Thu, 21 Nov 2013 12:30:35 -0800
Merge
1 /*
2 * Copyright (c) 1998, 2012, 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 //------------------------------implicit_null_check----------------------------
58 // Detect implicit-null-check opportunities. Basically, find NULL checks
59 // with suitable memory ops nearby. Use the memory op to do the NULL check.
60 // I can generate a memory op if there is not one nearby.
61 // The proj is the control projection for the not-null case.
62 // The val is the pointer being checked for nullness or
63 // decodeHeapOop_not_null node if it did not fold into address.
64 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) {
65 // Assume if null check need for 0 offset then always needed
66 // Intel solaris doesn't support any null checks yet and no
67 // mechanism exists (yet) to set the switches at an os_cpu level
68 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
70 // Make sure the ptr-is-null path appears to be uncommon!
71 float f = block->end()->as_MachIf()->_prob;
72 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
73 if( f > PROB_UNLIKELY_MAG(4) ) return;
75 uint bidx = 0; // Capture index of value into memop
76 bool was_store; // Memory op is a store op
78 // Get the successor block for if the test ptr is non-null
79 Block* not_null_block; // this one goes with the proj
80 Block* null_block;
81 if (block->get_node(block->number_of_nodes()-1) == proj) {
82 null_block = block->_succs[0];
83 not_null_block = block->_succs[1];
84 } else {
85 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other");
86 not_null_block = block->_succs[0];
87 null_block = block->_succs[1];
88 }
89 while (null_block->is_Empty() == Block::empty_with_goto) {
90 null_block = null_block->_succs[0];
91 }
93 // Search the exception block for an uncommon trap.
94 // (See Parse::do_if and Parse::do_ifnull for the reason
95 // we need an uncommon trap. Briefly, we need a way to
96 // detect failure of this optimization, as in 6366351.)
97 {
98 bool found_trap = false;
99 for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) {
100 Node* nn = null_block->get_node(i1);
101 if (nn->is_MachCall() &&
102 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
103 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
104 if (trtype->isa_int() && trtype->is_int()->is_con()) {
105 jint tr_con = trtype->is_int()->get_con();
106 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
107 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
108 assert((int)reason < (int)BitsPerInt, "recode bit map");
109 if (is_set_nth_bit(allowed_reasons, (int) reason)
110 && action != Deoptimization::Action_none) {
111 // This uncommon trap is sure to recompile, eventually.
112 // When that happens, C->too_many_traps will prevent
113 // this transformation from happening again.
114 found_trap = true;
115 }
116 }
117 break;
118 }
119 }
120 if (!found_trap) {
121 // We did not find an uncommon trap.
122 return;
123 }
124 }
126 // Check for decodeHeapOop_not_null node which did not fold into address
127 bool is_decoden = ((intptr_t)val) & 1;
128 val = (Node*)(((intptr_t)val) & ~1);
130 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() &&
131 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity");
133 // Search the successor block for a load or store who's base value is also
134 // the tested value. There may be several.
135 Node_List *out = new Node_List(Thread::current()->resource_area());
136 MachNode *best = NULL; // Best found so far
137 for (DUIterator i = val->outs(); val->has_out(i); i++) {
138 Node *m = val->out(i);
139 if( !m->is_Mach() ) continue;
140 MachNode *mach = m->as_Mach();
141 was_store = false;
142 int iop = mach->ideal_Opcode();
143 switch( iop ) {
144 case Op_LoadB:
145 case Op_LoadUB:
146 case Op_LoadUS:
147 case Op_LoadD:
148 case Op_LoadF:
149 case Op_LoadI:
150 case Op_LoadL:
151 case Op_LoadP:
152 case Op_LoadN:
153 case Op_LoadS:
154 case Op_LoadKlass:
155 case Op_LoadNKlass:
156 case Op_LoadRange:
157 case Op_LoadD_unaligned:
158 case Op_LoadL_unaligned:
159 assert(mach->in(2) == val, "should be address");
160 break;
161 case Op_StoreB:
162 case Op_StoreC:
163 case Op_StoreCM:
164 case Op_StoreD:
165 case Op_StoreF:
166 case Op_StoreI:
167 case Op_StoreL:
168 case Op_StoreP:
169 case Op_StoreN:
170 case Op_StoreNKlass:
171 was_store = true; // Memory op is a store op
172 // Stores will have their address in slot 2 (memory in slot 1).
173 // If the value being nul-checked is in another slot, it means we
174 // are storing the checked value, which does NOT check the value!
175 if( mach->in(2) != val ) continue;
176 break; // Found a memory op?
