Fri, 26 Jun 2009 13:03:29 -0700
6818666: G1: Type lost in g1 pre-barrier
Reviewed-by: kvn
1 /*
2 * Copyright 1997-2009 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 // Portions of code courtesy of Clifford Click
27 // Optimization - Graph Style
29 #include "incls/_precompiled.incl"
30 #include "incls/_gcm.cpp.incl"
32 // To avoid float value underflow
33 #define MIN_BLOCK_FREQUENCY 1.e-35f
35 //----------------------------schedule_node_into_block-------------------------
36 // Insert node n into block b. Look for projections of n and make sure they
37 // are in b also.
38 void PhaseCFG::schedule_node_into_block( Node *n, Block *b ) {
39 // Set basic block of n, Add n to b,
40 _bbs.map(n->_idx, b);
41 b->add_inst(n);
43 // After Matching, nearly any old Node may have projections trailing it.
44 // These are usually machine-dependent flags. In any case, they might
45 // float to another block below this one. Move them up.
46 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
47 Node* use = n->fast_out(i);
48 if (use->is_Proj()) {
49 Block* buse = _bbs[use->_idx];
50 if (buse != b) { // In wrong block?
51 if (buse != NULL)
52 buse->find_remove(use); // Remove from wrong block
53 _bbs.map(use->_idx, b); // Re-insert in this block
54 b->add_inst(use);
55 }
56 }
57 }
58 }
60 //----------------------------replace_block_proj_ctrl-------------------------
61 // Nodes that have is_block_proj() nodes as their control need to use
62 // the appropriate Region for their actual block as their control since
63 // the projection will be in a predecessor block.
64 void PhaseCFG::replace_block_proj_ctrl( Node *n ) {
65 const Node *in0 = n->in(0);
66 assert(in0 != NULL, "Only control-dependent");
67 const Node *p = in0->is_block_proj();
68 if (p != NULL && p != n) { // Control from a block projection?
69 assert(!n->pinned() || n->is_SafePointScalarObject(), "only SafePointScalarObject pinned node is expected here");
70 // Find trailing Region
71 Block *pb = _bbs[in0->_idx]; // Block-projection already has basic block
72 uint j = 0;
73 if (pb->_num_succs != 1) { // More then 1 successor?
74 // Search for successor
75 uint max = pb->_nodes.size();
76 assert( max > 1, "" );
77 uint start = max - pb->_num_succs;
78 // Find which output path belongs to projection
79 for (j = start; j < max; j++) {
80 if( pb->_nodes[j] == in0 )
81 break;
82 }
83 assert( j < max, "must find" );
84 // Change control to match head of successor basic block
85 j -= start;
86 }
87 n->set_req(0, pb->_succs[j]->head());
88 }
89 }
92 //------------------------------schedule_pinned_nodes--------------------------
93 // Set the basic block for Nodes pinned into blocks
94 void PhaseCFG::schedule_pinned_nodes( VectorSet &visited ) {
95 // Allocate node stack of size C->unique()+8 to avoid frequent realloc
96 GrowableArray <Node *> spstack(C->unique()+8);
97 spstack.push(_root);
98 while ( spstack.is_nonempty() ) {
99 Node *n = spstack.pop();
100 if( !visited.test_set(n->_idx) ) { // Test node and flag it as visited
101 if( n->pinned() && !_bbs.lookup(n->_idx) ) { // Pinned? Nail it down!
102 assert( n->in(0), "pinned Node must have Control" );
103 // Before setting block replace block_proj control edge
104 replace_block_proj_ctrl(n);
105 Node *input = n->in(0);
106 while( !input->is_block_start() )
107 input = input->in(0);
108 Block *b = _bbs[input->_idx]; // Basic block of controlling input
109 schedule_node_into_block(n, b);
110 }
111 for( int i = n->req() - 1; i >= 0; --i ) { // For all inputs
112 if( n->in(i) != NULL )
113 spstack.push(n->in(i));
114 }
115 }
116 }
117 }
119 #ifdef ASSERT
120 // Assert that new input b2 is dominated by all previous inputs.
121 // Check this by by seeing that it is dominated by b1, the deepest
122 // input observed until b2.
123 static void assert_dom(Block* b1, Block* b2, Node* n, Block_Array &bbs) {
124 if (b1 == NULL) return;
125 assert(b1->_dom_depth < b2->_dom_depth, "sanity");
126 Block* tmp = b2;
127 while (tmp != b1 && tmp != NULL) {
128 tmp = tmp->_idom;
129 }
130 if (tmp != b1) {
131 // Detected an unschedulable graph. Print some nice stuff and die.
