Mercurial > hg > truffle
annotate src/share/vm/opto/lcm.cpp @ 169:9148c65abefc
6695049: (coll) Create an x86 intrinsic for Arrays.equals
Summary: Intrinsify java/util/Arrays.equals(char[], char[])
Reviewed-by: kvn, never
author | rasbold |
---|---|
date | Thu, 29 May 2008 16:22:09 -0700 |
parents | c436414a719e |
children | d1605aabd0a1 6aae2f9d0294 |
rev | line source |
---|---|
0 | 1 /* |
2 * Copyright 1998-2007 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 */ | |
24 | |
25 // Optimization - Graph Style | |
26 | |
27 #include "incls/_precompiled.incl" | |
28 #include "incls/_lcm.cpp.incl" | |
29 | |
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; | |
41 | |
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; | |
46 | |
47 uint bidx = 0; // Capture index of value into memop | |
48 bool was_store; // Memory op is a store op | |
49 | |
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 | |
62 // Search the exception block for an uncommon trap. | |
63 // (See Parse::do_if and Parse::do_ifnull for the reason | |
64 // we need an uncommon trap. Briefly, we need a way to | |
65 // detect failure of this optimization, as in 6366351.) | |
66 { | |
67 bool found_trap = false; | |
68 for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) { | |
69 Node* nn = null_block->_nodes[i1]; | |
70 if (nn->is_MachCall() && | |
71 nn->as_MachCall()->entry_point() == | |
72 SharedRuntime::uncommon_trap_blob()->instructions_begin()) { | |
73 const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type(); | |
74 if (trtype->isa_int() && trtype->is_int()->is_con()) { | |
75 jint tr_con = trtype->is_int()->get_con(); | |
76 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); | |
77 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); | |
78 assert((int)reason < (int)BitsPerInt, "recode bit map"); | |
79 if (is_set_nth_bit(allowed_reasons, (int) reason) | |
80 && action != Deoptimization::Action_none) { | |
81 // This uncommon trap is sure to recompile, eventually. | |
82 // When that happens, C->too_many_traps will prevent | |
83 // this transformation from happening again. | |
84 found_trap = true; | |
85 } | |
86 } | |
87 break; | |
88 } | |
89 } | |
90 if (!found_trap) { | |
91 // We did not find an uncommon trap. | |
92 return; | |
93 } | |
94 } | |
95 | |
96 // Search the successor block for a load or store who's base value is also | |
97 // the tested value. There may be several. | |
98 Node_List *out = new Node_List(Thread::current()->resource_area()); | |
99 MachNode *best = NULL; // Best found so far | |
100 for (DUIterator i = val->outs(); val->has_out(i); i++) { | |
101 Node *m = val->out(i); | |
102 if( !m->is_Mach() ) continue; | |
103 MachNode *mach = m->as_Mach(); | |
104 was_store = false; | |
105 switch( mach->ideal_Opcode() ) { | |
106 case Op_LoadB: | |
107 case Op_LoadC: | |
108 case Op_LoadD: | |
109 case Op_LoadF: | |
110 case Op_LoadI: | |
111 case Op_LoadL: | |
112 case Op_LoadP: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
0
diff
changeset
|
113 case Op_LoadN: |
0 | 114 case Op_LoadS: |
115 case Op_LoadKlass: | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
125
diff
changeset
|
116 case Op_LoadNKlass: |
0 | 117 case Op_LoadRange: |
118 case Op_LoadD_unaligned: | |
119 case Op_LoadL_unaligned: | |
120 break; | |
121 case Op_StoreB: | |
122 case Op_StoreC: | |
123 case Op_StoreCM: | |
124 case Op_StoreD: | |
125 case Op_StoreF: | |
126 case Op_StoreI: | |
127 case Op_StoreL: | |
128 case Op_StoreP: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
0
diff
changeset
|
129 case Op_StoreN: |
0 | 130 was_store = true; // Memory op is a store op |
131 // Stores will have their address in slot 2 (memory in slot 1). | |
132 // If the value being nul-checked is in another slot, it means we | |
133 // are storing the checked value, which does NOT check the value! | |
134 if( mach->in(2) != val ) continue; | |
135 break; // Found a memory op? | |
136 case Op_StrComp: | |
169
9148c65abefc
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
164
diff
changeset
|
137 case Op_AryEq: |
0 | 138 // Not a legit memory op for implicit null check regardless of |
139 // embedded loads | |
140 continue; | |
141 default: // Also check for embedded loads | |
142 if( !mach->needs_anti_dependence_check() ) | |
143 continue; // Not an memory op; skip it | |
144 break; | |
145 } | |
146 // check if the offset is not too high for implicit exception | |
147 { | |
148 intptr_t offset = 0; | |
