1 | /* |
2 | * Copyright (c) 1998, 2018, 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 | */ |
24 | |
25 | #include "precompiled.hpp" |
26 | #include "asm/macroAssembler.inline.hpp" |
27 | #include "memory/allocation.inline.hpp" |
28 | #include "oops/compressedOops.hpp" |
29 | #include "opto/ad.hpp" |
30 | #include "opto/block.hpp" |
31 | #include "opto/c2compiler.hpp" |
32 | #include "opto/callnode.hpp" |
33 | #include "opto/cfgnode.hpp" |
34 | #include "opto/machnode.hpp" |
35 | #include "opto/runtime.hpp" |
36 | #include "opto/chaitin.hpp" |
37 | #include "runtime/sharedRuntime.hpp" |
38 | |
39 | // Optimization - Graph Style |
40 | |
41 | // Check whether val is not-null-decoded compressed oop, |
42 | // i.e. will grab into the base of the heap if it represents NULL. |
43 | static bool accesses_heap_base_zone(Node *val) { |
44 | if (CompressedOops::base() != NULL) { // Implies UseCompressedOops. |
45 | if (val && val->is_Mach()) { |
46 | if (val->as_Mach()->ideal_Opcode() == Op_DecodeN) { |
47 | // This assumes all Decodes with TypePtr::NotNull are matched to nodes that |
48 | // decode NULL to point to the heap base (Decode_NN). |
49 | if (val->bottom_type()->is_oopptr()->ptr() == TypePtr::NotNull) { |
50 | return true; |
51 | } |
52 | } |
53 | // Must recognize load operation with Decode matched in memory operand. |
54 | // We should not reach here exept for PPC/AIX, as os::zero_page_read_protected() |
55 | // returns true everywhere else. On PPC, no such memory operands |
56 | // exist, therefore we did not yet implement a check for such operands. |
57 | NOT_AIX(Unimplemented()); |
58 | } |
59 | } |
60 | return false; |
61 | } |
62 | |
63 | static bool needs_explicit_null_check_for_read(Node *val) { |
64 | // On some OSes (AIX) the page at address 0 is only write protected. |
65 | // If so, only Store operations will trap. |
66 | if (os::zero_page_read_protected()) { |
67 | return false; // Implicit null check will work. |
68 | } |
69 | // Also a read accessing the base of a heap-based compressed heap will trap. |
70 | if (accesses_heap_base_zone(val) && // Hits the base zone page. |
71 | CompressedOops::use_implicit_null_checks()) { // Base zone page is protected. |
72 | return false; |
73 | } |
74 | |
75 | return true; |
76 | } |
77 | |
78 | //------------------------------implicit_null_check---------------------------- |
79 | // Detect implicit-null-check opportunities. Basically, find NULL checks |
80 | // with suitable memory ops nearby. Use the memory op to do the NULL check. |
81 | // I can generate a memory op if there is not one nearby. |
82 | // The proj is the control projection for the not-null case. |
83 | // The val is the pointer being checked for nullness or |
84 | // decodeHeapOop_not_null node if it did not fold into address. |
85 | void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) { |
86 | // Assume if null check need for 0 offset then always needed |
87 | // Intel solaris doesn't support any null checks yet and no |
88 | // mechanism exists (yet) to set the switches at an os_cpu level |
89 | if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return; |
90 | |
91 | // Make sure the ptr-is-null path appears to be uncommon! |
92 | float f = block->end()->as_MachIf()->_prob; |
93 | if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f; |
94 | if( f > PROB_UNLIKELY_MAG(4) ) return; |
95 | |
96 | uint bidx = 0; // Capture index of value into memop |
97 | bool was_store; // Memory op is a store op |
98 | |
99 | // Get the successor block for if the test ptr is non-null |
100 | Block* not_null_block; // this one goes with the proj |
101 | Block* null_block; |
102 | if (block->get_node(block->number_of_nodes()-1) == proj) { |
103 | null_block = block->_succs[0]; |
104 | not_null_block = block->_succs[1]; |
105 | } else { |
106 | assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other" ); |
107 | not_null_block = block->_succs[0]; |
108 | null_block = block->_succs[1]; |
109 | } |
110 | while (null_block->is_Empty() == Block::empty_with_goto) { |
111 | null_block = null_block->_succs[0]; |
112 | } |
113 | |
114 | // Search the exception block for an uncommon trap. |
115 | // (See Parse::do_if and Parse::do_ifnull for the reason |
116 | // we need an uncommon trap. Briefly, we need a way to |
117 | // detect failure of this optimization, as in 6366351.) |
118 | { |
119 | bool found_trap = false; |
120 | for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) { |
121 | Node* nn = null_block->get_node(i1); |
122 | if (nn->is_MachCall() && |
123 | nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { |
124 | const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type(); |
125 | if (trtype->isa_int() && trtype->is_int()->is_con()) { |
126 | jint tr_con = trtype->is_int()->get_con(); |
127 | Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); |
128 | Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); |
129 | assert((int)reason < (int)BitsPerInt, "recode bit map" ); |
130 | if (is_set_nth_bit(allowed_reasons, (int) reason) |
131 | && action != Deoptimization::Action_none) { |
132 | // This uncommon trap is sure to recompile, eventually. |
133 | // When that happens, C->too_many_traps will prevent |
134 | // this transformation from happening again. |
135 | found_trap = true; |
136 | } |
137 | } |
138 | break; |
139 | } |
140 | } |
141 | if (!found_trap) { |
142 | // We did not find an uncommon trap. |
143 | return; |
144 | } |
145 | } |
146 | |
147 | // Check for decodeHeapOop_not_null node which did not fold into address |
148 | bool is_decoden = ((intptr_t)val) & 1; |
149 | val = (Node*)(((intptr_t)val) & ~1); |
150 | |
151 | assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() && |
152 | (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity" ); |
153 | |
154 | // Search the successor block for a load or store who's base value is also |
155 | // the tested value. There may be several. |
156 | Node_List *out = new Node_List(Thread::current()->resource_area()); |
157 | MachNode *best = NULL; // Best found so far |
158 | for (DUIterator i = val->outs(); val->has_out(i); i++) { |
159 | Node *m = val->out(i); |
160 | if( !m->is_Mach() ) continue; |
161 | MachNode *mach = m->as_Mach(); |
162 | was_store = false; |
163 | int iop = mach->ideal_Opcode(); |
164 | switch( iop ) { |
165 | case Op_LoadB: |
166 | case Op_LoadUB: |
167 | case Op_LoadUS: |
168 | case Op_LoadD: |
169 | case Op_LoadF: |
170 | case Op_LoadI: |
171 | case Op_LoadL: |
172 | case Op_LoadP: |
173 | case Op_LoadN: |
174 | case Op_LoadS: |
175 | case Op_LoadKlass: |
176 | case Op_LoadNKlass: |
177 | case Op_LoadRange: |
178 | case Op_LoadD_unaligned: |
179 | case Op_LoadL_unaligned: |
180 | assert(mach->in(2) == val, "should be address" ); |
181 | break; |
182 | case Op_StoreB: |
183 | case Op_StoreC: |
184 | case Op_StoreCM: |
185 | case Op_StoreD: |
186 | case Op_StoreF: |
187 | case Op_StoreI: |
188 | case Op_StoreL: |
189 | case Op_StoreP: |
190 | case Op_StoreN: |
191 | case Op_StoreNKlass: |
