Sun, 01 Sep 2013 19:21:05 +0200
8023988: Move local scheduling of nodes to the CFG creation and code motion phase (PhaseCFG)
Summary: Moved local scheduling code from class Block to class PhaseCFG
Reviewed-by: kvn, roland
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
49 # include "adfiles/ad_ppc.hpp"
50 #endif
52 // Optimization - Graph Style
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 PhaseCFG::implicit_null_check(Block* block, 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;
67 // Make sure the ptr-is-null path appears to be uncommon!
68 float f = block->end()->as_MachIf()->_prob;
69 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
70 if( f > PROB_UNLIKELY_MAG(4) ) return;
72 uint bidx = 0; // Capture index of value into memop
73 bool was_store; // Memory op is a store op
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 (block->get_node(block->number_of_nodes()-1) == proj) {
79 null_block = block->_succs[0];
80 not_null_block = block->_succs[1];
81 } else {
82 assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other");
83 not_null_block = block->_succs[0];
84 null_block = block->_succs[1];
85 }
86 while (null_block->is_Empty() == Block::empty_with_goto) {
87 null_block = null_block->_succs[0];
88 }
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->number_of_nodes(); i1++) {
97 Node* nn = null_block->get_node(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 }
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);
127 assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() &&
128 (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity");
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_LoadUB:
143 case Op_LoadUS:
144 case Op_LoadD:
145 case Op_LoadF:
146 case Op_LoadI:
147 case Op_LoadL:
148 case Op_LoadP:
149 case Op_LoadN:
150 case Op_LoadS:
151 case Op_LoadKlass:
152 case Op_LoadNKlass:
153 case Op_LoadRange:
154 case Op_LoadD_unaligned:
155 case Op_LoadL_unaligned:
156 assert(mach->in(2) == val, "should be address");
157 break;
158 case Op_StoreB:
159 case Op_StoreC:
160 case Op_StoreCM:
161 case Op_StoreD:
162 case Op_StoreF:
163 case Op_StoreI:
164 case Op_StoreL:
165 case Op_StoreP:
166 case Op_StoreN:
167 case Op_StoreNKlass:
168 was_store = true; // Memory op is a store op
169 // Stores will have their address in slot 2 (memory in slot 1).
170 // If the value being nul-checked is in another slot, it means we
171 // are storing the checked value, which does NOT check the value!
172 if( mach->in(2) != val ) continue;
173 break; // Found a memory op?
174 case Op_StrComp:
175 case Op_StrEquals:
176 case Op_StrIndexOf:
177 case Op_AryEq:
178 case Op_EncodeISOArray:
179 // Not a legit memory op for implicit null check regardless of
180 // embedded loads
181 continue;
182 default: // Also check for embedded loads
183 if( !mach->needs_anti_dependence_check() )
184 continue; // Not an memory op; skip it
185 if( must_clone[iop] ) {
186 // Do not move nodes which produce flags because
187 // RA will try to clone it to place near branch and
188 // it will cause recompilation, see clone_node().
189 continue;
190 }
191 {
192 // Check that value is used in memory address in
193 // instructions with embedded load (CmpP val1,(val2+off)).
194 Node* base;
195 Node* index;
196 const MachOper* oper = mach->memory_inputs(base, index);
197 if (oper == NULL || oper == (MachOper*)-1) {
198 continue; // Not an memory op; skip it
199 }
200 if (val == base ||
201 val == index && val->bottom_type()->isa_narrowoop()) {
202 break; // Found it
203 } else {
204 continue; // Skip it
205 }
206 }
207 break;
208 }
209 // check if the offset is not too high for implicit exception
210 {
211 intptr_t offset = 0;
212 const TypePtr *adr_type = NULL; // Do not need this return value here
213 const Node* base = mach->get_base_and_disp(offset, adr_type);
214 if (base == NULL || base == NodeSentinel) {
215 // Narrow oop address doesn't have base, only index
216 if( val->bottom_type()->isa_narrowoop() &&
217 MacroAssembler::needs_explicit_null_check(offset) )
218 continue; // Give up if offset is beyond page size
219 // cannot reason about it; is probably not implicit null exception
220 } else {
221 const TypePtr* tptr;
222 if (UseCompressedOops && (Universe::narrow_oop_shift() == 0 ||
223 Universe::narrow_klass_shift() == 0)) {
224 // 32-bits narrow oop can be the base of address expressions
225 tptr = base->get_ptr_type();
226 } else {
227 // only regular oops are expected here
228 tptr = base->bottom_type()->is_ptr();
229 }
230 // Give up if offset is not a compile-time constant
231 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
232 continue;
233 offset += tptr->_offset; // correct if base is offseted
234 if( MacroAssembler::needs_explicit_null_check(offset) )
235 continue; // Give up is reference is beyond 4K page size
236 }
237 }
239 // Check ctrl input to see if the null-check dominates the memory op
240 Block *cb = get_block_for_node(mach);
241 cb = cb->_idom; // Always hoist at least 1 block
242 if( !was_store ) { // Stores can be hoisted only one block
243 while( cb->_dom_depth > (block->_dom_depth + 1))
244 cb = cb->_idom; // Hoist loads as far as we want
245 // The non-null-block should dominate the memory op, too. Live
246 // range spilling will insert a spill in the non-null-block if it is
247 // needs to spill the memory op for an implicit null check.
