Thu, 20 Mar 2008 15:11:44 -0700
6674600: (Escape Analysis) Optimize memory graph for instance's fields
Summary: EA gives opportunite to do more aggressive memory optimizations.
Reviewed-by: never, jrose
1 /*
2 * Copyright 1997-2007 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 */
25 // Portions of code courtesy of Clifford Click
27 // Optimization - Graph Style
29 #include "incls/_precompiled.incl"
30 #include "incls/_gcm.cpp.incl"
32 //----------------------------schedule_node_into_block-------------------------
33 // Insert node n into block b. Look for projections of n and make sure they
34 // are in b also.
35 void PhaseCFG::schedule_node_into_block( Node *n, Block *b ) {
36 // Set basic block of n, Add n to b,
37 _bbs.map(n->_idx, b);
38 b->add_inst(n);
40 // After Matching, nearly any old Node may have projections trailing it.
41 // These are usually machine-dependent flags. In any case, they might
42 // float to another block below this one. Move them up.
43 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
44 Node* use = n->fast_out(i);
45 if (use->is_Proj()) {
46 Block* buse = _bbs[use->_idx];
47 if (buse != b) { // In wrong block?
48 if (buse != NULL)
49 buse->find_remove(use); // Remove from wrong block
50 _bbs.map(use->_idx, b); // Re-insert in this block
51 b->add_inst(use);
52 }
53 }
54 }
55 }
58 //------------------------------schedule_pinned_nodes--------------------------
59 // Set the basic block for Nodes pinned into blocks
60 void PhaseCFG::schedule_pinned_nodes( VectorSet &visited ) {
61 // Allocate node stack of size C->unique()+8 to avoid frequent realloc
62 GrowableArray <Node *> spstack(C->unique()+8);
63 spstack.push(_root);
64 while ( spstack.is_nonempty() ) {
65 Node *n = spstack.pop();
66 if( !visited.test_set(n->_idx) ) { // Test node and flag it as visited
67 if( n->pinned() && !_bbs.lookup(n->_idx) ) { // Pinned? Nail it down!
68 Node *input = n->in(0);
69 assert( input, "pinned Node must have Control" );
70 while( !input->is_block_start() )
71 input = input->in(0);
72 Block *b = _bbs[input->_idx]; // Basic block of controlling input
73 schedule_node_into_block(n, b);
74 }
75 for( int i = n->req() - 1; i >= 0; --i ) { // For all inputs
76 if( n->in(i) != NULL )
77 spstack.push(n->in(i));
78 }
79 }
80 }
81 }
83 #ifdef ASSERT
84 // Assert that new input b2 is dominated by all previous inputs.
85 // Check this by by seeing that it is dominated by b1, the deepest
86 // input observed until b2.
87 static void assert_dom(Block* b1, Block* b2, Node* n, Block_Array &bbs) {
88 if (b1 == NULL) return;
89 assert(b1->_dom_depth < b2->_dom_depth, "sanity");
90 Block* tmp = b2;
91 while (tmp != b1 && tmp != NULL) {
92 tmp = tmp->_idom;
93 }
94 if (tmp != b1) {
95 // Detected an unschedulable graph. Print some nice stuff and die.
96 tty->print_cr("!!! Unschedulable graph !!!");
97 for (uint j=0; j<n->len(); j++) { // For all inputs
98 Node* inn = n->in(j); // Get input
99 if (inn == NULL) continue; // Ignore NULL, missing inputs
100 Block* inb = bbs[inn->_idx];
101 tty->print("B%d idom=B%d depth=%2d ",inb->_pre_order,
102 inb->_idom ? inb->_idom->_pre_order : 0, inb->_dom_depth);
103 inn->dump();
104 }
105 tty->print("Failing node: ");
106 n->dump();
107 assert(false, "unscheduable graph");
108 }
109 }
110 #endif
112 static Block* find_deepest_input(Node* n, Block_Array &bbs) {
113 // Find the last input dominated by all other inputs.
114 Block* deepb = NULL; // Deepest block so far
115 int deepb_dom_depth = 0;
116 for (uint k = 0; k < n->len(); k++) { // For all inputs
117 Node* inn = n->in(k); // Get input
118 if (inn == NULL) continue; // Ignore NULL, missing inputs
119 Block* inb = bbs[inn->_idx];
120 assert(inb != NULL, "must already have scheduled this input");
121 if (deepb_dom_depth < (int) inb->_dom_depth) {
122 // The new inb must be dominated by the previous deepb.
123 // The various inputs must be linearly ordered in the dom
124 // tree, or else there will not be a unique deepest block.
125 DEBUG_ONLY(assert_dom(deepb, inb, n, bbs));
126 deepb = inb; // Save deepest block
127 deepb_dom_depth = deepb->_dom_depth;
128 }
129 }
130 assert(deepb != NULL, "must be at least one input to n");
131 return deepb;
132 }
135 //------------------------------schedule_early---------------------------------
136 // Find the earliest Block any instruction can be placed in. Some instructions
137 // are pinned into Blocks. Unpinned instructions can appear in last block in
138 // which all their inputs occur.
139 bool PhaseCFG::schedule_early(VectorSet &visited, Node_List &roots) {
140 // Allocate stack with enough space to avoid frequent realloc
141 Node_Stack nstack(roots.Size() + 8); // (unique >> 1) + 24 from Java2D stats
142 // roots.push(_root); _root will be processed among C->top() inputs
143 roots.push(C->top());
144 visited.set(C->top()->_idx);
146 while (roots.size() != 0) {
147 // Use local variables nstack_top_n & nstack_top_i to cache values
148 // on stack's top.
149 Node *nstack_top_n = roots.pop();
150 uint nstack_top_i = 0;
151 //while_nstack_nonempty:
152 while (true) {
153 // Get parent node and next input's index from stack's top.
154 Node *n = nstack_top_n;
155 uint i = nstack_top_i;
157 if (i == 0) {
158 // Special control input processing.
159 // While I am here, go ahead and look for Nodes which are taking control
160 // from a is_block_proj Node. After I inserted RegionNodes to make proper
161 // blocks, the control at a is_block_proj more properly comes from the
162 // Region being controlled by the block_proj Node.
163 const Node *in0 = n->in(0);
164 if (in0 != NULL) { // Control-dependent?
165 const Node *p = in0->is_block_proj();
166 if (p != NULL && p != n) { // Control from a block projection?
167 // Find trailing Region
168 Block *pb = _bbs[in0->_idx]; // Block-projection already has basic block
169 uint j = 0;
170 if (pb->_num_succs != 1) { // More then 1 successor?
171 // Search for successor
172 uint max = pb->_nodes.size();
173 assert( max > 1, "" );
174 uint start = max - pb->_num_succs;
175 // Find which output path belongs to projection
176 for (j = start; j < max; j++) {
177 if( pb->_nodes[j] == in0 )
178 break;
179 }
180 assert( j < max, "must find" );
181 // Change control to match head of successor basic block
182 j -= start;
183 }
184 n->set_req(0, pb->_succs[j]->head());
185 }
186 } else { // n->in(0) == NULL
187 if (n->req() == 1) { // This guy is a constant with NO inputs?
188 n->set_req(0, _root);
189 }
190 }
191 }
193 // First, visit all inputs and force them to get a block. If an
194 // input is already in a block we quit following inputs (to avoid
195 // cycles). Instead we put that Node on a worklist to be handled
196 // later (since IT'S inputs may not have a block yet).
197 bool done = true; // Assume all n's inputs will be processed
198 while (i < n->len()) { // For all inputs
199 Node *in = n->in(i); // Get input
200 ++i;
201 if (in == NULL) continue; // Ignore NULL, missing inputs
202 int is_visited = visited.test_set(in->_idx);
203 if (!_bbs.lookup(in->_idx)) { // Missing block selection?
204 if (is_visited) {
205 // assert( !visited.test(in->_idx), "did not schedule early" );
206 return false;
207 }
208 nstack.push(n, i); // Save parent node and next input's index.
209 nstack_top_n = in; // Process current input now.
210 nstack_top_i = 0;
211 done = false; // Not all n's inputs processed.
212 break; // continue while_nstack_nonempty;
213 } else if (!is_visited) { // Input not yet visited?
214 roots.push(in); // Visit this guy later, using worklist
215 }
216 }
217 if (done) {
218 // All of n's inputs have been processed, complete post-processing.
220 // Some instructions are pinned into a block. These include Region,
221 // Phi, Start, Return, and other control-dependent instructions and
222 // any projections which depend on them.
223 if (!n->pinned()) {
224 // Set earliest legal block.