192 | was_store = true; // Memory op is a store op |
193 | // Stores will have their address in slot 2 (memory in slot 1). |
194 | // If the value being nul-checked is in another slot, it means we |
195 | // are storing the checked value, which does NOT check the value! |
196 | if( mach->in(2) != val ) continue; |
197 | break; // Found a memory op? |
198 | case Op_StrComp: |
199 | case Op_StrEquals: |
200 | case Op_StrIndexOf: |
201 | case Op_StrIndexOfChar: |
202 | case Op_AryEq: |
203 | case Op_StrInflatedCopy: |
204 | case Op_StrCompressedCopy: |
205 | case Op_EncodeISOArray: |
206 | case Op_HasNegatives: |
207 | // Not a legit memory op for implicit null check regardless of |
208 | // embedded loads |
209 | continue; |
210 | default: // Also check for embedded loads |
211 | if( !mach->needs_anti_dependence_check() ) |
212 | continue; // Not an memory op; skip it |
213 | if( must_clone[iop] ) { |
214 | // Do not move nodes which produce flags because |
215 | // RA will try to clone it to place near branch and |
216 | // it will cause recompilation, see clone_node(). |
217 | continue; |
218 | } |
219 | { |
220 | // Check that value is used in memory address in |
221 | // instructions with embedded load (CmpP val1,(val2+off)). |
222 | Node* base; |
223 | Node* index; |
224 | const MachOper* oper = mach->memory_inputs(base, index); |
225 | if (oper == NULL || oper == (MachOper*)-1) { |
226 | continue; // Not an memory op; skip it |
227 | } |
228 | if (val == base || |
229 | (val == index && val->bottom_type()->isa_narrowoop())) { |
230 | break; // Found it |
231 | } else { |
232 | continue; // Skip it |
233 | } |
234 | } |
235 | break; |
236 | } |
237 | |
238 | // On some OSes (AIX) the page at address 0 is only write protected. |
239 | // If so, only Store operations will trap. |
240 | // But a read accessing the base of a heap-based compressed heap will trap. |
241 | if (!was_store && needs_explicit_null_check_for_read(val)) { |
242 | continue; |
243 | } |
244 | |
245 | // Check that node's control edge is not-null block's head or dominates it, |
246 | // otherwise we can't hoist it because there are other control dependencies. |
247 | Node* ctrl = mach->in(0); |
248 | if (ctrl != NULL && !(ctrl == not_null_block->head() || |
249 | get_block_for_node(ctrl)->dominates(not_null_block))) { |
250 | continue; |
251 | } |
252 | |
253 | // check if the offset is not too high for implicit exception |
254 | { |
255 | intptr_t offset = 0; |
256 | const TypePtr *adr_type = NULL; // Do not need this return value here |
257 | const Node* base = mach->get_base_and_disp(offset, adr_type); |
258 | if (base == NULL || base == NodeSentinel) { |
259 | // Narrow oop address doesn't have base, only index. |
260 | // Give up if offset is beyond page size or if heap base is not protected. |
261 | if (val->bottom_type()->isa_narrowoop() && |
262 | (MacroAssembler::needs_explicit_null_check(offset) || |
263 | !CompressedOops::use_implicit_null_checks())) |
264 | continue; |
265 | // cannot reason about it; is probably not implicit null exception |
266 | } else { |
267 | const TypePtr* tptr; |
268 | if (UseCompressedOops && (CompressedOops::shift() == 0 || |
269 | CompressedKlassPointers::shift() == 0)) { |
270 | // 32-bits narrow oop can be the base of address expressions |
271 | tptr = base->get_ptr_type(); |
272 | } else { |
273 | // only regular oops are expected here |
274 | tptr = base->bottom_type()->is_ptr(); |
275 | } |
276 | // Give up if offset is not a compile-time constant. |
277 | if (offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot) |
278 | continue; |
279 | offset += tptr->_offset; // correct if base is offseted |
280 | // Give up if reference is beyond page size. |
281 | if (MacroAssembler::needs_explicit_null_check(offset)) |
282 | continue; |
283 | // Give up if base is a decode node and the heap base is not protected. |
284 | if (base->is_Mach() && base->as_Mach()->ideal_Opcode() == Op_DecodeN && |
285 | !CompressedOops::use_implicit_null_checks()) |
286 | continue; |
287 | } |
288 | } |
289 | |
290 | // Check ctrl input to see if the null-check dominates the memory op |
291 | Block *cb = get_block_for_node(mach); |
292 | cb = cb->_idom; // Always hoist at least 1 block |
293 | if( !was_store ) { // Stores can be hoisted only one block |
294 | while( cb->_dom_depth > (block->_dom_depth + 1)) |
295 | cb = cb->_idom; // Hoist loads as far as we want |
296 | // The non-null-block should dominate the memory op, too. Live |
297 | // range spilling will insert a spill in the non-null-block if it is |
298 | // needs to spill the memory op for an implicit null check. |
299 | if (cb->_dom_depth == (block->_dom_depth + 1)) { |
300 | if (cb != not_null_block) continue; |
301 | cb = cb->_idom; |
302 | } |
303 | } |
304 | if( cb != block ) continue; |
305 | |
306 | // Found a memory user; see if it can be hoisted to check-block |
307 | uint vidx = 0; // Capture index of value into memop |
308 | uint j; |
309 | for( j = mach->req()-1; j > 0; j-- ) { |
310 | if( mach->in(j) == val ) { |
311 | vidx = j; |
312 | // Ignore DecodeN val which could be hoisted to where needed. |
313 | if( is_decoden ) continue; |
314 | } |
315 | // Block of memory-op input |
316 | Block *inb = get_block_for_node(mach->in(j)); |
317 | Block *b = block; // Start from nul check |
318 | while( b != inb && b->_dom_depth > inb->_dom_depth ) |
319 | b = b->_idom; // search upwards for input |
320 | // See if input dominates null check |
321 | if( b != inb ) |
322 | break; |
323 | } |
324 | if( j > 0 ) |
325 | continue; |
326 | Block *mb = get_block_for_node(mach); |
327 | // Hoisting stores requires more checks for the anti-dependence case. |
328 | // Give up hoisting if we have to move the store past any load. |
329 | if( was_store ) { |
330 | Block *b = mb; // Start searching here for a local load |
331 | // mach use (faulting) trying to hoist |
332 | // n might be blocker to hoisting |
333 | while( b != block ) { |
334 | uint k; |
335 | for( k = 1; k < b->number_of_nodes(); k++ ) { |
336 | Node *n = b->get_node(k); |
337 | if( n->needs_anti_dependence_check() && |
338 | n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) |
339 | break; // Found anti-dependent load |
340 | } |
341 | if( k < b->number_of_nodes() ) |
342 | break; // Found anti-dependent load |
343 | // Make sure control does not do a merge (would have to check allpaths) |
344 | if( b->num_preds() != 2 ) break; |
345 | b = get_block_for_node(b->pred(1)); // Move up to predecessor block |
346 | } |
347 | if( b != block ) continue; |
348 | } |
349 | |
350 | // Make sure this memory op is not already being used for a NullCheck |
351 | Node *e = mb->end(); |
352 | if( e->is_MachNullCheck() && e->in(1) == mach ) |
353 | continue; // Already being used as a NULL check |
354 | |
355 | // Found a candidate! Pick one with least dom depth - the highest |
356 | // in the dom tree should be closest to the null check. |
357 | if (best == NULL || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) { |
