Tue, 09 Mar 2010 20:16:19 +0100
6919934: JSR 292 needs to support x86 C1
Summary: This implements JSR 292 support for C1 x86.
Reviewed-by: never, jrose, kvn
1 /*
2 * Copyright 1998-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 // Optimization - Graph Style
27 #include "incls/_precompiled.incl"
28 #include "incls/_lcm.cpp.incl"
30 //------------------------------implicit_null_check----------------------------
31 // Detect implicit-null-check opportunities. Basically, find NULL checks
32 // with suitable memory ops nearby. Use the memory op to do the NULL check.
33 // I can generate a memory op if there is not one nearby.
34 // The proj is the control projection for the not-null case.
35 // The val is the pointer being checked for nullness.
36 void Block::implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons) {
37 // Assume if null check need for 0 offset then always needed
38 // Intel solaris doesn't support any null checks yet and no
39 // mechanism exists (yet) to set the switches at an os_cpu level
40 if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
42 // Make sure the ptr-is-null path appears to be uncommon!
43 float f = end()->as_MachIf()->_prob;
44 if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
45 if( f > PROB_UNLIKELY_MAG(4) ) return;
47 uint bidx = 0; // Capture index of value into memop
48 bool was_store; // Memory op is a store op
50 // Get the successor block for if the test ptr is non-null
51 Block* not_null_block; // this one goes with the proj
52 Block* null_block;
53 if (_nodes[_nodes.size()-1] == proj) {
54 null_block = _succs[0];
55 not_null_block = _succs[1];
56 } else {
57 assert(_nodes[_nodes.size()-2] == proj, "proj is one or the other");
58 not_null_block = _succs[0];
59 null_block = _succs[1];
60 }
61 while (null_block->is_Empty() == Block::empty_with_goto) {
62 null_block = null_block->_succs[0];
63 }
65 // Search the exception block for an uncommon trap.
66 // (See Parse::do_if and Parse::do_ifnull for the reason
67 // we need an uncommon trap. Briefly, we need a way to
68 // detect failure of this optimization, as in 6366351.)
69 {
70 bool found_trap = false;
71 for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) {
72 Node* nn = null_block->_nodes[i1];
73 if (nn->is_MachCall() &&
74 nn->as_MachCall()->entry_point() ==
75 SharedRuntime::uncommon_trap_blob()->instructions_begin()) {
76 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
77 if (trtype->isa_int() && trtype->is_int()->is_con()) {
78 jint tr_con = trtype->is_int()->get_con();
79 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
80 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
81 assert((int)reason < (int)BitsPerInt, "recode bit map");
82 if (is_set_nth_bit(allowed_reasons, (int) reason)
83 && action != Deoptimization::Action_none) {
84 // This uncommon trap is sure to recompile, eventually.
85 // When that happens, C->too_many_traps will prevent
86 // this transformation from happening again.
87 found_trap = true;
88 }
89 }
90 break;
91 }
92 }
93 if (!found_trap) {
94 // We did not find an uncommon trap.
95 return;
96 }
97 }
99 // Search the successor block for a load or store who's base value is also
100 // the tested value. There may be several.
101 Node_List *out = new Node_List(Thread::current()->resource_area());
102 MachNode *best = NULL; // Best found so far
103 for (DUIterator i = val->outs(); val->has_out(i); i++) {
104 Node *m = val->out(i);
105 if( !m->is_Mach() ) continue;
106 MachNode *mach = m->as_Mach();
107 was_store = false;
108 switch( mach->ideal_Opcode() ) {
109 case Op_LoadB:
110 case Op_LoadUS:
111 case Op_LoadD:
112 case Op_LoadF:
113 case Op_LoadI:
114 case Op_LoadL:
115 case Op_LoadP:
116 case Op_LoadN:
117 case Op_LoadS:
118 case Op_LoadKlass:
119 case Op_LoadNKlass:
120 case Op_LoadRange:
121 case Op_LoadD_unaligned:
122 case Op_LoadL_unaligned:
123 assert(mach->in(2) == val, "should be address");
124 break;
125 case Op_StoreB:
126 case Op_StoreC:
127 case Op_StoreCM:
128 case Op_StoreD:
129 case Op_StoreF:
130 case Op_StoreI:
131 case Op_StoreL:
132 case Op_StoreP:
133 case Op_StoreN:
134 was_store = true; // Memory op is a store op
135 // Stores will have their address in slot 2 (memory in slot 1).