132 tty->print_cr("!!! Unschedulable graph !!!");
133 for (uint j=0; j<n->len(); j++) { // For all inputs
134 Node* inn = n->in(j); // Get input
135 if (inn == NULL) continue; // Ignore NULL, missing inputs
136 Block* inb = bbs[inn->_idx];
137 tty->print("B%d idom=B%d depth=%2d ",inb->_pre_order,
138 inb->_idom ? inb->_idom->_pre_order : 0, inb->_dom_depth);
139 inn->dump();
140 }
141 tty->print("Failing node: ");
142 n->dump();
143 assert(false, "unscheduable graph");
144 }
145 }
146 #endif
148 static Block* find_deepest_input(Node* n, Block_Array &bbs) {
149 // Find the last input dominated by all other inputs.
150 Block* deepb = NULL; // Deepest block so far
151 int deepb_dom_depth = 0;
152 for (uint k = 0; k < n->len(); k++) { // For all inputs
153 Node* inn = n->in(k); // Get input
154 if (inn == NULL) continue; // Ignore NULL, missing inputs
155 Block* inb = bbs[inn->_idx];
156 assert(inb != NULL, "must already have scheduled this input");
157 if (deepb_dom_depth < (int) inb->_dom_depth) {
158 // The new inb must be dominated by the previous deepb.
159 // The various inputs must be linearly ordered in the dom
160 // tree, or else there will not be a unique deepest block.
161 DEBUG_ONLY(assert_dom(deepb, inb, n, bbs));
162 deepb = inb; // Save deepest block
163 deepb_dom_depth = deepb->_dom_depth;
164 }
165 }
166 assert(deepb != NULL, "must be at least one input to n");
167 return deepb;
168 }
171 //------------------------------schedule_early---------------------------------
172 // Find the earliest Block any instruction can be placed in. Some instructions
173 // are pinned into Blocks. Unpinned instructions can appear in last block in
174 // which all their inputs occur.
175 bool PhaseCFG::schedule_early(VectorSet &visited, Node_List &roots) {
176 // Allocate stack with enough space to avoid frequent realloc
177 Node_Stack nstack(roots.Size() + 8); // (unique >> 1) + 24 from Java2D stats
178 // roots.push(_root); _root will be processed among C->top() inputs
179 roots.push(C->top());
180 visited.set(C->top()->_idx);
182 while (roots.size() != 0) {
183 // Use local variables nstack_top_n & nstack_top_i to cache values
184 // on stack's top.
185 Node *nstack_top_n = roots.pop();
186 uint nstack_top_i = 0;
187 //while_nstack_nonempty:
188 while (true) {
189 // Get parent node and next input's index from stack's top.
190 Node *n = nstack_top_n;
191 uint i = nstack_top_i;
193 if (i == 0) {
194 // Fixup some control. Constants without control get attached
195 // to root and nodes that use is_block_proj() nodes should be attached
196 // to the region that starts their block.
197 const Node *in0 = n->in(0);
198 if (in0 != NULL) { // Control-dependent?
199 replace_block_proj_ctrl(n);
200 } else { // n->in(0) == NULL
201 if (n->req() == 1) { // This guy is a constant with NO inputs?
202 n->set_req(0, _root);
203 }
204 }
205 }
207 // First, visit all inputs and force them to get a block. If an
208 // input is already in a block we quit following inputs (to avoid
209 // cycles). Instead we put that Node on a worklist to be handled
210 // later (since IT'S inputs may not have a block yet).
211 bool done = true; // Assume all n's inputs will be processed
212 while (i < n->len()) { // For all inputs
213 Node *in = n->in(i); // Get input
214 ++i;
215 if (in == NULL) continue; // Ignore NULL, missing inputs
216 int is_visited = visited.test_set(in->_idx);
217 if (!_bbs.lookup(in->_idx)) { // Missing block selection?
218 if (is_visited) {
219 // assert( !visited.test(in->_idx), "did not schedule early" );
220 return false;
221 }
222 nstack.push(n, i); // Save parent node and next input's index.
223 nstack_top_n = in; // Process current input now.
224 nstack_top_i = 0;
225 done = false; // Not all n's inputs processed.
226 break; // continue while_nstack_nonempty;
227 } else if (!is_visited) { // Input not yet visited?
228 roots.push(in); // Visit this guy later, using worklist
229 }
230 }
231 if (done) {
232 // All of n's inputs have been processed, complete post-processing.
234 // Some instructions are pinned into a block. These include Region,
235 // Phi, Start, Return, and other control-dependent instructions and
236 // any projections which depend on them.