358 | best = mach; |
359 | bidx = vidx; |
360 | } |
361 | } |
362 | // No candidate! |
363 | if (best == NULL) { |
364 | return; |
365 | } |
366 | |
367 | // ---- Found an implicit null check |
368 | #ifndef PRODUCT |
369 | extern int implicit_null_checks; |
370 | implicit_null_checks++; |
371 | #endif |
372 | |
373 | if( is_decoden ) { |
374 | // Check if we need to hoist decodeHeapOop_not_null first. |
375 | Block *valb = get_block_for_node(val); |
376 | if( block != valb && block->_dom_depth < valb->_dom_depth ) { |
377 | // Hoist it up to the end of the test block. |
378 | valb->find_remove(val); |
379 | block->add_inst(val); |
380 | map_node_to_block(val, block); |
381 | // DecodeN on x86 may kill flags. Check for flag-killing projections |
382 | // that also need to be hoisted. |
383 | for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { |
384 | Node* n = val->fast_out(j); |
385 | if( n->is_MachProj() ) { |
386 | get_block_for_node(n)->find_remove(n); |
387 | block->add_inst(n); |
388 | map_node_to_block(n, block); |
389 | } |
390 | } |
391 | } |
392 | } |
393 | // Hoist the memory candidate up to the end of the test block. |
394 | Block *old_block = get_block_for_node(best); |
395 | old_block->find_remove(best); |
396 | block->add_inst(best); |
397 | map_node_to_block(best, block); |
398 | |
399 | // Move the control dependence if it is pinned to not-null block. |
400 | // Don't change it in other cases: NULL or dominating control. |
401 | if (best->in(0) == not_null_block->head()) { |
402 | // Set it to control edge of null check. |
403 | best->set_req(0, proj->in(0)->in(0)); |
404 | } |
405 | |
406 | // Check for flag-killing projections that also need to be hoisted |
407 | // Should be DU safe because no edge updates. |
408 | for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { |
409 | Node* n = best->fast_out(j); |
410 | if( n->is_MachProj() ) { |
411 | get_block_for_node(n)->find_remove(n); |
412 | block->add_inst(n); |
413 | map_node_to_block(n, block); |
414 | } |
415 | } |
416 | |
417 | // proj==Op_True --> ne test; proj==Op_False --> eq test. |
418 | // One of two graph shapes got matched: |
419 | // (IfTrue (If (Bool NE (CmpP ptr NULL)))) |
420 | // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) |
421 | // NULL checks are always branch-if-eq. If we see a IfTrue projection |
422 | // then we are replacing a 'ne' test with a 'eq' NULL check test. |
423 | // We need to flip the projections to keep the same semantics. |
424 | if( proj->Opcode() == Op_IfTrue ) { |
425 | // Swap order of projections in basic block to swap branch targets |
426 | Node *tmp1 = block->get_node(block->end_idx()+1); |
427 | Node *tmp2 = block->get_node(block->end_idx()+2); |
428 | block->map_node(tmp2, block->end_idx()+1); |
429 | block->map_node(tmp1, block->end_idx()+2); |
430 | Node *tmp = new Node(C->top()); // Use not NULL input |
431 | tmp1->replace_by(tmp); |
432 | tmp2->replace_by(tmp1); |
433 | tmp->replace_by(tmp2); |
434 | tmp->destruct(); |
435 | } |
436 | |
437 | // Remove the existing null check; use a new implicit null check instead. |
438 | // Since schedule-local needs precise def-use info, we need to correct |
439 | // it as well. |
440 | Node *old_tst = proj->in(0); |
441 | MachNode *nul_chk = new MachNullCheckNode(old_tst->in(0),best,bidx); |
442 | block->map_node(nul_chk, block->end_idx()); |
443 | map_node_to_block(nul_chk, block); |
444 | // Redirect users of old_test to nul_chk |
445 | for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) |
446 | old_tst->last_out(i2)->set_req(0, nul_chk); |
447 | // Clean-up any dead code |
448 | for (uint i3 = 0; i3 < old_tst->req(); i3++) { |
449 | Node* in = old_tst->in(i3); |
450 | old_tst->set_req(i3, NULL); |
451 | if (in->outcnt() == 0) { |
452 | // Remove dead input node |
453 | in->disconnect_inputs(NULL, C); |
454 | block->find_remove(in); |
455 | } |
456 | } |
457 | |
458 | latency_from_uses(nul_chk); |
459 | latency_from_uses(best); |
460 | |
461 | // insert anti-dependences to defs in this block |
462 | if (! best->needs_anti_dependence_check()) { |
463 | for (uint k = 1; k < block->number_of_nodes(); k++) { |
464 | Node *n = block->get_node(k); |
465 | if (n->needs_anti_dependence_check() && |
466 | n->in(LoadNode::Memory) == best->in(StoreNode::Memory)) { |
467 | // Found anti-dependent load |
468 | insert_anti_dependences(block, n); |
469 | } |
470 | } |
471 | } |
472 | } |
473 | |
474 | |
475 | //------------------------------select----------------------------------------- |
476 | // Select a nice fellow from the worklist to schedule next. If there is only |
477 | // one choice, then use it. Projections take top priority for correctness |
478 | // reasons - if I see a projection, then it is next. There are a number of |
479 | // other special cases, for instructions that consume condition codes, et al. |
480 | // These are chosen immediately. Some instructions are required to immediately |
481 | // precede the last instruction in the block, and these are taken last. Of the |
482 | // remaining cases (most), choose the instruction with the greatest latency |
483 | // (that is, the most number of pseudo-cycles required to the end of the |
484 | // routine). If there is a tie, choose the instruction with the most inputs. |
485 | Node* PhaseCFG::select( |
486 | Block* block, |
487 | Node_List &worklist, |
488 | GrowableArray<int> &ready_cnt, |
489 | VectorSet &next_call, |
490 | uint sched_slot, |
491 | intptr_t* recalc_pressure_nodes) { |
492 | |
493 | // If only a single entry on the stack, use it |
494 | uint cnt = worklist.size(); |
495 | if (cnt == 1) { |
496 | Node *n = worklist[0]; |
497 | worklist.map(0,worklist.pop()); |
498 | return n; |
499 | } |
500 | |
501 | uint choice = 0; // Bigger is most important |
502 | uint latency = 0; // Bigger is scheduled first |
503 | uint score = 0; // Bigger is better |
504 | int idx = -1; // Index in worklist |
505 | int cand_cnt = 0; // Candidate count |
506 | bool block_size_threshold_ok = (block->number_of_nodes() > 10) ? true : false; |
507 | |
508 | for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist |
509 | // Order in worklist is used to break ties. |
510 | // See caller for how this is used to delay scheduling |
511 | // of induction variable increments to after the other |
512 | // uses of the phi are scheduled. |
513 | Node *n = worklist[i]; // Get Node on worklist |
514 | |
515 | int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; |
516 | if( n->is_Proj() || // Projections always win |
517 | n->Opcode()== Op_Con || // So does constant 'Top' |
518 | iop == Op_CreateEx || // Create-exception must start block |
519 | iop == Op_CheckCastPP |
520 | ) { |
521 | worklist.map(i,worklist.pop()); |
522 | return n; |
523 | } |
524 | |
525 | // Final call in a block must be adjacent to 'catch' |
526 | Node *e = block->end(); |
527 | if( e->is_Catch() && e->in(0)->in(0) == n ) |
528 | continue; |
529 | |
530 | // Memory op for an implicit null check has to be at the end of the block |