136 // If the value being nul-checked is in another slot, it means we
137 // are storing the checked value, which does NOT check the value!
138 if( mach->in(2) != val ) continue;
139 break; // Found a memory op?
140 case Op_StrComp:
141 case Op_StrEquals:
142 case Op_StrIndexOf:
143 case Op_AryEq:
144 // Not a legit memory op for implicit null check regardless of
145 // embedded loads
146 continue;
147 default: // Also check for embedded loads
148 if( !mach->needs_anti_dependence_check() )
149 continue; // Not an memory op; skip it
150 {
151 // Check that value is used in memory address.
152 Node* base;
153 Node* index;
154 const MachOper* oper = mach->memory_inputs(base, index);
155 if (oper == NULL || oper == (MachOper*)-1) {
156 continue; // Not an memory op; skip it
157 }
158 if (val == base ||
159 val == index && val->bottom_type()->isa_narrowoop()) {
160 break; // Found it
161 } else {
162 continue; // Skip it
163 }
164 }
165 break;
166 }
167 // check if the offset is not too high for implicit exception
168 {
169 intptr_t offset = 0;
170 const TypePtr *adr_type = NULL; // Do not need this return value here
171 const Node* base = mach->get_base_and_disp(offset, adr_type);
172 if (base == NULL || base == NodeSentinel) {
173 // Narrow oop address doesn't have base, only index
174 if( val->bottom_type()->isa_narrowoop() &&
175 MacroAssembler::needs_explicit_null_check(offset) )
176 continue; // Give up if offset is beyond page size
177 // cannot reason about it; is probably not implicit null exception
178 } else {
179 const TypePtr* tptr;
180 if (UseCompressedOops && Universe::narrow_oop_shift() == 0) {
181 // 32-bits narrow oop can be the base of address expressions
182 tptr = base->bottom_type()->make_ptr();
183 } else {
184 // only regular oops are expected here
185 tptr = base->bottom_type()->is_ptr();
186 }
187 // Give up if offset is not a compile-time constant
188 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
189 continue;
190 offset += tptr->_offset; // correct if base is offseted
191 if( MacroAssembler::needs_explicit_null_check(offset) )
192 continue; // Give up is reference is beyond 4K page size
193 }
194 }
196 // Check ctrl input to see if the null-check dominates the memory op
197 Block *cb = cfg->_bbs[mach->_idx];
198 cb = cb->_idom; // Always hoist at least 1 block
199 if( !was_store ) { // Stores can be hoisted only one block
200 while( cb->_dom_depth > (_dom_depth + 1))
201 cb = cb->_idom; // Hoist loads as far as we want
202 // The non-null-block should dominate the memory op, too. Live
203 // range spilling will insert a spill in the non-null-block if it is
204 // needs to spill the memory op for an implicit null check.
205 if (cb->_dom_depth == (_dom_depth + 1)) {
206 if (cb != not_null_block) continue;
207 cb = cb->_idom;
208 }
209 }
210 if( cb != this ) continue;
212 // Found a memory user; see if it can be hoisted to check-block
213 uint vidx = 0; // Capture index of value into memop
214 uint j;
215 for( j = mach->req()-1; j > 0; j-- ) {
216 if( mach->in(j) == val ) vidx = j;
217 // Block of memory-op input
218 Block *inb = cfg->_bbs[mach->in(j)->_idx];
219 Block *b = this; // Start from nul check
220 while( b != inb && b->_dom_depth > inb->_dom_depth )
221 b = b->_idom; // search upwards for input
222 // See if input dominates null check
223 if( b != inb )
224 break;
225 }
226 if( j > 0 )
227 continue;
228 Block *mb = cfg->_bbs[mach->_idx];
229 // Hoisting stores requires more checks for the anti-dependence case.
230 // Give up hoisting if we have to move the store past any load.