177 case Op_StrComp:
178 case Op_StrEquals:
179 case Op_StrIndexOf:
180 case Op_AryEq:
181 case Op_EncodeISOArray:
182 // Not a legit memory op for implicit null check regardless of
183 // embedded loads
184 continue;
185 default: // Also check for embedded loads
186 if( !mach->needs_anti_dependence_check() )
187 continue; // Not an memory op; skip it
188 if( must_clone[iop] ) {
189 // Do not move nodes which produce flags because
190 // RA will try to clone it to place near branch and
191 // it will cause recompilation, see clone_node().
192 continue;
193 }
194 {
195 // Check that value is used in memory address in
196 // instructions with embedded load (CmpP val1,(val2+off)).
197 Node* base;
198 Node* index;
199 const MachOper* oper = mach->memory_inputs(base, index);
200 if (oper == NULL || oper == (MachOper*)-1) {
201 continue; // Not an memory op; skip it
202 }
203 if (val == base ||
204 val == index && val->bottom_type()->isa_narrowoop()) {
205 break; // Found it
206 } else {
207 continue; // Skip it
208 }
209 }
210 break;
211 }
212 // check if the offset is not too high for implicit exception
213 {
214 intptr_t offset = 0;
215 const TypePtr *adr_type = NULL; // Do not need this return value here
216 const Node* base = mach->get_base_and_disp(offset, adr_type);
217 if (base == NULL || base == NodeSentinel) {
218 // Narrow oop address doesn't have base, only index
219 if( val->bottom_type()->isa_narrowoop() &&
220 MacroAssembler::needs_explicit_null_check(offset) )
221 continue; // Give up if offset is beyond page size
222 // cannot reason about it; is probably not implicit null exception
223 } else {
224 const TypePtr* tptr;
225 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 ||
226 Universe::narrow_klass_shift() == 0)) {
227 // 32-bits narrow oop can be the base of address expressions
228 tptr = base->get_ptr_type();
229 } else {
230 // only regular oops are expected here
231 tptr = base->bottom_type()->is_ptr();
232 }
233 // Give up if offset is not a compile-time constant
234 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
235 continue;
236 offset += tptr->_offset; // correct if base is offseted
237 if( MacroAssembler::needs_explicit_null_check(offset) )
238 continue; // Give up is reference is beyond 4K page size
239 }
240 }
242 // Check ctrl input to see if the null-check dominates the memory op
243 Block *cb = get_block_for_node(mach);
244 cb = cb->_idom; // Always hoist at least 1 block
245 if( !was_store ) { // Stores can be hoisted only one block
246 while( cb->_dom_depth > (block->_dom_depth + 1))
247 cb = cb->_idom; // Hoist loads as far as we want
248 // The non-null-block should dominate the memory op, too. Live
249 // range spilling will insert a spill in the non-null-block if it is
250 // needs to spill the memory op for an implicit null check.
251 if (cb->_dom_depth == (block->_dom_depth + 1)) {
252 if (cb != not_null_block) continue;
253 cb = cb->_idom;
254 }
255 }
256 if( cb != block ) continue;
258 // Found a memory user; see if it can be hoisted to check-block
259 uint vidx = 0; // Capture index of value into memop
260 uint j;
261 for( j = mach->req()-1; j > 0; j-- ) {
262 if( mach->in(j) == val ) {
263 vidx = j;
264 // Ignore DecodeN val which could be hoisted to where needed.
265 if( is_decoden ) continue;
266 }
267 // Block of memory-op input
268 Block *inb = get_block_for_node(mach->in(j));
269 Block *b = block; // Start from nul check
270 while( b != inb && b->_dom_depth > inb->_dom_depth )
271 b = b->_idom; // search upwards for input
272 // See if input dominates null check
273 if( b != inb )
274 break;
275 }
276 if( j > 0 )
277 continue;
278 Block *mb = get_block_for_node(mach);
279 // Hoisting stores requires more checks for the anti-dependence case.
280 // Give up hoisting if we have to move the store past any load.
281 if( was_store ) {
282 Block *b = mb; // Start searching here for a local load
283 // mach use (faulting) trying to hoist
284 // n might be blocker to hoisting
285 while( b != block ) {
286 uint k;
287 for( k = 1; k < b->number_of_nodes(); k++ ) {
288 Node *n = b->get_node(k);
289 if( n->needs_anti_dependence_check() &&
290 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
291 break; // Found anti-dependent load
292 }
293 if( k < b->number_of_nodes() )
294 break; // Found anti-dependent load
295 // Make sure control does not do a merge (would have to check allpaths)
296 if( b->num_preds() != 2 ) break;
297 b = get_block_for_node(b->pred(1)); // Move up to predecessor block
298 }
299 if( b != block ) continue;
300 }
302 // Make sure this memory op is not already being used for a NullCheck
303 Node *e = mb->end();
304 if( e->is_MachNullCheck() && e->in(1) == mach )
305 continue; // Already being used as a NULL check
307 // Found a candidate! Pick one with least dom depth - the highest
308 // in the dom tree should be closest to the null check.