531 | if( e->is_MachNullCheck() && e->in(1) == n ) |
532 | continue; |
533 | |
534 | // Schedule IV increment last. |
535 | if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd) { |
536 | // Cmp might be matched into CountedLoopEnd node. |
537 | Node *cmp = (e->in(1)->ideal_reg() == Op_RegFlags) ? e->in(1) : e; |
538 | if (cmp->req() > 1 && cmp->in(1) == n && n->is_iteratively_computed()) { |
539 | continue; |
540 | } |
541 | } |
542 | |
543 | uint n_choice = 2; |
544 | |
545 | // See if this instruction is consumed by a branch. If so, then (as the |
546 | // branch is the last instruction in the basic block) force it to the |
547 | // end of the basic block |
548 | if ( must_clone[iop] ) { |
549 | // See if any use is a branch |
550 | bool found_machif = false; |
551 | |
552 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
553 | Node* use = n->fast_out(j); |
554 | |
555 | // The use is a conditional branch, make them adjacent |
556 | if (use->is_MachIf() && get_block_for_node(use) == block) { |
557 | found_machif = true; |
558 | break; |
559 | } |
560 | |
561 | // More than this instruction pending for successor to be ready, |
562 | // don't choose this if other opportunities are ready |
563 | if (ready_cnt.at(use->_idx) > 1) |
564 | n_choice = 1; |
565 | } |
566 | |
567 | // loop terminated, prefer not to use this instruction |
568 | if (found_machif) |
569 | continue; |
570 | } |
571 | |
572 | // See if this has a predecessor that is "must_clone", i.e. sets the |
573 | // condition code. If so, choose this first |
574 | for (uint j = 0; j < n->req() ; j++) { |
575 | Node *inn = n->in(j); |
576 | if (inn) { |
577 | if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { |
578 | n_choice = 3; |
579 | break; |
580 | } |
581 | } |
582 | } |
583 | |
584 | // MachTemps should be scheduled last so they are near their uses |
585 | if (n->is_MachTemp()) { |
586 | n_choice = 1; |
587 | } |
588 | |
589 | uint n_latency = get_latency_for_node(n); |
590 | uint n_score = n->req(); // Many inputs get high score to break ties |
591 | |
592 | if (OptoRegScheduling && block_size_threshold_ok) { |
593 | if (recalc_pressure_nodes[n->_idx] == 0x7fff7fff) { |
594 | _regalloc->_scratch_int_pressure.init(_regalloc->_sched_int_pressure.high_pressure_limit()); |
595 | _regalloc->_scratch_float_pressure.init(_regalloc->_sched_float_pressure.high_pressure_limit()); |
596 | // simulate the notion that we just picked this node to schedule |
597 | n->add_flag(Node::Flag_is_scheduled); |
598 | // now caculate its effect upon the graph if we did |
599 | adjust_register_pressure(n, block, recalc_pressure_nodes, false); |
600 | // return its state for finalize in case somebody else wins |
601 | n->remove_flag(Node::Flag_is_scheduled); |
602 | // now save the two final pressure components of register pressure, limiting pressure calcs to short size |
603 | short int_pressure = (short)_regalloc->_scratch_int_pressure.current_pressure(); |
604 | short float_pressure = (short)_regalloc->_scratch_float_pressure.current_pressure(); |
605 | recalc_pressure_nodes[n->_idx] = int_pressure; |
606 | recalc_pressure_nodes[n->_idx] |= (float_pressure << 16); |
607 | } |
608 | |
609 | if (_scheduling_for_pressure) { |
610 | latency = n_latency; |
611 | if (n_choice != 3) { |
612 | // Now evaluate each register pressure component based on threshold in the score. |
613 | // In general the defining register type will dominate the score, ergo we will not see register pressure grow on both banks |
614 | // on a single instruction, but we might see it shrink on both banks. |
615 | // For each use of register that has a register class that is over the high pressure limit, we build n_score up for |
616 | // live ranges that terminate on this instruction. |
617 | if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) { |
618 | short int_pressure = (short)recalc_pressure_nodes[n->_idx]; |
619 | n_score = (int_pressure < 0) ? ((score + n_score) - int_pressure) : (int_pressure > 0) ? 1 : n_score; |
620 | } |
621 | if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) { |
622 | short float_pressure = (short)(recalc_pressure_nodes[n->_idx] >> 16); |
623 | n_score = (float_pressure < 0) ? ((score + n_score) - float_pressure) : (float_pressure > 0) ? 1 : n_score; |
624 | } |
625 | } else { |
626 | // make sure we choose these candidates |
627 | score = 0; |
628 | } |
629 | } |
630 | } |
631 | |
632 | // Keep best latency found |
633 | cand_cnt++; |
634 | if (choice < n_choice || |
635 | (choice == n_choice && |
636 | ((StressLCM && Compile::randomized_select(cand_cnt)) || |
637 | (!StressLCM && |
638 | (latency < n_latency || |
639 | (latency == n_latency && |
640 | (score < n_score))))))) { |
641 | choice = n_choice; |
642 | latency = n_latency; |
643 | score = n_score; |
644 | idx = i; // Also keep index in worklist |
645 | } |
646 | } // End of for all ready nodes in worklist |
647 | |
648 | guarantee(idx >= 0, "index should be set" ); |
649 | Node *n = worklist[(uint)idx]; // Get the winner |
650 | |
651 | worklist.map((uint)idx, worklist.pop()); // Compress worklist |
652 | return n; |
653 | } |
654 | |
655 | //-------------------------adjust_register_pressure---------------------------- |
656 | void PhaseCFG::adjust_register_pressure(Node* n, Block* block, intptr_t* recalc_pressure_nodes, bool finalize_mode) { |
657 | PhaseLive* liveinfo = _regalloc->get_live(); |
658 | IndexSet* liveout = liveinfo->live(block); |
659 | // first adjust the register pressure for the sources |
660 | for (uint i = 1; i < n->req(); i++) { |
661 | bool lrg_ends = false; |
662 | Node *src_n = n->in(i); |
663 | if (src_n == NULL) continue; |
664 | if (!src_n->is_Mach()) continue; |
665 | uint src = _regalloc->_lrg_map.find(src_n); |
666 | if (src == 0) continue; |
667 | LRG& lrg_src = _regalloc->lrgs(src); |
668 | // detect if the live range ends or not |
669 | if (liveout->member(src) == false) { |
670 | lrg_ends = true; |
671 | for (DUIterator_Fast jmax, j = src_n->fast_outs(jmax); j < jmax; j++) { |
672 | Node* m = src_n->fast_out(j); // Get user |
673 | if (m == n) continue; |
674 | if (!m->is_Mach()) continue; |
675 | MachNode *mach = m->as_Mach(); |
676 | bool src_matches = false; |
677 | int iop = mach->ideal_Opcode(); |
678 | |
679 | switch (iop) { |
680 | case Op_StoreB: |
681 | case Op_StoreC: |
682 | case Op_StoreCM: |
683 | case Op_StoreD: |
684 | case Op_StoreF: |
685 | case Op_StoreI: |
686 | case Op_StoreL: |
687 | case Op_StoreP: |
688 | case Op_StoreN: |
689 | case Op_StoreVector: |
690 | case Op_StoreNKlass: |
691 | for (uint k = 1; k < m->req(); k++) { |
692 | Node *in = m->in(k); |
693 | if (in == src_n) { |
694 | src_matches = true; |
695 | break; |
696 | } |
697 | } |
698 | break; |
699 | |
700 | default: |
701 | src_matches = true; |
702 | break; |
703 | } |
704 | |
705 | // If we have a store as our use, ignore the non source operands |
706 | if (src_matches == false) continue; |
707 | |
708 | // Mark every unscheduled use which is not n with a recalculation |