231 if( was_store ) {
232 Block *b = mb; // Start searching here for a local load
233 // mach use (faulting) trying to hoist
234 // n might be blocker to hoisting
235 while( b != this ) {
236 uint k;
237 for( k = 1; k < b->_nodes.size(); k++ ) {
238 Node *n = b->_nodes[k];
239 if( n->needs_anti_dependence_check() &&
240 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
241 break; // Found anti-dependent load
242 }
243 if( k < b->_nodes.size() )
244 break; // Found anti-dependent load
245 // Make sure control does not do a merge (would have to check allpaths)
246 if( b->num_preds() != 2 ) break;
247 b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block
248 }
249 if( b != this ) continue;
250 }
252 // Make sure this memory op is not already being used for a NullCheck
253 Node *e = mb->end();
254 if( e->is_MachNullCheck() && e->in(1) == mach )
255 continue; // Already being used as a NULL check
257 // Found a candidate! Pick one with least dom depth - the highest
258 // in the dom tree should be closest to the null check.
259 if( !best ||
260 cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) {
261 best = mach;
262 bidx = vidx;
264 }
265 }
266 // No candidate!
267 if( !best ) return;
269 // ---- Found an implicit null check
270 extern int implicit_null_checks;
271 implicit_null_checks++;
273 // Hoist the memory candidate up to the end of the test block.
274 Block *old_block = cfg->_bbs[best->_idx];
275 old_block->find_remove(best);
276 add_inst(best);
277 cfg->_bbs.map(best->_idx,this);
279 // Move the control dependence
280 if (best->in(0) && best->in(0) == old_block->_nodes[0])
281 best->set_req(0, _nodes[0]);
283 // Check for flag-killing projections that also need to be hoisted
284 // Should be DU safe because no edge updates.
285 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
286 Node* n = best->fast_out(j);
287 if( n->Opcode() == Op_MachProj ) {
288 cfg->_bbs[n->_idx]->find_remove(n);
289 add_inst(n);
290 cfg->_bbs.map(n->_idx,this);
291 }
292 }
294 Compile *C = cfg->C;
295 // proj==Op_True --> ne test; proj==Op_False --> eq test.
296 // One of two graph shapes got matched:
297 // (IfTrue (If (Bool NE (CmpP ptr NULL))))
298 // (IfFalse (If (Bool EQ (CmpP ptr NULL))))
299 // NULL checks are always branch-if-eq. If we see a IfTrue projection
300 // then we are replacing a 'ne' test with a 'eq' NULL check test.
301 // We need to flip the projections to keep the same semantics.
302 if( proj->Opcode() == Op_IfTrue ) {
303 // Swap order of projections in basic block to swap branch targets
304 Node *tmp1 = _nodes[end_idx()+1];
305 Node *tmp2 = _nodes[end_idx()+2];
306 _nodes.map(end_idx()+1, tmp2);
307 _nodes.map(end_idx()+2, tmp1);
308 Node *tmp = new (C, 1) Node(C->top()); // Use not NULL input
309 tmp1->replace_by(tmp);
310 tmp2->replace_by(tmp1);
311 tmp->replace_by(tmp2);
312 tmp->destruct();
313 }
315 // Remove the existing null check; use a new implicit null check instead.
316 // Since schedule-local needs precise def-use info, we need to correct
317 // it as well.
318 Node *old_tst = proj->in(0);
319 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
320 _nodes.map(end_idx(),nul_chk);
321 cfg->_bbs.map(nul_chk->_idx,this);
322 // Redirect users of old_test to nul_chk
323 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
324 old_tst->last_out(i2)->set_req(0, nul_chk);
325 // Clean-up any dead code
326 for (uint i3 = 0; i3 < old_tst->req(); i3++)
327 old_tst->set_req(i3, NULL);
329 cfg->latency_from_uses(nul_chk);
330 cfg->latency_from_uses(best);
331 }
334 //------------------------------select-----------------------------------------
335 // Select a nice fellow from the worklist to schedule next. If there is only
336 // one choice, then use it. Projections take top priority for correctness
337 // reasons - if I see a projection, then it is next. There are a number of
338 // other special cases, for instructions that consume condition codes, et al.
339 // These are chosen immediately. Some instructions are required to immediately
340 // precede the last instruction in the block, and these are taken last. Of the
341 // remaining cases (most), choose the instruction with the greatest latency
342 // (that is, the most number of pseudo-cycles required to the end of the
343 // routine). If there is a tie, choose the instruction with the most inputs.
344 Node *Block::select(PhaseCFG *cfg, Node_List &worklist, int *ready_cnt, VectorSet &next_call, uint sched_slot) {
346 // If only a single entry on the stack, use it
347 uint cnt = worklist.size();
348 if (cnt == 1) {
349 Node *n = worklist[0];
350 worklist.map(0,worklist.pop());
351 return n;
352 }
354 uint choice = 0; // Bigger is most important
355 uint latency = 0; // Bigger is scheduled first
356 uint score = 0; // Bigger is better
357 int idx = -1; // Index in worklist
359 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
360 // Order in worklist is used to break ties.