225 _bbs.map(n->_idx, find_deepest_input(n, _bbs));
226 }
228 if (nstack.is_empty()) {
229 // Finished all nodes on stack.
230 // Process next node on the worklist 'roots'.
231 break;
232 }
233 // Get saved parent node and next input's index.
234 nstack_top_n = nstack.node();
235 nstack_top_i = nstack.index();
236 nstack.pop();
237 } // if (done)
238 } // while (true)
239 } // while (roots.size() != 0)
240 return true;
241 }
243 //------------------------------dom_lca----------------------------------------
244 // Find least common ancestor in dominator tree
245 // LCA is a current notion of LCA, to be raised above 'this'.
246 // As a convenient boundary condition, return 'this' if LCA is NULL.
247 // Find the LCA of those two nodes.
248 Block* Block::dom_lca(Block* LCA) {
249 if (LCA == NULL || LCA == this) return this;
251 Block* anc = this;
252 while (anc->_dom_depth > LCA->_dom_depth)
253 anc = anc->_idom; // Walk up till anc is as high as LCA
255 while (LCA->_dom_depth > anc->_dom_depth)
256 LCA = LCA->_idom; // Walk up till LCA is as high as anc
258 while (LCA != anc) { // Walk both up till they are the same
259 LCA = LCA->_idom;
260 anc = anc->_idom;
261 }
263 return LCA;
264 }
266 //--------------------------raise_LCA_above_use--------------------------------
267 // We are placing a definition, and have been given a def->use edge.
268 // The definition must dominate the use, so move the LCA upward in the
269 // dominator tree to dominate the use. If the use is a phi, adjust
270 // the LCA only with the phi input paths which actually use this def.
271 static Block* raise_LCA_above_use(Block* LCA, Node* use, Node* def, Block_Array &bbs) {
272 Block* buse = bbs[use->_idx];
273 if (buse == NULL) return LCA; // Unused killing Projs have no use block
274 if (!use->is_Phi()) return buse->dom_lca(LCA);
275 uint pmax = use->req(); // Number of Phi inputs
276 // Why does not this loop just break after finding the matching input to
277 // the Phi? Well...it's like this. I do not have true def-use/use-def
278 // chains. Means I cannot distinguish, from the def-use direction, which
279 // of many use-defs lead from the same use to the same def. That is, this
280 // Phi might have several uses of the same def. Each use appears in a
281 // different predecessor block. But when I enter here, I cannot distinguish
282 // which use-def edge I should find the predecessor block for. So I find
283 // them all. Means I do a little extra work if a Phi uses the same value
284 // more than once.
285 for (uint j=1; j<pmax; j++) { // For all inputs
286 if (use->in(j) == def) { // Found matching input?
287 Block* pred = bbs[buse->pred(j)->_idx];
288 LCA = pred->dom_lca(LCA);
289 }
290 }
291 return LCA;
292 }
294 //----------------------------raise_LCA_above_marks----------------------------
295 // Return a new LCA that dominates LCA and any of its marked predecessors.
296 // Search all my parents up to 'early' (exclusive), looking for predecessors
297 // which are marked with the given index. Return the LCA (in the dom tree)
298 // of all marked blocks. If there are none marked, return the original
299 // LCA.
300 static Block* raise_LCA_above_marks(Block* LCA, node_idx_t mark,
301 Block* early, Block_Array &bbs) {
302 Block_List worklist;
303 worklist.push(LCA);
304 while (worklist.size() > 0) {
305 Block* mid = worklist.pop();
306 if (mid == early) continue; // stop searching here
308 // Test and set the visited bit.
309 if (mid->raise_LCA_visited() == mark) continue; // already visited
310 mid->set_raise_LCA_visited(mark);
312 // Don't process the current LCA, otherwise the search may terminate early
313 if (mid != LCA && mid->raise_LCA_mark() == mark) {
314 // Raise the LCA.
315 LCA = mid->dom_lca(LCA);
316 if (LCA == early) break; // stop searching everywhere
317 assert(early->dominates(LCA), "early is high enough");
318 // Resume searching at that point, skipping intermediate levels.
319 worklist.push(LCA);
320 } else {
321 // Keep searching through this block's predecessors.
322 for (uint j = 1, jmax = mid->num_preds(); j < jmax; j++) {
323 Block* mid_parent = bbs[ mid->pred(j)->_idx ];
324 worklist.push(mid_parent);
325 }
326 }
327 }
328 return LCA;
329 }
331 //--------------------------memory_early_block--------------------------------
332 // This is a variation of find_deepest_input, the heart of schedule_early.
333 // Find the "early" block for a load, if we considered only memory and
334 // address inputs, that is, if other data inputs were ignored.
335 //
336 // Because a subset of edges are considered, the resulting block will
337 // be earlier (at a shallower dom_depth) than the true schedule_early
338 // point of the node. We compute this earlier block as a more permissive
339 // site for anti-dependency insertion, but only if subsume_loads is enabled.
340 static Block* memory_early_block(Node* load, Block* early, Block_Array &bbs) {
341 Node* base;
342 Node* index;
343 Node* store = load->in(MemNode::Memory);
344 load->as_Mach()->memory_inputs(base, index);
346 assert(base != NodeSentinel && index != NodeSentinel,
347 "unexpected base/index inputs");
349 Node* mem_inputs[4];
350 int mem_inputs_length = 0;
351 if (base != NULL) mem_inputs[mem_inputs_length++] = base;
352 if (index != NULL) mem_inputs[mem_inputs_length++] = index;
353 if (store != NULL) mem_inputs[mem_inputs_length++] = store;
355 // In the comparision below, add one to account for the control input,
356 // which may be null, but always takes up a spot in the in array.
357 if (mem_inputs_length + 1 < (int) load->req()) {
358 // This "load" has more inputs than just the memory, base and index inputs.
359 // For purposes of checking anti-dependences, we need to start
360 // from the early block of only the address portion of the instruction,
361 // and ignore other blocks that may have factored into the wider
362 // schedule_early calculation.
363 if (load->in(0) != NULL) mem_inputs[mem_inputs_length++] = load->in(0);
365 Block* deepb = NULL; // Deepest block so far
366 int deepb_dom_depth = 0;
367 for (int i = 0; i < mem_inputs_length; i++) {
368 Block* inb = bbs[mem_inputs[i]->_idx];
369 if (deepb_dom_depth < (int) inb->_dom_depth) {
370 // The new inb must be dominated by the previous deepb.
371 // The various inputs must be linearly ordered in the dom
372 // tree, or else there will not be a unique deepest block.
373 DEBUG_ONLY(assert_dom(deepb, inb, load, bbs));
374 deepb = inb; // Save deepest block
375 deepb_dom_depth = deepb->_dom_depth;
376 }
377 }
378 early = deepb;
379 }
381 return early;
382 }
384 //--------------------------insert_anti_dependences---------------------------
385 // A load may need to witness memory that nearby stores can overwrite.
386 // For each nearby store, either insert an "anti-dependence" edge
387 // from the load to the store, or else move LCA upward to force the
388 // load to (eventually) be scheduled in a block above the store.
389 //
390 // Do not add edges to stores on distinct control-flow paths;
391 // only add edges to stores which might interfere.
392 //
393 // Return the (updated) LCA. There will not be any possibly interfering
394 // store between the load's "early block" and the updated LCA.
395 // Any stores in the updated LCA will have new precedence edges
396 // back to the load. The caller is expected to schedule the load
397 // in the LCA, in which case the precedence edges will make LCM
398 // preserve anti-dependences. The caller may also hoist the load
399 // above the LCA, if it is not the early block.
400 Block* PhaseCFG::insert_anti_dependences(Block* LCA, Node* load, bool verify) {
401 assert(load->needs_anti_dependence_check(), "must be a load of some sort");
402 assert(LCA != NULL, "");
403 DEBUG_ONLY(Block* LCA_orig = LCA);
405 // Compute the alias index. Loads and stores with different alias indices
406 // do not need anti-dependence edges.
407 uint load_alias_idx = C->get_alias_index(load->adr_type());
408 #ifdef ASSERT
409 if (load_alias_idx == Compile::AliasIdxBot && C->AliasLevel() > 0 &&
410 (PrintOpto || VerifyAliases ||
411 PrintMiscellaneous && (WizardMode || Verbose))) {
412 // Load nodes should not consume all of memory.
413 // Reporting a bottom type indicates a bug in adlc.
414 // If some particular type of node validly consumes all of memory,
415 // sharpen the preceding "if" to exclude it, so we can catch bugs here.