709 | if ((get_block_for_node(m) == block) && (!m->is_scheduled())) { |
710 | if (finalize_mode && !m->is_Phi()) { |
711 | recalc_pressure_nodes[m->_idx] = 0x7fff7fff; |
712 | } |
713 | lrg_ends = false; |
714 | } |
715 | } |
716 | } |
717 | // if none, this live range ends and we can adjust register pressure |
718 | if (lrg_ends) { |
719 | if (finalize_mode) { |
720 | _regalloc->lower_pressure(block, 0, lrg_src, NULL, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure); |
721 | } else { |
722 | _regalloc->lower_pressure(block, 0, lrg_src, NULL, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure); |
723 | } |
724 | } |
725 | } |
726 | |
727 | // now add the register pressure from the dest and evaluate which heuristic we should use: |
728 | // 1.) The default, latency scheduling |
729 | // 2.) Register pressure scheduling based on the high pressure limit threshold for int or float register stacks |
730 | uint dst = _regalloc->_lrg_map.find(n); |
731 | if (dst != 0) { |
732 | LRG& lrg_dst = _regalloc->lrgs(dst); |
733 | if (finalize_mode) { |
734 | _regalloc->raise_pressure(block, lrg_dst, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure); |
735 | // check to see if we fall over the register pressure cliff here |
736 | if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) { |
737 | _scheduling_for_pressure = true; |
738 | } else if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) { |
739 | _scheduling_for_pressure = true; |
740 | } else { |
741 | // restore latency scheduling mode |
742 | _scheduling_for_pressure = false; |
743 | } |
744 | } else { |
745 | _regalloc->raise_pressure(block, lrg_dst, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure); |
746 | } |
747 | } |
748 | } |
749 | |
750 | //------------------------------set_next_call---------------------------------- |
751 | void PhaseCFG::set_next_call(Block* block, Node* n, VectorSet& next_call) { |
752 | if( next_call.test_set(n->_idx) ) return; |
753 | for( uint i=0; i<n->len(); i++ ) { |
754 | Node *m = n->in(i); |
755 | if( !m ) continue; // must see all nodes in block that precede call |
756 | if (get_block_for_node(m) == block) { |
757 | set_next_call(block, m, next_call); |
758 | } |
759 | } |
760 | } |
761 | |
762 | //------------------------------needed_for_next_call--------------------------- |
763 | // Set the flag 'next_call' for each Node that is needed for the next call to |
764 | // be scheduled. This flag lets me bias scheduling so Nodes needed for the |
765 | // next subroutine call get priority - basically it moves things NOT needed |
766 | // for the next call till after the call. This prevents me from trying to |
767 | // carry lots of stuff live across a call. |
768 | void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) { |
769 | // Find the next control-defining Node in this block |
770 | Node* call = NULL; |
771 | for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { |
772 | Node* m = this_call->fast_out(i); |
773 | if (get_block_for_node(m) == block && // Local-block user |
774 | m != this_call && // Not self-start node |
775 | m->is_MachCall()) { |
776 | call = m; |
777 | break; |
778 | } |
779 | } |
780 | if (call == NULL) return; // No next call (e.g., block end is near) |
781 | // Set next-call for all inputs to this call |
782 | set_next_call(block, call, next_call); |
783 | } |
784 | |
785 | //------------------------------add_call_kills------------------------------------- |
786 | // helper function that adds caller save registers to MachProjNode |
787 | static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) { |
788 | // Fill in the kill mask for the call |
789 | for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { |
790 | if( !regs.Member(r) ) { // Not already defined by the call |
791 | // Save-on-call register? |
792 | if ((save_policy[r] == 'C') || |
793 | (save_policy[r] == 'A') || |
794 | ((save_policy[r] == 'E') && exclude_soe)) { |
795 | proj->_rout.Insert(r); |
796 | } |
797 | } |
798 | } |
799 | } |
800 | |
801 | |
802 | //------------------------------sched_call------------------------------------- |
803 | uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) { |
804 | RegMask regs; |
805 | |
806 | // Schedule all the users of the call right now. All the users are |
807 | // projection Nodes, so they must be scheduled next to the call. |
808 | // Collect all the defined registers. |
809 | for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { |
810 | Node* n = mcall->fast_out(i); |
811 | assert( n->is_MachProj(), "" ); |
812 | int n_cnt = ready_cnt.at(n->_idx)-1; |
813 | ready_cnt.at_put(n->_idx, n_cnt); |
814 | assert( n_cnt == 0, "" ); |
815 | // Schedule next to call |
816 | block->map_node(n, node_cnt++); |
817 | // Collect defined registers |
818 | regs.OR(n->out_RegMask()); |
819 | // Check for scheduling the next control-definer |
820 | if( n->bottom_type() == Type::CONTROL ) |
821 | // Warm up next pile of heuristic bits |
822 | needed_for_next_call(block, n, next_call); |
823 | |
824 | // Children of projections are now all ready |
825 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
826 | Node* m = n->fast_out(j); // Get user |
827 | if(get_block_for_node(m) != block) { |
828 | continue; |
829 | } |
830 | if( m->is_Phi() ) continue; |
831 | int m_cnt = ready_cnt.at(m->_idx) - 1; |
832 | ready_cnt.at_put(m->_idx, m_cnt); |
833 | if( m_cnt == 0 ) |
834 | worklist.push(m); |
835 | } |
836 | |
837 | } |
838 | |
839 | // Act as if the call defines the Frame Pointer. |
840 | // Certainly the FP is alive and well after the call. |
841 | regs.Insert(_matcher.c_frame_pointer()); |
842 | |
843 | // Set all registers killed and not already defined by the call. |
844 | uint r_cnt = mcall->tf()->range()->cnt(); |
845 | int op = mcall->ideal_Opcode(); |
846 | MachProjNode *proj = new MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); |
847 | map_node_to_block(proj, block); |
848 | block->insert_node(proj, node_cnt++); |
849 | |
850 | // Select the right register save policy. |
851 | const char *save_policy = NULL; |
852 | switch (op) { |
853 | case Op_CallRuntime: |
854 | case Op_CallLeaf: |
855 | case Op_CallLeafNoFP: |
856 | // Calling C code so use C calling convention |
857 | save_policy = _matcher._c_reg_save_policy; |
858 | break; |
859 | |
860 | case Op_CallStaticJava: |
861 | case Op_CallDynamicJava: |
862 | // Calling Java code so use Java calling convention |
863 | save_policy = _matcher._register_save_policy; |
864 | break; |
865 | |
866 | default: |
867 | ShouldNotReachHere(); |
868 | } |
869 | |
870 | // When using CallRuntime mark SOE registers as killed by the call |
871 | // so values that could show up in the RegisterMap aren't live in a |
872 | // callee saved register since the register wouldn't know where to |
873 | // find them. CallLeaf and CallLeafNoFP are ok because they can't |
874 | // have debug info on them. Strictly speaking this only needs to be |
875 | // done for oops since idealreg2debugmask takes care of debug info |