309 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) {
310 best = mach;
311 bidx = vidx;
312 }
313 }
314 // No candidate!
315 if (best == NULL) {
316 return;
317 }
319 // ---- Found an implicit null check
320 extern int implicit_null_checks;
321 implicit_null_checks++;
323 if( is_decoden ) {
324 // Check if we need to hoist decodeHeapOop_not_null first.
325 Block *valb = get_block_for_node(val);
326 if( block != valb && block->_dom_depth < valb->_dom_depth ) {
327 // Hoist it up to the end of the test block.
328 valb->find_remove(val);
329 block->add_inst(val);
330 map_node_to_block(val, block);
331 // DecodeN on x86 may kill flags. Check for flag-killing projections
332 // that also need to be hoisted.
333 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
334 Node* n = val->fast_out(j);
335 if( n->is_MachProj() ) {
336 get_block_for_node(n)->find_remove(n);
337 block->add_inst(n);
338 map_node_to_block(n, block);
339 }
340 }
341 }
342 }
343 // Hoist the memory candidate up to the end of the test block.
344 Block *old_block = get_block_for_node(best);
345 old_block->find_remove(best);
346 block->add_inst(best);
347 map_node_to_block(best, block);
349 // Move the control dependence
350 if (best->in(0) && best->in(0) == old_block->head())
351 best->set_req(0, block->head());
353 // Check for flag-killing projections that also need to be hoisted
354 // Should be DU safe because no edge updates.
355 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
356 Node* n = best->fast_out(j);
357 if( n->is_MachProj() ) {
358 get_block_for_node(n)->find_remove(n);
359 block->add_inst(n);
360 map_node_to_block(n, block);
361 }
362 }
364 // proj==Op_True --> ne test; proj==Op_False --> eq test.
365 // One of two graph shapes got matched:
366 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
367 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
368 // NULL checks are always branch-if-eq. If we see a IfTrue projection
369 // then we are replacing a 'ne' test with a 'eq' NULL check test.
370 // We need to flip the projections to keep the same semantics.
371 if( proj->Opcode() == Op_IfTrue ) {
372 // Swap order of projections in basic block to swap branch targets
373 Node *tmp1 = block->get_node(block->end_idx()+1);
374 Node *tmp2 = block->get_node(block->end_idx()+2);
375 block->map_node(tmp2, block->end_idx()+1);
376 block->map_node(tmp1, block->end_idx()+2);
377 Node *tmp = new (C) Node(C->top()); // Use not NULL input
378 tmp1->replace_by(tmp);
379 tmp2->replace_by(tmp1);
380 tmp->replace_by(tmp2);
381 tmp->destruct();
382 }
384 // Remove the existing null check; use a new implicit null check instead.
385 // Since schedule-local needs precise def-use info, we need to correct
386 // it as well.
387 Node *old_tst = proj->in(0);
388 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
389 block->map_node(nul_chk, block->end_idx());
390 map_node_to_block(nul_chk, block);
391 // Redirect users of old_test to nul_chk
392 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
393 old_tst->last_out(i2)->set_req(0, nul_chk);
394 // Clean-up any dead code
395 for (uint i3 = 0; i3 < old_tst->req(); i3++)
396 old_tst->set_req(i3, NULL);
398 latency_from_uses(nul_chk);
399 latency_from_uses(best);
400 }
403 //------------------------------select-----------------------------------------
404 // Select a nice fellow from the worklist to schedule next. If there is only
405 // one choice, then use it. Projections take top priority for correctness
406 // reasons - if I see a projection, then it is next. There are a number of
407 // other special cases, for instructions that consume condition codes, et al.
408 // These are chosen immediately. Some instructions are required to immediately
409 // precede the last instruction in the block, and these are taken last. Of the
410 // remaining cases (most), choose the instruction with the greatest latency
411 // (that is, the most number of pseudo-cycles required to the end of the
412 // routine). If there is a tie, choose the instruction with the most inputs.
413 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) {
415 // If only a single entry on the stack, use it
416 uint cnt = worklist.size();
417 if (cnt == 1) {
418 Node *n = worklist[0];
419 worklist.map(0,worklist.pop());
420 return n;
421 }
423 uint choice = 0; // Bigger is most important
424 uint latency = 0; // Bigger is scheduled first
425 uint score = 0; // Bigger is better
426 int idx = -1; // Index in worklist
427 int cand_cnt = 0; // Candidate count
429 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
430 // Order in worklist is used to break ties.
