Tue, 09 Oct 2012 10:11:38 +0200
7054512: Compress class pointers after perm gen removal
Summary: support of compress class pointers in the compilers.
Reviewed-by: kvn, twisti
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 Block::implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons) {
62 // Assume if null check need for 0 offset then always needed
63 // Intel solaris doesn't support any null checks yet and no
64 // mechanism exists (yet) to set the switches at an os_cpu level
65 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
67 // Make sure the ptr-is-null path appears to be uncommon!
68 float f = end()->as_MachIf()->_prob;
69 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
70 if( f > PROB_UNLIKELY_MAG(4) ) return;
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 (_nodes[_nodes.size()-1] == proj) {
79 null_block = _succs[0];
80 not_null_block = _succs[1];
81 } else {
82 assert(_nodes[_nodes.size()-2] == proj, "proj is one or the other");
83 not_null_block = _succs[0];
84 null_block = _succs[1];
85 }
86 while (null_block->is_Empty() == Block::empty_with_goto) {
87 null_block = null_block->_succs[0];
88 }
90 // Search the exception block for an uncommon trap.
91 // (See Parse::do_if and Parse::do_ifnull for the reason
92 // we need an uncommon trap. Briefly, we need a way to
93 // detect failure of this optimization, as in 6366351.)
94 {
95 bool found_trap = false;
96 for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) {
97 Node* nn = null_block->_nodes[i1];
98 if (nn->is_MachCall() &&
99 nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) {
100 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
101 if (trtype->isa_int() && trtype->is_int()->is_con()) {
102 jint tr_con = trtype->is_int()->get_con();
103 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
104 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
105 assert((int)reason < (int)BitsPerInt, "recode bit map");
106 if (is_set_nth_bit(allowed_reasons, (int) reason)
107 && action != Deoptimization::Action_none) {
108 // This uncommon trap is sure to recompile, eventually.
109 // When that happens, C->too_many_traps will prevent
110 // this transformation from happening again.
111 found_trap = true;
112 }
113 }
114 break;
115 }
116 }
117 if (!found_trap) {
118 // We did not find an uncommon trap.
119 return;
120 }
121 }
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 // Not a legit memory op for implicit null check regardless of
179 // embedded loads
180 continue;
181 default: // Also check for embedded loads
182 if( !mach->needs_anti_dependence_check() )
183 continue; // Not an memory op; skip it
184 if( must_clone[iop] ) {
185 // Do not move nodes which produce flags because
186 // RA will try to clone it to place near branch and
187 // it will cause recompilation, see clone_node().
188 continue;
189 }
190 {
191 // Check that value is used in memory address in
192 // instructions with embedded load (CmpP val1,(val2+off)).
193 Node* base;
194 Node* index;
195 const MachOper* oper = mach->memory_inputs(base, index);
196 if (oper == NULL || oper == (MachOper*)-1) {
197 continue; // Not an memory op; skip it
198 }
199 if (val == base ||
200 val == index && val->bottom_type()->isa_narrowoop()) {
201 break; // Found it
202 } else {
203 continue; // Skip it
204 }
205 }
206 break;
207 }
208 // check if the offset is not too high for implicit exception
209 {
210 intptr_t offset = 0;
211 const TypePtr *adr_type = NULL; // Do not need this return value here
212 const Node* base = mach->get_base_and_disp(offset, adr_type);
213 if (base == NULL || base == NodeSentinel) {
214 // Narrow oop address doesn't have base, only index
215 if( val->bottom_type()->isa_narrowoop() &&
216 MacroAssembler::needs_explicit_null_check(offset) )
217 continue; // Give up if offset is beyond page size
218 // cannot reason about it; is probably not implicit null exception
219 } else {
220 const TypePtr* tptr;
221 if (UseCompressedOops && Universe::narrow_oop_shift() == 0) {
222 // 32-bits narrow oop can be the base of address expressions
223 tptr = base->bottom_type()->make_ptr();
224 } else {
225 // only regular oops are expected here
226 tptr = base->bottom_type()->is_ptr();
227 }
228 // Give up if offset is not a compile-time constant
229 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
230 continue;
231 offset += tptr->_offset; // correct if base is offseted
232 if( MacroAssembler::needs_explicit_null_check(offset) )
233 continue; // Give up is reference is beyond 4K page size
234 }
235 }
237 // Check ctrl input to see if the null-check dominates the memory op
238 Block *cb = cfg->_bbs[mach->_idx];
239 cb = cb->_idom; // Always hoist at least 1 block
240 if( !was_store ) { // Stores can be hoisted only one block
241 while( cb->_dom_depth > (_dom_depth + 1))
242 cb = cb->_idom; // Hoist loads as far as we want
243 // The non-null-block should dominate the memory op, too. Live
244 // range spilling will insert a spill in the non-null-block if it is
245 // needs to spill the memory op for an implicit null check.