149 const TypePtr *adr_type = NULL; // Do not need this return value here | |
150 const Node* base = mach->get_base_and_disp(offset, adr_type); | |
151 if (base == NULL || base == NodeSentinel) { | |
152 // cannot reason about it; is probably not implicit null exception | |
153 } else { | |
154 const TypePtr* tptr = base->bottom_type()->is_ptr(); | |
155 // Give up if offset is not a compile-time constant | |
156 if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot ) | |
157 continue; | |
158 offset += tptr->_offset; // correct if base is offseted | |
159 if( MacroAssembler::needs_explicit_null_check(offset) ) | |
160 continue; // Give up is reference is beyond 4K page size | |
161 } | |
162 } | |
163 | |
164 // Check ctrl input to see if the null-check dominates the memory op | |
165 Block *cb = cfg->_bbs[mach->_idx]; | |
166 cb = cb->_idom; // Always hoist at least 1 block | |
167 if( !was_store ) { // Stores can be hoisted only one block | |
168 while( cb->_dom_depth > (_dom_depth + 1)) | |
169 cb = cb->_idom; // Hoist loads as far as we want | |
170 // The non-null-block should dominate the memory op, too. Live | |
171 // range spilling will insert a spill in the non-null-block if it is | |
172 // needs to spill the memory op for an implicit null check. | |
173 if (cb->_dom_depth == (_dom_depth + 1)) { | |
174 if (cb != not_null_block) continue; | |
175 cb = cb->_idom; | |
176 } | |
177 } | |
178 if( cb != this ) continue; | |
179 | |
180 // Found a memory user; see if it can be hoisted to check-block | |
181 uint vidx = 0; // Capture index of value into memop | |
182 uint j; | |
183 for( j = mach->req()-1; j > 0; j-- ) { | |
184 if( mach->in(j) == val ) vidx = j; | |
185 // Block of memory-op input | |
186 Block *inb = cfg->_bbs[mach->in(j)->_idx]; | |
187 Block *b = this; // Start from nul check | |
188 while( b != inb && b->_dom_depth > inb->_dom_depth ) | |
189 b = b->_idom; // search upwards for input | |
190 // See if input dominates null check | |
191 if( b != inb ) | |
192 break; | |
193 } | |
194 if( j > 0 ) | |
195 continue; | |
196 Block *mb = cfg->_bbs[mach->_idx]; | |
197 // Hoisting stores requires more checks for the anti-dependence case. | |
198 // Give up hoisting if we have to move the store past any load. | |
199 if( was_store ) { | |
200 Block *b = mb; // Start searching here for a local load | |
201 // mach use (faulting) trying to hoist | |
202 // n might be blocker to hoisting | |
203 while( b != this ) { | |
204 uint k; | |
205 for( k = 1; k < b->_nodes.size(); k++ ) { | |
206 Node *n = b->_nodes[k]; | |
207 if( n->needs_anti_dependence_check() && | |
208 n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) | |
209 break; // Found anti-dependent load | |
210 } | |
211 if( k < b->_nodes.size() ) | |
212 break; // Found anti-dependent load | |
213 // Make sure control does not do a merge (would have to check allpaths) | |
214 if( b->num_preds() != 2 ) break; | |
215 b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block | |
216 } | |
217 if( b != this ) continue; | |
218 } | |
219 | |
220 // Make sure this memory op is not already being used for a NullCheck | |
221 Node *e = mb->end(); | |
222 if( e->is_MachNullCheck() && e->in(1) == mach ) | |
223 continue; // Already being used as a NULL check | |
224 | |
225 // Found a candidate! Pick one with least dom depth - the highest | |
226 // in the dom tree should be closest to the null check. | |
227 if( !best || | |
228 cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) { | |
229 best = mach; | |
230 bidx = vidx; | |
231 | |
232 } | |
233 } | |
234 // No candidate! | |
235 if( !best ) return; | |
236 | |
237 // ---- Found an implicit null check | |
238 extern int implicit_null_checks; | |
239 implicit_null_checks++; | |
240 | |
241 // Hoist the memory candidate up to the end of the test block. | |
242 Block *old_block = cfg->_bbs[best->_idx]; | |
243 old_block->find_remove(best); | |
244 add_inst(best); | |
245 cfg->_bbs.map(best->_idx,this); | |
246 | |
247 // Move the control dependence | |
248 if (best->in(0) && best->in(0) == old_block->_nodes[0]) | |
249 best->set_req(0, _nodes[0]); | |
250 | |
251 // Check for flag-killing projections that also need to be hoisted | |
252 // Should be DU safe because no edge updates. | |
253 for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { | |
254 Node* n = best->fast_out(j); | |
255 if( n->Opcode() == Op_MachProj ) { | |
256 cfg->_bbs[n->_idx]->find_remove(n); | |
257 add_inst(n); | |
258 cfg->_bbs.map(n->_idx,this); | |
259 } | |
260 } | |
