449 bool found_machif = false;
450
451 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
452 Node* use = n->fast_out(j);
453
454 // The use is a conditional branch, make them adjacent
455 if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) {
456 found_machif = true;
457 break;
458 }
459
460 // More than this instruction pending for successor to be ready,
461 // don't choose this if other opportunities are ready
462 if (ready_cnt[use->_idx] > 1)
463 n_choice = 1;
464 }
465
466 // loop terminated, prefer not to use this instruction
467 if (found_machif)
468 continue;
469 }
470
471 // See if this has a predecessor that is "must_clone", i.e. sets the
472 // condition code. If so, choose this first
473 for (uint j = 0; j < n->req() ; j++) {
474 Node *inn = n->in(j);
475 if (inn) {
476 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
477 n_choice = 3;
478 break;
479 }
480 }
481 }
482
483 // MachTemps should be scheduled last so they are near their uses
484 if (n->is_MachTemp()) {
485 n_choice = 1;
486 }
487
488 uint n_latency = cfg->_node_latency->at_grow(n->_idx);
489 uint n_score = n->req(); // Many inputs get high score to break ties
490
491 // Keep best latency found
492 if( choice < n_choice ||
493 ( choice == n_choice &&
494 ( latency < n_latency ||
495 ( latency == n_latency &&
496 ( score < n_score ))))) {
497 choice = n_choice;
498 latency = n_latency;
499 score = n_score;
500 idx = i; // Also keep index in worklist
501 }
502 } // End of for all ready nodes in worklist
503
504 assert(idx >= 0, "index should be set");
505 Node *n = worklist[(uint)idx]; // Get the winner
506
507 worklist.map((uint)idx, worklist.pop()); // Compress worklist
508 return n;
509 }
510
511
512 //------------------------------set_next_call----------------------------------
513 void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) {
514 if( next_call.test_set(n->_idx) ) return;
515 for( uint i=0; i<n->len(); i++ ) {
516 Node *m = n->in(i);
517 if( !m ) continue; // must see all nodes in block that precede call
518 if( bbs[m->_idx] == this )
519 set_next_call( m, next_call, bbs );
520 }
521 }
522
523 //------------------------------needed_for_next_call---------------------------
524 // Set the flag 'next_call' for each Node that is needed for the next call to
525 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
526 // next subroutine call get priority - basically it moves things NOT needed
527 // for the next call till after the call. This prevents me from trying to
528 // carry lots of stuff live across a call.
529 void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) {
530 // Find the next control-defining Node in this block
531 Node* call = NULL;
532 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
533 Node* m = this_call->fast_out(i);
534 if( bbs[m->_idx] == this && // Local-block user
535 m != this_call && // Not self-start node
536 m->is_Call() )
537 call = m;
538 break;
539 }
540 if (call == NULL) return; // No next call (e.g., block end is near)
541 // Set next-call for all inputs to this call
542 set_next_call(call, next_call, bbs);
543 }
544
545 //------------------------------sched_call-------------------------------------
546 uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, int *ready_cnt, MachCallNode *mcall, VectorSet &next_call ) {
547 RegMask regs;
548
549 // Schedule all the users of the call right now. All the users are
550 // projection Nodes, so they must be scheduled next to the call.
551 // Collect all the defined registers.
552 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
553 Node* n = mcall->fast_out(i);
554 assert( n->Opcode()==Op_MachProj, "" );
555 --ready_cnt[n->_idx];
556 assert( !ready_cnt[n->_idx], "" );
557 // Schedule next to call
558 _nodes.map(node_cnt++, n);
559 // Collect defined registers
560 regs.OR(n->out_RegMask());
561 // Check for scheduling the next control-definer
562 if( n->bottom_type() == Type::CONTROL )
563 // Warm up next pile of heuristic bits
564 needed_for_next_call(n, next_call, bbs);
565
566 // Children of projections are now all ready
567 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
568 Node* m = n->fast_out(j); // Get user
569 if( bbs[m->_idx] != this ) continue;
570 if( m->is_Phi() ) continue;
571 if( !--ready_cnt[m->_idx] )
572 worklist.push(m);
573 }
574
575 }
576
577 // Act as if the call defines the Frame Pointer.