248 if (cb->_dom_depth == (block->_dom_depth + 1)) {
249 if (cb != not_null_block) continue;
250 cb = cb->_idom;
251 }
252 }
253 if( cb != block ) continue;
255 // Found a memory user; see if it can be hoisted to check-block
256 uint vidx = 0; // Capture index of value into memop
257 uint j;
258 for( j = mach->req()-1; j > 0; j-- ) {
259 if( mach->in(j) == val ) {
260 vidx = j;
261 // Ignore DecodeN val which could be hoisted to where needed.
262 if( is_decoden ) continue;
263 }
264 // Block of memory-op input
265 Block *inb = get_block_for_node(mach->in(j));
266 Block *b = block; // Start from nul check
267 while( b != inb && b->_dom_depth > inb->_dom_depth )
268 b = b->_idom; // search upwards for input
269 // See if input dominates null check
270 if( b != inb )
271 break;
272 }
273 if( j > 0 )
274 continue;
275 Block *mb = get_block_for_node(mach);
276 // Hoisting stores requires more checks for the anti-dependence case.
277 // Give up hoisting if we have to move the store past any load.
278 if( was_store ) {
279 Block *b = mb; // Start searching here for a local load
280 // mach use (faulting) trying to hoist
281 // n might be blocker to hoisting
282 while( b != block ) {
283 uint k;
284 for( k = 1; k < b->number_of_nodes(); k++ ) {
285 Node *n = b->get_node(k);
286 if( n->needs_anti_dependence_check() &&
287 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
288 break; // Found anti-dependent load
289 }
290 if( k < b->number_of_nodes() )
291 break; // Found anti-dependent load
292 // Make sure control does not do a merge (would have to check allpaths)
293 if( b->num_preds() != 2 ) break;
294 b = get_block_for_node(b->pred(1)); // Move up to predecessor block
295 }
296 if( b != block ) continue;
297 }
299 // Make sure this memory op is not already being used for a NullCheck
300 Node *e = mb->end();
301 if( e->is_MachNullCheck() && e->in(1) == mach )
302 continue; // Already being used as a NULL check
304 // Found a candidate! Pick one with least dom depth - the highest
305 // in the dom tree should be closest to the null check.
306 if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) {
307 best = mach;
308 bidx = vidx;
309 }
310 }
311 // No candidate!
312 if (best == NULL) {
313 return;
314 }
316 // ---- Found an implicit null check
317 extern int implicit_null_checks;
318 implicit_null_checks++;
320 if( is_decoden ) {
321 // Check if we need to hoist decodeHeapOop_not_null first.
322 Block *valb = get_block_for_node(val);
323 if( block != valb && block->_dom_depth < valb->_dom_depth ) {
324 // Hoist it up to the end of the test block.
325 valb->find_remove(val);
326 block->add_inst(val);
327 map_node_to_block(val, block);
328 // DecodeN on x86 may kill flags. Check for flag-killing projections
329 // that also need to be hoisted.
330 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
331 Node* n = val->fast_out(j);
332 if( n->is_MachProj() ) {
333 get_block_for_node(n)->find_remove(n);
334 block->add_inst(n);
335 map_node_to_block(n, block);
336 }
337 }
338 }
339 }
340 // Hoist the memory candidate up to the end of the test block.