246 if (cb->_dom_depth == (_dom_depth + 1)) {
247 if (cb != not_null_block) continue;
248 cb = cb->_idom;
249 }
250 }
251 if( cb != this ) continue;
253 // Found a memory user; see if it can be hoisted to check-block
254 uint vidx = 0; // Capture index of value into memop
255 uint j;
256 for( j = mach->req()-1; j > 0; j-- ) {
257 if( mach->in(j) == val ) {
258 vidx = j;
259 // Ignore DecodeN val which could be hoisted to where needed.
260 if( is_decoden ) continue;
261 }
262 // Block of memory-op input
263 Block *inb = cfg->_bbs[mach->in(j)->_idx];
264 Block *b = this; // Start from nul check
265 while( b != inb && b->_dom_depth > inb->_dom_depth )
266 b = b->_idom; // search upwards for input
267 // See if input dominates null check
268 if( b != inb )
269 break;
270 }
271 if( j > 0 )
272 continue;
273 Block *mb = cfg->_bbs[mach->_idx];
274 // Hoisting stores requires more checks for the anti-dependence case.
275 // Give up hoisting if we have to move the store past any load.
276 if( was_store ) {
277 Block *b = mb; // Start searching here for a local load
278 // mach use (faulting) trying to hoist
279 // n might be blocker to hoisting
280 while( b != this ) {
281 uint k;
282 for( k = 1; k < b->_nodes.size(); k++ ) {
283 Node *n = b->_nodes[k];
284 if( n->needs_anti_dependence_check() &&
285 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
286 break; // Found anti-dependent load
287 }
288 if( k < b->_nodes.size() )
289 break; // Found anti-dependent load
290 // Make sure control does not do a merge (would have to check allpaths)
291 if( b->num_preds() != 2 ) break;
292 b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block
293 }
294 if( b != this ) continue;
295 }
297 // Make sure this memory op is not already being used for a NullCheck
298 Node *e = mb->end();
299 if( e->is_MachNullCheck() && e->in(1) == mach )
300 continue; // Already being used as a NULL check
302 // Found a candidate! Pick one with least dom depth - the highest
303 // in the dom tree should be closest to the null check.
304 if( !best ||
305 cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) {
306 best = mach;
307 bidx = vidx;
309 }
310 }
311 // No candidate!
312 if( !best ) return;
314 // ---- Found an implicit null check
315 extern int implicit_null_checks;
316 implicit_null_checks++;
318 if( is_decoden ) {
319 // Check if we need to hoist decodeHeapOop_not_null first.
320 Block *valb = cfg->_bbs[val->_idx];
321 if( this != valb && this->_dom_depth < valb->_dom_depth ) {
322 // Hoist it up to the end of the test block.
323 valb->find_remove(val);
324 this->add_inst(val);
325 cfg->_bbs.map(val->_idx,this);
326 // DecodeN on x86 may kill flags. Check for flag-killing projections
327 // that also need to be hoisted.
328 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
329 Node* n = val->fast_out(j);
330 if( n->is_MachProj() ) {
331 cfg->_bbs[n->_idx]->find_remove(n);
332 this->add_inst(n);
333 cfg->_bbs.map(n->_idx,this);
334 }
335 }
336 }
337 }
338 // Hoist the memory candidate up to the end of the test block.
339 Block *old_block = cfg->_bbs[best->_idx];
340 old_block->find_remove(best);
341 add_inst(best);
342 cfg->_bbs.map(best->_idx,this);
344 // Move the control dependence
345 if (best->in(0) && best->in(0) == old_block->_nodes[0])
346 best->set_req(0, _nodes[0]);
348 // Check for flag-killing projections that also need to be hoisted
349 // Should be DU safe because no edge updates.