237 if (!n->pinned()) {
238 // Set earliest legal block.
239 _bbs.map(n->_idx, find_deepest_input(n, _bbs));
240 } else {
241 assert(_bbs[n->_idx] == _bbs[n->in(0)->_idx], "Pinned Node should be at the same block as its control edge");
242 }
244 if (nstack.is_empty()) {
245 // Finished all nodes on stack.
246 // Process next node on the worklist 'roots'.
247 break;
248 }
249 // Get saved parent node and next input's index.
250 nstack_top_n = nstack.node();
251 nstack_top_i = nstack.index();
252 nstack.pop();
253 } // if (done)
254 } // while (true)
255 } // while (roots.size() != 0)
256 return true;
257 }
259 //------------------------------dom_lca----------------------------------------
260 // Find least common ancestor in dominator tree
261 // LCA is a current notion of LCA, to be raised above 'this'.
262 // As a convenient boundary condition, return 'this' if LCA is NULL.
263 // Find the LCA of those two nodes.
264 Block* Block::dom_lca(Block* LCA) {
265 if (LCA == NULL || LCA == this) return this;
267 Block* anc = this;
268 while (anc->_dom_depth > LCA->_dom_depth)
269 anc = anc->_idom; // Walk up till anc is as high as LCA
271 while (LCA->_dom_depth > anc->_dom_depth)
272 LCA = LCA->_idom; // Walk up till LCA is as high as anc
274 while (LCA != anc) { // Walk both up till they are the same
275 LCA = LCA->_idom;
276 anc = anc->_idom;
277 }
279 return LCA;
280 }
282 //--------------------------raise_LCA_above_use--------------------------------
283 // We are placing a definition, and have been given a def->use edge.
284 // The definition must dominate the use, so move the LCA upward in the
285 // dominator tree to dominate the use. If the use is a phi, adjust
286 // the LCA only with the phi input paths which actually use this def.
287 static Block* raise_LCA_above_use(Block* LCA, Node* use, Node* def, Block_Array &bbs) {
288 Block* buse = bbs[use->_idx];
289 if (buse == NULL) return LCA; // Unused killing Projs have no use block
290 if (!use->is_Phi()) return buse->dom_lca(LCA);
291 uint pmax = use->req(); // Number of Phi inputs
292 // Why does not this loop just break after finding the matching input to
293 // the Phi? Well...it's like this. I do not have true def-use/use-def
294 // chains. Means I cannot distinguish, from the def-use direction, which
295 // of many use-defs lead from the same use to the same def. That is, this
296 // Phi might have several uses of the same def. Each use appears in a
297 // different predecessor block. But when I enter here, I cannot distinguish
298 // which use-def edge I should find the predecessor block for. So I find
299 // them all. Means I do a little extra work if a Phi uses the same value
300 // more than once.
301 for (uint j=1; j<pmax; j++) { // For all inputs
302 if (use->in(j) == def) { // Found matching input?
303 Block* pred = bbs[buse->pred(j)->_idx];
304 LCA = pred->dom_lca(LCA);
305 }
306 }
307 return LCA;
308 }
310 //----------------------------raise_LCA_above_marks----------------------------
311 // Return a new LCA that dominates LCA and any of its marked predecessors.
312 // Search all my parents up to 'early' (exclusive), looking for predecessors
313 // which are marked with the given index. Return the LCA (in the dom tree)
314 // of all marked blocks. If there are none marked, return the original
315 // LCA.
316 static Block* raise_LCA_above_marks(Block* LCA, node_idx_t mark,
317 Block* early, Block_Array &bbs) {
318 Block_List worklist;
319 worklist.push(LCA);
320 while (worklist.size() > 0) {
321 Block* mid = worklist.pop();
322 if (mid == early) continue; // stop searching here
324 // Test and set the visited bit.
325 if (mid->raise_LCA_visited() == mark) continue; // already visited
327 // Don't process the current LCA, otherwise the search may terminate early
328 if (mid != LCA && mid->raise_LCA_mark() == mark) {
329 // Raise the LCA.
330 LCA = mid->dom_lca(LCA);
331 if (LCA == early) break; // stop searching everywhere
332 assert(early->dominates(LCA), "early is high enough");
333 // Resume searching at that point, skipping intermediate levels.
334 worklist.push(LCA);
335 if (LCA == mid)
336 continue; // Don't mark as visited to avoid early termination.
337 } else {
338 // Keep searching through this block's predecessors.