361 // See caller for how this is used to delay scheduling
362 // of induction variable increments to after the other
363 // uses of the phi are scheduled.
364 Node *n = worklist[i]; // Get Node on worklist
366 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
367 if( n->is_Proj() || // Projections always win
368 n->Opcode()== Op_Con || // So does constant 'Top'
369 iop == Op_CreateEx || // Create-exception must start block
370 iop == Op_CheckCastPP
371 ) {
372 worklist.map(i,worklist.pop());
373 return n;
374 }
376 // Final call in a block must be adjacent to 'catch'
377 Node *e = end();
378 if( e->is_Catch() && e->in(0)->in(0) == n )
379 continue;
381 // Memory op for an implicit null check has to be at the end of the block
382 if( e->is_MachNullCheck() && e->in(1) == n )
383 continue;
385 uint n_choice = 2;
387 // See if this instruction is consumed by a branch. If so, then (as the
388 // branch is the last instruction in the basic block) force it to the
389 // end of the basic block
390 if ( must_clone[iop] ) {
391 // See if any use is a branch
392 bool found_machif = false;
394 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
395 Node* use = n->fast_out(j);
397 // The use is a conditional branch, make them adjacent
398 if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) {
399 found_machif = true;
400 break;
401 }
403 // More than this instruction pending for successor to be ready,
404 // don't choose this if other opportunities are ready
405 if (ready_cnt[use->_idx] > 1)
406 n_choice = 1;
407 }
409 // loop terminated, prefer not to use this instruction
410 if (found_machif)
411 continue;
412 }
414 // See if this has a predecessor that is "must_clone", i.e. sets the
415 // condition code. If so, choose this first
416 for (uint j = 0; j < n->req() ; j++) {
417 Node *inn = n->in(j);
418 if (inn) {
419 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
420 n_choice = 3;
421 break;
422 }
423 }
424 }
426 // MachTemps should be scheduled last so they are near their uses
427 if (n->is_MachTemp()) {
428 n_choice = 1;
429 }
431 uint n_latency = cfg->_node_latency.at_grow(n->_idx);
432 uint n_score = n->req(); // Many inputs get high score to break ties
434 // Keep best latency found
435 if( choice < n_choice ||
436 ( choice == n_choice &&
437 ( latency < n_latency ||
438 ( latency == n_latency &&
439 ( score < n_score ))))) {
440 choice = n_choice;
441 latency = n_latency;
442 score = n_score;
443 idx = i; // Also keep index in worklist
444 }
445 } // End of for all ready nodes in worklist
447 assert(idx >= 0, "index should be set");
448 Node *n = worklist[(uint)idx]; // Get the winner
450 worklist.map((uint)idx, worklist.pop()); // Compress worklist
451 return n;
452 }
455 //------------------------------set_next_call----------------------------------
456 void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) {
457 if( next_call.test_set(n->_idx) ) return;
458 for( uint i=0; i<n->len(); i++ ) {
459 Node *m = n->in(i);
460 if( !m ) continue; // must see all nodes in block that precede call
461 if( bbs[m->_idx] == this )
462 set_next_call( m, next_call, bbs );
463 }
464 }
466 //------------------------------needed_for_next_call---------------------------
467 // Set the flag 'next_call' for each Node that is needed for the next call to
468 // be scheduled. This flag lets me bias scheduling so Nodes needed for the
469 // next subroutine call get priority - basically it moves things NOT needed
470 // for the next call till after the call. This prevents me from trying to
471 // carry lots of stuff live across a call.
472 void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) {
473 // Find the next control-defining Node in this block
474 Node* call = NULL;
475 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
476 Node* m = this_call->fast_out(i);
477 if( bbs[m->_idx] == this && // Local-block user
478 m != this_call && // Not self-start node
479 m->is_Call() )
480 call = m;
481 break;
482 }
483 if (call == NULL) return; // No next call (e.g., block end is near)
484 // Set next-call for all inputs to this call
485 set_next_call(call, next_call, bbs);
486 }
488 //------------------------------sched_call-------------------------------------
489 uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, int *ready_cnt, MachCallNode *mcall, VectorSet &next_call ) {
490 RegMask regs;
492 // Schedule all the users of the call right now. All the users are
493 // projection Nodes, so they must be scheduled next to the call.