261 | |
262 Compile *C = cfg->C; | |
263 // proj==Op_True --> ne test; proj==Op_False --> eq test. | |
264 // One of two graph shapes got matched: | |
265 // (IfTrue (If (Bool NE (CmpP ptr NULL)))) | |
266 // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) | |
267 // NULL checks are always branch-if-eq. If we see a IfTrue projection | |
268 // then we are replacing a 'ne' test with a 'eq' NULL check test. | |
269 // We need to flip the projections to keep the same semantics. | |
270 if( proj->Opcode() == Op_IfTrue ) { | |
271 // Swap order of projections in basic block to swap branch targets | |
272 Node *tmp1 = _nodes[end_idx()+1]; | |
273 Node *tmp2 = _nodes[end_idx()+2]; | |
274 _nodes.map(end_idx()+1, tmp2); | |
275 _nodes.map(end_idx()+2, tmp1); | |
276 Node *tmp = new (C, 1) Node(C->top()); // Use not NULL input | |
277 tmp1->replace_by(tmp); | |
278 tmp2->replace_by(tmp1); | |
279 tmp->replace_by(tmp2); | |
280 tmp->destruct(); | |
281 } | |
282 | |
283 // Remove the existing null check; use a new implicit null check instead. | |
284 // Since schedule-local needs precise def-use info, we need to correct | |
285 // it as well. | |
286 Node *old_tst = proj->in(0); | |
287 MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx); | |
288 _nodes.map(end_idx(),nul_chk); | |
289 cfg->_bbs.map(nul_chk->_idx,this); | |
290 // Redirect users of old_test to nul_chk | |
291 for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) | |
292 old_tst->last_out(i2)->set_req(0, nul_chk); | |
293 // Clean-up any dead code | |
294 for (uint i3 = 0; i3 < old_tst->req(); i3++) | |
295 old_tst->set_req(i3, NULL); | |
296 | |
297 cfg->latency_from_uses(nul_chk); | |
298 cfg->latency_from_uses(best); | |
299 } | |
300 | |
301 | |
302 //------------------------------select----------------------------------------- | |
303 // Select a nice fellow from the worklist to schedule next. If there is only | |
304 // one choice, then use it. Projections take top priority for correctness | |
305 // reasons - if I see a projection, then it is next. There are a number of | |
306 // other special cases, for instructions that consume condition codes, et al. | |
307 // These are chosen immediately. Some instructions are required to immediately | |
308 // precede the last instruction in the block, and these are taken last. Of the | |
309 // remaining cases (most), choose the instruction with the greatest latency | |
310 // (that is, the most number of pseudo-cycles required to the end of the | |
311 // routine). If there is a tie, choose the instruction with the most inputs. | |
312 Node *Block::select(PhaseCFG *cfg, Node_List &worklist, int *ready_cnt, VectorSet &next_call, uint sched_slot) { | |
313 | |
314 // If only a single entry on the stack, use it | |
315 uint cnt = worklist.size(); | |
316 if (cnt == 1) { | |
317 Node *n = worklist[0]; | |
318 worklist.map(0,worklist.pop()); | |
319 return n; | |
320 } | |
321 | |
322 uint choice = 0; // Bigger is most important | |
323 uint latency = 0; // Bigger is scheduled first | |
324 uint score = 0; // Bigger is better | |
325 uint idx; // Index in worklist | |
326 | |
327 for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist | |
328 // Order in worklist is used to break ties. | |
329 // See caller for how this is used to delay scheduling | |
330 // of induction variable increments to after the other | |
331 // uses of the phi are scheduled. | |
332 Node *n = worklist[i]; // Get Node on worklist | |
333 | |
334 int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; | |
335 if( n->is_Proj() || // Projections always win | |
336 n->Opcode()== Op_Con || // So does constant 'Top' | |
337 iop == Op_CreateEx || // Create-exception must start block | |
338 iop == Op_CheckCastPP | |
339 ) { | |
340 worklist.map(i,worklist.pop()); | |
341 return n; | |
342 } | |
343 | |
344 // Final call in a block must be adjacent to 'catch' | |
345 Node *e = end(); | |
346 if( e->is_Catch() && e->in(0)->in(0) == n ) | |
347 continue; | |
348 | |
349 // Memory op for an implicit null check has to be at the end of the block | |
350 if( e->is_MachNullCheck() && e->in(1) == n ) | |
351 continue; | |
352 | |
353 uint n_choice = 2; | |
354 | |
355 // See if this instruction is consumed by a branch. If so, then (as the | |
356 // branch is the last instruction in the basic block) force it to the | |
357 // end of the basic block | |
358 if ( must_clone[iop] ) { | |
359 // See if any use is a branch | |
360 bool found_machif = false; | |
361 | |