341 Block *old_block = get_block_for_node(best);
342 old_block->find_remove(best);
343 block->add_inst(best);
344 map_node_to_block(best, block);
346 // Move the control dependence
347 if (best->in(0) && best->in(0) == old_block->head())
348 best->set_req(0, block->head());
350 // Check for flag-killing projections that also need to be hoisted
351 // Should be DU safe because no edge updates.
352 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
353 Node* n = best->fast_out(j);
354 if( n->is_MachProj() ) {
355 get_block_for_node(n)->find_remove(n);
356 block->add_inst(n);
357 map_node_to_block(n, block);
358 }
359 }
361 // proj==Op_True --> ne test; proj==Op_False --> eq test.
362 // One of two graph shapes got matched:
363 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
364 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
365 // NULL checks are always branch-if-eq. If we see a IfTrue projection
366 // then we are replacing a 'ne' test with a 'eq' NULL check test.
367 // We need to flip the projections to keep the same semantics.
368 if( proj->Opcode() == Op_IfTrue ) {
369 // Swap order of projections in basic block to swap branch targets
370 Node *tmp1 = block->get_node(block->end_idx()+1);
371 Node *tmp2 = block->get_node(block->end_idx()+2);
372 block->map_node(tmp2, block->end_idx()+1);
373 block->map_node(tmp1, block->end_idx()+2);
374 Node *tmp = new (C) Node(C->top()); // Use not NULL input
375 tmp1->replace_by(tmp);
376 tmp2->replace_by(tmp1);
377 tmp->replace_by(tmp2);
378 tmp->destruct();
379 }
381 // Remove the existing null check; use a new implicit null check instead.
382 // Since schedule-local needs precise def-use info, we need to correct
383 // it as well.
384 Node *old_tst = proj->in(0);
385 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
386 block->map_node(nul_chk, block->end_idx());
387 map_node_to_block(nul_chk, block);
388 // Redirect users of old_test to nul_chk
389 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
390 old_tst->last_out(i2)->set_req(0, nul_chk);
391 // Clean-up any dead code
392 for (uint i3 = 0; i3 < old_tst->req(); i3++)
393 old_tst->set_req(i3, NULL);
395 latency_from_uses(nul_chk);
396 latency_from_uses(best);
397 }
400 //------------------------------select-----------------------------------------
401 // Select a nice fellow from the worklist to schedule next. If there is only
402 // one choice, then use it. Projections take top priority for correctness
403 // reasons - if I see a projection, then it is next. There are a number of
404 // other special cases, for instructions that consume condition codes, et al.
405 // These are chosen immediately. Some instructions are required to immediately
406 // precede the last instruction in the block, and these are taken last. Of the
407 // remaining cases (most), choose the instruction with the greatest latency
408 // (that is, the most number of pseudo-cycles required to the end of the
409 // routine). If there is a tie, choose the instruction with the most inputs.
410 Node* PhaseCFG::select(Block* block, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) {
412 // If only a single entry on the stack, use it
413 uint cnt = worklist.size();
414 if (cnt == 1) {
415 Node *n = worklist[0];
416 worklist.map(0,worklist.pop());
417 return n;
418 }
420 uint choice = 0; // Bigger is most important
421 uint latency = 0; // Bigger is scheduled first
422 uint score = 0; // Bigger is better
423 int idx = -1; // Index in worklist
424 int cand_cnt = 0; // Candidate count
426 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
427 // Order in worklist is used to break ties.
428 // See caller for how this is used to delay scheduling
429 // of induction variable increments to after the other
430 // uses of the phi are scheduled.
431 Node *n = worklist[i]; // Get Node on worklist
433 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
434 if( n->is_Proj() || // Projections always win
435 n->Opcode()== Op_Con || // So does constant 'Top'
436 iop == Op_CreateEx || // Create-exception must start block
437 iop == Op_CheckCastPP
438 ) {
439 worklist.map(i,worklist.pop());
440 return n;
441 }
443 // Final call in a block must be adjacent to 'catch'
444 Node *e = block->end();
445 if( e->is_Catch() && e->in(0)->in(0) == n )
446 continue;
448 // Memory op for an implicit null check has to be at the end of the block
449 if( e->is_MachNullCheck() && e->in(1) == n )
450 continue;
452 // Schedule IV increment last.