494 // Collect all the defined registers.
495 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
496 Node* n = mcall->fast_out(i);
497 assert( n->Opcode()==Op_MachProj, "" );
498 --ready_cnt[n->_idx];
499 assert( !ready_cnt[n->_idx], "" );
500 // Schedule next to call
501 _nodes.map(node_cnt++, n);
502 // Collect defined registers
503 regs.OR(n->out_RegMask());
504 // Check for scheduling the next control-definer
505 if( n->bottom_type() == Type::CONTROL )
506 // Warm up next pile of heuristic bits
507 needed_for_next_call(n, next_call, bbs);
509 // Children of projections are now all ready
510 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
511 Node* m = n->fast_out(j); // Get user
512 if( bbs[m->_idx] != this ) continue;
513 if( m->is_Phi() ) continue;
514 if( !--ready_cnt[m->_idx] )
515 worklist.push(m);
516 }
518 }
520 // Act as if the call defines the Frame Pointer.
521 // Certainly the FP is alive and well after the call.
522 regs.Insert(matcher.c_frame_pointer());
524 // Set all registers killed and not already defined by the call.
525 uint r_cnt = mcall->tf()->range()->cnt();
526 int op = mcall->ideal_Opcode();
527 MachProjNode *proj = new (matcher.C, 1) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
528 bbs.map(proj->_idx,this);
529 _nodes.insert(node_cnt++, proj);
531 // Select the right register save policy.
532 const char * save_policy;
533 switch (op) {
534 case Op_CallRuntime:
535 case Op_CallLeaf:
536 case Op_CallLeafNoFP:
537 // Calling C code so use C calling convention
538 save_policy = matcher._c_reg_save_policy;
539 break;
541 case Op_CallStaticJava:
542 case Op_CallDynamicJava:
543 // Calling Java code so use Java calling convention
544 save_policy = matcher._register_save_policy;
545 break;
547 default:
548 ShouldNotReachHere();
549 }
551 // When using CallRuntime mark SOE registers as killed by the call
552 // so values that could show up in the RegisterMap aren't live in a
553 // callee saved register since the register wouldn't know where to
554 // find them. CallLeaf and CallLeafNoFP are ok because they can't
555 // have debug info on them. Strictly speaking this only needs to be
556 // done for oops since idealreg2debugmask takes care of debug info
557 // references but there no way to handle oops differently than other
558 // pointers as far as the kill mask goes.
559 bool exclude_soe = op == Op_CallRuntime;
561 // If the call is a MethodHandle invoke, we need to exclude the
562 // register which is used to save the SP value over MH invokes from
563 // the mask. Otherwise this register could be used for
564 // deoptimization information.
565 if (op == Op_CallStaticJava) {
566 MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall;
567 if (mcallstaticjava->_method_handle_invoke)
568 proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
569 }
571 // Fill in the kill mask for the call
572 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
573 if( !regs.Member(r) ) { // Not already defined by the call
574 // Save-on-call register?
575 if ((save_policy[r] == 'C') ||
576 (save_policy[r] == 'A') ||
577 ((save_policy[r] == 'E') && exclude_soe)) {
578 proj->_rout.Insert(r);
579 }
580 }
581 }
583 return node_cnt;
584 }
587 //------------------------------schedule_local---------------------------------
588 // Topological sort within a block. Someday become a real scheduler.
589 bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, int *ready_cnt, VectorSet &next_call) {
590 // Already "sorted" are the block start Node (as the first entry), and
591 // the block-ending Node and any trailing control projections. We leave
592 // these alone. PhiNodes and ParmNodes are made to follow the block start
593 // Node. Everything else gets topo-sorted.