339 for (uint j = 1, jmax = mid->num_preds(); j < jmax; j++) {
340 Block* mid_parent = bbs[ mid->pred(j)->_idx ];
341 worklist.push(mid_parent);
342 }
343 }
344 mid->set_raise_LCA_visited(mark);
345 }
346 return LCA;
347 }
349 //--------------------------memory_early_block--------------------------------
350 // This is a variation of find_deepest_input, the heart of schedule_early.
351 // Find the "early" block for a load, if we considered only memory and
352 // address inputs, that is, if other data inputs were ignored.
353 //
354 // Because a subset of edges are considered, the resulting block will
355 // be earlier (at a shallower dom_depth) than the true schedule_early
356 // point of the node. We compute this earlier block as a more permissive
357 // site for anti-dependency insertion, but only if subsume_loads is enabled.
358 static Block* memory_early_block(Node* load, Block* early, Block_Array &bbs) {
359 Node* base;
360 Node* index;
361 Node* store = load->in(MemNode::Memory);
362 load->as_Mach()->memory_inputs(base, index);
364 assert(base != NodeSentinel && index != NodeSentinel,
365 "unexpected base/index inputs");
367 Node* mem_inputs[4];
368 int mem_inputs_length = 0;
369 if (base != NULL) mem_inputs[mem_inputs_length++] = base;
370 if (index != NULL) mem_inputs[mem_inputs_length++] = index;
371 if (store != NULL) mem_inputs[mem_inputs_length++] = store;
373 // In the comparision below, add one to account for the control input,
374 // which may be null, but always takes up a spot in the in array.
375 if (mem_inputs_length + 1 < (int) load->req()) {
376 // This "load" has more inputs than just the memory, base and index inputs.
377 // For purposes of checking anti-dependences, we need to start
378 // from the early block of only the address portion of the instruction,
379 // and ignore other blocks that may have factored into the wider
380 // schedule_early calculation.
381 if (load->in(0) != NULL) mem_inputs[mem_inputs_length++] = load->in(0);
383 Block* deepb = NULL; // Deepest block so far
384 int deepb_dom_depth = 0;
385 for (int i = 0; i < mem_inputs_length; i++) {
386 Block* inb = bbs[mem_inputs[i]->_idx];
387 if (deepb_dom_depth < (int) inb->_dom_depth) {
388 // The new inb must be dominated by the previous deepb.
389 // The various inputs must be linearly ordered in the dom
390 // tree, or else there will not be a unique deepest block.
391 DEBUG_ONLY(assert_dom(deepb, inb, load, bbs));
392 deepb = inb; // Save deepest block
393 deepb_dom_depth = deepb->_dom_depth;
394 }
395 }
396 early = deepb;
397 }
399 return early;
400 }
402 //--------------------------insert_anti_dependences---------------------------
403 // A load may need to witness memory that nearby stores can overwrite.
404 // For each nearby store, either insert an "anti-dependence" edge
405 // from the load to the store, or else move LCA upward to force the
406 // load to (eventually) be scheduled in a block above the store.
407 //
408 // Do not add edges to stores on distinct control-flow paths;
409 // only add edges to stores which might interfere.
410 //
411 // Return the (updated) LCA. There will not be any possibly interfering
412 // store between the load's "early block" and the updated LCA.
413 // Any stores in the updated LCA will have new precedence edges
414 // back to the load. The caller is expected to schedule the load
415 // in the LCA, in which case the precedence edges will make LCM
416 // preserve anti-dependences. The caller may also hoist the load
417 // above the LCA, if it is not the early block.
418 Block* PhaseCFG::insert_anti_dependences(Block* LCA, Node* load, bool verify) {
419 assert(load->needs_anti_dependence_check(), "must be a load of some sort");
420 assert(LCA != NULL, "");
421 DEBUG_ONLY(Block* LCA_orig = LCA);
423 // Compute the alias index. Loads and stores with different alias indices
424 // do not need anti-dependence edges.
425 uint load_alias_idx = C->get_alias_index(load->adr_type());
426 #ifdef ASSERT
427 if (load_alias_idx == Compile::AliasIdxBot && C->AliasLevel() > 0 &&
428 (PrintOpto || VerifyAliases ||
429 PrintMiscellaneous && (WizardMode || Verbose))) {
430 // Load nodes should not consume all of memory.
431 // Reporting a bottom type indicates a bug in adlc.
432 // If some particular type of node validly consumes all of memory,
433 // sharpen the preceding "if" to exclude it, so we can catch bugs here.