350 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
351 Node* n = best->fast_out(j);
352 if( n->is_MachProj() ) {
353 cfg->_bbs[n->_idx]->find_remove(n);
354 add_inst(n);
355 cfg->_bbs.map(n->_idx,this);
356 }
357 }
359 Compile *C = cfg->C;
360 // proj==Op_True --> ne test; proj==Op_False --> eq test.
361 // One of two graph shapes got matched:
362 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
363 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
364 // NULL checks are always branch-if-eq. If we see a IfTrue projection
365 // then we are replacing a 'ne' test with a 'eq' NULL check test.
366 // We need to flip the projections to keep the same semantics.
367 if( proj->Opcode() == Op_IfTrue ) {
368 // Swap order of projections in basic block to swap branch targets
369 Node *tmp1 = _nodes[end_idx()+1];
370 Node *tmp2 = _nodes[end_idx()+2];
371 _nodes.map(end_idx()+1, tmp2);
372 _nodes.map(end_idx()+2, tmp1);
373 Node *tmp = new (C) Node(C->top()); // Use not NULL input
374 tmp1->replace_by(tmp);
375 tmp2->replace_by(tmp1);
376 tmp->replace_by(tmp2);
377 tmp->destruct();
378 }
380 // Remove the existing null check; use a new implicit null check instead.
381 // Since schedule-local needs precise def-use info, we need to correct
382 // it as well.
383 Node *old_tst = proj->in(0);
384 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
385 _nodes.map(end_idx(),nul_chk);
386 cfg->_bbs.map(nul_chk->_idx,this);
387 // Redirect users of old_test to nul_chk
388 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
389 old_tst->last_out(i2)->set_req(0, nul_chk);
390 // Clean-up any dead code
391 for (uint i3 = 0; i3 < old_tst->req(); i3++)
392 old_tst->set_req(i3, NULL);
394 cfg->latency_from_uses(nul_chk);
395 cfg->latency_from_uses(best);
396 }
399 //------------------------------select-----------------------------------------
400 // Select a nice fellow from the worklist to schedule next. If there is only
401 // one choice, then use it. Projections take top priority for correctness
402 // reasons - if I see a projection, then it is next. There are a number of
403 // other special cases, for instructions that consume condition codes, et al.
404 // These are chosen immediately. Some instructions are required to immediately
405 // precede the last instruction in the block, and these are taken last. Of the
406 // remaining cases (most), choose the instruction with the greatest latency
407 // (that is, the most number of pseudo-cycles required to the end of the
408 // routine). If there is a tie, choose the instruction with the most inputs.
409 Node *Block::select(PhaseCFG *cfg, Node_List &worklist, GrowableArray<int> &ready_cnt, VectorSet &next_call, uint sched_slot) {
411 // If only a single entry on the stack, use it
412 uint cnt = worklist.size();
413 if (cnt == 1) {
414 Node *n = worklist[0];
415 worklist.map(0,worklist.pop());
416 return n;
417 }
419 uint choice = 0; // Bigger is most important
420 uint latency = 0; // Bigger is scheduled first
421 uint score = 0; // Bigger is better
422 int idx = -1; // Index in worklist
424 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
425 // Order in worklist is used to break ties.
426 // See caller for how this is used to delay scheduling
427 // of induction variable increments to after the other
428 // uses of the phi are scheduled.
429 Node *n = worklist[i]; // Get Node on worklist
431 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
432 if( n->is_Proj() || // Projections always win
433 n->Opcode()== Op_Con || // So does constant 'Top'
434 iop == Op_CreateEx || // Create-exception must start block
435 iop == Op_CheckCastPP
436 ) {
437 worklist.map(i,worklist.pop());
438 return n;
439 }
441 // Final call in a block must be adjacent to 'catch'
442 Node *e = end();
443 if( e->is_Catch() && e->in(0)->in(0) == n )
444 continue;
446 // Memory op for an implicit null check has to be at the end of the block
447 if( e->is_MachNullCheck() && e->in(1) == n )
448 continue;
450 // Schedule IV increment last.