453 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd &&
454 e->in(1)->in(1) == n && n->is_iteratively_computed())
455 continue;
457 uint n_choice = 2;
459 // See if this instruction is consumed by a branch. If so, then (as the
460 // branch is the last instruction in the basic block) force it to the
461 // end of the basic block
462 if ( must_clone[iop] ) {
463 // See if any use is a branch
464 bool found_machif = false;
466 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
467 Node* use = n->fast_out(j);
469 // The use is a conditional branch, make them adjacent
470 if (use->is_MachIf() && get_block_for_node(use) == block) {
471 found_machif = true;
472 break;
473 }
475 // More than this instruction pending for successor to be ready,
476 // don't choose this if other opportunities are ready
477 if (ready_cnt.at(use->_idx) > 1)
478 n_choice = 1;
479 }
481 // loop terminated, prefer not to use this instruction
482 if (found_machif)
483 continue;
484 }
486 // See if this has a predecessor that is "must_clone", i.e. sets the
487 // condition code. If so, choose this first
488 for (uint j = 0; j < n->req() ; j++) {
489 Node *inn = n->in(j);
490 if (inn) {
491 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
492 n_choice = 3;
493 break;
494 }
495 }
496 }
498 // MachTemps should be scheduled last so they are near their uses
499 if (n->is_MachTemp()) {
500 n_choice = 1;
501 }
503 uint n_latency = get_latency_for_node(n);
504 uint n_score = n->req(); // Many inputs get high score to break ties
506 // Keep best latency found
507 cand_cnt++;
508 if (choice < n_choice ||
509 (choice == n_choice &&
510 ((StressLCM && Compile::randomized_select(cand_cnt)) ||
511 (!StressLCM &&
512 (latency < n_latency ||
513 (latency == n_latency &&
514 (score < n_score))))))) {
515 choice = n_choice;
516 latency = n_latency;
517 score = n_score;
518 idx = i; // Also keep index in worklist
519 }
520 } // End of for all ready nodes in worklist
522 assert(idx >= 0, "index should be set");
523 Node *n = worklist[(uint)idx]; // Get the winner
525 worklist.map((uint)idx, worklist.pop()); // Compress worklist
526 return n;
527 }
530 //------------------------------set_next_call----------------------------------
531 void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) {
532 if( next_call.test_set(n->_idx) ) return;
533 for( uint i=0; i<n->len(); i++ ) {
534 Node *m = n->in(i);
535 if( !m ) continue; // must see all nodes in block that precede call
536 if (get_block_for_node(m) == block) {
537 set_next_call(block, m, next_call);
538 }
539 }
540 }
542 //------------------------------needed_for_next_call---------------------------
543 // Set the flag 'next_call' for each Node that is needed for the next call to
544 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
545 // next subroutine call get priority - basically it moves things NOT needed
546 // for the next call till after the call. This prevents me from trying to
547 // carry lots of stuff live across a call.
548 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) {
549 // Find the next control-defining Node in this block
550 Node* call = NULL;
551 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
552 Node* m = this_call->fast_out(i);
553 if(get_block_for_node(m) == block && // Local-block user
554 m != this_call && // Not self-start node
555 m->is_MachCall() )
556 call = m;
557 break;
558 }
559 if (call == NULL) return; // No next call (e.g., block end is near)
560 // Set next-call for all inputs to this call
561 set_next_call(block, call, next_call);
562 }
564 //------------------------------add_call_kills-------------------------------------
565 // helper function that adds caller save registers to MachProjNode
566 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
567 // Fill in the kill mask for the call
568 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
569 if( !regs.Member(r) ) { // Not already defined by the call
570 // Save-on-call register?
571 if ((save_policy[r] == 'C') ||
572 (save_policy[r] == 'A') ||
573 ((save_policy[r] == 'E') && exclude_soe)) {
574 proj->_rout.Insert(r);
575 }
576 }
577 }
578 }
581 //------------------------------sched_call-------------------------------------
582 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) {
583 RegMask regs;
585 // Schedule all the users of the call right now. All the users are
586 // projection Nodes, so they must be scheduled next to the call.
587 // Collect all the defined registers.