416 tty->print_cr("*** Possible Anti-Dependence Bug: Load consumes all of memory.");
417 load->dump(2);
418 if (VerifyAliases) assert(load_alias_idx != Compile::AliasIdxBot, "");
419 }
420 #endif
421 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrComp),
422 "String compare is only known 'load' that does not conflict with any stores");
424 if (!C->alias_type(load_alias_idx)->is_rewritable()) {
425 // It is impossible to spoil this load by putting stores before it,
426 // because we know that the stores will never update the value
427 // which 'load' must witness.
428 return LCA;
429 }
431 node_idx_t load_index = load->_idx;
433 // Note the earliest legal placement of 'load', as determined by
434 // by the unique point in the dom tree where all memory effects
435 // and other inputs are first available. (Computed by schedule_early.)
436 // For normal loads, 'early' is the shallowest place (dom graph wise)
437 // to look for anti-deps between this load and any store.
438 Block* early = _bbs[load_index];
440 // If we are subsuming loads, compute an "early" block that only considers
441 // memory or address inputs. This block may be different than the
442 // schedule_early block in that it could be at an even shallower depth in the
443 // dominator tree, and allow for a broader discovery of anti-dependences.
444 if (C->subsume_loads()) {
445 early = memory_early_block(load, early, _bbs);
446 }
448 ResourceArea *area = Thread::current()->resource_area();
449 Node_List worklist_mem(area); // prior memory state to store
450 Node_List worklist_store(area); // possible-def to explore
451 Node_List worklist_visited(area); // visited mergemem nodes
452 Node_List non_early_stores(area); // all relevant stores outside of early
453 bool must_raise_LCA = false;
455 #ifdef TRACK_PHI_INPUTS
456 // %%% This extra checking fails because MergeMem nodes are not GVNed.
457 // Provide "phi_inputs" to check if every input to a PhiNode is from the
458 // original memory state. This indicates a PhiNode for which should not
459 // prevent the load from sinking. For such a block, set_raise_LCA_mark
460 // may be overly conservative.
461 // Mechanism: count inputs seen for each Phi encountered in worklist_store.
462 DEBUG_ONLY(GrowableArray<uint> phi_inputs(area, C->unique(),0,0));
463 #endif
465 // 'load' uses some memory state; look for users of the same state.
466 // Recurse through MergeMem nodes to the stores that use them.
468 // Each of these stores is a possible definition of memory
469 // that 'load' needs to use. We need to force 'load'
470 // to occur before each such store. When the store is in
471 // the same block as 'load', we insert an anti-dependence
472 // edge load->store.
474 // The relevant stores "nearby" the load consist of a tree rooted
475 // at initial_mem, with internal nodes of type MergeMem.
476 // Therefore, the branches visited by the worklist are of this form:
477 // initial_mem -> (MergeMem ->)* store
478 // The anti-dependence constraints apply only to the fringe of this tree.
480 Node* initial_mem = load->in(MemNode::Memory);
481 worklist_store.push(initial_mem);
482 worklist_visited.push(initial_mem);
483 worklist_mem.push(NULL);
484 while (worklist_store.size() > 0) {
485 // Examine a nearby store to see if it might interfere with our load.
486 Node* mem = worklist_mem.pop();
487 Node* store = worklist_store.pop();
488 uint op = store->Opcode();
490 // MergeMems do not directly have anti-deps.
491 // Treat them as internal nodes in a forward tree of memory states,
492 // the leaves of which are each a 'possible-def'.
493 if (store == initial_mem // root (exclusive) of tree we are searching
494 || op == Op_MergeMem // internal node of tree we are searching
495 ) {
496 mem = store; // It's not a possibly interfering store.
497 if (store == initial_mem)
498 initial_mem = NULL; // only process initial memory once
500 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
501 store = mem->fast_out(i);
502 if (store->is_MergeMem()) {
503 // Be sure we don't get into combinatorial problems.
504 // (Allow phis to be repeated; they can merge two relevant states.)
505 uint j = worklist_visited.size();
506 for (; j > 0; j--) {
507 if (worklist_visited.at(j-1) == store) break;
508 }
509 if (j > 0) continue; // already on work list; do not repeat
510 worklist_visited.push(store);
511 }
512 worklist_mem.push(mem);
513 worklist_store.push(store);
514 }
515 continue;
516 }
518 if (op == Op_MachProj || op == Op_Catch) continue;
519 if (store->needs_anti_dependence_check()) continue; // not really a store
521 // Compute the alias index. Loads and stores with different alias
522 // indices do not need anti-dependence edges. Wide MemBar's are
523 // anti-dependent on everything (except immutable memories).
524 const TypePtr* adr_type = store->adr_type();
525 if (!C->can_alias(adr_type, load_alias_idx)) continue;
527 // Most slow-path runtime calls do NOT modify Java memory, but
528 // they can block and so write Raw memory.
529 if (store->is_Mach()) {
530 MachNode* mstore = store->as_Mach();
531 if (load_alias_idx != Compile::AliasIdxRaw) {
532 // Check for call into the runtime using the Java calling
533 // convention (and from there into a wrapper); it has no
534 // _method. Can't do this optimization for Native calls because
535 // they CAN write to Java memory.
536 if (mstore->ideal_Opcode() == Op_CallStaticJava) {
537 assert(mstore->is_MachSafePoint(), "");
538 MachSafePointNode* ms = (MachSafePointNode*) mstore;
539 assert(ms->is_MachCallJava(), "");
540 MachCallJavaNode* mcj = (MachCallJavaNode*) ms;
541 if (mcj->_method == NULL) {
542 // These runtime calls do not write to Java visible memory
543 // (other than Raw) and so do not require anti-dependence edges.
544 continue;
545 }
546 }
547 // Same for SafePoints: they read/write Raw but only read otherwise.
548 // This is basically a workaround for SafePoints only defining control
549 // instead of control + memory.
550 if (mstore->ideal_Opcode() == Op_SafePoint)
551 continue;
552 } else {
553 // Some raw memory, such as the load of "top" at an allocation,
554 // can be control dependent on the previous safepoint. See
555 // comments in GraphKit::allocate_heap() about control input.
556 // Inserting an anti-dep between such a safepoint and a use
557 // creates a cycle, and will cause a subsequent failure in
558 // local scheduling. (BugId 4919904)
559 // (%%% How can a control input be a safepoint and not a projection??)
560 if (mstore->ideal_Opcode() == Op_SafePoint && load->in(0) == mstore)
561 continue;
562 }
563 }
565 // Identify a block that the current load must be above,
566 // or else observe that 'store' is all the way up in the
567 // earliest legal block for 'load'. In the latter case,
568 // immediately insert an anti-dependence edge.
569 Block* store_block = _bbs[store->_idx];
570 assert(store_block != NULL, "unused killing projections skipped above");
572 if (store->is_Phi()) {
573 // 'load' uses memory which is one (or more) of the Phi's inputs.
574 // It must be scheduled not before the Phi, but rather before
575 // each of the relevant Phi inputs.
576 //
577 // Instead of finding the LCA of all inputs to a Phi that match 'mem',
578 // we mark each corresponding predecessor block and do a combined
579 // hoisting operation later (raise_LCA_above_marks).
580 //
581 // Do not assert(store_block != early, "Phi merging memory after access")
582 // PhiNode may be at start of block 'early' with backedge to 'early'
583 DEBUG_ONLY(bool found_match = false);
584 for (uint j = PhiNode::Input, jmax = store->req(); j < jmax; j++) {
585 if (store->in(j) == mem) { // Found matching input?
586 DEBUG_ONLY(found_match = true);
587 Block* pred_block = _bbs[store_block->pred(j)->_idx];
588 if (pred_block != early) {
589 // If any predecessor of the Phi matches the load's "early block",
590 // we do not need a precedence edge between the Phi and 'load'
591 // since the load will be forced into a block preceeding the Phi.
592 pred_block->set_raise_LCA_mark(load_index);
593 assert(!LCA_orig->dominates(pred_block) ||
594 early->dominates(pred_block), "early is high enough");
595 must_raise_LCA = true;
596 }
597 }
598 }
599 assert(found_match, "no worklist bug");
600 #ifdef TRACK_PHI_INPUTS
601 #ifdef ASSERT
602 // This assert asks about correct handling of PhiNodes, which may not
603 // have all input edges directly from 'mem'. See BugId 4621264
604 int num_mem_inputs = phi_inputs.at_grow(store->_idx,0) + 1;
605 // Increment by exactly one even if there are multiple copies of 'mem'
606 // coming into the phi, because we will run this block several times
607 // if there are several copies of 'mem'. (That's how DU iterators work.)