451 if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd &&
452 e->in(1)->in(1) == n && n->is_iteratively_computed())
453 continue;
455 uint n_choice = 2;
457 // See if this instruction is consumed by a branch. If so, then (as the
458 // branch is the last instruction in the basic block) force it to the
459 // end of the basic block
460 if ( must_clone[iop] ) {
461 // See if any use is a branch
462 bool found_machif = false;
464 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
465 Node* use = n->fast_out(j);
467 // The use is a conditional branch, make them adjacent
468 if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) {
469 found_machif = true;
470 break;
471 }
473 // More than this instruction pending for successor to be ready,
474 // don't choose this if other opportunities are ready
475 if (ready_cnt.at(use->_idx) > 1)
476 n_choice = 1;
477 }
479 // loop terminated, prefer not to use this instruction
480 if (found_machif)
481 continue;
482 }
484 // See if this has a predecessor that is "must_clone", i.e. sets the
485 // condition code. If so, choose this first
486 for (uint j = 0; j < n->req() ; j++) {
487 Node *inn = n->in(j);
488 if (inn) {
489 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
490 n_choice = 3;
491 break;
492 }
493 }
494 }
496 // MachTemps should be scheduled last so they are near their uses
497 if (n->is_MachTemp()) {
498 n_choice = 1;
499 }
501 uint n_latency = cfg->_node_latency->at_grow(n->_idx);
502 uint n_score = n->req(); // Many inputs get high score to break ties
504 // Keep best latency found
505 if( choice < n_choice ||
506 ( choice == n_choice &&
507 ( latency < n_latency ||
508 ( latency == n_latency &&
509 ( score < n_score ))))) {
510 choice = n_choice;
511 latency = n_latency;
512 score = n_score;
513 idx = i; // Also keep index in worklist
514 }
515 } // End of for all ready nodes in worklist
517 assert(idx >= 0, "index should be set");
518 Node *n = worklist[(uint)idx]; // Get the winner
520 worklist.map((uint)idx, worklist.pop()); // Compress worklist
521 return n;
522 }
525 //------------------------------set_next_call----------------------------------
526 void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) {
527 if( next_call.test_set(n->_idx) ) return;
528 for( uint i=0; i<n->len(); i++ ) {
529 Node *m = n->in(i);
530 if( !m ) continue; // must see all nodes in block that precede call
531 if( bbs[m->_idx] == this )
532 set_next_call( m, next_call, bbs );
533 }
534 }
536 //------------------------------needed_for_next_call---------------------------
537 // Set the flag 'next_call' for each Node that is needed for the next call to
538 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
539 // next subroutine call get priority - basically it moves things NOT needed
540 // for the next call till after the call. This prevents me from trying to
541 // carry lots of stuff live across a call.
542 void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) {
543 // Find the next control-defining Node in this block
544 Node* call = NULL;
545 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
546 Node* m = this_call->fast_out(i);
547 if( bbs[m->_idx] == this && // Local-block user
548 m != this_call && // Not self-start node
549 m->is_MachCall() )
550 call = m;
551 break;
552 }
553 if (call == NULL) return; // No next call (e.g., block end is near)
554 // Set next-call for all inputs to this call
555 set_next_call(call, next_call, bbs);
556 }
558 //------------------------------add_call_kills-------------------------------------
559 void Block::add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
560 // Fill in the kill mask for the call
561 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
562 if( !regs.Member(r) ) { // Not already defined by the call
563 // Save-on-call register?
564 if ((save_policy[r] == 'C') ||
565 (save_policy[r] == 'A') ||
566 ((save_policy[r] == 'E') && exclude_soe)) {
567 proj->_rout.Insert(r);
568 }
569 }
570 }
571 }
574 //------------------------------sched_call-------------------------------------
575 uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, GrowableArray<int> &ready_cnt, MachCallNode *mcall, VectorSet &next_call ) {
576 RegMask regs;
578 // Schedule all the users of the call right now. All the users are
579 // projection Nodes, so they must be scheduled next to the call.
580 // Collect all the defined registers.
581 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
582 Node* n = mcall->fast_out(i);
583 assert( n->is_MachProj(), "" );
584 int n_cnt = ready_cnt.at(n->_idx)-1;
585 ready_cnt.at_put(n->_idx, n_cnt);
586 assert( n_cnt == 0, "" );
587 // Schedule next to call
588 _nodes.map(node_cnt++, n);
589 // Collect defined registers
590 regs.OR(n->out_RegMask());
591 // Check for scheduling the next control-definer
592 if( n->bottom_type() == Type::CONTROL )
593 // Warm up next pile of heuristic bits
594 needed_for_next_call(n, next_call, bbs);
596 // Children of projections are now all ready
597 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
598 Node* m = n->fast_out(j); // Get user
599 if( bbs[m->_idx] != this ) continue;
600 if( m->is_Phi() ) continue;
601 int m_cnt = ready_cnt.at(m->_idx)-1;
602 ready_cnt.at_put(m->_idx, m_cnt);
603 if( m_cnt == 0 )
604 worklist.push(m);
605 }
607 }
609 // Act as if the call defines the Frame Pointer.