588 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
589 Node* n = mcall->fast_out(i);
590 assert( n->is_MachProj(), "" );
591 int n_cnt = ready_cnt.at(n->_idx)-1;
592 ready_cnt.at_put(n->_idx, n_cnt);
593 assert( n_cnt == 0, "" );
594 // Schedule next to call
595 block->map_node(n, node_cnt++);
596 // Collect defined registers
597 regs.OR(n->out_RegMask());
598 // Check for scheduling the next control-definer
599 if( n->bottom_type() == Type::CONTROL )
600 // Warm up next pile of heuristic bits
601 needed_for_next_call(block, n, next_call);
603 // Children of projections are now all ready
604 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
605 Node* m = n->fast_out(j); // Get user
606 if(get_block_for_node(m) != block) {
607 continue;
608 }
609 if( m->is_Phi() ) continue;
610 int m_cnt = ready_cnt.at(m->_idx)-1;
611 ready_cnt.at_put(m->_idx, m_cnt);
612 if( m_cnt == 0 )
613 worklist.push(m);
614 }
616 }
618 // Act as if the call defines the Frame Pointer.
619 // Certainly the FP is alive and well after the call.
620 regs.Insert(_matcher.c_frame_pointer());
622 // Set all registers killed and not already defined by the call.
623 uint r_cnt = mcall->tf()->range()->cnt();
624 int op = mcall->ideal_Opcode();
625 MachProjNode *proj = new (C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
626 map_node_to_block(proj, block);
627 block->insert_node(proj, node_cnt++);
629 // Select the right register save policy.
630 const char * save_policy;
631 switch (op) {
632 case Op_CallRuntime:
633 case Op_CallLeaf:
634 case Op_CallLeafNoFP:
635 // Calling C code so use C calling convention
636 save_policy = _matcher._c_reg_save_policy;
637 break;
639 case Op_CallStaticJava:
640 case Op_CallDynamicJava:
641 // Calling Java code so use Java calling convention
642 save_policy = _matcher._register_save_policy;
643 break;
645 default:
646 ShouldNotReachHere();
647 }
649 // When using CallRuntime mark SOE registers as killed by the call
650 // so values that could show up in the RegisterMap aren't live in a
651 // callee saved register since the register wouldn't know where to
652 // find them. CallLeaf and CallLeafNoFP are ok because they can't
653 // have debug info on them. Strictly speaking this only needs to be
654 // done for oops since idealreg2debugmask takes care of debug info
655 // references but there no way to handle oops differently than other
656 // pointers as far as the kill mask goes.
657 bool exclude_soe = op == Op_CallRuntime;
659 // If the call is a MethodHandle invoke, we need to exclude the
660 // register which is used to save the SP value over MH invokes from
661 // the mask. Otherwise this register could be used for
662 // deoptimization information.
663 if (op == Op_CallStaticJava) {
664 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
665 if (mcallstaticjava->_method_handle_invoke)
666 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
667 }
669 add_call_kills(proj, regs, save_policy, exclude_soe);
671 return node_cnt;
672 }
675 //------------------------------schedule_local---------------------------------
676 // Topological sort within a block. Someday become a real scheduler.
677 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call) {
678 // Already "sorted" are the block start Node (as the first entry), and
679 // the block-ending Node and any trailing control projections. We leave
680 // these alone. PhiNodes and ParmNodes are made to follow the block start
681 // Node. Everything else gets topo-sorted.
683 #ifndef PRODUCT
684 if (trace_opto_pipelining()) {
685 tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order);
686 for (uint i = 0;i < block->number_of_nodes(); i++) {
687 tty->print("# ");
688 block->get_node(i)->fast_dump();
689 }
690 tty->print_cr("#");
691 }
692 #endif
694 // RootNode is already sorted
695 if (block->number_of_nodes() == 1) {
696 return true;
697 }
699 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
700 uint node_cnt = block->end_idx();
701 uint phi_cnt = 1;
702 uint i;
703 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
704 Node *n = block->get_node(i);
705 if( n->is_Phi() || // Found a PhiNode or ParmNode
706 (n->is_Proj() && n->in(0) == block->head()) ) {
707 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
708 block->map_node(block->get_node(phi_cnt), i);
709 block->map_node(n, phi_cnt++); // swap Phi/Parm up front
710 } else { // All others
711 // Count block-local inputs to 'n'
712 uint cnt = n->len(); // Input count
713 uint local = 0;
714 for( uint j=0; j<cnt; j++ ) {
715 Node *m = n->in(j);
716 if( m && get_block_for_node(m) == block && !m->is_top() )
717 local++; // One more block-local input
718 }
719 ready_cnt.at_put(n->_idx, local); // Count em up
721 #ifdef ASSERT
722 if( UseConcMarkSweepGC || UseG1GC ) {
723 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
724 // Check the precedence edges
725 for (uint prec = n->req(); prec < n->len(); prec++) {
726 Node* oop_store = n->in(prec);
727 if (oop_store != NULL) {
728 assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark");
729 }
730 }
731 }
732 }
733 #endif
735 // A few node types require changing a required edge to a precedence edge
736 // before allocation.