431 // See caller for how this is used to delay scheduling
432 // of induction variable increments to after the other
433 // uses of the phi are scheduled.
434 Node *n = worklist[i]; // Get Node on worklist
436 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
437 if( n->is_Proj() || // Projections always win
438 n->Opcode()== Op_Con || // So does constant 'Top'
439 iop == Op_CreateEx || // Create-exception must start block
440 iop == Op_CheckCastPP
441 ) {
442 worklist.map(i,worklist.pop());
443 return n;
444 }
446 // Final call in a block must be adjacent to 'catch'
447 Node *e = block->end();
448 if( e->is_Catch() && e->in(0)->in(0) == n )
449 continue;
451 // Memory op for an implicit null check has to be at the end of the block
452 if( e->is_MachNullCheck() && e->in(1) == n )
453 continue;
455 // Schedule IV increment last.
456 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd &&
457 e->in(1)->in(1) == n && n->is_iteratively_computed())
458 continue;
460 uint n_choice = 2;
462 // See if this instruction is consumed by a branch. If so, then (as the
463 // branch is the last instruction in the basic block) force it to the
464 // end of the basic block
465 if ( must_clone[iop] ) {
466 // See if any use is a branch
467 bool found_machif = false;
469 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
470 Node* use = n->fast_out(j);
472 // The use is a conditional branch, make them adjacent
473 if (use->is_MachIf() && get_block_for_node(use) == block) {
474 found_machif = true;
475 break;
476 }
478 // For nodes that produce a FlagsProj, make the node adjacent to the
479 // use of the FlagsProj
480 if (use->is_FlagsProj() && get_block_for_node(use) == block) {
481 found_machif = true;
482 break;
483 }
485 // More than this instruction pending for successor to be ready,
486 // don't choose this if other opportunities are ready
487 if (ready_cnt.at(use->_idx) > 1)
488 n_choice = 1;
489 }
491 // loop terminated, prefer not to use this instruction
492 if (found_machif)
493 continue;
494 }
496 // See if this has a predecessor that is "must_clone", i.e. sets the
497 // condition code. If so, choose this first
498 for (uint j = 0; j < n->req() ; j++) {
499 Node *inn = n->in(j);
500 if (inn) {
501 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
502 n_choice = 3;
503 break;
504 }
505 }
506 }
508 // MachTemps should be scheduled last so they are near their uses
509 if (n->is_MachTemp()) {
510 n_choice = 1;
511 }
513 uint n_latency = get_latency_for_node(n);
514 uint n_score = n->req(); // Many inputs get high score to break ties
516 // Keep best latency found
517 cand_cnt++;
518 if (choice < n_choice ||
519 (choice == n_choice &&
520 ((StressLCM && Compile::randomized_select(cand_cnt)) ||
521 (!StressLCM &&
522 (latency < n_latency ||
523 (latency == n_latency &&
524 (score < n_score))))))) {
525 choice = n_choice;
526 latency = n_latency;
527 score = n_score;
528 idx = i; // Also keep index in worklist
529 }
530 } // End of for all ready nodes in worklist
532 assert(idx >= 0, "index should be set");
533 Node *n = worklist[(uint)idx]; // Get the winner
535 worklist.map((uint)idx, worklist.pop()); // Compress worklist
536 return n;
537 }
540 //------------------------------set_next_call----------------------------------
541 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) {
542 if( next_call.test_set(n->_idx) ) return;
543 for( uint i=0; i<n->len(); i++ ) {
544 Node *m = n->in(i);
545 if( !m ) continue; // must see all nodes in block that precede call
546 if (get_block_for_node(m) == block) {
547 set_next_call(block, m, next_call);
548 }
549 }
550 }
552 //------------------------------needed_for_next_call---------------------------
553 // Set the flag 'next_call' for each Node that is needed for the next call to
554 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
555 // next subroutine call get priority - basically it moves things NOT needed
556 // for the next call till after the call. This prevents me from trying to
557 // carry lots of stuff live across a call.
558 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) {
559 // Find the next control-defining Node in this block
560 Node* call = NULL;
561 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
562 Node* m = this_call->fast_out(i);
563 if (get_block_for_node(m) == block && // Local-block user
564 m != this_call && // Not self-start node
565 m->is_MachCall()) {
566 call = m;
567 break;
568 }
569 }
570 if (call == NULL) return; // No next call (e.g., block end is near)
571 // Set next-call for all inputs to this call
572 set_next_call(block, call, next_call);
573 }
575 //------------------------------add_call_kills-------------------------------------
576 // helper function that adds caller save registers to MachProjNode
577 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
578 // Fill in the kill mask for the call
579 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
580 if( !regs.Member(r) ) { // Not already defined by the call
581 // Save-on-call register?