610 // Certainly the FP is alive and well after the call.
611 regs.Insert(matcher.c_frame_pointer());
613 // Set all registers killed and not already defined by the call.
614 uint r_cnt = mcall->tf()->range()->cnt();
615 int op = mcall->ideal_Opcode();
616 MachProjNode *proj = new (matcher.C) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
617 bbs.map(proj->_idx,this);
618 _nodes.insert(node_cnt++, proj);
620 // Select the right register save policy.
621 const char * save_policy;
622 switch (op) {
623 case Op_CallRuntime:
624 case Op_CallLeaf:
625 case Op_CallLeafNoFP:
626 // Calling C code so use C calling convention
627 save_policy = matcher._c_reg_save_policy;
628 break;
630 case Op_CallStaticJava:
631 case Op_CallDynamicJava:
632 // Calling Java code so use Java calling convention
633 save_policy = matcher._register_save_policy;
634 break;
636 default:
637 ShouldNotReachHere();
638 }
640 // When using CallRuntime mark SOE registers as killed by the call
641 // so values that could show up in the RegisterMap aren't live in a
642 // callee saved register since the register wouldn't know where to
643 // find them. CallLeaf and CallLeafNoFP are ok because they can't
644 // have debug info on them. Strictly speaking this only needs to be
645 // done for oops since idealreg2debugmask takes care of debug info
646 // references but there no way to handle oops differently than other
647 // pointers as far as the kill mask goes.
648 bool exclude_soe = op == Op_CallRuntime;
650 // If the call is a MethodHandle invoke, we need to exclude the
651 // register which is used to save the SP value over MH invokes from
652 // the mask. Otherwise this register could be used for
653 // deoptimization information.
654 if (op == Op_CallStaticJava) {
655 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
656 if (mcallstaticjava->_method_handle_invoke)
657 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
658 }
660 add_call_kills(proj, regs, save_policy, exclude_soe);
662 return node_cnt;
663 }
666 //------------------------------schedule_local---------------------------------
667 // Topological sort within a block. Someday become a real scheduler.
668 bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, GrowableArray<int> &ready_cnt, VectorSet &next_call) {
669 // Already "sorted" are the block start Node (as the first entry), and
670 // the block-ending Node and any trailing control projections. We leave
671 // these alone. PhiNodes and ParmNodes are made to follow the block start
672 // Node. Everything else gets topo-sorted.
674 #ifndef PRODUCT
675 if (cfg->trace_opto_pipelining()) {
676 tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order);
677 for (uint i = 0;i < _nodes.size();i++) {
678 tty->print("# ");
679 _nodes[i]->fast_dump();
680 }
681 tty->print_cr("#");
682 }
683 #endif
685 // RootNode is already sorted
686 if( _nodes.size() == 1 ) return true;
688 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
689 uint node_cnt = end_idx();
690 uint phi_cnt = 1;
691 uint i;
692 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
693 Node *n = _nodes[i];
694 if( n->is_Phi() || // Found a PhiNode or ParmNode
695 (n->is_Proj() && n->in(0) == head()) ) {
696 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
697 _nodes.map(i,_nodes[phi_cnt]);
698 _nodes.map(phi_cnt++,n); // swap Phi/Parm up front
699 } else { // All others
700 // Count block-local inputs to 'n'
701 uint cnt = n->len(); // Input count
702 uint local = 0;
703 for( uint j=0; j<cnt; j++ ) {
704 Node *m = n->in(j);
705 if( m && cfg->_bbs[m->_idx] == this && !m->is_top() )
706 local++; // One more block-local input
707 }
708 ready_cnt.at_put(n->_idx, local); // Count em up
710 #ifdef ASSERT
711 if( UseConcMarkSweepGC || UseG1GC ) {
712 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
713 // Check the precedence edges
714 for (uint prec = n->req(); prec < n->len(); prec++) {
715 Node* oop_store = n->in(prec);
716 if (oop_store != NULL) {
717 assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark");
718 }
719 }
720 }
721 }
722 #endif
724 // A few node types require changing a required edge to a precedence edge
725 // before allocation.