608 phi_inputs.at_put(store->_idx, num_mem_inputs);
609 assert(PhiNode::Input + num_mem_inputs < store->req(),
610 "Expect at least one phi input will not be from original memory state");
611 #endif //ASSERT
612 #endif //TRACK_PHI_INPUTS
613 } else if (store_block != early) {
614 // 'store' is between the current LCA and earliest possible block.
615 // Label its block, and decide later on how to raise the LCA
616 // to include the effect on LCA of this store.
617 // If this store's block gets chosen as the raised LCA, we
618 // will find him on the non_early_stores list and stick him
619 // with a precedence edge.
620 // (But, don't bother if LCA is already raised all the way.)
621 if (LCA != early) {
622 store_block->set_raise_LCA_mark(load_index);
623 must_raise_LCA = true;
624 non_early_stores.push(store);
625 }
626 } else {
627 // Found a possibly-interfering store in the load's 'early' block.
628 // This means 'load' cannot sink at all in the dominator tree.
629 // Add an anti-dep edge, and squeeze 'load' into the highest block.
630 assert(store != load->in(0), "dependence cycle found");
631 if (verify) {
632 assert(store->find_edge(load) != -1, "missing precedence edge");
633 } else {
634 store->add_prec(load);
635 }
636 LCA = early;
637 // This turns off the process of gathering non_early_stores.
638 }
639 }
640 // (Worklist is now empty; all nearby stores have been visited.)
642 // Finished if 'load' must be scheduled in its 'early' block.
643 // If we found any stores there, they have already been given
644 // precedence edges.
645 if (LCA == early) return LCA;
647 // We get here only if there are no possibly-interfering stores
648 // in the load's 'early' block. Move LCA up above all predecessors
649 // which contain stores we have noted.
650 //
651 // The raised LCA block can be a home to such interfering stores,
652 // but its predecessors must not contain any such stores.
653 //
654 // The raised LCA will be a lower bound for placing the load,
655 // preventing the load from sinking past any block containing
656 // a store that may invalidate the memory state required by 'load'.
657 if (must_raise_LCA)
658 LCA = raise_LCA_above_marks(LCA, load->_idx, early, _bbs);
659 if (LCA == early) return LCA;
661 // Insert anti-dependence edges from 'load' to each store
662 // in the non-early LCA block.
663 // Mine the non_early_stores list for such stores.
664 if (LCA->raise_LCA_mark() == load_index) {
665 while (non_early_stores.size() > 0) {
666 Node* store = non_early_stores.pop();
667 Block* store_block = _bbs[store->_idx];
668 if (store_block == LCA) {
669 // add anti_dependence from store to load in its own block
670 assert(store != load->in(0), "dependence cycle found");
671 if (verify) {
672 assert(store->find_edge(load) != -1, "missing precedence edge");
673 } else {
674 store->add_prec(load);
675 }
676 } else {
677 assert(store_block->raise_LCA_mark() == load_index, "block was marked");
678 // Any other stores we found must be either inside the new LCA
679 // or else outside the original LCA. In the latter case, they
680 // did not interfere with any use of 'load'.
681 assert(LCA->dominates(store_block)
682 || !LCA_orig->dominates(store_block), "no stray stores");
683 }
684 }
685 }
687 // Return the highest block containing stores; any stores
688 // within that block have been given anti-dependence edges.
689 return LCA;
690 }
692 // This class is used to iterate backwards over the nodes in the graph.
694 class Node_Backward_Iterator {
696 private:
697 Node_Backward_Iterator();
699 public:
700 // Constructor for the iterator
701 Node_Backward_Iterator(Node *root, VectorSet &visited, Node_List &stack, Block_Array &bbs);
703 // Postincrement operator to iterate over the nodes
704 Node *next();
706 private:
707 VectorSet &_visited;
708 Node_List &_stack;
709 Block_Array &_bbs;
710 };
712 // Constructor for the Node_Backward_Iterator
713 Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_List &stack, Block_Array &bbs )
714 : _visited(visited), _stack(stack), _bbs(bbs) {
715 // The stack should contain exactly the root
716 stack.clear();
717 stack.push(root);
719 // Clear the visited bits
720 visited.Clear();
721 }
723 // Iterator for the Node_Backward_Iterator
724 Node *Node_Backward_Iterator::next() {
726 // If the _stack is empty, then just return NULL: finished.
727 if ( !_stack.size() )
728 return NULL;
730 // '_stack' is emulating a real _stack. The 'visit-all-users' loop has been
731 // made stateless, so I do not need to record the index 'i' on my _stack.
732 // Instead I visit all users each time, scanning for unvisited users.
733 // I visit unvisited not-anti-dependence users first, then anti-dependent
734 // children next.
735 Node *self = _stack.pop();
737 // I cycle here when I am entering a deeper level of recursion.
738 // The key variable 'self' was set prior to jumping here.
739 while( 1 ) {
741 _visited.set(self->_idx);
743 // Now schedule all uses as late as possible.
744 uint src = self->is_Proj() ? self->in(0)->_idx : self->_idx;
745 uint src_rpo = _bbs[src]->_rpo;
747 // Schedule all nodes in a post-order visit
748 Node *unvisited = NULL; // Unvisited anti-dependent Node, if any
750 // Scan for unvisited nodes
751 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
752 // For all uses, schedule late
753 Node* n = self->fast_out(i); // Use
755 // Skip already visited children
756 if ( _visited.test(n->_idx) )
757 continue;
759 // do not traverse backward control edges
760 Node *use = n->is_Proj() ? n->in(0) : n;
761 uint use_rpo = _bbs[use->_idx]->_rpo;
763 if ( use_rpo < src_rpo )
764 continue;
766 // Phi nodes always precede uses in a basic block
767 if ( use_rpo == src_rpo && use->is_Phi() )
768 continue;
770 unvisited = n; // Found unvisited
772 // Check for possible-anti-dependent
773 if( !n->needs_anti_dependence_check() )
774 break; // Not visited, not anti-dep; schedule it NOW
775 }
777 // Did I find an unvisited not-anti-dependent Node?
778 if ( !unvisited )
779 break; // All done with children; post-visit 'self'
781 // Visit the unvisited Node. Contains the obvious push to
782 // indicate I'm entering a deeper level of recursion. I push the
783 // old state onto the _stack and set a new state and loop (recurse).
784 _stack.push(self);
785 self = unvisited;
786 } // End recursion loop
788 return self;
789 }
791 //------------------------------ComputeLatenciesBackwards----------------------
792 // Compute the latency of all the instructions.
793 void PhaseCFG::ComputeLatenciesBackwards(VectorSet &visited, Node_List &stack) {
794 #ifndef PRODUCT
795 if (trace_opto_pipelining())
796 tty->print("\n#---- ComputeLatenciesBackwards ----\n");
797 #endif
799 Node_Backward_Iterator iter((Node *)_root, visited, stack, _bbs);
800 Node *n;
802 // Walk over all the nodes from last to first
803 while (n = iter.next()) {
804 // Set the latency for the definitions of this instruction
805 partial_latency_of_defs(n);
806 }
807 } // end ComputeLatenciesBackwards
809 //------------------------------partial_latency_of_defs------------------------
810 // Compute the latency impact of this node on all defs. This computes
811 // a number that increases as we approach the beginning of the routine.
812 void PhaseCFG::partial_latency_of_defs(Node *n) {
813 // Set the latency for this instruction
814 #ifndef PRODUCT
815 if (trace_opto_pipelining()) {
816 tty->print("# latency_to_inputs: node_latency[%d] = %d for node",
817 n->_idx, _node_latency.at_grow(n->_idx));
818 dump();
819 }
820 #endif
822 if (n->is_Proj())
823 n = n->in(0);
825 if (n->is_Root())
826 return;
828 uint nlen = n->len();
829 uint use_latency = _node_latency.at_grow(n->_idx);
830 uint use_pre_order = _bbs[n->_idx]->_pre_order;
832 for ( uint j=0; j<nlen; j++ ) {
833 Node *def = n->in(j);
835 if (!def || def == n)
836 continue;
838 // Walk backwards thru projections
839 if (def->is_Proj())
840 def = def->in(0);
842 #ifndef PRODUCT
843 if (trace_opto_pipelining()) {
844 tty->print("# in(%2d): ", j);
845 def->dump();
846 }
847 #endif
849 // If the defining block is not known, assume it is ok
850 Block *def_block = _bbs[def->_idx];
851 uint def_pre_order = def_block ? def_block->_pre_order : 0;
853 if ( (use_pre_order < def_pre_order) ||
854 (use_pre_order == def_pre_order && n->is_Phi()) )
855 continue;
857 uint delta_latency = n->latency(j);
858 uint current_latency = delta_latency + use_latency;
860 if (_node_latency.at_grow(def->_idx) < current_latency) {
861 _node_latency.at_put_grow(def->_idx, current_latency);
862 }
864 #ifndef PRODUCT
865 if (trace_opto_pipelining()) {
866 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, node_latency[%d] = %d",
867 use_latency, j, delta_latency, current_latency, def->_idx,
868 _node_latency.at_grow(def->_idx));
869 }
870 #endif
871 }
872 }
874 //------------------------------latency_from_use-------------------------------
875 // Compute the latency of a specific use
876 int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) {
877 // If self-reference, return no latency
878 if (use == n || use->is_Root())
879 return 0;
881 uint def_pre_order = _bbs[def->_idx]->_pre_order;
882 uint latency = 0;
884 // If the use is not a projection, then it is simple...