582 if ((save_policy[r] == 'C') ||
583 (save_policy[r] == 'A') ||
584 ((save_policy[r] == 'E') && exclude_soe)) {
585 proj->_rout.Insert(r);
586 }
587 }
588 }
589 }
592 //------------------------------sched_call-------------------------------------
593 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) {
594 RegMask regs;
596 // Schedule all the users of the call right now. All the users are
597 // projection Nodes, so they must be scheduled next to the call.
598 // Collect all the defined registers.
599 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
600 Node* n = mcall->fast_out(i);
601 assert( n->is_MachProj(), "" );
602 int n_cnt = ready_cnt.at(n->_idx)-1;
603 ready_cnt.at_put(n->_idx, n_cnt);
604 assert( n_cnt == 0, "" );
605 // Schedule next to call
606 block->map_node(n, node_cnt++);
607 // Collect defined registers
608 regs.OR(n->out_RegMask());
609 // Check for scheduling the next control-definer
610 if( n->bottom_type() == Type::CONTROL )
611 // Warm up next pile of heuristic bits
612 needed_for_next_call(block, n, next_call);
614 // Children of projections are now all ready
615 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
616 Node* m = n->fast_out(j); // Get user
617 if(get_block_for_node(m) != block) {
618 continue;
619 }
620 if( m->is_Phi() ) continue;
621 int m_cnt = ready_cnt.at(m->_idx)-1;
622 ready_cnt.at_put(m->_idx, m_cnt);
623 if( m_cnt == 0 )
624 worklist.push(m);
625 }
627 }
629 // Act as if the call defines the Frame Pointer.
630 // Certainly the FP is alive and well after the call.
631 regs.Insert(_matcher.c_frame_pointer());
633 // Set all registers killed and not already defined by the call.
634 uint r_cnt = mcall->tf()->range()->cnt();
635 int op = mcall->ideal_Opcode();
636 MachProjNode *proj = new (C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
637 map_node_to_block(proj, block);
638 block->insert_node(proj, node_cnt++);
640 // Select the right register save policy.
641 const char * save_policy;
642 switch (op) {
643 case Op_CallRuntime:
644 case Op_CallLeaf:
645 case Op_CallLeafNoFP:
646 // Calling C code so use C calling convention
647 save_policy = _matcher._c_reg_save_policy;
648 break;
650 case Op_CallStaticJava:
651 case Op_CallDynamicJava:
652 // Calling Java code so use Java calling convention
653 save_policy = _matcher._register_save_policy;
654 break;
656 default:
657 ShouldNotReachHere();
658 }
660 // When using CallRuntime mark SOE registers as killed by the call
661 // so values that could show up in the RegisterMap aren't live in a
662 // callee saved register since the register wouldn't know where to
663 // find them. CallLeaf and CallLeafNoFP are ok because they can't
664 // have debug info on them. Strictly speaking this only needs to be
665 // done for oops since idealreg2debugmask takes care of debug info
666 // references but there no way to handle oops differently than other
667 // pointers as far as the kill mask goes.
668 bool exclude_soe = op == Op_CallRuntime;
670 // If the call is a MethodHandle invoke, we need to exclude the
671 // register which is used to save the SP value over MH invokes from
672 // the mask. Otherwise this register could be used for
673 // deoptimization information.
674 if (op == Op_CallStaticJava) {
675 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
676 if (mcallstaticjava->_method_handle_invoke)
677 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
678 }
680 add_call_kills(proj, regs, save_policy, exclude_soe);
682 return node_cnt;
683 }
686 //------------------------------schedule_local---------------------------------
687 // Topological sort within a block. Someday become a real scheduler.
688 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) {
689 // Already "sorted" are the block start Node (as the first entry), and
690 // the block-ending Node and any trailing control projections. We leave
691 // these alone. PhiNodes and ParmNodes are made to follow the block start
692 // Node. Everything else gets topo-sorted.