726 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
727 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
728 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
729 // MemBarAcquire could be created without Precedent edge.
730 // del_req() replaces the specified edge with the last input edge
731 // and then removes the last edge. If the specified edge > number of
732 // edges the last edge will be moved outside of the input edges array
733 // and the edge will be lost. This is why this code should be
734 // executed only when Precedent (== TypeFunc::Parms) edge is present.
735 Node *x = n->in(TypeFunc::Parms);
736 n->del_req(TypeFunc::Parms);
737 n->add_prec(x);
738 }
739 }
740 }
741 for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count
742 ready_cnt.at_put(_nodes[i2]->_idx, 0);
744 // All the prescheduled guys do not hold back internal nodes
745 uint i3;
746 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
747 Node *n = _nodes[i3]; // Get pre-scheduled
748 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
749 Node* m = n->fast_out(j);
750 if( cfg->_bbs[m->_idx] ==this ) { // Local-block user
751 int m_cnt = ready_cnt.at(m->_idx)-1;
752 ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count
753 }
754 }
755 }
757 Node_List delay;
758 // Make a worklist
759 Node_List worklist;
760 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
761 Node *m = _nodes[i4];
762 if( !ready_cnt.at(m->_idx) ) { // Zero ready count?
763 if (m->is_iteratively_computed()) {
764 // Push induction variable increments last to allow other uses
765 // of the phi to be scheduled first. The select() method breaks
766 // ties in scheduling by worklist order.
767 delay.push(m);
768 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
769 // Force the CreateEx to the top of the list so it's processed
770 // first and ends up at the start of the block.
771 worklist.insert(0, m);
772 } else {
773 worklist.push(m); // Then on to worklist!
774 }
775 }
776 }
777 while (delay.size()) {
778 Node* d = delay.pop();
779 worklist.push(d);
780 }
782 // Warm up the 'next_call' heuristic bits
783 needed_for_next_call(_nodes[0], next_call, cfg->_bbs);
785 #ifndef PRODUCT
786 if (cfg->trace_opto_pipelining()) {
787 for (uint j=0; j<_nodes.size(); j++) {
788 Node *n = _nodes[j];
789 int idx = n->_idx;
790 tty->print("# ready cnt:%3d ", ready_cnt.at(idx));
791 tty->print("latency:%3d ", cfg->_node_latency->at_grow(idx));
792 tty->print("%4d: %s\n", idx, n->Name());
793 }
794 }
795 #endif
797 uint max_idx = (uint)ready_cnt.length();
798 // Pull from worklist and schedule
799 while( worklist.size() ) { // Worklist is not ready
801 #ifndef PRODUCT
802 if (cfg->trace_opto_pipelining()) {
803 tty->print("# ready list:");
804 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
805 Node *n = worklist[i]; // Get Node on worklist
806 tty->print(" %d", n->_idx);
807 }
808 tty->cr();
809 }
810 #endif
812 // Select and pop a ready guy from worklist
813 Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt);
814 _nodes.map(phi_cnt++,n); // Schedule him next
816 #ifndef PRODUCT
817 if (cfg->trace_opto_pipelining()) {
818 tty->print("# select %d: %s", n->_idx, n->Name());
819 tty->print(", latency:%d", cfg->_node_latency->at_grow(n->_idx));
820 n->dump();
821 if (Verbose) {
822 tty->print("# ready list:");
823 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
824 Node *n = worklist[i]; // Get Node on worklist
825 tty->print(" %d", n->_idx);
826 }
827 tty->cr();
828 }
829 }
831 #endif
832 if( n->is_MachCall() ) {
833 MachCallNode *mcall = n->as_MachCall();
834 phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call);
835 continue;
836 }
838 if (n->is_Mach() && n->as_Mach()->has_call()) {
839 RegMask regs;
840 regs.Insert(matcher.c_frame_pointer());
841 regs.OR(n->out_RegMask());
843 MachProjNode *proj = new (matcher.C) MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj );
844 cfg->_bbs.map(proj->_idx,this);
845 _nodes.insert(phi_cnt++, proj);
847 add_call_kills(proj, regs, matcher._c_reg_save_policy, false);
848 }
850 // Children are now all ready
851 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
852 Node* m = n->fast_out(i5); // Get user
853 if( cfg->_bbs[m->_idx] != this ) continue;
854 if( m->is_Phi() ) continue;
855 if (m->_idx >= max_idx) { // new node, skip it
856 assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
857 continue;
858 }
859 int m_cnt = ready_cnt.at(m->_idx)-1;
860 ready_cnt.at_put(m->_idx, m_cnt);
861 if( m_cnt == 0 )
862 worklist.push(m);
863 }
864 }
866 if( phi_cnt != end_idx() ) {
867 // did not schedule all. Retry, Bailout, or Die
868 Compile* C = matcher.C;
869 if (C->subsume_loads() == true && !C->failing()) {
870 // Retry with subsume_loads == false
871 // If this is the first failure, the sentinel string will "stick"
872 // to the Compile object, and the C2Compiler will see it and retry.