876 | // references but there no way to handle oops differently than other |
877 | // pointers as far as the kill mask goes. |
878 | bool exclude_soe = op == Op_CallRuntime; |
879 | |
880 | // If the call is a MethodHandle invoke, we need to exclude the |
881 | // register which is used to save the SP value over MH invokes from |
882 | // the mask. Otherwise this register could be used for |
883 | // deoptimization information. |
884 | if (op == Op_CallStaticJava) { |
885 | MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; |
886 | if (mcallstaticjava->_method_handle_invoke) |
887 | proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); |
888 | } |
889 | |
890 | add_call_kills(proj, regs, save_policy, exclude_soe); |
891 | |
892 | return node_cnt; |
893 | } |
894 | |
895 | |
896 | //------------------------------schedule_local--------------------------------- |
897 | // Topological sort within a block. Someday become a real scheduler. |
898 | bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call, intptr_t *recalc_pressure_nodes) { |
899 | // Already "sorted" are the block start Node (as the first entry), and |
900 | // the block-ending Node and any trailing control projections. We leave |
901 | // these alone. PhiNodes and ParmNodes are made to follow the block start |
902 | // Node. Everything else gets topo-sorted. |
903 | |
904 | #ifndef PRODUCT |
905 | if (trace_opto_pipelining()) { |
906 | tty->print_cr("# --- schedule_local B%d, before: ---" , block->_pre_order); |
907 | for (uint i = 0;i < block->number_of_nodes(); i++) { |
908 | tty->print("# " ); |
909 | block->get_node(i)->fast_dump(); |
910 | } |
911 | tty->print_cr("#" ); |
912 | } |
913 | #endif |
914 | |
915 | // RootNode is already sorted |
916 | if (block->number_of_nodes() == 1) { |
917 | return true; |
918 | } |
919 | |
920 | bool block_size_threshold_ok = (block->number_of_nodes() > 10) ? true : false; |
921 | |
922 | // We track the uses of local definitions as input dependences so that |
923 | // we know when a given instruction is avialable to be scheduled. |
924 | uint i; |
925 | if (OptoRegScheduling && block_size_threshold_ok) { |
926 | for (i = 1; i < block->number_of_nodes(); i++) { // setup nodes for pressure calc |
927 | Node *n = block->get_node(i); |
928 | n->remove_flag(Node::Flag_is_scheduled); |
929 | if (!n->is_Phi()) { |
930 | recalc_pressure_nodes[n->_idx] = 0x7fff7fff; |
931 | } |
932 | } |
933 | } |
934 | |
935 | // Move PhiNodes and ParmNodes from 1 to cnt up to the start |
936 | uint node_cnt = block->end_idx(); |
937 | uint phi_cnt = 1; |
938 | for( i = 1; i<node_cnt; i++ ) { // Scan for Phi |
939 | Node *n = block->get_node(i); |
940 | if( n->is_Phi() || // Found a PhiNode or ParmNode |
941 | (n->is_Proj() && n->in(0) == block->head()) ) { |
942 | // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt |
943 | block->map_node(block->get_node(phi_cnt), i); |
944 | block->map_node(n, phi_cnt++); // swap Phi/Parm up front |
945 | if (OptoRegScheduling && block_size_threshold_ok) { |
946 | // mark n as scheduled |
947 | n->add_flag(Node::Flag_is_scheduled); |
948 | } |
949 | } else { // All others |
950 | // Count block-local inputs to 'n' |
951 | uint cnt = n->len(); // Input count |
952 | uint local = 0; |
953 | for( uint j=0; j<cnt; j++ ) { |
954 | Node *m = n->in(j); |
955 | if( m && get_block_for_node(m) == block && !m->is_top() ) |
956 | local++; // One more block-local input |
957 | } |
958 | ready_cnt.at_put(n->_idx, local); // Count em up |
959 | |
960 | #ifdef ASSERT |
961 | if( UseConcMarkSweepGC || UseG1GC ) { |
962 | if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { |
963 | // Check the precedence edges |
964 | for (uint prec = n->req(); prec < n->len(); prec++) { |
965 | Node* oop_store = n->in(prec); |
966 | if (oop_store != NULL) { |
967 | assert(get_block_for_node(oop_store)->_dom_depth <= block->_dom_depth, "oop_store must dominate card-mark" ); |
968 | } |
969 | } |
970 | } |
971 | } |
972 | #endif |
973 | |
974 | // A few node types require changing a required edge to a precedence edge |
975 | // before allocation. |
976 | if( n->is_Mach() && n->req() > TypeFunc::Parms && |
977 | (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || |
978 | n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { |
979 | // MemBarAcquire could be created without Precedent edge. |
980 | // del_req() replaces the specified edge with the last input edge |
981 | // and then removes the last edge. If the specified edge > number of |
982 | // edges the last edge will be moved outside of the input edges array |
983 | // and the edge will be lost. This is why this code should be |
984 | // executed only when Precedent (== TypeFunc::Parms) edge is present. |
985 | Node *x = n->in(TypeFunc::Parms); |
986 | if (x != NULL && get_block_for_node(x) == block && n->find_prec_edge(x) != -1) { |
987 | // Old edge to node within same block will get removed, but no precedence |
988 | // edge will get added because it already exists. Update ready count. |
989 | int cnt = ready_cnt.at(n->_idx); |
990 | assert(cnt > 1, "MemBar node %d must not get ready here" , n->_idx); |
991 | ready_cnt.at_put(n->_idx, cnt-1); |
992 | } |
993 | n->del_req(TypeFunc::Parms); |
994 | n->add_prec(x); |
995 | } |
996 | } |
997 | } |
998 | for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count |
999 | ready_cnt.at_put(block->get_node(i2)->_idx, 0); |
1000 | |
1001 | // All the prescheduled guys do not hold back internal nodes |
1002 | uint i3; |
1003 | for (i3 = 0; i3 < phi_cnt; i3++) { // For all pre-scheduled |
1004 | Node *n = block->get_node(i3); // Get pre-scheduled |
1005 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
1006 | Node* m = n->fast_out(j); |
1007 | if (get_block_for_node(m) == block) { // Local-block user |
1008 | int m_cnt = ready_cnt.at(m->_idx)-1; |
1009 | if (OptoRegScheduling && block_size_threshold_ok) { |
1010 | // mark m as scheduled |
1011 | if (m_cnt < 0) { |
1012 | m->add_flag(Node::Flag_is_scheduled); |
1013 | } |
1014 | } |
1015 | ready_cnt.at_put(m->_idx, m_cnt); // Fix ready count |
1016 | } |
1017 | } |
1018 | } |
1019 | |
1020 | Node_List delay; |
1021 | // Make a worklist |
1022 | Node_List worklist; |
1023 | for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist |
1024 | Node *m = block->get_node(i4); |
1025 | if( !ready_cnt.at(m->_idx) ) { // Zero ready count? |
1026 | if (m->is_iteratively_computed()) { |
1027 | // Push induction variable increments last to allow other uses |
1028 | // of the phi to be scheduled first. The select() method breaks |
1029 | // ties in scheduling by worklist order. |
1030 | delay.push(m); |
1031 | } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { |
1032 | // Force the CreateEx to the top of the list so it's processed |
1033 | // first and ends up at the start of the block. |
1034 | worklist.insert(0, m); |
1035 | } else { |
1036 | worklist.push(m); // Then on to worklist! |
1037 | } |
1038 | } |
1039 | } |
1040 | while (delay.size()) { |