694 #ifndef PRODUCT
695 if (trace_opto_pipelining()) {
696 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order);
697 for (uint i = 0;i < block->number_of_nodes(); i++) {
698 tty->print("# ");
699 block->get_node(i)->fast_dump();
700 }
701 tty->print_cr("#");
702 }
703 #endif
705 // RootNode is already sorted
706 if (block->number_of_nodes() == 1) {
707 return true;
708 }
710 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
711 uint node_cnt = block->end_idx();
712 uint phi_cnt = 1;
713 uint i;
714 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
715 Node *n = block->get_node(i);
716 if( n->is_Phi() || // Found a PhiNode or ParmNode
717 (n->is_Proj() && n->in(0) == block->head()) ) {
718 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
719 block->map_node(block->get_node(phi_cnt), i);
720 block->map_node(n, phi_cnt++); // swap Phi/Parm up front
721 } else { // All others
722 // Count block-local inputs to 'n'
723 uint cnt = n->len(); // Input count
724 uint local = 0;
725 for( uint j=0; j<cnt; j++ ) {
726 Node *m = n->in(j);
727 if( m && get_block_for_node(m) == block && !m->is_top() )
728 local++; // One more block-local input
729 }
730 ready_cnt.at_put(n->_idx, local); // Count em up
732 #ifdef ASSERT
733 if( UseConcMarkSweepGC || UseG1GC ) {
734 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
735 // Check the precedence edges
736 for (uint prec = n->req(); prec < n->len(); prec++) {
737 Node* oop_store = n->in(prec);
738 if (oop_store != NULL) {
739 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark");
740 }
741 }
742 }
743 }
744 #endif
746 // A few node types require changing a required edge to a precedence edge
747 // before allocation.
748 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
749 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
750 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
751 // MemBarAcquire could be created without Precedent edge.
752 // del_req() replaces the specified edge with the last input edge
753 // and then removes the last edge. If the specified edge > number of
754 // edges the last edge will be moved outside of the input edges array
755 // and the edge will be lost. This is why this code should be
756 // executed only when Precedent (== TypeFunc::Parms) edge is present.
757 Node *x = n->in(TypeFunc::Parms);
758 n->del_req(TypeFunc::Parms);
759 n->add_prec(x);
760 }
761 }
762 }
763 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count
764 ready_cnt.at_put(block->get_node(i2)->_idx, 0);
766 // All the prescheduled guys do not hold back internal nodes
767 uint i3;
768 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
769 Node *n = block->get_node(i3); // Get pre-scheduled
770 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
771 Node* m = n->fast_out(j);
772 if (get_block_for_node(m) == block) { // Local-block user
773 int m_cnt = ready_cnt.at(m->_idx)-1;
774 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
775 }
776 }
777 }
779 Node_List delay;
780 // Make a worklist
781 Node_List worklist;
782 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
783 Node *m = block->get_node(i4);
784 if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
785 if (m->is_iteratively_computed()) {
786 // Push induction variable increments last to allow other uses
787 // of the phi to be scheduled first. The select() method breaks
788 // ties in scheduling by worklist order.
789 delay.push(m);
790 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
791 // Force the CreateEx to the top of the list so it's processed
792 // first and ends up at the start of the block.
793 worklist.insert(0, m);
794 } else {
795 worklist.push(m); // Then on to worklist!
796 }
797 }
798 }
799 while (delay.size()) {
800 Node* d = delay.pop();
801 worklist.push(d);
802 }
804 // Warm up the 'next_call' heuristic bits
805 needed_for_next_call(block, block->head(), next_call);
807 #ifndef PRODUCT
808 if (trace_opto_pipelining()) {
809 for (uint j=0; j< block->number_of_nodes(); j++) {
810 Node *n = block->get_node(j);
811 int idx = n->_idx;
812 tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
813 tty->print("latency:%3d ", get_latency_for_node(n));
814 tty->print("%4d: %s\n", idx, n->Name());
815 }
816 }
817 #endif
819 uint max_idx = (uint)ready_cnt.length();
820 // Pull from worklist and schedule
821 while( worklist.size() ) { // Worklist is not ready
823 #ifndef PRODUCT
824 if (trace_opto_pipelining()) {
825 tty->print("# ready list:");
826 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
827 Node *n = worklist[i]; // Get Node on worklist
828 tty->print(" %d", n->_idx);
829 }
830 tty->cr();
831 }
832 #endif
834 // Select and pop a ready guy from worklist
835 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt);
836 block->map_node(n, phi_cnt++); // Schedule him next
838 #ifndef PRODUCT
839 if (trace_opto_pipelining()) {
840 tty->print("# select %d: %s", n->_idx, n->Name());