885 if (!use->is_Proj()) {
886 #ifndef PRODUCT
887 if (trace_opto_pipelining()) {
888 tty->print("# out(): ");
889 use->dump();
890 }
891 #endif
893 uint use_pre_order = _bbs[use->_idx]->_pre_order;
895 if (use_pre_order < def_pre_order)
896 return 0;
898 if (use_pre_order == def_pre_order && use->is_Phi())
899 return 0;
901 uint nlen = use->len();
902 uint nl = _node_latency.at_grow(use->_idx);
904 for ( uint j=0; j<nlen; j++ ) {
905 if (use->in(j) == n) {
906 // Change this if we want local latencies
907 uint ul = use->latency(j);
908 uint l = ul + nl;
909 if (latency < l) latency = l;
910 #ifndef PRODUCT
911 if (trace_opto_pipelining()) {
912 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, latency = %d",
913 nl, j, ul, l, latency);
914 }
915 #endif
916 }
917 }
918 } else {
919 // This is a projection, just grab the latency of the use(s)
920 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
921 uint l = latency_from_use(use, def, use->fast_out(j));
922 if (latency < l) latency = l;
923 }
924 }
926 return latency;
927 }
929 //------------------------------latency_from_uses------------------------------
930 // Compute the latency of this instruction relative to all of it's uses.
931 // This computes a number that increases as we approach the beginning of the
932 // routine.
933 void PhaseCFG::latency_from_uses(Node *n) {
934 // Set the latency for this instruction
935 #ifndef PRODUCT
936 if (trace_opto_pipelining()) {
937 tty->print("# latency_from_outputs: node_latency[%d] = %d for node",
938 n->_idx, _node_latency.at_grow(n->_idx));
939 dump();
940 }
941 #endif
942 uint latency=0;
943 const Node *def = n->is_Proj() ? n->in(0): n;
945 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
946 uint l = latency_from_use(n, def, n->fast_out(i));
948 if (latency < l) latency = l;
949 }
951 _node_latency.at_put_grow(n->_idx, latency);
952 }
954 //------------------------------hoist_to_cheaper_block-------------------------
955 // Pick a block for node self, between early and LCA, that is a cheaper
956 // alternative to LCA.
957 Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) {
958 const double delta = 1+PROB_UNLIKELY_MAG(4);
959 Block* least = LCA;
960 double least_freq = least->_freq;
961 uint target = _node_latency.at_grow(self->_idx);
962 uint start_latency = _node_latency.at_grow(LCA->_nodes[0]->_idx);
963 uint end_latency = _node_latency.at_grow(LCA->_nodes[LCA->end_idx()]->_idx);
964 bool in_latency = (target <= start_latency);
965 const Block* root_block = _bbs[_root->_idx];
967 // Turn off latency scheduling if scheduling is just plain off
968 if (!C->do_scheduling())
969 in_latency = true;
971 // Do not hoist (to cover latency) instructions which target a
972 // single register. Hoisting stretches the live range of the
973 // single register and may force spilling.
974 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
975 if (mach && mach->out_RegMask().is_bound1() && mach->out_RegMask().is_NotEmpty())
976 in_latency = true;
978 #ifndef PRODUCT
979 if (trace_opto_pipelining()) {
980 tty->print("# Find cheaper block for latency %d: ",
981 _node_latency.at_grow(self->_idx));
982 self->dump();
983 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
984 LCA->_pre_order,
985 LCA->_nodes[0]->_idx,
986 start_latency,
987 LCA->_nodes[LCA->end_idx()]->_idx,
988 end_latency,
989 least_freq);
990 }
991 #endif
993 // Walk up the dominator tree from LCA (Lowest common ancestor) to
994 // the earliest legal location. Capture the least execution frequency.
995 while (LCA != early) {
996 LCA = LCA->_idom; // Follow up the dominator tree
998 if (LCA == NULL) {
999 // Bailout without retry
1000 C->record_method_not_compilable("late schedule failed: LCA == NULL");
1001 return least;
1002 }
1004 // Don't hoist machine instructions to the root basic block
1005 if (mach && LCA == root_block)
1006 break;
1008 uint start_lat = _node_latency.at_grow(LCA->_nodes[0]->_idx);
1009 uint end_idx = LCA->end_idx();
1010 uint end_lat = _node_latency.at_grow(LCA->_nodes[end_idx]->_idx);
1011 double LCA_freq = LCA->_freq;
1012 #ifndef PRODUCT
1013 if (trace_opto_pipelining()) {
1014 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1015 LCA->_pre_order, LCA->_nodes[0]->_idx, start_lat, end_idx, end_lat, LCA_freq);
1016 }
1017 #endif
1018 if (LCA_freq < least_freq || // Better Frequency
1019 ( !in_latency && // No block containing latency
1020 LCA_freq < least_freq * delta && // No worse frequency
1021 target >= end_lat && // within latency range
1022 !self->is_iteratively_computed() ) // But don't hoist IV increments
1023 // because they may end up above other uses of their phi forcing
1024 // their result register to be different from their input.
1025 ) {
1026 least = LCA; // Found cheaper block
1027 least_freq = LCA_freq;
1028 start_latency = start_lat;
1029 end_latency = end_lat;
1030 if (target <= start_lat)
1031 in_latency = true;
1032 }
1033 }
1035 #ifndef PRODUCT
1036 if (trace_opto_pipelining()) {
1037 tty->print_cr("# Choose block B%d with start latency=%d and freq=%g",
1038 least->_pre_order, start_latency, least_freq);
1039 }
1040 #endif
1042 // See if the latency needs to be updated
1043 if (target < end_latency) {
1044 #ifndef PRODUCT
1045 if (trace_opto_pipelining()) {
1046 tty->print_cr("# Change latency for [%4d] from %d to %d", self->_idx, target, end_latency);
1047 }
1048 #endif
1049 _node_latency.at_put_grow(self->_idx, end_latency);
1050 partial_latency_of_defs(self);
1051 }
1053 return least;
1054 }
1057 //------------------------------schedule_late-----------------------------------
1058 // Now schedule all codes as LATE as possible. This is the LCA in the
1059 // dominator tree of all USES of a value. Pick the block with the least
1060 // loop nesting depth that is lowest in the dominator tree.
1061 extern const char must_clone[];
1062 void PhaseCFG::schedule_late(VectorSet &visited, Node_List &stack) {
1063 #ifndef PRODUCT
1064 if (trace_opto_pipelining())
1065 tty->print("\n#---- schedule_late ----\n");
1066 #endif
1068 Node_Backward_Iterator iter((Node *)_root, visited, stack, _bbs);
1069 Node *self;
1071 // Walk over all the nodes from last to first
1072 while (self = iter.next()) {
1073 Block* early = _bbs[self->_idx]; // Earliest legal placement
1075 if (self->is_top()) {
1076 // Top node goes in bb #2 with other constants.
1077 // It must be special-cased, because it has no out edges.
1078 early->add_inst(self);
1079 continue;
1080 }
1082 // No uses, just terminate
1083 if (self->outcnt() == 0) {
1084 assert(self->Opcode() == Op_MachProj, "sanity");
1085 continue; // Must be a dead machine projection
1086 }
1088 // If node is pinned in the block, then no scheduling can be done.
1089 if( self->pinned() ) // Pinned in block?
1090 continue;
1092 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1093 if (mach) {
1094 switch (mach->ideal_Opcode()) {
1095 case Op_CreateEx:
1096 // Don't move exception creation
1097 early->add_inst(self);
1098 continue;
1099 break;
1100 case Op_CheckCastPP:
1101 // Don't move CheckCastPP nodes away from their input, if the input
1102 // is a rawptr (5071820).