595 #ifndef PRODUCT
596 if (cfg->trace_opto_pipelining()) {
597 tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order);
598 for (uint i = 0;i < _nodes.size();i++) {
599 tty->print("# ");
600 _nodes[i]->fast_dump();
601 }
602 tty->print_cr("#");
603 }
604 #endif
606 // RootNode is already sorted
607 if( _nodes.size() == 1 ) return true;
609 // Move PhiNodes and ParmNodes from 1 to cnt up to the start
610 uint node_cnt = end_idx();
611 uint phi_cnt = 1;
612 uint i;
613 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
614 Node *n = _nodes[i];
615 if( n->is_Phi() || // Found a PhiNode or ParmNode
616 (n->is_Proj() && n->in(0) == head()) ) {
617 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
618 _nodes.map(i,_nodes[phi_cnt]);
619 _nodes.map(phi_cnt++,n); // swap Phi/Parm up front
620 } else { // All others
621 // Count block-local inputs to 'n'
622 uint cnt = n->len(); // Input count
623 uint local = 0;
624 for( uint j=0; j<cnt; j++ ) {
625 Node *m = n->in(j);
626 if( m && cfg->_bbs[m->_idx] == this && !m->is_top() )
627 local++; // One more block-local input
628 }
629 ready_cnt[n->_idx] = local; // Count em up
631 // A few node types require changing a required edge to a precedence edge
632 // before allocation.
633 if( UseConcMarkSweepGC || UseG1GC ) {
634 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
635 // Note: Required edges with an index greater than oper_input_base
636 // are not supported by the allocator.
637 // Note2: Can only depend on unmatched edge being last,
638 // can not depend on its absolute position.
639 Node *oop_store = n->in(n->req() - 1);
640 n->del_req(n->req() - 1);
641 n->add_prec(oop_store);
642 assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark");
643 }
644 }
645 if( n->is_Mach() && n->req() > TypeFunc::Parms &&
646 (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
647 n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
648 // MemBarAcquire could be created without Precedent edge.
649 // del_req() replaces the specified edge with the last input edge
650 // and then removes the last edge. If the specified edge > number of
651 // edges the last edge will be moved outside of the input edges array
652 // and the edge will be lost. This is why this code should be
653 // executed only when Precedent (== TypeFunc::Parms) edge is present.
654 Node *x = n->in(TypeFunc::Parms);
655 n->del_req(TypeFunc::Parms);
656 n->add_prec(x);
657 }
658 }
659 }
660 for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count
661 ready_cnt[_nodes[i2]->_idx] = 0;
663 // All the prescheduled guys do not hold back internal nodes
664 uint i3;
665 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled
666 Node *n = _nodes[i3]; // Get pre-scheduled
667 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
668 Node* m = n->fast_out(j);
669 if( cfg->_bbs[m->_idx] ==this ) // Local-block user
670 ready_cnt[m->_idx]--; // Fix ready count
671 }
672 }
674 Node_List delay;
675 // Make a worklist
676 Node_List worklist;
677 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist
678 Node *m = _nodes[i4];
679 if( !ready_cnt[m->_idx] ) { // Zero ready count?
680 if (m->is_iteratively_computed()) {
681 // Push induction variable increments last to allow other uses
682 // of the phi to be scheduled first. The select() method breaks
683 // ties in scheduling by worklist order.
684 delay.push(m);
685 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
686 // Force the CreateEx to the top of the list so it's processed
687 // first and ends up at the start of the block.
688 worklist.insert(0, m);
689 } else {
690 worklist.push(m); // Then on to worklist!
691 }
692 }
693 }
694 while (delay.size()) {
695 Node* d = delay.pop();
696 worklist.push(d);
697 }
699 // Warm up the 'next_call' heuristic bits
700 needed_for_next_call(_nodes[0], next_call, cfg->_bbs);
702 #ifndef PRODUCT
703 if (cfg->trace_opto_pipelining()) {
704 for (uint j=0; j<_nodes.size(); j++) {
705 Node *n = _nodes[j];
706 int idx = n->_idx;
707 tty->print("# ready cnt:%3d ", ready_cnt[idx]);
708 tty->print("latency:%3d ", cfg->_node_latency.at_grow(idx));
709 tty->print("%4d: %s\n", idx, n->Name());
710 }
711 }
712 #endif
714 // Pull from worklist and schedule
715 while( worklist.size() ) { // Worklist is not ready
717 #ifndef PRODUCT
718 if (cfg->trace_opto_pipelining()) {
719 tty->print("# ready list:");
720 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
721 Node *n = worklist[i]; // Get Node on worklist
722 tty->print(" %d", n->_idx);
723 }
724 tty->cr();
725 }
726 #endif
728 // Select and pop a ready guy from worklist
729 Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt);
730 _nodes.map(phi_cnt++,n); // Schedule him next
732 #ifndef PRODUCT
733 if (cfg->trace_opto_pipelining()) {
734 tty->print("# select %d: %s", n->_idx, n->Name());
735 tty->print(", latency:%d", cfg->_node_latency.at_grow(n->_idx));
736 n->dump();
737 if (Verbose) {
738 tty->print("# ready list:");
739 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
740 Node *n = worklist[i]; // Get Node on worklist
741 tty->print(" %d", n->_idx);
742 }
743 tty->cr();
744 }
745 }
747 #endif
748 if( n->is_MachCall() ) {
749 MachCallNode *mcall = n->as_MachCall();
750 phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call);
751 continue;
752 }
753 // Children are now all ready
754 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
755 Node* m = n->fast_out(i5); // Get user
756 if( cfg->_bbs[m->_idx] != this ) continue;
757 if( m->is_Phi() ) continue;
758 if( !--ready_cnt[m->_idx] )
759 worklist.push(m);
760 }
761 }
763 if( phi_cnt != end_idx() ) {
764 // did not schedule all. Retry, Bailout, or Die
765 Compile* C = matcher.C;
766 if (C->subsume_loads() == true && !C->failing()) {
767 // Retry with subsume_loads == false
768 // If this is the first failure, the sentinel string will "stick"
769 // to the Compile object, and the C2Compiler will see it and retry.