841 tty->print(", latency:%d", get_latency_for_node(n));
842 n->dump();
843 if (Verbose) {
844 tty->print("# ready list:");
845 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
846 Node *n = worklist[i]; // Get Node on worklist
847 tty->print(" %d", n->_idx);
848 }
849 tty->cr();
850 }
851 }
853 #endif
854 if( n->is_MachCall() ) {
855 MachCallNode *mcall = n->as_MachCall();
856 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call);
857 continue;
858 }
860 if (n->is_Mach() && n->as_Mach()->has_call()) {
861 RegMask regs;
862 regs.Insert(_matcher.c_frame_pointer());
863 regs.OR(n->out_RegMask());
865 MachProjNode *proj = new (C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
866 map_node_to_block(proj, block);
867 block->insert_node(proj, phi_cnt++);
869 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false);
870 }
872 // Children are now all ready
873 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
874 Node* m = n->fast_out(i5); // Get user
875 if (get_block_for_node(m) != block) {
876 continue;
877 }
878 if( m->is_Phi() ) continue;
879 if (m->_idx >= max_idx) { // new node, skip it
880 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
881 continue;
882 }
883 int m_cnt = ready_cnt.at(m->_idx)-1;
884 ready_cnt.at_put(m->_idx, m_cnt);
885 if( m_cnt == 0 )
886 worklist.push(m);
887 }
888 }
890 if( phi_cnt != block->end_idx() ) {
891 // did not schedule all. Retry, Bailout, or Die
892 if (C->subsume_loads() == true && !C->failing()) {
893 // Retry with subsume_loads == false
894 // If this is the first failure, the sentinel string will "stick"
895 // to the Compile object, and the C2Compiler will see it and retry.
896 C->record_failure(C2Compiler::retry_no_subsuming_loads());
897 }
898 // assert( phi_cnt == end_idx(), "did not schedule all" );
899 return false;
900 }
902 #ifndef PRODUCT
903 if (trace_opto_pipelining()) {
904 tty->print_cr("#");
905 tty->print_cr("# after schedule_local");
906 for (uint i = 0;i < block->number_of_nodes();i++) {
907 tty->print("# ");
908 block->get_node(i)->fast_dump();
909 }
910 tty->cr();
911 }
912 #endif
915 return true;
916 }
918 //--------------------------catch_cleanup_fix_all_inputs-----------------------
919 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
920 for (uint l = 0; l < use->len(); l++) {
921 if (use->in(l) == old_def) {
922 if (l < use->req()) {
923 use->set_req(l, new_def);
924 } else {
925 use->rm_prec(l);
926 use->add_prec(new_def);
927 l--;
928 }
929 }
930 }
931 }
933 //------------------------------catch_cleanup_find_cloned_def------------------
934 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
935 assert( use_blk != def_blk, "Inter-block cleanup only");
937 // The use is some block below the Catch. Find and return the clone of the def
938 // that dominates the use. If there is no clone in a dominating block, then
939 // create a phi for the def in a dominating block.
941 // Find which successor block dominates this use. The successor
942 // blocks must all be single-entry (from the Catch only; I will have
943 // split blocks to make this so), hence they all dominate.
944 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
945 use_blk = use_blk->_idom;
947 // Find the successor
948 Node *fixup = NULL;
950 uint j;
951 for( j = 0; j < def_blk->_num_succs; j++ )
952 if( use_blk == def_blk->_succs[j] )
953 break;
955 if( j == def_blk->_num_succs ) {
956 // Block at same level in dom-tree is not a successor. It needs a
957 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
958 Node_Array inputs = new Node_List(Thread::current()->resource_area());
959 for(uint k = 1; k < use_blk->num_preds(); k++) {
960 Block* block = get_block_for_node(use_blk->pred(k));
961 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx));
962 }
964 // Check to see if the use_blk already has an identical phi inserted.
965 // If it exists, it will be at the first position since all uses of a
966 // def are processed together.
967 Node *phi = use_blk->get_node(1);
968 if( phi->is_Phi() ) {
969 fixup = phi;
970 for (uint k = 1; k < use_blk->num_preds(); k++) {
971 if (phi->in(k) != inputs[k]) {
972 // Not a match
973 fixup = NULL;
974 break;
975 }
976 }
977 }
979 // If an existing PhiNode was not found, make a new one.
980 if (fixup == NULL) {
981 Node *new_phi = PhiNode::make(use_blk->head(), def);
982 use_blk->insert_node(new_phi, 1);
983 map_node_to_block(new_phi, use_blk);
984 for (uint k = 1; k < use_blk->num_preds(); k++) {
985 new_phi->set_req(k, inputs[k]);
986 }
987 fixup = new_phi;
988 }
990 } else {
991 // Found the use just below the Catch. Make it use the clone.
992 fixup = use_blk->get_node(n_clone_idx);
993 }
995 return fixup;
996 }
998 //--------------------------catch_cleanup_intra_block--------------------------
999 // Fix all input edges in use that reference "def". The use is in the same
1000 // block as the def and both have been cloned in each successor block.