578 // Certainly the FP is alive and well after the call.
579 regs.Insert(matcher.c_frame_pointer());
580
581 // Set all registers killed and not already defined by the call.
582 uint r_cnt = mcall->tf()->range()->cnt();
583 int op = mcall->ideal_Opcode();
584 MachProjNode *proj = new (matcher.C, 1) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
585 bbs.map(proj->_idx,this);
586 _nodes.insert(node_cnt++, proj);
587
588 // Select the right register save policy.
589 const char * save_policy;
590 switch (op) {
591 case Op_CallRuntime:
592 case Op_CallLeaf:
593 case Op_CallLeafNoFP:
594 // Calling C code so use C calling convention
595 save_policy = matcher._c_reg_save_policy;
596 break;
597
598 case Op_CallStaticJava:
599 case Op_CallDynamicJava:
600 // Calling Java code so use Java calling convention
601 save_policy = matcher._register_save_policy;
602 break;
603
604 default:
605 ShouldNotReachHere();
606 }
607
608 // When using CallRuntime mark SOE registers as killed by the call
609 // so values that could show up in the RegisterMap aren't live in a
610 // callee saved register since the register wouldn't know where to
611 // find them. CallLeaf and CallLeafNoFP are ok because they can't
612 // have debug info on them. Strictly speaking this only needs to be
613 // done for oops since idealreg2debugmask takes care of debug info
614 // references but there no way to handle oops differently than other
615 // pointers as far as the kill mask goes.
616 bool exclude_soe = op == Op_CallRuntime;
617
618 // If the call is a MethodHandle invoke, we need to exclude the
619 // register which is used to save the SP value over MH invokes from
620 // the mask. Otherwise this register could be used for
621 // deoptimization information.
622 if (op == Op_CallStaticJava) {
623 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
624 if (mcallstaticjava->_method_handle_invoke)
625 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
626 }
627
628 // Fill in the kill mask for the call
629 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
630 if( !regs.Member(r) ) { // Not already defined by the call
631 // Save-on-call register?
632 if ((save_policy[r] == 'C') ||
633 (save_policy[r] == 'A') ||
634 ((save_policy[r] == 'E') && exclude_soe)) {
635 proj->_rout.Insert(r);
636 }
637 }
638 }
639
640 return node_cnt;
641 }
642
643
644 //------------------------------schedule_local---------------------------------
645 // Topological sort within a block. Someday become a real scheduler.
646 bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, int *ready_cnt, VectorSet &next_call) {
647 // Already "sorted" are the block start Node (as the first entry), and
648 // the block-ending Node and any trailing control projections. We leave
649 // these alone. PhiNodes and ParmNodes are made to follow the block start
650 // Node. Everything else gets topo-sorted.
651
652 #ifndef PRODUCT
653 if (cfg->trace_opto_pipelining()) {
654 tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order);
655 for (uint i = 0;i < _nodes.size();i++) {
656 tty->print("# ");
657 _nodes[i]->fast_dump();
658 }
659 tty->print_cr("#");
660 }
661 #endif
662
663 // RootNode is already sorted
664 if( _nodes.size() == 1 ) return true;
665
666 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
667 uint node_cnt = end_idx();
668 uint phi_cnt = 1;
669 uint i;
670 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
671 Node *n = _nodes[i];
672 if( n->is_Phi() || // Found a PhiNode or ParmNode
673 (n->is_Proj() && n->in(0) == head()) ) {
674 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
675 _nodes.map(i,_nodes[phi_cnt]);
676 _nodes.map(phi_cnt++,n); // swap Phi/Parm up front
677 } else { // All others
678 // Count block-local inputs to 'n'
679 uint cnt = n->len(); // Input count
680 uint local = 0;
681 for( uint j=0; j<cnt; j++ ) {
682 Node *m = n->in(j);
683 if( m && cfg->_bbs[m->_idx] == this && !m->is_top() )
684 local++; // One more block-local input
685 }
686 ready_cnt[n->_idx] = local; // Count em up
687
688 // A few node types require changing a required edge to a precedence edge
689 // before allocation.
690 if( UseConcMarkSweepGC || UseG1GC ) {
691 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
692 // Note: Required edges with an index greater than oper_input_base
693 // are not supported by the allocator.