1041 | Node* d = delay.pop(); |
1042 | worklist.push(d); |
1043 | } |
1044 | |
1045 | if (OptoRegScheduling && block_size_threshold_ok) { |
1046 | // To stage register pressure calculations we need to examine the live set variables |
1047 | // breaking them up by register class to compartmentalize the calculations. |
1048 | uint float_pressure = Matcher::float_pressure(FLOATPRESSURE); |
1049 | _regalloc->_sched_int_pressure.init(INTPRESSURE); |
1050 | _regalloc->_sched_float_pressure.init(float_pressure); |
1051 | _regalloc->_scratch_int_pressure.init(INTPRESSURE); |
1052 | _regalloc->_scratch_float_pressure.init(float_pressure); |
1053 | |
1054 | _regalloc->compute_entry_block_pressure(block); |
1055 | } |
1056 | |
1057 | // Warm up the 'next_call' heuristic bits |
1058 | needed_for_next_call(block, block->head(), next_call); |
1059 | |
1060 | #ifndef PRODUCT |
1061 | if (trace_opto_pipelining()) { |
1062 | for (uint j=0; j< block->number_of_nodes(); j++) { |
1063 | Node *n = block->get_node(j); |
1064 | int idx = n->_idx; |
1065 | tty->print("# ready cnt:%3d " , ready_cnt.at(idx)); |
1066 | tty->print("latency:%3d " , get_latency_for_node(n)); |
1067 | tty->print("%4d: %s\n" , idx, n->Name()); |
1068 | } |
1069 | } |
1070 | #endif |
1071 | |
1072 | uint max_idx = (uint)ready_cnt.length(); |
1073 | // Pull from worklist and schedule |
1074 | while( worklist.size() ) { // Worklist is not ready |
1075 | |
1076 | #ifndef PRODUCT |
1077 | if (trace_opto_pipelining()) { |
1078 | tty->print("# ready list:" ); |
1079 | for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist |
1080 | Node *n = worklist[i]; // Get Node on worklist |
1081 | tty->print(" %d" , n->_idx); |
1082 | } |
1083 | tty->cr(); |
1084 | } |
1085 | #endif |
1086 | |
1087 | // Select and pop a ready guy from worklist |
1088 | Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt, recalc_pressure_nodes); |
1089 | block->map_node(n, phi_cnt++); // Schedule him next |
1090 | |
1091 | if (OptoRegScheduling && block_size_threshold_ok) { |
1092 | n->add_flag(Node::Flag_is_scheduled); |
1093 | |
1094 | // Now adjust the resister pressure with the node we selected |
1095 | if (!n->is_Phi()) { |
1096 | adjust_register_pressure(n, block, recalc_pressure_nodes, true); |
1097 | } |
1098 | } |
1099 | |
1100 | #ifndef PRODUCT |
1101 | if (trace_opto_pipelining()) { |
1102 | tty->print("# select %d: %s" , n->_idx, n->Name()); |
1103 | tty->print(", latency:%d" , get_latency_for_node(n)); |
1104 | n->dump(); |
1105 | if (Verbose) { |
1106 | tty->print("# ready list:" ); |
1107 | for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist |
1108 | Node *n = worklist[i]; // Get Node on worklist |
1109 | tty->print(" %d" , n->_idx); |
1110 | } |
1111 | tty->cr(); |
1112 | } |
1113 | } |
1114 | |
1115 | #endif |
1116 | if( n->is_MachCall() ) { |
1117 | MachCallNode *mcall = n->as_MachCall(); |
1118 | phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call); |
1119 | continue; |
1120 | } |
1121 | |
1122 | if (n->is_Mach() && n->as_Mach()->has_call()) { |
1123 | RegMask regs; |
1124 | regs.Insert(_matcher.c_frame_pointer()); |
1125 | regs.OR(n->out_RegMask()); |
1126 | |
1127 | MachProjNode *proj = new MachProjNode( n, 1, RegMask::Empty, MachProjNode::fat_proj ); |
1128 | map_node_to_block(proj, block); |
1129 | block->insert_node(proj, phi_cnt++); |
1130 | |
1131 | add_call_kills(proj, regs, _matcher._c_reg_save_policy, false); |
1132 | } |
1133 | |
1134 | // Children are now all ready |
1135 | for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { |
1136 | Node* m = n->fast_out(i5); // Get user |
1137 | if (get_block_for_node(m) != block) { |
1138 | continue; |
1139 | } |
1140 | if( m->is_Phi() ) continue; |
1141 | if (m->_idx >= max_idx) { // new node, skip it |
1142 | assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types" ); |
1143 | continue; |
1144 | } |
1145 | int m_cnt = ready_cnt.at(m->_idx) - 1; |
1146 | ready_cnt.at_put(m->_idx, m_cnt); |
1147 | if( m_cnt == 0 ) |
1148 | worklist.push(m); |
1149 | } |
1150 | } |
1151 | |
1152 | if( phi_cnt != block->end_idx() ) { |
1153 | // did not schedule all. Retry, Bailout, or Die |
1154 | if (C->subsume_loads() == true && !C->failing()) { |
1155 | // Retry with subsume_loads == false |
1156 | // If this is the first failure, the sentinel string will "stick" |
1157 | // to the Compile object, and the C2Compiler will see it and retry. |
1158 | C->record_failure(C2Compiler::retry_no_subsuming_loads()); |
1159 | } else { |
1160 | assert(false, "graph should be schedulable" ); |
1161 | } |
1162 | // assert( phi_cnt == end_idx(), "did not schedule all" ); |
1163 | return false; |
1164 | } |
1165 | |
1166 | if (OptoRegScheduling && block_size_threshold_ok) { |
1167 | _regalloc->compute_exit_block_pressure(block); |
1168 | block->_reg_pressure = _regalloc->_sched_int_pressure.final_pressure(); |
1169 | block->_freg_pressure = _regalloc->_sched_float_pressure.final_pressure(); |
1170 | } |
1171 | |
1172 | #ifndef PRODUCT |
1173 | if (trace_opto_pipelining()) { |
1174 | tty->print_cr("#" ); |
1175 | tty->print_cr("# after schedule_local" ); |
1176 | for (uint i = 0;i < block->number_of_nodes();i++) { |
1177 | tty->print("# " ); |
1178 | block->get_node(i)->fast_dump(); |
1179 | } |
1180 | tty->print_cr("# " ); |
1181 | |
1182 | if (OptoRegScheduling && block_size_threshold_ok) { |
1183 | tty->print_cr("# pressure info : %d" , block->_pre_order); |
1184 | _regalloc->print_pressure_info(_regalloc->_sched_int_pressure, "int register info" ); |
1185 | _regalloc->print_pressure_info(_regalloc->_sched_float_pressure, "float register info" ); |
1186 | } |
1187 | tty->cr(); |
1188 | } |
1189 | #endif |
1190 | |
1191 | return true; |
1192 | } |
1193 | |
1194 | //--------------------------catch_cleanup_fix_all_inputs----------------------- |
1195 | static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { |
1196 | for (uint l = 0; l < use->len(); l++) { |
1197 | if (use->in(l) == old_def) { |
1198 | if (l < use->req()) { |
1199 | use->set_req(l, new_def); |
1200 | } else { |
1201 | use->rm_prec(l); |
1202 | use->add_prec(new_def); |
1203 | l--; |
1204 | } |
1205 | } |
1206 | } |
1207 | } |
1208 | |
1209 | //------------------------------catch_cleanup_find_cloned_def------------------ |
1210 | Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { |
1211 | assert( use_blk != def_blk, "Inter-block cleanup only" ); |
1212 | |
1213 | // The use is some block below the Catch. Find and return the clone of the def |
1214 | // that dominates the use. If there is no clone in a dominating block, then |
1215 | // create a phi for the def in a dominating block. |
1216 | |
1217 | // Find which successor block dominates this use. The successor |
1218 | // blocks must all be single-entry (from the Catch only; I will have |
1219 | // split blocks to make this so), hence they all dominate. |
1220 | while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) |
1221 | use_blk = use_blk->_idom; |
1222 | |
1223 | // Find the successor |
1224 | Node *fixup = NULL; |