1103 Node *def = self->in(1);
1104 if (def != NULL && def->bottom_type()->base() == Type::RawPtr) {
1105 early->add_inst(self);
1106 continue;
1107 }
1108 break;
1109 }
1110 }
1112 // Gather LCA of all uses
1113 Block *LCA = NULL;
1114 {
1115 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
1116 // For all uses, find LCA
1117 Node* use = self->fast_out(i);
1118 LCA = raise_LCA_above_use(LCA, use, self, _bbs);
1119 }
1120 } // (Hide defs of imax, i from rest of block.)
1122 // Place temps in the block of their use. This isn't a
1123 // requirement for correctness but it reduces useless
1124 // interference between temps and other nodes.
1125 if (mach != NULL && mach->is_MachTemp()) {
1126 _bbs.map(self->_idx, LCA);
1127 LCA->add_inst(self);
1128 continue;
1129 }
1131 // Check if 'self' could be anti-dependent on memory
1132 if (self->needs_anti_dependence_check()) {
1133 // Hoist LCA above possible-defs and insert anti-dependences to
1134 // defs in new LCA block.
1135 LCA = insert_anti_dependences(LCA, self);
1136 }
1138 if (early->_dom_depth > LCA->_dom_depth) {
1139 // Somehow the LCA has moved above the earliest legal point.
1140 // (One way this can happen is via memory_early_block.)
1141 if (C->subsume_loads() == true && !C->failing()) {
1142 // Retry with subsume_loads == false
1143 // If this is the first failure, the sentinel string will "stick"
1144 // to the Compile object, and the C2Compiler will see it and retry.
1145 C->record_failure(C2Compiler::retry_no_subsuming_loads());
1146 } else {
1147 // Bailout without retry when (early->_dom_depth > LCA->_dom_depth)
1148 C->record_method_not_compilable("late schedule failed: incorrect graph");
1149 }
1150 return;
1151 }
1153 // If there is no opportunity to hoist, then we're done.
1154 bool try_to_hoist = (LCA != early);
1156 // Must clone guys stay next to use; no hoisting allowed.
1157 // Also cannot hoist guys that alter memory or are otherwise not
1158 // allocatable (hoisting can make a value live longer, leading to
1159 // anti and output dependency problems which are normally resolved
1160 // by the register allocator giving everyone a different register).
1161 if (mach != NULL && must_clone[mach->ideal_Opcode()])
1162 try_to_hoist = false;
1164 Block* late = NULL;
1165 if (try_to_hoist) {
1166 // Now find the block with the least execution frequency.
1167 // Start at the latest schedule and work up to the earliest schedule
1168 // in the dominator tree. Thus the Node will dominate all its uses.
1169 late = hoist_to_cheaper_block(LCA, early, self);
1170 } else {
1171 // Just use the LCA of the uses.
1172 late = LCA;
1173 }
1175 // Put the node into target block
1176 schedule_node_into_block(self, late);
1178 #ifdef ASSERT
1179 if (self->needs_anti_dependence_check()) {
1180 // since precedence edges are only inserted when we're sure they
1181 // are needed make sure that after placement in a block we don't
1182 // need any new precedence edges.
1183 verify_anti_dependences(late, self);
1184 }
1185 #endif
1186 } // Loop until all nodes have been visited
1188 } // end ScheduleLate
1190 //------------------------------GlobalCodeMotion-------------------------------
1191 void PhaseCFG::GlobalCodeMotion( Matcher &matcher, uint unique, Node_List &proj_list ) {
1192 ResourceMark rm;
1194 #ifndef PRODUCT
1195 if (trace_opto_pipelining()) {
1196 tty->print("\n---- Start GlobalCodeMotion ----\n");
1197 }
1198 #endif
1200 // Initialize the bbs.map for things on the proj_list
1201 uint i;
1202 for( i=0; i < proj_list.size(); i++ )
1203 _bbs.map(proj_list[i]->_idx, NULL);
1205 // Set the basic block for Nodes pinned into blocks
1206 Arena *a = Thread::current()->resource_area();
1207 VectorSet visited(a);
1208 schedule_pinned_nodes( visited );
1210 // Find the earliest Block any instruction can be placed in. Some
1211 // instructions are pinned into Blocks. Unpinned instructions can
1212 // appear in last block in which all their inputs occur.
1213 visited.Clear();
1214 Node_List stack(a);
1215 stack.map( (unique >> 1) + 16, NULL); // Pre-grow the list
1216 if (!schedule_early(visited, stack)) {
1217 // Bailout without retry
1218 C->record_method_not_compilable("early schedule failed");
1219 return;
1220 }
1222 // Build Def-Use edges.
1223 proj_list.push(_root); // Add real root as another root
1224 proj_list.pop();
1226 // Compute the latency information (via backwards walk) for all the
1227 // instructions in the graph
1228 GrowableArray<uint> node_latency;
1229 _node_latency = node_latency;
1231 if( C->do_scheduling() )
1232 ComputeLatenciesBackwards(visited, stack);
1234 // Now schedule all codes as LATE as possible. This is the LCA in the
1235 // dominator tree of all USES of a value. Pick the block with the least
1236 // loop nesting depth that is lowest in the dominator tree.
1237 // ( visited.Clear() called in schedule_late()->Node_Backward_Iterator() )
1238 schedule_late(visited, stack);
1239 if( C->failing() ) {
1240 // schedule_late fails only when graph is incorrect.
1241 assert(!VerifyGraphEdges, "verification should have failed");
1242 return;
1243 }
1245 unique = C->unique();
1247 #ifndef PRODUCT
1248 if (trace_opto_pipelining()) {
1249 tty->print("\n---- Detect implicit null checks ----\n");
1250 }
1251 #endif
1253 // Detect implicit-null-check opportunities. Basically, find NULL checks
1254 // with suitable memory ops nearby. Use the memory op to do the NULL check.
1255 // I can generate a memory op if there is not one nearby.
1256 if (C->is_method_compilation()) {
1257 // Don't do it for natives, adapters, or runtime stubs
1258 int allowed_reasons = 0;
1259 // ...and don't do it when there have been too many traps, globally.
1260 for (int reason = (int)Deoptimization::Reason_none+1;
1261 reason < Compile::trapHistLength; reason++) {
1262 assert(reason < BitsPerInt, "recode bit map");
1263 if (!C->too_many_traps((Deoptimization::DeoptReason) reason))
1264 allowed_reasons |= nth_bit(reason);
1265 }
1266 // By reversing the loop direction we get a very minor gain on mpegaudio.
1267 // Feel free to revert to a forward loop for clarity.
1268 // for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) {
1269 for( int i= matcher._null_check_tests.size()-2; i>=0; i-=2 ) {
1270 Node *proj = matcher._null_check_tests[i ];
1271 Node *val = matcher._null_check_tests[i+1];
1272 _bbs[proj->_idx]->implicit_null_check(this, proj, val, allowed_reasons);
1273 // The implicit_null_check will only perform the transformation
1274 // if the null branch is truly uncommon, *and* it leads to an
1275 // uncommon trap. Combined with the too_many_traps guards
1276 // above, this prevents SEGV storms reported in 6366351,
1277 // by recompiling offending methods without this optimization.
1278 }
1279 }
1281 #ifndef PRODUCT
1282 if (trace_opto_pipelining()) {
1283 tty->print("\n---- Start Local Scheduling ----\n");
1284 }
1285 #endif
1287 // Schedule locally. Right now a simple topological sort.
1288 // Later, do a real latency aware scheduler.
1289 int *ready_cnt = NEW_RESOURCE_ARRAY(int,C->unique());
1290 memset( ready_cnt, -1, C->unique() * sizeof(int) );
1291 visited.Clear();
1292 for (i = 0; i < _num_blocks; i++) {
1293 if (!_blocks[i]->schedule_local(this, matcher, ready_cnt, visited)) {
1294 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
1295 C->record_method_not_compilable("local schedule failed");
1296 }
1297 return;
1298 }
1299 }
1301 // If we inserted any instructions between a Call and his CatchNode,
1302 // clone the instructions on all paths below the Catch.
1303 for( i=0; i < _num_blocks; i++ )
1304 _blocks[i]->call_catch_cleanup(_bbs);
1306 #ifndef PRODUCT
1307 if (trace_opto_pipelining()) {
1308 tty->print("\n---- After GlobalCodeMotion ----\n");
1309 for (uint i = 0; i < _num_blocks; i++) {
1310 _blocks[i]->dump();
1311 }
1312 }
1313 #endif
1314 }
1317 //------------------------------Estimate_Block_Frequency-----------------------
1318 // Estimate block frequencies based on IfNode probabilities.