770 C->record_failure(C2Compiler::retry_no_subsuming_loads());
771 }
772 // assert( phi_cnt == end_idx(), "did not schedule all" );
773 return false;
774 }
776 #ifndef PRODUCT
777 if (cfg->trace_opto_pipelining()) {
778 tty->print_cr("#");
779 tty->print_cr("# after schedule_local");
780 for (uint i = 0;i < _nodes.size();i++) {
781 tty->print("# ");
782 _nodes[i]->fast_dump();
783 }
784 tty->cr();
785 }
786 #endif
789 return true;
790 }
792 //--------------------------catch_cleanup_fix_all_inputs-----------------------
793 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
794 for (uint l = 0; l < use->len(); l++) {
795 if (use->in(l) == old_def) {
796 if (l < use->req()) {
797 use->set_req(l, new_def);
798 } else {
799 use->rm_prec(l);
800 use->add_prec(new_def);
801 l--;
802 }
803 }
804 }
805 }
807 //------------------------------catch_cleanup_find_cloned_def------------------
808 static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
809 assert( use_blk != def_blk, "Inter-block cleanup only");
811 // The use is some block below the Catch. Find and return the clone of the def
812 // that dominates the use. If there is no clone in a dominating block, then
813 // create a phi for the def in a dominating block.
815 // Find which successor block dominates this use. The successor
816 // blocks must all be single-entry (from the Catch only; I will have
817 // split blocks to make this so), hence they all dominate.
818 while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
819 use_blk = use_blk->_idom;
821 // Find the successor
822 Node *fixup = NULL;
824 uint j;
825 for( j = 0; j < def_blk->_num_succs; j++ )
826 if( use_blk == def_blk->_succs[j] )
827 break;
829 if( j == def_blk->_num_succs ) {
830 // Block at same level in dom-tree is not a successor. It needs a
831 // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
832 Node_Array inputs = new Node_List(Thread::current()->resource_area());
833 for(uint k = 1; k < use_blk->num_preds(); k++) {
834 inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx));
835 }
837 // Check to see if the use_blk already has an identical phi inserted.
838 // If it exists, it will be at the first position since all uses of a
839 // def are processed together.
840 Node *phi = use_blk->_nodes[1];
841 if( phi->is_Phi() ) {
842 fixup = phi;
843 for (uint k = 1; k < use_blk->num_preds(); k++) {
844 if (phi->in(k) != inputs[k]) {
845 // Not a match
846 fixup = NULL;
847 break;
848 }
849 }
850 }
852 // If an existing PhiNode was not found, make a new one.
853 if (fixup == NULL) {
854 Node *new_phi = PhiNode::make(use_blk->head(), def);
855 use_blk->_nodes.insert(1, new_phi);
856 bbs.map(new_phi->_idx, use_blk);
857 for (uint k = 1; k < use_blk->num_preds(); k++) {
858 new_phi->set_req(k, inputs[k]);
859 }
860 fixup = new_phi;
861 }
863 } else {
864 // Found the use just below the Catch. Make it use the clone.
865 fixup = use_blk->_nodes[n_clone_idx];
866 }
868 return fixup;
869 }
871 //--------------------------catch_cleanup_intra_block--------------------------
872 // Fix all input edges in use that reference "def". The use is in the same
873 // block as the def and both have been cloned in each successor block.