737 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
738 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
739 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
740 // MemBarAcquire could be created without Precedent edge.
741 // del_req() replaces the specified edge with the last input edge
742 // and then removes the last edge. If the specified edge > number of
743 // edges the last edge will be moved outside of the input edges array
744 // and the edge will be lost. This is why this code should be
745 // executed only when Precedent (== TypeFunc::Parms) edge is present.
746 Node *x = n->in(TypeFunc::Parms);
747 n->del_req(TypeFunc::Parms);
748 n->add_prec(x);
749 }
750 }
751 }
752 for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count
753 ready_cnt.at_put(block->get_node(i2)->_idx, 0);
755 // All the prescheduled guys do not hold back internal nodes
756 uint i3;
757 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
758 Node *n = block->get_node(i3); // Get pre-scheduled
759 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
760 Node* m = n->fast_out(j);
761 if (get_block_for_node(m) == block) { // Local-block user
762 int m_cnt = ready_cnt.at(m->_idx)-1;
763 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
764 }
765 }
766 }
768 Node_List delay;
769 // Make a worklist
770 Node_List worklist;
771 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
772 Node *m = block->get_node(i4);
773 if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
774 if (m->is_iteratively_computed()) {
775 // Push induction variable increments last to allow other uses
776 // of the phi to be scheduled first. The select() method breaks
777 // ties in scheduling by worklist order.
778 delay.push(m);
779 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
780 // Force the CreateEx to the top of the list so it's processed
781 // first and ends up at the start of the block.
782 worklist.insert(0, m);
783 } else {
784 worklist.push(m); // Then on to worklist!
785 }
786 }
787 }
788 while (delay.size()) {
789 Node* d = delay.pop();
790 worklist.push(d);
791 }
793 // Warm up the 'next_call' heuristic bits
794 needed_for_next_call(block, block->head(), next_call);
796 #ifndef PRODUCT
797 if (trace_opto_pipelining()) {
798 for (uint j=0; j< block->number_of_nodes(); j++) {
799 Node *n = block->get_node(j);
800 int idx = n->_idx;
801 tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
802 tty->print("latency:%3d ", get_latency_for_node(n));
803 tty->print("%4d: %s\n", idx, n->Name());
804 }
805 }
806 #endif
808 uint max_idx = (uint)ready_cnt.length();
809 // Pull from worklist and schedule
810 while( worklist.size() ) { // Worklist is not ready
812 #ifndef PRODUCT
813 if (trace_opto_pipelining()) {
814 tty->print("# ready list:");
815 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
816 Node *n = worklist[i]; // Get Node on worklist
817 tty->print(" %d", n->_idx);
818 }
819 tty->cr();
820 }
821 #endif
823 // Select and pop a ready guy from worklist
824 Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt);
825 block->map_node(n, phi_cnt++); // Schedule him next
827 #ifndef PRODUCT
828 if (trace_opto_pipelining()) {
829 tty->print("# select %d: %s", n->_idx, n->Name());
830 tty->print(", latency:%d", get_latency_for_node(n));
831 n->dump();
832 if (Verbose) {
833 tty->print("# ready list:");
834 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
835 Node *n = worklist[i]; // Get Node on worklist
836 tty->print(" %d", n->_idx);
837 }
838 tty->cr();
839 }
840 }
842 #endif
843 if( n->is_MachCall() ) {
844 MachCallNode *mcall = n->as_MachCall();
845 phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call);
846 continue;
847 }
849 if (n->is_Mach() && n->as_Mach()->has_call()) {
850 RegMask regs;
851 regs.Insert(_matcher.c_frame_pointer());
852 regs.OR(n->out_RegMask());
854 MachProjNode *proj = new (C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
855 map_node_to_block(proj, block);
856 block->insert_node(proj, phi_cnt++);
858 add_call_kills(proj, regs, _matcher._c_reg_save_policy, false);
859 }
861 // Children are now all ready
862 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
863 Node* m = n->fast_out(i5); // Get user
864 if (get_block_for_node(m) != block) {
865 continue;
866 }
867 if( m->is_Phi() ) continue;
868 if (m->_idx >= max_idx) { // new node, skip it
869 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
870 continue;
871 }
872 int m_cnt = ready_cnt.at(m->_idx)-1;
873 ready_cnt.at_put(m->_idx, m_cnt);
874 if( m_cnt == 0 )
875 worklist.push(m);
876 }
877 }
879 if( phi_cnt != block->end_idx() ) {
880 // did not schedule all. Retry, Bailout, or Die
881 if (C->subsume_loads() == true && !C->failing()) {
882 // Retry with subsume_loads == false
883 // If this is the first failure, the sentinel string will "stick"
884 // to the Compile object, and the C2Compiler will see it and retry.