1319 void PhaseCFG::Estimate_Block_Frequency() {
1320 int cnts = C->method() ? C->method()->interpreter_invocation_count() : 1;
1321 // Most of our algorithms will die horribly if frequency can become
1322 // negative so make sure cnts is a sane value.
1323 if( cnts <= 0 ) cnts = 1;
1324 float f = (float)cnts/(float)FreqCountInvocations;
1326 // Create the loop tree and calculate loop depth.
1327 _root_loop = create_loop_tree();
1328 _root_loop->compute_loop_depth(0);
1330 // Compute block frequency of each block, relative to a single loop entry.
1331 _root_loop->compute_freq();
1333 // Adjust all frequencies to be relative to a single method entry
1334 _root_loop->_freq = f * 1.0;
1335 _root_loop->scale_freq();
1337 // force paths ending at uncommon traps to be infrequent
1338 Block_List worklist;
1339 Block* root_blk = _blocks[0];
1340 for (uint i = 0; i < root_blk->num_preds(); i++) {
1341 Block *pb = _bbs[root_blk->pred(i)->_idx];
1342 if (pb->has_uncommon_code()) {
1343 worklist.push(pb);
1344 }
1345 }
1346 while (worklist.size() > 0) {
1347 Block* uct = worklist.pop();
1348 uct->_freq = PROB_MIN;
1349 for (uint i = 0; i < uct->num_preds(); i++) {
1350 Block *pb = _bbs[uct->pred(i)->_idx];
1351 if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) {
1352 worklist.push(pb);
1353 }
1354 }
1355 }
1357 #ifndef PRODUCT
1358 if (PrintCFGBlockFreq) {
1359 tty->print_cr("CFG Block Frequencies");
1360 _root_loop->dump_tree();
1361 if (Verbose) {
1362 tty->print_cr("PhaseCFG dump");
1363 dump();
1364 tty->print_cr("Node dump");
1365 _root->dump(99999);
1366 }
1367 }
1368 #endif
1369 }
1371 //----------------------------create_loop_tree--------------------------------
1372 // Create a loop tree from the CFG
1373 CFGLoop* PhaseCFG::create_loop_tree() {
1375 #ifdef ASSERT
1376 assert( _blocks[0] == _broot, "" );
1377 for (uint i = 0; i < _num_blocks; i++ ) {
1378 Block *b = _blocks[i];
1379 // Check that _loop field are clear...we could clear them if not.
1380 assert(b->_loop == NULL, "clear _loop expected");
1381 // Sanity check that the RPO numbering is reflected in the _blocks array.
1382 // It doesn't have to be for the loop tree to be built, but if it is not,
1383 // then the blocks have been reordered since dom graph building...which
1384 // may question the RPO numbering
1385 assert(b->_rpo == i, "unexpected reverse post order number");
1386 }
1387 #endif
1389 int idct = 0;
1390 CFGLoop* root_loop = new CFGLoop(idct++);
1392 Block_List worklist;
1394 // Assign blocks to loops
1395 for(uint i = _num_blocks - 1; i > 0; i-- ) { // skip Root block
1396 Block *b = _blocks[i];
1398 if (b->head()->is_Loop()) {
1399 Block* loop_head = b;
1400 assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1401 Node* tail_n = loop_head->pred(LoopNode::LoopBackControl);
1402 Block* tail = _bbs[tail_n->_idx];
1404 // Defensively filter out Loop nodes for non-single-entry loops.
1405 // For all reasonable loops, the head occurs before the tail in RPO.
1406 if (i <= tail->_rpo) {
1408 // The tail and (recursive) predecessors of the tail
1409 // are made members of a new loop.
1411 assert(worklist.size() == 0, "nonempty worklist");
1412 CFGLoop* nloop = new CFGLoop(idct++);
1413 assert(loop_head->_loop == NULL, "just checking");
1414 loop_head->_loop = nloop;
1415 // Add to nloop so push_pred() will skip over inner loops
1416 nloop->add_member(loop_head);
1417 nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, _bbs);
1419 while (worklist.size() > 0) {
1420 Block* member = worklist.pop();
1421 if (member != loop_head) {
1422 for (uint j = 1; j < member->num_preds(); j++) {
1423 nloop->push_pred(member, j, worklist, _bbs);
1424 }
1425 }
1426 }
1427 }
1428 }
1429 }
1431 // Create a member list for each loop consisting
1432 // of both blocks and (immediate child) loops.
1433 for (uint i = 0; i < _num_blocks; i++) {
1434 Block *b = _blocks[i];
1435 CFGLoop* lp = b->_loop;
1436 if (lp == NULL) {
1437 // Not assigned to a loop. Add it to the method's pseudo loop.
1438 b->_loop = root_loop;
1439 lp = root_loop;
1440 }
1441 if (lp == root_loop || b != lp->head()) { // loop heads are already members
1442 lp->add_member(b);
1443 }
1444 if (lp != root_loop) {
1445 if (lp->parent() == NULL) {
1446 // Not a nested loop. Make it a child of the method's pseudo loop.
1447 root_loop->add_nested_loop(lp);
1448 }
1449 if (b == lp->head()) {
1450 // Add nested loop to member list of parent loop.
1451 lp->parent()->add_member(lp);
1452 }
1453 }
1454 }
1456 return root_loop;
1457 }
1459 //------------------------------push_pred--------------------------------------
1460 void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, Block_Array& node_to_blk) {
1461 Node* pred_n = blk->pred(i);
1462 Block* pred = node_to_blk[pred_n->_idx];
1463 CFGLoop *pred_loop = pred->_loop;
1464 if (pred_loop == NULL) {
1465 // Filter out blocks for non-single-entry loops.
1466 // For all reasonable loops, the head occurs before the tail in RPO.
1467 if (pred->_rpo > head()->_rpo) {
1468 pred->_loop = this;
1469 worklist.push(pred);
1470 }
1471 } else if (pred_loop != this) {
1472 // Nested loop.
1473 while (pred_loop->_parent != NULL && pred_loop->_parent != this) {
1474 pred_loop = pred_loop->_parent;
1475 }
1476 // Make pred's loop be a child
1477 if (pred_loop->_parent == NULL) {
1478 add_nested_loop(pred_loop);
1479 // Continue with loop entry predecessor.
1480 Block* pred_head = pred_loop->head();
1481 assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1482 assert(pred_head != head(), "loop head in only one loop");
1483 push_pred(pred_head, LoopNode::EntryControl, worklist, node_to_blk);
1484 } else {
1485 assert(pred_loop->_parent == this && _parent == NULL, "just checking");
1486 }
1487 }
1488 }
1490 //------------------------------add_nested_loop--------------------------------
1491 // Make cl a child of the current loop in the loop tree.
1492 void CFGLoop::add_nested_loop(CFGLoop* cl) {
1493 assert(_parent == NULL, "no parent yet");
1494 assert(cl != this, "not my own parent");
1495 cl->_parent = this;
1496 CFGLoop* ch = _child;
1497 if (ch == NULL) {
1498 _child = cl;
1499 } else {
1500 while (ch->_sibling != NULL) { ch = ch->_sibling; }
1501 ch->_sibling = cl;
1502 }
1503 }
1505 //------------------------------compute_loop_depth-----------------------------
1506 // Store the loop depth in each CFGLoop object.
1507 // Recursively walk the children to do the same for them.
1508 void CFGLoop::compute_loop_depth(int depth) {
1509 _depth = depth;
1510 CFGLoop* ch = _child;
1511 while (ch != NULL) {
1512 ch->compute_loop_depth(depth + 1);
1513 ch = ch->_sibling;
1514 }
1515 }
1517 //------------------------------compute_freq-----------------------------------
1518 // Compute the frequency of each block and loop, relative to a single entry
1519 // into the dominating loop head.
1520 void CFGLoop::compute_freq() {
1521 // Bottom up traversal of loop tree (visit inner loops first.)
1522 // Set loop head frequency to 1.0, then transitively
1523 // compute frequency for all successors in the loop,
1524 // as well as for each exit edge. Inner loops are
1525 // treated as single blocks with loop exit targets
1526 // as the successor blocks.