362 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { | |
363 Node* use = n->fast_out(j); | |
364 | |
365 // The use is a conditional branch, make them adjacent | |
366 if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) { | |
367 found_machif = true; | |
368 break; | |
369 } | |
370 | |
371 // More than this instruction pending for successor to be ready, | |
372 // don't choose this if other opportunities are ready | |
373 if (ready_cnt[use->_idx] > 1) | |
374 n_choice = 1; | |
375 } | |
376 | |
377 // loop terminated, prefer not to use this instruction | |
378 if (found_machif) | |
379 continue; | |
380 } | |
381 | |
382 // See if this has a predecessor that is "must_clone", i.e. sets the | |
383 // condition code. If so, choose this first | |
384 for (uint j = 0; j < n->req() ; j++) { | |
385 Node *inn = n->in(j); | |
386 if (inn) { | |
387 if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { | |
388 n_choice = 3; | |
389 break; | |
390 } | |
391 } | |
392 } | |
393 | |
394 // MachTemps should be scheduled last so they are near their uses | |
395 if (n->is_MachTemp()) { | |
396 n_choice = 1; | |
397 } | |
398 | |
399 uint n_latency = cfg->_node_latency.at_grow(n->_idx); | |
400 uint n_score = n->req(); // Many inputs get high score to break ties | |
401 | |
402 // Keep best latency found | |
403 if( choice < n_choice || | |
404 ( choice == n_choice && | |
405 ( latency < n_latency || | |
406 ( latency == n_latency && | |
407 ( score < n_score ))))) { | |
408 choice = n_choice; | |
409 latency = n_latency; | |
410 score = n_score; | |
411 idx = i; // Also keep index in worklist | |
412 } | |
413 } // End of for all ready nodes in worklist | |
414 | |
415 Node *n = worklist[idx]; // Get the winner | |
416 | |
417 worklist.map(idx,worklist.pop()); // Compress worklist | |
418 return n; | |
419 } | |
420 | |
421 | |
422 //------------------------------set_next_call---------------------------------- | |
423 void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) { | |
424 if( next_call.test_set(n->_idx) ) return; | |
425 for( uint i=0; i<n->len(); i++ ) { | |
426 Node *m = n->in(i); | |
427 if( !m ) continue; // must see all nodes in block that precede call | |
428 if( bbs[m->_idx] == this ) | |
429 set_next_call( m, next_call, bbs ); | |
430 } | |
431 } | |
432 | |
433 //------------------------------needed_for_next_call--------------------------- | |
434 // Set the flag 'next_call' for each Node that is needed for the next call to | |
435 // be scheduled. This flag lets me bias scheduling so Nodes needed for the | |
436 // next subroutine call get priority - basically it moves things NOT needed | |
437 // for the next call till after the call. This prevents me from trying to | |
438 // carry lots of stuff live across a call. | |
439 void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) { | |
440 // Find the next control-defining Node in this block | |
441 Node* call = NULL; | |
442 for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { | |
443 Node* m = this_call->fast_out(i); | |
444 if( bbs[m->_idx] == this && // Local-block user | |
445 m != this_call && // Not self-start node | |
446 m->is_Call() ) | |
447 call = m; | |
448 break; | |
449 } | |
450 if (call == NULL) return; // No next call (e.g., block end is near) | |
451 // Set next-call for all inputs to this call | |
452 set_next_call(call, next_call, bbs); | |
453 } | |
454 | |
455 //------------------------------sched_call------------------------------------- | |
456 uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, int *ready_cnt, MachCallNode *mcall, VectorSet &next_call ) { | |
457 RegMask regs; | |
458 | |
459 // Schedule all the users of the call right now. All the users are | |
460 // projection Nodes, so they must be scheduled next to the call. | |
461 // Collect all the defined registers. | |
462 for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { | |
463 Node* n = mcall->fast_out(i); | |
464 assert( n->Opcode()==Op_MachProj, "" ); | |
465 --ready_cnt[n->_idx]; | |
466 assert( !ready_cnt[n->_idx], "" ); | |
467 // Schedule next to call | |
468 _nodes.map(node_cnt++, n); | |
469 // Collect defined registers | |
470 regs.OR(n->out_RegMask()); | |
471 // Check for scheduling the next control-definer | |
472 if( n->bottom_type() == Type::CONTROL ) | |
473 // Warm up next pile of heuristic bits | |
474 needed_for_next_call(n, next_call, bbs); | |
475 | |
476 // Children of projections are now all ready | |
477 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { | |
478 Node* m = n->fast_out(j); // Get user | |
479 if( bbs[m->_idx] != this ) continue; | |
480 if( m->is_Phi() ) continue; | |
481 if( !--ready_cnt[m->_idx] ) | |
482 worklist.push(m); | |
483 } | |
484 | |
485 } | |
486 | |
487 // Act as if the call defines the Frame Pointer. | |
488 // Certainly the FP is alive and well after the call. | |
489 regs.Insert(matcher.c_frame_pointer()); | |
490 | |
491 // Set all registers killed and not already defined by the call. | |
492 uint r_cnt = mcall->tf()->range()->cnt(); | |
493 int op = mcall->ideal_Opcode(); | |
494 MachProjNode *proj = new (matcher.C, 1) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); | |
495 bbs.map(proj->_idx,this); | |
496 _nodes.insert(node_cnt++, proj); | |
497 | |
498 // Select the right register save policy. | |
499 const char * save_policy; | |
500 switch (op) { | |
501 case Op_CallRuntime: | |
502 case Op_CallLeaf: | |
503 case Op_CallLeafNoFP: | |
504 // Calling C code so use C calling convention | |
505 save_policy = matcher._c_reg_save_policy; | |
506 break; | |
507 | |
508 case Op_CallStaticJava: | |
509 case Op_CallDynamicJava: | |
510 // Calling Java code so use Java calling convention | |
511 save_policy = matcher._register_save_policy; | |
512 break; | |
513 | |
514 default: | |
515 ShouldNotReachHere(); | |
516 } | |
517 | |
518 // When using CallRuntime mark SOE registers as killed by the call | |
519 // so values that could show up in the RegisterMap aren't live in a | |
520 // callee saved register since the register wouldn't know where to | |
521 // find them. CallLeaf and CallLeafNoFP are ok because they can't | |
522 // have debug info on them. Strictly speaking this only needs to be | |
523 // done for oops since idealreg2debugmask takes care of debug info | |
524 // references but there no way to handle oops differently than other | |
525 // pointers as far as the kill mask goes. | |
526 bool exclude_soe = op == Op_CallRuntime; | |
527 | |
528 // Fill in the kill mask for the call | |
529 for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { | |
530 if( !regs.Member(r) ) { // Not already defined by the call | |
531 // Save-on-call register? | |
532 if ((save_policy[r] == 'C') || | |
533 (save_policy[r] == 'A') || | |
534 ((save_policy[r] == 'E') && exclude_soe)) { | |
535 proj->_rout.Insert(r); | |
536 } | |
537 } | |
538 } | |
539 | |
540 return node_cnt; | |
541 } | |
542 | |
543 | |
544 //------------------------------schedule_local--------------------------------- | |
545 // Topological sort within a block. Someday become a real scheduler. | |
546 bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, int *ready_cnt, VectorSet &next_call) { | |
547 // Already "sorted" are the block start Node (as the first entry), and | |
548 // the block-ending Node and any trailing control projections. We leave | |
549 // these alone. PhiNodes and ParmNodes are made to follow the block start | |
550 // Node. Everything else gets topo-sorted. | |
551 | |
552 #ifndef PRODUCT | |
553 if (cfg->trace_opto_pipelining()) { | |
554 tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order); | |
555 for (uint i = 0;i < _nodes.size();i++) { | |
556 tty->print("# "); | |
557 _nodes[i]->fast_dump(); | |
558 } | |
559 tty->print_cr("#"); | |
560 } | |
561 #endif | |
562 | |
563 // RootNode is already sorted | |
564 if( _nodes.size() == 1 ) return true; | |
565 | |
566 // Move PhiNodes and ParmNodes from 1 to cnt up to the start | |
567 uint node_cnt = end_idx(); | |
568 uint phi_cnt = 1; | |
569 uint i; | |
570 for( i = 1; i<node_cnt; i++ ) { // Scan for Phi | |
571 Node *n = _nodes[i]; | |
572 if( n->is_Phi() || // Found a PhiNode or ParmNode | |
573 (n->is_Proj() && n->in(0) == head()) ) { | |
574 // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt | |
575 _nodes.map(i,_nodes[phi_cnt]); | |
576 _nodes.map(phi_cnt++,n); // swap Phi/Parm up front | |
577 } else { // All others | |
578 // Count block-local inputs to 'n' | |
579 uint cnt = n->len(); // Input count | |
580 uint local = 0; | |
581 for( uint j=0; j<cnt; j++ ) { | |
582 Node *m = n->in(j); | |
583 if( m && cfg->_bbs[m->_idx] == this && !m->is_top() ) | |
584 local++; // One more block-local input | |
585 } | |
586 ready_cnt[n->_idx] = local; // Count em up | |
587 | |
588 // A few node types require changing a required edge to a precedence edge | |
589 // before allocation. | |
590 if( UseConcMarkSweepGC ) { | |
591 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { | |
592 // Note: Required edges with an index greater than oper_input_base | |
593 // are not supported by the allocator. | |
594 // Note2: Can only depend on unmatched edge being last, | |
595 // can not depend on its absolute position. | |
596 Node *oop_store = n->in(n->req() - 1); | |
597 n->del_req(n->req() - 1); | |
598 n->add_prec(oop_store); | |
599 assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark"); | |
600 } | |
601 } | |
602 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ) { | |
603 Node *x = n->in(TypeFunc::Parms); | |
604 n->del_req(TypeFunc::Parms); | |
605 n->add_prec(x); | |
606 } | |
607 } | |
608 } | |
609 for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count | |
610 ready_cnt[_nodes[i2]->_idx] = 0; | |
611 | |
612 // All the prescheduled guys do not hold back internal nodes | |
613 uint i3; | |
614 for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled | |
615 Node *n = _nodes[i3]; // Get pre-scheduled | |
616 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { | |
617 Node* m = n->fast_out(j); | |
618 if( cfg->_bbs[m->_idx] ==this ) // Local-block user | |
619 ready_cnt[m->_idx]--; // Fix ready count | |
620 } | |
621 } | |
622 | |
623 Node_List delay; | |
624 // Make a worklist | |
625 Node_List worklist; | |
626 for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist | |
627 Node *m = _nodes[i4]; | |
628 if( !ready_cnt[m->_idx] ) { // Zero ready count? | |
629 if (m->is_iteratively_computed()) { | |
630 // Push induction variable increments last to allow other uses | |
631 // of the phi to be scheduled first. The select() method breaks | |
632 // ties in scheduling by worklist order. | |
633 delay.push(m); | |
125
d942c7e64bd9
6601321: Assert(j == 1 || b->_nodes[j-1]->is_Phi(),"CreateEx must be first instruction in block")
never
parents:
113
diff
changeset
|
634 } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { |
d942c7e64bd9
6601321: Assert(j == 1 || b->_nodes[j-1]->is_Phi(),"CreateEx must be first instruction in block")
never
parents:
113
diff
changeset
|
635 // Force the CreateEx to the top of the list so it's processed |
d942c7e64bd9
6601321: Assert(j == 1 || b->_nodes[j-1]->is_Phi(),"CreateEx must be first instruction in block")
never
parents:
113
diff
changeset
|
636 // first and ends up at the start of the block. |
d942c7e64bd9
6601321: Assert(j == 1 || b->_nodes[j-1]->is_Phi(),"CreateEx must be first instruction in block")
never
parents:
113
diff
changeset
|
637 worklist.insert(0, m); |
0 | 638 } else { |
639 worklist.push(m); // Then on to worklist! | |
640 } | |
641 } | |
642 } | |
643 while (delay.size()) { | |
644 Node* d = delay.pop(); | |
645 worklist.push(d); | |
646 } | |
647 | |
648 // Warm up the 'next_call' heuristic bits | |
649 needed_for_next_call(_nodes[0], next_call, cfg->_bbs); | |
650 | |
651 #ifndef PRODUCT | |
652 if (cfg->trace_opto_pipelining()) { | |
653 for (uint j=0; j<_nodes.size(); j++) { | |
654 Node *n = _nodes[j]; | |
655 int idx = n->_idx; | |
656 tty->print("# ready cnt:%3d ", ready_cnt[idx]); | |
657 tty->print("latency:%3d ", cfg->_node_latency.at_grow(idx)); | |
658 tty->print("%4d: %s\n", idx, n->Name()); | |
659 } | |
660 } | |
661 #endif | |
662 | |
663 // Pull from worklist and schedule | |
664 while( worklist.size() ) { // Worklist is not ready | |
665 | |
666 #ifndef PRODUCT | |
667 if (cfg->trace_opto_pipelining()) { | |
668 tty->print("# ready list:"); | |
669 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist | |
670 Node *n = worklist[i]; // Get Node on worklist | |
671 tty->print(" %d", n->_idx); | |
672 } | |
673 tty->cr(); | |
674 } | |
675 #endif | |
676 | |
677 // Select and pop a ready guy from worklist | |
678 Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt); | |
679 _nodes.map(phi_cnt++,n); // Schedule him next | |
680 | |
681 #ifndef PRODUCT | |
682 if (cfg->trace_opto_pipelining()) { | |
683 tty->print("# select %d: %s", n->_idx, n->Name()); | |
684 tty->print(", latency:%d", cfg->_node_latency.at_grow(n->_idx)); | |
685 n->dump(); | |
686 if (Verbose) { | |
687 tty->print("# ready list:"); | |
688 for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist | |
689 Node *n = worklist[i]; // Get Node on worklist | |
690 tty->print(" %d", n->_idx); | |
691 } | |
692 tty->cr(); | |
693 } | |
694 } | |
695 | |
696 #endif | |
697 if( n->is_MachCall() ) { | |
698 MachCallNode *mcall = n->as_MachCall(); | |
699 phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call); | |
700 continue; | |
701 } | |