694 // Note2: Can only depend on unmatched edge being last,
695 // can not depend on its absolute position.
696 Node *oop_store = n->in(n->req() - 1);
697 n->del_req(n->req() - 1);
698 n->add_prec(oop_store);
699 assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark");
700 }
701 }
702 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
703 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
704 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
705 // MemBarAcquire could be created without Precedent edge.
706 // del_req() replaces the specified edge with the last input edge
707 // and then removes the last edge. If the specified edge > number of
708 // edges the last edge will be moved outside of the input edges array
709 // and the edge will be lost. This is why this code should be
710 // executed only when Precedent (== TypeFunc::Parms) edge is present.
711 Node *x = n->in(TypeFunc::Parms);
712 n->del_req(TypeFunc::Parms);
713 n->add_prec(x);
714 }
715 }
716 }
717 for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count
718 ready_cnt[_nodes[i2]->_idx] = 0;
719
720 // All the prescheduled guys do not hold back internal nodes
721 uint i3;
722 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
723 Node *n = _nodes[i3]; // Get pre-scheduled
724 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
725 Node* m = n->fast_out(j);
726 if( cfg->_bbs[m->_idx] ==this ) // Local-block user
727 ready_cnt[m->_idx]--; // Fix ready count
728 }
729 }
730
731 Node_List delay;
732 // Make a worklist
733 Node_List worklist;
734 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
735 Node *m = _nodes[i4];
736 if( !ready_cnt[m->_idx] ) { // Zero ready count?
737 if (m->is_iteratively_computed()) {
738 // Push induction variable increments last to allow other uses
739 // of the phi to be scheduled first. The select() method breaks
740 // ties in scheduling by worklist order.
741 delay.push(m);
742 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
743 // Force the CreateEx to the top of the list so it's processed
744 // first and ends up at the start of the block.
745 worklist.insert(0, m);
746 } else {
747 worklist.push(m); // Then on to worklist!
748 }
749 }
750 }
751 while (delay.size()) {
752 Node* d = delay.pop();
753 worklist.push(d);
754 }
755
756 // Warm up the 'next_call' heuristic bits
757 needed_for_next_call(_nodes[0], next_call, cfg->_bbs);
758
759 #ifndef PRODUCT
760 if (cfg->trace_opto_pipelining()) {
761 for (uint j=0; j<_nodes.size(); j++) {
762 Node *n = _nodes[j];
763 int idx = n->_idx;
764 tty->print("# ready cnt:%3d ", ready_cnt[idx]);
765 tty->print("latency:%3d ", cfg->_node_latency->at_grow(idx));
766 tty->print("%4d: %s\n", idx, n->Name());
767 }
768 }
769 #endif
770
771 // Pull from worklist and schedule
772 while( worklist.size() ) { // Worklist is not ready
773
774 #ifndef PRODUCT
775 if (cfg->trace_opto_pipelining()) {
776 tty->print("# ready list:");
777 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
778 Node *n = worklist[i]; // Get Node on worklist
779 tty->print(" %d", n->_idx);
780 }
781 tty->cr();
782 }
783 #endif
784
785 // Select and pop a ready guy from worklist
786 Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt);
787 _nodes.map(phi_cnt++,n); // Schedule him next
788
789 #ifndef PRODUCT
790 if (cfg->trace_opto_pipelining()) {
791 tty->print("# select %d: %s", n->_idx, n->Name());
792 tty->print(", latency:%d", cfg->_node_latency->at_grow(n->_idx));
793 n->dump();
794 if (Verbose) {
795 tty->print("# ready list:");
796 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
797 Node *n = worklist[i]; // Get Node on worklist
798 tty->print(" %d", n->_idx);
799 }
800 tty->cr();
801 }
802 }
803
804 #endif
805 if( n->is_MachCall() ) {
806 MachCallNode *mcall = n->as_MachCall();
807 phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call);
808 continue;
809 }
810 // Children are now all ready
811 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
812 Node* m = n->fast_out(i5); // Get user
813 if( cfg->_bbs[m->_idx] != this ) continue;
814 if( m->is_Phi() ) continue;
815 if( !--ready_cnt[m->_idx] )
816 worklist.push(m);
817 }
818 }
819
820 if( phi_cnt != end_idx() ) {
821 // did not schedule all. Retry, Bailout, or Die
822 Compile* C = matcher.C;
823 if (C->subsume_loads() == true && !C->failing()) {
824 // Retry with subsume_loads == false
825 // If this is the first failure, the sentinel string will "stick"
826 // to the Compile object, and the C2Compiler will see it and retry.