885 C->record_failure(C2Compiler::retry_no_subsuming_loads());
886 }
887 // assert( phi_cnt == end_idx(), "did not schedule all" );
888 return false;
889 }
891 #ifndef PRODUCT
892 if (trace_opto_pipelining()) {
893 tty->print_cr("#");
894 tty->print_cr("# after schedule_local");
895 for (uint i = 0;i < block->number_of_nodes();i++) {
896 tty->print("# ");
897 block->get_node(i)->fast_dump();
898 }
899 tty->cr();
900 }
901 #endif
904 return true;
905 }
907 //--------------------------catch_cleanup_fix_all_inputs-----------------------
908 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
909 for (uint l = 0; l < use->len(); l++) {
910 if (use->in(l) == old_def) {
911 if (l < use->req()) {
912 use->set_req(l, new_def);
913 } else {
914 use->rm_prec(l);
915 use->add_prec(new_def);
916 l--;
917 }
918 }
919 }
920 }
922 //------------------------------catch_cleanup_find_cloned_def------------------
923 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
924 assert( use_blk != def_blk, "Inter-block cleanup only");
926 // The use is some block below the Catch. Find and return the clone of the def
927 // that dominates the use. If there is no clone in a dominating block, then
928 // create a phi for the def in a dominating block.
930 // Find which successor block dominates this use. The successor
931 // blocks must all be single-entry (from the Catch only; I will have
932 // split blocks to make this so), hence they all dominate.
933 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
934 use_blk = use_blk->_idom;
936 // Find the successor
937 Node *fixup = NULL;
939 uint j;
940 for( j = 0; j < def_blk->_num_succs; j++ )
941 if( use_blk == def_blk->_succs[j] )
942 break;
944 if( j == def_blk->_num_succs ) {
945 // Block at same level in dom-tree is not a successor. It needs a
946 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
947 Node_Array inputs = new Node_List(Thread::current()->resource_area());
948 for(uint k = 1; k < use_blk->num_preds(); k++) {
949 Block* block = get_block_for_node(use_blk->pred(k));
950 inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx));
951 }
953 // Check to see if the use_blk already has an identical phi inserted.
954 // If it exists, it will be at the first position since all uses of a
955 // def are processed together.
956 Node *phi = use_blk->get_node(1);
957 if( phi->is_Phi() ) {
958 fixup = phi;
959 for (uint k = 1; k < use_blk->num_preds(); k++) {
960 if (phi->in(k) != inputs[k]) {
961 // Not a match
962 fixup = NULL;
963 break;
964 }
965 }
966 }
968 // If an existing PhiNode was not found, make a new one.
969 if (fixup == NULL) {
970 Node *new_phi = PhiNode::make(use_blk->head(), def);
971 use_blk->insert_node(new_phi, 1);
972 map_node_to_block(new_phi, use_blk);
973 for (uint k = 1; k < use_blk->num_preds(); k++) {
974 new_phi->set_req(k, inputs[k]);
975 }
976 fixup = new_phi;
977 }
979 } else {
980 // Found the use just below the Catch. Make it use the clone.
981 fixup = use_blk->get_node(n_clone_idx);
982 }
984 return fixup;
985 }
987 //--------------------------catch_cleanup_intra_block--------------------------
988 // Fix all input edges in use that reference "def". The use is in the same
989 // block as the def and both have been cloned in each successor block.
990 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
992 // Both the use and def have been cloned. For each successor block,
993 // get the clone of the use, and make its input the clone of the def
994 // found in that block.