702 // Children are now all ready | |
703 for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { | |
704 Node* m = n->fast_out(i5); // Get user | |
705 if( cfg->_bbs[m->_idx] != this ) continue; | |
706 if( m->is_Phi() ) continue; | |
707 if( !--ready_cnt[m->_idx] ) | |
708 worklist.push(m); | |
709 } | |
710 } | |
711 | |
712 if( phi_cnt != end_idx() ) { | |
713 // did not schedule all. Retry, Bailout, or Die | |
714 Compile* C = matcher.C; | |
715 if (C->subsume_loads() == true && !C->failing()) { | |
716 // Retry with subsume_loads == false | |
717 // If this is the first failure, the sentinel string will "stick" | |
718 // to the Compile object, and the C2Compiler will see it and retry. | |
719 C->record_failure(C2Compiler::retry_no_subsuming_loads()); | |
720 } | |
721 // assert( phi_cnt == end_idx(), "did not schedule all" ); | |
722 return false; | |
723 } | |
724 | |
725 #ifndef PRODUCT | |
726 if (cfg->trace_opto_pipelining()) { | |
727 tty->print_cr("#"); | |
728 tty->print_cr("# after schedule_local"); | |
729 for (uint i = 0;i < _nodes.size();i++) { | |
730 tty->print("# "); | |
731 _nodes[i]->fast_dump(); | |
732 } | |
733 tty->cr(); | |
734 } | |
735 #endif | |
736 | |
737 | |
738 return true; | |
739 } | |
740 | |
741 //--------------------------catch_cleanup_fix_all_inputs----------------------- | |
742 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { | |
743 for (uint l = 0; l < use->len(); l++) { | |
744 if (use->in(l) == old_def) { | |
745 if (l < use->req()) { | |
746 use->set_req(l, new_def); | |
747 } else { | |
748 use->rm_prec(l); | |
749 use->add_prec(new_def); | |
750 l--; | |
751 } | |
752 } | |
753 } | |
754 } | |
755 | |
756 //------------------------------catch_cleanup_find_cloned_def------------------ | |
757 static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) { | |
758 assert( use_blk != def_blk, "Inter-block cleanup only"); | |
759 | |
760 // The use is some block below the Catch. Find and return the clone of the def | |
761 // that dominates the use. If there is no clone in a dominating block, then | |
762 // create a phi for the def in a dominating block. | |
763 | |
764 // Find which successor block dominates this use. The successor | |
765 // blocks must all be single-entry (from the Catch only; I will have | |
766 // split blocks to make this so), hence they all dominate. | |
767 while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) | |
768 use_blk = use_blk->_idom; | |
769 | |
770 // Find the successor | |
771 Node *fixup = NULL; | |
772 | |
773 uint j; | |
774 for( j = 0; j < def_blk->_num_succs; j++ ) | |
775 if( use_blk == def_blk->_succs[j] ) | |
776 break; | |
777 | |
778 if( j == def_blk->_num_succs ) { | |
779 // Block at same level in dom-tree is not a successor. It needs a | |
780 // PhiNode, the PhiNode uses from the def and IT's uses need fixup. | |
781 Node_Array inputs = new Node_List(Thread::current()->resource_area()); | |
782 for(uint k = 1; k < use_blk->num_preds(); k++) { | |
783 inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx)); | |
784 } | |
785 | |
786 // Check to see if the use_blk already has an identical phi inserted. | |
787 // If it exists, it will be at the first position since all uses of a | |
788 // def are processed together. | |
789 Node *phi = use_blk->_nodes[1]; | |
790 if( phi->is_Phi() ) { | |
791 fixup = phi; | |
792 for (uint k = 1; k < use_blk->num_preds(); k++) { | |
793 if (phi->in(k) != inputs[k]) { | |
794 // Not a match | |
795 fixup = NULL; | |
796 break; | |
797 } | |
798 } | |
799 } | |
800 | |
801 // If an existing PhiNode was not found, make a new one. | |
802 if (fixup == NULL) { | |
803 Node *new_phi = PhiNode::make(use_blk->head(), def); | |
804 use_blk->_nodes.insert(1, new_phi); | |
805 bbs.map(new_phi->_idx, use_blk); | |
806 for (uint k = 1; k < use_blk->num_preds(); k++) { | |
807 new_phi->set_req(k, inputs[k]); | |
808 } | |
809 fixup = new_phi; | |
810 } | |
811 | |
812 } else { | |
813 // Found the use just below the Catch. Make it use the clone. | |
814 fixup = use_blk->_nodes[n_clone_idx]; | |
815 } | |
816 | |
817 return fixup; | |
818 } | |
819 | |
820 //--------------------------catch_cleanup_intra_block-------------------------- | |
821 // Fix all input edges in use that reference "def". The use is in the same | |
822 // block as the def and both have been cloned in each successor block. | |
823 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { | |
824 | |
825 // Both the use and def have been cloned. For each successor block, | |