873 C->record_failure(C2Compiler::retry_no_subsuming_loads());
874 }
875 // assert( phi_cnt == end_idx(), "did not schedule all" );
876 return false;
877 }
879 #ifndef PRODUCT
880 if (cfg->trace_opto_pipelining()) {
881 tty->print_cr("#");
882 tty->print_cr("# after schedule_local");
883 for (uint i = 0;i < _nodes.size();i++) {
884 tty->print("# ");
885 _nodes[i]->fast_dump();
886 }
887 tty->cr();
888 }
889 #endif
892 return true;
893 }
895 //--------------------------catch_cleanup_fix_all_inputs-----------------------
896 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
897 for (uint l = 0; l < use->len(); l++) {
898 if (use->in(l) == old_def) {
899 if (l < use->req()) {
900 use->set_req(l, new_def);
901 } else {
902 use->rm_prec(l);
903 use->add_prec(new_def);
904 l--;
905 }
906 }
907 }
908 }
910 //------------------------------catch_cleanup_find_cloned_def------------------
911 static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
912 assert( use_blk != def_blk, "Inter-block cleanup only");
914 // The use is some block below the Catch. Find and return the clone of the def
915 // that dominates the use. If there is no clone in a dominating block, then
916 // create a phi for the def in a dominating block.
918 // Find which successor block dominates this use. The successor
919 // blocks must all be single-entry (from the Catch only; I will have
920 // split blocks to make this so), hence they all dominate.
921 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
922 use_blk = use_blk->_idom;
924 // Find the successor
925 Node *fixup = NULL;
927 uint j;
928 for( j = 0; j < def_blk->_num_succs; j++ )
929 if( use_blk == def_blk->_succs[j] )
930 break;
932 if( j == def_blk->_num_succs ) {
933 // Block at same level in dom-tree is not a successor. It needs a
934 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
935 Node_Array inputs = new Node_List(Thread::current()->resource_area());
936 for(uint k = 1; k < use_blk->num_preds(); k++) {
937 inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx));
938 }
940 // Check to see if the use_blk already has an identical phi inserted.
941 // If it exists, it will be at the first position since all uses of a
942 // def are processed together.
943 Node *phi = use_blk->_nodes[1];
944 if( phi->is_Phi() ) {
945 fixup = phi;
946 for (uint k = 1; k < use_blk->num_preds(); k++) {
947 if (phi->in(k) != inputs[k]) {
948 // Not a match
949 fixup = NULL;
950 break;
951 }
952 }
953 }
955 // If an existing PhiNode was not found, make a new one.
956 if (fixup == NULL) {
957 Node *new_phi = PhiNode::make(use_blk->head(), def);
958 use_blk->_nodes.insert(1, new_phi);
959 bbs.map(new_phi->_idx, use_blk);
960 for (uint k = 1; k < use_blk->num_preds(); k++) {
961 new_phi->set_req(k, inputs[k]);
962 }
963 fixup = new_phi;
964 }
966 } else {
967 // Found the use just below the Catch. Make it use the clone.
968 fixup = use_blk->_nodes[n_clone_idx];
969 }
971 return fixup;
972 }
974 //--------------------------catch_cleanup_intra_block--------------------------
975 // Fix all input edges in use that reference "def". The use is in the same
976 // block as the def and both have been cloned in each successor block.
977 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
979 // Both the use and def have been cloned. For each successor block,
980 // get the clone of the use, and make its input the clone of the def
981 // found in that block.