827 C->record_failure(C2Compiler::retry_no_subsuming_loads());
828 }
829 // assert( phi_cnt == end_idx(), "did not schedule all" );
830 return false;
831 }
832
833 #ifndef PRODUCT
834 if (cfg->trace_opto_pipelining()) {
835 tty->print_cr("#");
836 tty->print_cr("# after schedule_local");
837 for (uint i = 0;i < _nodes.size();i++) {
838 tty->print("# ");
839 _nodes[i]->fast_dump();
840 }
841 tty->cr();
842 }
843 #endif
844
845
846 return true;
847 }
848
849 //--------------------------catch_cleanup_fix_all_inputs-----------------------
850 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
851 for (uint l = 0; l < use->len(); l++) {
852 if (use->in(l) == old_def) {
853 if (l < use->req()) {
854 use->set_req(l, new_def);
855 } else {
856 use->rm_prec(l);
857 use->add_prec(new_def);
858 l--;
859 }
860 }
861 }
862 }
863
864 //------------------------------catch_cleanup_find_cloned_def------------------
865 static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
866 assert( use_blk != def_blk, "Inter-block cleanup only");
867
868 // The use is some block below the Catch. Find and return the clone of the def
869 // that dominates the use. If there is no clone in a dominating block, then
870 // create a phi for the def in a dominating block.
871
872 // Find which successor block dominates this use. The successor
873 // blocks must all be single-entry (from the Catch only; I will have
874 // split blocks to make this so), hence they all dominate.
875 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
876 use_blk = use_blk->_idom;
877
878 // Find the successor
879 Node *fixup = NULL;
880
881 uint j;
882 for( j = 0; j < def_blk->_num_succs; j++ )
883 if( use_blk == def_blk->_succs[j] )
884 break;
885
886 if( j == def_blk->_num_succs ) {
887 // Block at same level in dom-tree is not a successor. It needs a
888 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
889 Node_Array inputs = new Node_List(Thread::current()->resource_area());
890 for(uint k = 1; k < use_blk->num_preds(); k++) {
891 inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx));
892 }
893
894 // Check to see if the use_blk already has an identical phi inserted.
895 // If it exists, it will be at the first position since all uses of a
896 // def are processed together.
897 Node *phi = use_blk->_nodes[1];
898 if( phi->is_Phi() ) {
899 fixup = phi;
900 for (uint k = 1; k < use_blk->num_preds(); k++) {
901 if (phi->in(k) != inputs[k]) {
902 // Not a match
903 fixup = NULL;
904 break;
905 }
906 }
907 }
908
909 // If an existing PhiNode was not found, make a new one.
910 if (fixup == NULL) {
911 Node *new_phi = PhiNode::make(use_blk->head(), def);
912 use_blk->_nodes.insert(1, new_phi);
913 bbs.map(new_phi->_idx, use_blk);
914 for (uint k = 1; k < use_blk->num_preds(); k++) {
915 new_phi->set_req(k, inputs[k]);
916 }
917 fixup = new_phi;
918 }
919
920 } else {
921 // Found the use just below the Catch. Make it use the clone.
922 fixup = use_blk->_nodes[n_clone_idx];
923 }
924
925 return fixup;
926 }
927
928 //--------------------------catch_cleanup_intra_block--------------------------
929 // Fix all input edges in use that reference "def". The use is in the same
930 // block as the def and both have been cloned in each successor block.
931 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
932
933 // Both the use and def have been cloned. For each successor block,
934 // get the clone of the use, and make its input the clone of the def
935 // found in that block.