826 // get the clone of the use, and make its input the clone of the def | |
827 // found in that block. | |
828 | |
829 uint use_idx = blk->find_node(use); | |
830 uint offset_idx = use_idx - beg; | |
831 for( uint k = 0; k < blk->_num_succs; k++ ) { | |
832 // Get clone in each successor block | |
833 Block *sb = blk->_succs[k]; | |
834 Node *clone = sb->_nodes[offset_idx+1]; | |
835 assert( clone->Opcode() == use->Opcode(), "" ); | |
836 | |
837 // Make use-clone reference the def-clone | |
838 catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]); | |
839 } | |
840 } | |
841 | |
842 //------------------------------catch_cleanup_inter_block--------------------- | |
843 // Fix all input edges in use that reference "def". The use is in a different | |
844 // block than the def. | |
845 static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) { | |
846 if( !use_blk ) return; // Can happen if the use is a precedence edge | |
847 | |
848 Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx); | |
849 catch_cleanup_fix_all_inputs(use, def, new_def); | |
850 } | |
851 | |
852 //------------------------------call_catch_cleanup----------------------------- | |
853 // If we inserted any instructions between a Call and his CatchNode, | |
854 // clone the instructions on all paths below the Catch. | |
855 void Block::call_catch_cleanup(Block_Array &bbs) { | |
856 | |
857 // End of region to clone | |
858 uint end = end_idx(); | |
859 if( !_nodes[end]->is_Catch() ) return; | |
860 // Start of region to clone | |
861 uint beg = end; | |
862 while( _nodes[beg-1]->Opcode() != Op_MachProj || | |
863 !_nodes[beg-1]->in(0)->is_Call() ) { | |
864 beg--; | |
865 assert(beg > 0,"Catch cleanup walking beyond block boundary"); | |
866 } | |
867 // Range of inserted instructions is [beg, end) | |
868 if( beg == end ) return; | |
869 | |
870 // Clone along all Catch output paths. Clone area between the 'beg' and | |
871 // 'end' indices. | |
872 for( uint i = 0; i < _num_succs; i++ ) { | |
873 Block *sb = _succs[i]; | |
874 // Clone the entire area; ignoring the edge fixup for now. | |
875 for( uint j = end; j > beg; j-- ) { | |
876 Node *clone = _nodes[j-1]->clone(); | |
877 sb->_nodes.insert( 1, clone ); | |
878 bbs.map(clone->_idx,sb); | |
879 } | |
880 } | |
881 | |
882 | |
883 // Fixup edges. Check the def-use info per cloned Node | |
884 for(uint i2 = beg; i2 < end; i2++ ) { | |
885 uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block | |
886 Node *n = _nodes[i2]; // Node that got cloned | |
887 // Need DU safe iterator because of edge manipulation in calls. | |
888 Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); | |
889 for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { | |
890 out->push(n->fast_out(j1)); | |
891 } | |
892 uint max = out->size(); | |
893 for (uint j = 0; j < max; j++) {// For all users | |
894 Node *use = out->pop(); | |
895 Block *buse = bbs[use->_idx]; | |
896 if( use->is_Phi() ) { | |
897 for( uint k = 1; k < use->req(); k++ ) | |
898 if( use->in(k) == n ) { | |
899 Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx); | |
900 use->set_req(k, fixup); | |
901 } | |
902 } else { | |
903 if (this == buse) { | |
904 catch_cleanup_intra_block(use, n, this, beg, n_clone_idx); | |
905 } else { | |
906 catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx); | |
907 } | |
908 } | |
909 } // End for all users | |
910 | |
911 } // End of for all Nodes in cloned area | |
912 | |
913 // Remove the now-dead cloned ops | |
914 for(uint i3 = beg; i3 < end; i3++ ) { | |
915 _nodes[beg]->disconnect_inputs(NULL); | |
916 _nodes.remove(beg); | |
917 } | |
918 | |
919 // If the successor blocks have a CreateEx node, move it back to the top | |
920 for(uint i4 = 0; i4 < _num_succs; i4++ ) { | |
921 Block *sb = _succs[i4]; | |
922 uint new_cnt = end - beg; | |
923 // Remove any newly created, but dead, nodes. | |
924 for( uint j = new_cnt; j > 0; j-- ) { | |
925 Node *n = sb->_nodes[j]; | |
926 if (n->outcnt() == 0 && | |
927 (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ | |
928 n->disconnect_inputs(NULL); | |
929 sb->_nodes.remove(j); | |
930 new_cnt--; | |
931 } | |
932 } | |
933 // If any newly created nodes remain, move the CreateEx node to the top | |
934 if (new_cnt > 0) { | |
935 Node *cex = sb->_nodes[1+new_cnt]; | |
936 if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { | |
937 sb->_nodes.remove(1+new_cnt); | |
938 sb->_nodes.insert(1,cex); | |
939 } | |
940 } | |
941 } | |
942 } |