1528 // Nested loops first
1529 CFGLoop* ch = _child;
1530 while (ch != NULL) {
1531 ch->compute_freq();
1532 ch = ch->_sibling;
1533 }
1534 assert (_members.length() > 0, "no empty loops");
1535 Block* hd = head();
1536 hd->_freq = 1.0f;
1537 for (int i = 0; i < _members.length(); i++) {
1538 CFGElement* s = _members.at(i);
1539 float freq = s->_freq;
1540 if (s->is_block()) {
1541 Block* b = s->as_Block();
1542 for (uint j = 0; j < b->_num_succs; j++) {
1543 Block* sb = b->_succs[j];
1544 update_succ_freq(sb, freq * b->succ_prob(j));
1545 }
1546 } else {
1547 CFGLoop* lp = s->as_CFGLoop();
1548 assert(lp->_parent == this, "immediate child");
1549 for (int k = 0; k < lp->_exits.length(); k++) {
1550 Block* eb = lp->_exits.at(k).get_target();
1551 float prob = lp->_exits.at(k).get_prob();
1552 update_succ_freq(eb, freq * prob);
1553 }
1554 }
1555 }
1557 #if 0
1558 // Raise frequency of the loop backedge block, in an effort
1559 // to keep it empty. Skip the method level "loop".
1560 if (_parent != NULL) {
1561 CFGElement* s = _members.at(_members.length() - 1);
1562 if (s->is_block()) {
1563 Block* bk = s->as_Block();
1564 if (bk->_num_succs == 1 && bk->_succs[0] == hd) {
1565 // almost any value >= 1.0f works
1566 // FIXME: raw constant
1567 bk->_freq = 1.05f;
1568 }
1569 }
1570 }
1571 #endif
1573 // For all loops other than the outer, "method" loop,
1574 // sum and normalize the exit probability. The "method" loop
1575 // should keep the initial exit probability of 1, so that
1576 // inner blocks do not get erroneously scaled.
1577 if (_depth != 0) {
1578 // Total the exit probabilities for this loop.
1579 float exits_sum = 0.0f;
1580 for (int i = 0; i < _exits.length(); i++) {
1581 exits_sum += _exits.at(i).get_prob();
1582 }
1584 // Normalize the exit probabilities. Until now, the
1585 // probabilities estimate the possibility of exit per
1586 // a single loop iteration; afterward, they estimate
1587 // the probability of exit per loop entry.
1588 for (int i = 0; i < _exits.length(); i++) {
1589 Block* et = _exits.at(i).get_target();
1590 float new_prob = _exits.at(i).get_prob() / exits_sum;
1591 BlockProbPair bpp(et, new_prob);
1592 _exits.at_put(i, bpp);
1593 }
1595 // Save the total, but guard against unreasoable probability,
1596 // as the value is used to estimate the loop trip count.
1597 // An infinite trip count would blur relative block
1598 // frequencies.
1599 if (exits_sum > 1.0f) exits_sum = 1.0;
1600 if (exits_sum < PROB_MIN) exits_sum = PROB_MIN;
1601 _exit_prob = exits_sum;
1602 }
1603 }
1605 //------------------------------succ_prob-------------------------------------
1606 // Determine the probability of reaching successor 'i' from the receiver block.
1607 float Block::succ_prob(uint i) {
1608 int eidx = end_idx();
1609 Node *n = _nodes[eidx]; // Get ending Node
1610 int op = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : n->Opcode();
1612 // Switch on branch type
1613 switch( op ) {
1614 case Op_CountedLoopEnd:
1615 case Op_If: {
1616 assert (i < 2, "just checking");
1617 // Conditionals pass on only part of their frequency
1618 float prob = n->as_MachIf()->_prob;
1619 assert(prob >= 0.0 && prob <= 1.0, "out of range probability");
1620 // If succ[i] is the FALSE branch, invert path info
1621 if( _nodes[i + eidx + 1]->Opcode() == Op_IfFalse ) {
1622 return 1.0f - prob; // not taken
1623 } else {
1624 return prob; // taken
1625 }
1626 }
1628 case Op_Jump:
1629 // Divide the frequency between all successors evenly
1630 return 1.0f/_num_succs;
1632 case Op_Catch: {
1633 const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
1634 if (ci->_con == CatchProjNode::fall_through_index) {
1635 // Fall-thru path gets the lion's share.
1636 return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs;
1637 } else {
1638 // Presume exceptional paths are equally unlikely
1639 return PROB_UNLIKELY_MAG(5);
1640 }
1641 }
1643 case Op_Root:
1644 case Op_Goto:
1645 // Pass frequency straight thru to target
1646 return 1.0f;
1648 case Op_NeverBranch:
1649 return 0.0f;
1651 case Op_TailCall:
1652 case Op_TailJump:
1653 case Op_Return:
1654 case Op_Halt:
1655 case Op_Rethrow:
1656 // Do not push out freq to root block
1657 return 0.0f;
1659 default:
1660 ShouldNotReachHere();
1661 }
1663 return 0.0f;
1664 }
1666 //------------------------------update_succ_freq-------------------------------
1667 // Update the appropriate frequency associated with block 'b', a succesor of
1668 // a block in this loop.
1669 void CFGLoop::update_succ_freq(Block* b, float freq) {
1670 if (b->_loop == this) {
1671 if (b == head()) {
1672 // back branch within the loop
1673 // Do nothing now, the loop carried frequency will be
1674 // adjust later in scale_freq().
1675 } else {
1676 // simple branch within the loop
1677 b->_freq += freq;
1678 }
1679 } else if (!in_loop_nest(b)) {
1680 // branch is exit from this loop
1681 BlockProbPair bpp(b, freq);
1682 _exits.append(bpp);
1683 } else {
1684 // branch into nested loop
1685 CFGLoop* ch = b->_loop;
1686 ch->_freq += freq;
1687 }
1688 }
1690 //------------------------------in_loop_nest-----------------------------------
1691 // Determine if block b is in the receiver's loop nest.
1692 bool CFGLoop::in_loop_nest(Block* b) {
1693 int depth = _depth;
1694 CFGLoop* b_loop = b->_loop;
1695 int b_depth = b_loop->_depth;
1696 if (depth == b_depth) {
1697 return true;
1698 }
1699 while (b_depth > depth) {
1700 b_loop = b_loop->_parent;
1701 b_depth = b_loop->_depth;
1702 }
1703 return b_loop == this;
1704 }
1706 //------------------------------scale_freq-------------------------------------
1707 // Scale frequency of loops and blocks by trip counts from outer loops
1708 // Do a top down traversal of loop tree (visit outer loops first.)
1709 void CFGLoop::scale_freq() {
1710 float loop_freq = _freq * trip_count();
1711 for (int i = 0; i < _members.length(); i++) {
1712 CFGElement* s = _members.at(i);
1713 s->_freq *= loop_freq;
1714 }
1715 CFGLoop* ch = _child;
1716 while (ch != NULL) {
1717 ch->scale_freq();
1718 ch = ch->_sibling;
1719 }
1720 }
1722 #ifndef PRODUCT
1723 //------------------------------dump_tree--------------------------------------
1724 void CFGLoop::dump_tree() const {
1725 dump();
1726 if (_child != NULL) _child->dump_tree();
1727 if (_sibling != NULL) _sibling->dump_tree();
1728 }
1730 //------------------------------dump-------------------------------------------
1731 void CFGLoop::dump() const {
1732 for (int i = 0; i < _depth; i++) tty->print(" ");
1733 tty->print("%s: %d trip_count: %6.0f freq: %6.0f\n",
1734 _depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq);
1735 for (int i = 0; i < _depth; i++) tty->print(" ");
1736 tty->print(" members:", _id);
1737 int k = 0;
1738 for (int i = 0; i < _members.length(); i++) {
1739 if (k++ >= 6) {
1740 tty->print("\n ");
1741 for (int j = 0; j < _depth+1; j++) tty->print(" ");
1742 k = 0;
1743 }
1744 CFGElement *s = _members.at(i);
1745 if (s->is_block()) {
1746 Block *b = s->as_Block();
1747 tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq);
1748 } else {
1749 CFGLoop* lp = s->as_CFGLoop();
1750 tty->print(" L%d(%6.3f)", lp->_id, lp->_freq);
1751 }
1752 }
1753 tty->print("\n");
1754 for (int i = 0; i < _depth; i++) tty->print(" ");
1755 tty->print(" exits: ");
1756 k = 0;
1757 for (int i = 0; i < _exits.length(); i++) {
1758 if (k++ >= 7) {
1759 tty->print("\n ");
1760 for (int j = 0; j < _depth+1; j++) tty->print(" ");
1761 k = 0;
1762 }
1763 Block *blk = _exits.at(i).get_target();
1764 float prob = _exits.at(i).get_prob();
1765 tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100));
1766 }
1767 tty->print("\n");
1768 }
1769 #endif