936
937 uint use_idx = blk->find_node(use);
938 uint offset_idx = use_idx - beg;
939 for( uint k = 0; k < blk->_num_succs; k++ ) {
940 // Get clone in each successor block
941 Block *sb = blk->_succs[k];
942 Node *clone = sb->_nodes[offset_idx+1];
943 assert( clone->Opcode() == use->Opcode(), "" );
944
945 // Make use-clone reference the def-clone
946 catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]);
947 }
948 }
949
950 //------------------------------catch_cleanup_inter_block---------------------
951 // Fix all input edges in use that reference "def". The use is in a different
952 // block than the def.
953 static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
954 if( !use_blk ) return; // Can happen if the use is a precedence edge
955
956 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx);
957 catch_cleanup_fix_all_inputs(use, def, new_def);
958 }
959
960 //------------------------------call_catch_cleanup-----------------------------
961 // If we inserted any instructions between a Call and his CatchNode,
962 // clone the instructions on all paths below the Catch.
963 void Block::call_catch_cleanup(Block_Array &bbs) {
964
965 // End of region to clone
966 uint end = end_idx();
967 if( !_nodes[end]->is_Catch() ) return;
968 // Start of region to clone
969 uint beg = end;
970 while( _nodes[beg-1]->Opcode() != Op_MachProj ||
971 !_nodes[beg-1]->in(0)->is_Call() ) {
972 beg--;
973 assert(beg > 0,"Catch cleanup walking beyond block boundary");
974 }
975 // Range of inserted instructions is [beg, end)
976 if( beg == end ) return;
977
978 // Clone along all Catch output paths. Clone area between the 'beg' and
979 // 'end' indices.
980 for( uint i = 0; i < _num_succs; i++ ) {
981 Block *sb = _succs[i];
982 // Clone the entire area; ignoring the edge fixup for now.
983 for( uint j = end; j > beg; j-- ) {
984 // It is safe here to clone a node with anti_dependence
985 // since clones dominate on each path.
986 Node *clone = _nodes[j-1]->clone();
987 sb->_nodes.insert( 1, clone );
988 bbs.map(clone->_idx,sb);
989 }
990 }
991
992
993 // Fixup edges. Check the def-use info per cloned Node
994 for(uint i2 = beg; i2 < end; i2++ ) {
995 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
996 Node *n = _nodes[i2]; // Node that got cloned
997 // Need DU safe iterator because of edge manipulation in calls.
998 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
999 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1000 out->push(n->fast_out(j1));
1001 }
1002 uint max = out->size();
1003 for (uint j = 0; j < max; j++) {// For all users
1004 Node *use = out->pop();
1005 Block *buse = bbs[use->_idx];
1006 if( use->is_Phi() ) {
1007 for( uint k = 1; k < use->req(); k++ )
1008 if( use->in(k) == n ) {
1009 Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx);
1010 use->set_req(k, fixup);
1011 }
1012 } else {
1013 if (this == buse) {
1014 catch_cleanup_intra_block(use, n, this, beg, n_clone_idx);
1015 } else {
1016 catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx);
1017 }
1018 }
1019 } // End for all users
1020
1021 } // End of for all Nodes in cloned area
1022
1023 // Remove the now-dead cloned ops
1024 for(uint i3 = beg; i3 < end; i3++ ) {
1025 _nodes[beg]->disconnect_inputs(NULL);
1026 _nodes.remove(beg);
1027 }
1028
1029 // If the successor blocks have a CreateEx node, move it back to the top
1030 for(uint i4 = 0; i4 < _num_succs; i4++ ) {
1031 Block *sb = _succs[i4];
1032 uint new_cnt = end - beg;
1033 // Remove any newly created, but dead, nodes.
1034 for( uint j = new_cnt; j > 0; j-- ) {
1035 Node *n = sb->_nodes[j];
1036 if (n->outcnt() == 0 &&
1037 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1038 n->disconnect_inputs(NULL);
1039 sb->_nodes.remove(j);
1040 new_cnt--;
1041 }
1042 }
1043 // If any newly created nodes remain, move the CreateEx node to the top
1044 if (new_cnt > 0) {
1045 Node *cex = sb->_nodes[1+new_cnt];
1046 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1047 sb->_nodes.remove(1+new_cnt);
1048 sb->_nodes.insert(1,cex);
1049 }
1050 }
1051 }
1052 }