983 uint use_idx = blk->find_node(use);
984 uint offset_idx = use_idx - beg;
985 for( uint k = 0; k < blk->_num_succs; k++ ) {
986 // Get clone in each successor block
987 Block *sb = blk->_succs[k];
988 Node *clone = sb->_nodes[offset_idx+1];
989 assert( clone->Opcode() == use->Opcode(), "" );
991 // Make use-clone reference the def-clone
992 catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]);
993 }
994 }
996 //------------------------------catch_cleanup_inter_block---------------------
997 // Fix all input edges in use that reference "def". The use is in a different
998 // block than the def.
999 static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
1000 if( !use_blk ) return; // Can happen if the use is a precedence edge
1002 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx);
1003 catch_cleanup_fix_all_inputs(use, def, new_def);
1004 }
1006 //------------------------------call_catch_cleanup-----------------------------
1007 // If we inserted any instructions between a Call and his CatchNode,
1008 // clone the instructions on all paths below the Catch.
1009 void Block::call_catch_cleanup(Block_Array &bbs) {
1011 // End of region to clone
1012 uint end = end_idx();
1013 if( !_nodes[end]->is_Catch() ) return;
1014 // Start of region to clone
1015 uint beg = end;
1016 while(!_nodes[beg-1]->is_MachProj() ||
1017 !_nodes[beg-1]->in(0)->is_MachCall() ) {
1018 beg--;
1019 assert(beg > 0,"Catch cleanup walking beyond block boundary");
1020 }
1021 // Range of inserted instructions is [beg, end)
1022 if( beg == end ) return;
1024 // Clone along all Catch output paths. Clone area between the 'beg' and
1025 // 'end' indices.
1026 for( uint i = 0; i < _num_succs; i++ ) {
1027 Block *sb = _succs[i];
1028 // Clone the entire area; ignoring the edge fixup for now.
1029 for( uint j = end; j > beg; j-- ) {
1030 // It is safe here to clone a node with anti_dependence
1031 // since clones dominate on each path.
1032 Node *clone = _nodes[j-1]->clone();
1033 sb->_nodes.insert( 1, clone );
1034 bbs.map(clone->_idx,sb);
1035 }
1036 }
1039 // Fixup edges. Check the def-use info per cloned Node
1040 for(uint i2 = beg; i2 < end; i2++ ) {
1041 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1042 Node *n = _nodes[i2]; // Node that got cloned
1043 // Need DU safe iterator because of edge manipulation in calls.
1044 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
1045 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1046 out->push(n->fast_out(j1));
1047 }
1048 uint max = out->size();
1049 for (uint j = 0; j < max; j++) {// For all users
1050 Node *use = out->pop();
1051 Block *buse = bbs[use->_idx];
1052 if( use->is_Phi() ) {
1053 for( uint k = 1; k < use->req(); k++ )
1054 if( use->in(k) == n ) {
1055 Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx);
1056 use->set_req(k, fixup);
1057 }
1058 } else {
1059 if (this == buse) {
1060 catch_cleanup_intra_block(use, n, this, beg, n_clone_idx);
1061 } else {
1062 catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx);
1063 }
1064 }
1065 } // End for all users
1067 } // End of for all Nodes in cloned area
1069 // Remove the now-dead cloned ops
1070 for(uint i3 = beg; i3 < end; i3++ ) {
1071 _nodes[beg]->disconnect_inputs(NULL);
1072 _nodes.remove(beg);
1073 }
1075 // If the successor blocks have a CreateEx node, move it back to the top
1076 for(uint i4 = 0; i4 < _num_succs; i4++ ) {
1077 Block *sb = _succs[i4];
1078 uint new_cnt = end - beg;
1079 // Remove any newly created, but dead, nodes.
1080 for( uint j = new_cnt; j > 0; j-- ) {
1081 Node *n = sb->_nodes[j];
1082 if (n->outcnt() == 0 &&
1083 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
1084 n->disconnect_inputs(NULL);
1085 sb->_nodes.remove(j);
1086 new_cnt--;
1087 }
1088 }
1089 // If any newly created nodes remain, move the CreateEx node to the top
1090 if (new_cnt > 0) {
1091 Node *cex = sb->_nodes[1+new_cnt];
1092 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1093 sb->_nodes.remove(1+new_cnt);
1094 sb->_nodes.insert(1,cex);
1095 }
1096 }
1097 }
1098 }