1225 | |
1226 | uint j; |
1227 | for( j = 0; j < def_blk->_num_succs; j++ ) |
1228 | if( use_blk == def_blk->_succs[j] ) |
1229 | break; |
1230 | |
1231 | if( j == def_blk->_num_succs ) { |
1232 | // Block at same level in dom-tree is not a successor. It needs a |
1233 | // PhiNode, the PhiNode uses from the def and IT's uses need fixup. |
1234 | Node_Array inputs = new Node_List(Thread::current()->resource_area()); |
1235 | for(uint k = 1; k < use_blk->num_preds(); k++) { |
1236 | Block* block = get_block_for_node(use_blk->pred(k)); |
1237 | inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx)); |
1238 | } |
1239 | |
1240 | // Check to see if the use_blk already has an identical phi inserted. |
1241 | // If it exists, it will be at the first position since all uses of a |
1242 | // def are processed together. |
1243 | Node *phi = use_blk->get_node(1); |
1244 | if( phi->is_Phi() ) { |
1245 | fixup = phi; |
1246 | for (uint k = 1; k < use_blk->num_preds(); k++) { |
1247 | if (phi->in(k) != inputs[k]) { |
1248 | // Not a match |
1249 | fixup = NULL; |
1250 | break; |
1251 | } |
1252 | } |
1253 | } |
1254 | |
1255 | // If an existing PhiNode was not found, make a new one. |
1256 | if (fixup == NULL) { |
1257 | Node *new_phi = PhiNode::make(use_blk->head(), def); |
1258 | use_blk->insert_node(new_phi, 1); |
1259 | map_node_to_block(new_phi, use_blk); |
1260 | for (uint k = 1; k < use_blk->num_preds(); k++) { |
1261 | new_phi->set_req(k, inputs[k]); |
1262 | } |
1263 | fixup = new_phi; |
1264 | } |
1265 | |
1266 | } else { |
1267 | // Found the use just below the Catch. Make it use the clone. |
1268 | fixup = use_blk->get_node(n_clone_idx); |
1269 | } |
1270 | |
1271 | return fixup; |
1272 | } |
1273 | |
1274 | //--------------------------catch_cleanup_intra_block-------------------------- |
1275 | // Fix all input edges in use that reference "def". The use is in the same |
1276 | // block as the def and both have been cloned in each successor block. |
1277 | static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { |
1278 | |
1279 | // Both the use and def have been cloned. For each successor block, |
1280 | // get the clone of the use, and make its input the clone of the def |
1281 | // found in that block. |
1282 | |
1283 | uint use_idx = blk->find_node(use); |
1284 | uint offset_idx = use_idx - beg; |
1285 | for( uint k = 0; k < blk->_num_succs; k++ ) { |
1286 | // Get clone in each successor block |
1287 | Block *sb = blk->_succs[k]; |
1288 | Node *clone = sb->get_node(offset_idx+1); |
1289 | assert( clone->Opcode() == use->Opcode(), "" ); |
1290 | |
1291 | // Make use-clone reference the def-clone |
1292 | catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx)); |
1293 | } |
1294 | } |
1295 | |
1296 | //------------------------------catch_cleanup_inter_block--------------------- |
1297 | // Fix all input edges in use that reference "def". The use is in a different |
1298 | // block than the def. |
1299 | void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) { |
1300 | if( !use_blk ) return; // Can happen if the use is a precedence edge |
1301 | |
1302 | Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx); |
1303 | catch_cleanup_fix_all_inputs(use, def, new_def); |
1304 | } |
1305 | |
1306 | //------------------------------call_catch_cleanup----------------------------- |
1307 | // If we inserted any instructions between a Call and his CatchNode, |
1308 | // clone the instructions on all paths below the Catch. |
1309 | void PhaseCFG::call_catch_cleanup(Block* block) { |
1310 | |
1311 | // End of region to clone |
1312 | uint end = block->end_idx(); |
1313 | if( !block->get_node(end)->is_Catch() ) return; |
1314 | // Start of region to clone |
1315 | uint beg = end; |
1316 | while(!block->get_node(beg-1)->is_MachProj() || |
1317 | !block->get_node(beg-1)->in(0)->is_MachCall() ) { |
1318 | beg--; |
1319 | assert(beg > 0,"Catch cleanup walking beyond block boundary" ); |
1320 | } |
1321 | // Range of inserted instructions is [beg, end) |
1322 | if( beg == end ) return; |
1323 | |
1324 | // Clone along all Catch output paths. Clone area between the 'beg' and |
1325 | // 'end' indices. |
1326 | for( uint i = 0; i < block->_num_succs; i++ ) { |
1327 | Block *sb = block->_succs[i]; |
1328 | // Clone the entire area; ignoring the edge fixup for now. |
1329 | for( uint j = end; j > beg; j-- ) { |
1330 | Node *clone = block->get_node(j-1)->clone(); |
1331 | sb->insert_node(clone, 1); |
1332 | map_node_to_block(clone, sb); |
1333 | if (clone->needs_anti_dependence_check()) { |
1334 | insert_anti_dependences(sb, clone); |
1335 | } |
1336 | } |
1337 | } |
1338 | |
1339 | |
1340 | // Fixup edges. Check the def-use info per cloned Node |
1341 | for(uint i2 = beg; i2 < end; i2++ ) { |
1342 | uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block |
1343 | Node *n = block->get_node(i2); // Node that got cloned |
1344 | // Need DU safe iterator because of edge manipulation in calls. |
1345 | Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); |
1346 | for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { |
1347 | out->push(n->fast_out(j1)); |
1348 | } |
1349 | uint max = out->size(); |
1350 | for (uint j = 0; j < max; j++) {// For all users |
1351 | Node *use = out->pop(); |
1352 | Block *buse = get_block_for_node(use); |
1353 | if( use->is_Phi() ) { |
1354 | for( uint k = 1; k < use->req(); k++ ) |
1355 | if( use->in(k) == n ) { |
1356 | Block* b = get_block_for_node(buse->pred(k)); |
1357 | Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx); |
1358 | use->set_req(k, fixup); |
1359 | } |
1360 | } else { |
1361 | if (block == buse) { |
1362 | catch_cleanup_intra_block(use, n, block, beg, n_clone_idx); |
1363 | } else { |
1364 | catch_cleanup_inter_block(use, buse, n, block, n_clone_idx); |
1365 | } |
1366 | } |
1367 | } // End for all users |
1368 | |
1369 | } // End of for all Nodes in cloned area |
1370 | |
1371 | // Remove the now-dead cloned ops |
1372 | for(uint i3 = beg; i3 < end; i3++ ) { |
1373 | block->get_node(beg)->disconnect_inputs(NULL, C); |
1374 | block->remove_node(beg); |
1375 | } |
1376 | |
1377 | // If the successor blocks have a CreateEx node, move it back to the top |
1378 | for(uint i4 = 0; i4 < block->_num_succs; i4++ ) { |
1379 | Block *sb = block->_succs[i4]; |
1380 | uint new_cnt = end - beg; |
1381 | // Remove any newly created, but dead, nodes. |
1382 | for( uint j = new_cnt; j > 0; j-- ) { |
1383 | Node *n = sb->get_node(j); |
1384 | if (n->outcnt() == 0 && |
1385 | (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ |
1386 | n->disconnect_inputs(NULL, C); |
1387 | sb->remove_node(j); |
1388 | new_cnt--; |
1389 | } |
1390 | } |
1391 | // If any newly created nodes remain, move the CreateEx node to the top |
1392 | if (new_cnt > 0) { |
1393 | Node *cex = sb->get_node(1+new_cnt); |
1394 | if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { |
1395 | sb->remove_node(1+new_cnt); |
1396 | sb->insert_node(cex, 1); |
1397 | } |
1398 | } |
1399 | } |
1400 | } |
1401 | |