874 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
876 // Both the use and def have been cloned. For each successor block,
877 // get the clone of the use, and make its input the clone of the def
878 // found in that block.
880 uint use_idx = blk->find_node(use);
881 uint offset_idx = use_idx - beg;
882 for( uint k = 0; k < blk->_num_succs; k++ ) {
883 // Get clone in each successor block
884 Block *sb = blk->_succs[k];
885 Node *clone = sb->_nodes[offset_idx+1];
886 assert( clone->Opcode() == use->Opcode(), "" );
888 // Make use-clone reference the def-clone
889 catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]);
890 }
891 }
893 //------------------------------catch_cleanup_inter_block---------------------
894 // Fix all input edges in use that reference "def". The use is in a different
895 // block than the def.
896 static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
897 if( !use_blk ) return; // Can happen if the use is a precedence edge
899 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx);
900 catch_cleanup_fix_all_inputs(use, def, new_def);
901 }
903 //------------------------------call_catch_cleanup-----------------------------
904 // If we inserted any instructions between a Call and his CatchNode,
905 // clone the instructions on all paths below the Catch.
906 void Block::call_catch_cleanup(Block_Array &bbs) {
908 // End of region to clone
909 uint end = end_idx();
910 if( !_nodes[end]->is_Catch() ) return;
911 // Start of region to clone
912 uint beg = end;
913 while( _nodes[beg-1]->Opcode() != Op_MachProj ||
914 !_nodes[beg-1]->in(0)->is_Call() ) {
915 beg--;
916 assert(beg > 0,"Catch cleanup walking beyond block boundary");
917 }
918 // Range of inserted instructions is [beg, end)
919 if( beg == end ) return;
921 // Clone along all Catch output paths. Clone area between the 'beg' and
922 // 'end' indices.
923 for( uint i = 0; i < _num_succs; i++ ) {
924 Block *sb = _succs[i];
925 // Clone the entire area; ignoring the edge fixup for now.
926 for( uint j = end; j > beg; j-- ) {
927 Node *clone = _nodes[j-1]->clone();
928 sb->_nodes.insert( 1, clone );
929 bbs.map(clone->_idx,sb);
930 }
931 }
934 // Fixup edges. Check the def-use info per cloned Node
935 for(uint i2 = beg; i2 < end; i2++ ) {
936 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
937 Node *n = _nodes[i2]; // Node that got cloned
938 // Need DU safe iterator because of edge manipulation in calls.
939 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
940 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
941 out->push(n->fast_out(j1));
942 }
943 uint max = out->size();
944 for (uint j = 0; j < max; j++) {// For all users
945 Node *use = out->pop();
946 Block *buse = bbs[use->_idx];
947 if( use->is_Phi() ) {
948 for( uint k = 1; k < use->req(); k++ )
949 if( use->in(k) == n ) {
950 Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx);
951 use->set_req(k, fixup);
952 }
953 } else {
954 if (this == buse) {
955 catch_cleanup_intra_block(use, n, this, beg, n_clone_idx);
956 } else {
957 catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx);
958 }
959 }
960 } // End for all users
962 } // End of for all Nodes in cloned area
964 // Remove the now-dead cloned ops
965 for(uint i3 = beg; i3 < end; i3++ ) {
966 _nodes[beg]->disconnect_inputs(NULL);
967 _nodes.remove(beg);
968 }
970 // If the successor blocks have a CreateEx node, move it back to the top
971 for(uint i4 = 0; i4 < _num_succs; i4++ ) {
972 Block *sb = _succs[i4];
973 uint new_cnt = end - beg;
974 // Remove any newly created, but dead, nodes.
975 for( uint j = new_cnt; j > 0; j-- ) {
976 Node *n = sb->_nodes[j];
977 if (n->outcnt() == 0 &&
978 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
979 n->disconnect_inputs(NULL);
980 sb->_nodes.remove(j);
981 new_cnt--;
982 }
983 }
984 // If any newly created nodes remain, move the CreateEx node to the top
985 if (new_cnt > 0) {
986 Node *cex = sb->_nodes[1+new_cnt];
987 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
988 sb->_nodes.remove(1+new_cnt);
989 sb->_nodes.insert(1,cex);
990 }
991 }
992 }
993 }