996 uint use_idx = blk->find_node(use);
997 uint offset_idx = use_idx - beg;
998 for( uint k = 0; k < blk->_num_succs; k++ ) {
999 // Get clone in each successor block
1000 Block *sb = blk->_succs[k];
1001 Node *clone = sb->get_node(offset_idx+1);
1002 assert( clone->Opcode() == use->Opcode(), "" );
1004 // Make use-clone reference the def-clone
1005 catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx));
1006 }
1007 }
1009 //------------------------------catch_cleanup_inter_block---------------------
1010 // Fix all input edges in use that reference "def". The use is in a different
1011 // block than the def.
1012 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1013 if( !use_blk ) return; // Can happen if the use is a precedence edge
1015 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx);
1016 catch_cleanup_fix_all_inputs(use, def, new_def);
1017 }
1019 //------------------------------call_catch_cleanup-----------------------------
1020 // If we inserted any instructions between a Call and his CatchNode,
1021 // clone the instructions on all paths below the Catch.
1022 void PhaseCFG::call_catch_cleanup(Block* block) {
1024 // End of region to clone
1025 uint end = block->end_idx();
1026 if( !block->get_node(end)->is_Catch() ) return;
1027 // Start of region to clone
1028 uint beg = end;
1029 while(!block->get_node(beg-1)->is_MachProj() ||
1030 !block->get_node(beg-1)->in(0)->is_MachCall() ) {
1031 beg--;
1032 assert(beg > 0,"Catch cleanup walking beyond block boundary");
1033 }
1034 // Range of inserted instructions is [beg, end)
1035 if( beg == end ) return;
1037 // Clone along all Catch output paths. Clone area between the 'beg' and
1038 // 'end' indices.
1039 for( uint i = 0; i < block->_num_succs; i++ ) {
1040 Block *sb = block->_succs[i];
1041 // Clone the entire area; ignoring the edge fixup for now.
1042 for( uint j = end; j > beg; j-- ) {
1043 // It is safe here to clone a node with anti_dependence
1044 // since clones dominate on each path.
1045 Node *clone = block->get_node(j-1)->clone();
1046 sb->insert_node(clone, 1);
1047 map_node_to_block(clone, sb);
1048 }
1049 }
1052 // Fixup edges. Check the def-use info per cloned Node
1053 for(uint i2 = beg; i2 < end; i2++ ) {
1054 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1055 Node *n = block->get_node(i2); // Node that got cloned
1056 // Need DU safe iterator because of edge manipulation in calls.
1057 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1058 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1059 out->push(n->fast_out(j1));
1060 }
1061 uint max = out->size();
1062 for (uint j = 0; j < max; j++) {// For all users
1063 Node *use = out->pop();
1064 Block *buse = get_block_for_node(use);
1065 if( use->is_Phi() ) {
1066 for( uint k = 1; k < use->req(); k++ )
1067 if( use->in(k) == n ) {
1068 Block* b = get_block_for_node(buse->pred(k));
1069 Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx);
1070 use->set_req(k, fixup);
1071 }
1072 } else {
1073 if (block == buse) {
1074 catch_cleanup_intra_block(use, n, block, beg, n_clone_idx);
1075 } else {
1076 catch_cleanup_inter_block(use, buse, n, block, n_clone_idx);
1077 }
1078 }
1079 } // End for all users
1081 } // End of for all Nodes in cloned area
1083 // Remove the now-dead cloned ops
1084 for(uint i3 = beg; i3 < end; i3++ ) {
1085 block->get_node(beg)->disconnect_inputs(NULL, C);
1086 block->remove_node(beg);
1087 }
1089 // If the successor blocks have a CreateEx node, move it back to the top
1090 for(uint i4 = 0; i4 < block->_num_succs; i4++ ) {
1091 Block *sb = block->_succs[i4];
1092 uint new_cnt = end - beg;
1093 // Remove any newly created, but dead, nodes.
1094 for( uint j = new_cnt; j > 0; j-- ) {
1095 Node *n = sb->get_node(j);
1096 if (n->outcnt() == 0 &&
1097 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1098 n->disconnect_inputs(NULL, C);
1099 sb->remove_node(j);
1100 new_cnt--;
1101 }
1102 }
1103 // If any newly created nodes remain, move the CreateEx node to the top
1104 if (new_cnt > 0) {
1105 Node *cex = sb->get_node(1+new_cnt);
1106 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1107 sb->remove_node(1+new_cnt);
1108 sb->insert_node(cex, 1);
1109 }
1110 }
1111 }
1112 }