1001 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
1003 // Both the use and def have been cloned. For each successor block,
1004 // get the clone of the use, and make its input the clone of the def
1005 // found in that block.
1007 uint use_idx = blk->find_node(use);
1008 uint offset_idx = use_idx - beg;
1009 for( uint k = 0; k < blk->_num_succs; k++ ) {
1010 // Get clone in each successor block
1011 Block *sb = blk->_succs[k];
1012 Node *clone = sb->get_node(offset_idx+1);
1013 assert( clone->Opcode() == use->Opcode(), "" );
1015 // Make use-clone reference the def-clone
1016 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx));
1017 }
1018 }
1020 //------------------------------catch_cleanup_inter_block---------------------
1021 // Fix all input edges in use that reference "def". The use is in a different
1022 // block than the def.
1023 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1024 if( !use_blk ) return; // Can happen if the use is a precedence edge
1026 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx);
1027 catch_cleanup_fix_all_inputs(use, def, new_def);
1028 }
1030 //------------------------------call_catch_cleanup-----------------------------
1031 // If we inserted any instructions between a Call and his CatchNode,
1032 // clone the instructions on all paths below the Catch.
1033 void PhaseCFG::call_catch_cleanup(Block* block) {
1035 // End of region to clone
1036 uint end = block->end_idx();
1037 if( !block->get_node(end)->is_Catch() ) return;
1038 // Start of region to clone
1039 uint beg = end;
1040 while(!block->get_node(beg-1)->is_MachProj() ||
1041 !block->get_node(beg-1)->in(0)->is_MachCall() ) {
1042 beg--;
1043 assert(beg > 0,"Catch cleanup walking beyond block boundary");
1044 }
1045 // Range of inserted instructions is [beg, end)
1046 if( beg == end ) return;
1048 // Clone along all Catch output paths. Clone area between the 'beg' and
1049 // 'end' indices.
1050 for( uint i = 0; i < block->_num_succs; i++ ) {
1051 Block *sb = block->_succs[i];
1052 // Clone the entire area; ignoring the edge fixup for now.
1053 for( uint j = end; j > beg; j-- ) {
1054 // It is safe here to clone a node with anti_dependence
1055 // since clones dominate on each path.
1056 Node *clone = block->get_node(j-1)->clone();
1057 sb->insert_node(clone, 1);
1058 map_node_to_block(clone, sb);
1059 }
1060 }
1063 // Fixup edges. Check the def-use info per cloned Node
1064 for(uint i2 = beg; i2 < end; i2++ ) {
1065 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1066 Node *n = block->get_node(i2); // Node that got cloned
1067 // Need DU safe iterator because of edge manipulation in calls.
1068 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1069 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1070 out->push(n->fast_out(j1));
1071 }
1072 uint max = out->size();
1073 for (uint j = 0; j < max; j++) {// For all users
1074 Node *use = out->pop();
1075 Block *buse = get_block_for_node(use);
1076 if( use->is_Phi() ) {
1077 for( uint k = 1; k < use->req(); k++ )
1078 if( use->in(k) == n ) {
1079 Block* b = get_block_for_node(buse->pred(k));
1080 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx);
1081 use->set_req(k, fixup);
1082 }
1083 } else {
1084 if (block == buse) {
1085 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx);
1086 } else {
1087 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx);
1088 }
1089 }
1090 } // End for all users
1092 } // End of for all Nodes in cloned area
1094 // Remove the now-dead cloned ops
1095 for(uint i3 = beg; i3 < end; i3++ ) {
1096 block->get_node(beg)->disconnect_inputs(NULL, C);
1097 block->remove_node(beg);
1098 }
1100 // If the successor blocks have a CreateEx node, move it back to the top
1101 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) {
1102 Block *sb = block->_succs[i4];
1103 uint new_cnt = end - beg;
1104 // Remove any newly created, but dead, nodes.
1105 for( uint j = new_cnt; j > 0; j-- ) {
1106 Node *n = sb->get_node(j);
1107 if (n->outcnt() == 0 &&
1108 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1109 n->disconnect_inputs(NULL, C);
1110 sb->remove_node(j);
1111 new_cnt--;
1112 }
1113 }
1114 // If any newly created nodes remain, move the CreateEx node to the top
1115 if (new_cnt > 0) {
1116 Node *cex = sb->get_node(1+new_cnt);
1117 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1118 sb->remove_node(1+new_cnt);
1119 sb->insert_node(cex, 1);
1120 }
1121 }
1122 }
1123 }