434 tty->print_cr("*** Possible Anti-Dependence Bug: Load consumes all of memory.");
435 load->dump(2);
436 if (VerifyAliases) assert(load_alias_idx != Compile::AliasIdxBot, "");
437 }
438 #endif
439 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrComp),
440 "String compare is only known 'load' that does not conflict with any stores");
441 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrEquals),
442 "String equals is a 'load' that does not conflict with any stores");
443 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrIndexOf),
444 "String indexOf is a 'load' that does not conflict with any stores");
445 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_AryEq),
446 "Arrays equals is a 'load' that do not conflict with any stores");
448 if (!C->alias_type(load_alias_idx)->is_rewritable()) {
449 // It is impossible to spoil this load by putting stores before it,
450 // because we know that the stores will never update the value
451 // which 'load' must witness.
452 return LCA;
453 }
455 node_idx_t load_index = load->_idx;
457 // Note the earliest legal placement of 'load', as determined by
458 // by the unique point in the dom tree where all memory effects
459 // and other inputs are first available. (Computed by schedule_early.)
460 // For normal loads, 'early' is the shallowest place (dom graph wise)
461 // to look for anti-deps between this load and any store.
462 Block* early = _bbs[load_index];
464 // If we are subsuming loads, compute an "early" block that only considers
465 // memory or address inputs. This block may be different than the
466 // schedule_early block in that it could be at an even shallower depth in the
467 // dominator tree, and allow for a broader discovery of anti-dependences.
468 if (C->subsume_loads()) {
469 early = memory_early_block(load, early, _bbs);
470 }
472 ResourceArea *area = Thread::current()->resource_area();
473 Node_List worklist_mem(area); // prior memory state to store
474 Node_List worklist_store(area); // possible-def to explore
475 Node_List worklist_visited(area); // visited mergemem nodes
476 Node_List non_early_stores(area); // all relevant stores outside of early
477 bool must_raise_LCA = false;
479 #ifdef TRACK_PHI_INPUTS
480 // %%% This extra checking fails because MergeMem nodes are not GVNed.
481 // Provide "phi_inputs" to check if every input to a PhiNode is from the
482 // original memory state. This indicates a PhiNode for which should not
483 // prevent the load from sinking. For such a block, set_raise_LCA_mark
484 // may be overly conservative.
485 // Mechanism: count inputs seen for each Phi encountered in worklist_store.
486 DEBUG_ONLY(GrowableArray<uint> phi_inputs(area, C->unique(),0,0));
487 #endif
489 // 'load' uses some memory state; look for users of the same state.
490 // Recurse through MergeMem nodes to the stores that use them.
492 // Each of these stores is a possible definition of memory
493 // that 'load' needs to use. We need to force 'load'
494 // to occur before each such store. When the store is in
495 // the same block as 'load', we insert an anti-dependence
496 // edge load->store.
498 // The relevant stores "nearby" the load consist of a tree rooted
499 // at initial_mem, with internal nodes of type MergeMem.
500 // Therefore, the branches visited by the worklist are of this form:
501 // initial_mem -> (MergeMem ->)* store
502 // The anti-dependence constraints apply only to the fringe of this tree.
504 Node* initial_mem = load->in(MemNode::Memory);
505 worklist_store.push(initial_mem);
506 worklist_visited.push(initial_mem);
507 worklist_mem.push(NULL);
508 while (worklist_store.size() > 0) {
509 // Examine a nearby store to see if it might interfere with our load.
510 Node* mem = worklist_mem.pop();
511 Node* store = worklist_store.pop();
512 uint op = store->Opcode();
514 // MergeMems do not directly have anti-deps.
515 // Treat them as internal nodes in a forward tree of memory states,
516 // the leaves of which are each a 'possible-def'.
517 if (store == initial_mem // root (exclusive) of tree we are searching
518 || op == Op_MergeMem // internal node of tree we are searching
519 ) {
520 mem = store; // It's not a possibly interfering store.
521 if (store == initial_mem)
522 initial_mem = NULL; // only process initial memory once
524 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
525 store = mem->fast_out(i);
526 if (store->is_MergeMem()) {
527 // Be sure we don't get into combinatorial problems.
528 // (Allow phis to be repeated; they can merge two relevant states.)
529 uint j = worklist_visited.size();
530 for (; j > 0; j--) {
531 if (worklist_visited.at(j-1) == store) break;
532 }
533 if (j > 0) continue; // already on work list; do not repeat
534 worklist_visited.push(store);
535 }
536 worklist_mem.push(mem);
537 worklist_store.push(store);
538 }
539 continue;
540 }
542 if (op == Op_MachProj || op == Op_Catch) continue;
543 if (store->needs_anti_dependence_check()) continue; // not really a store
545 // Compute the alias index. Loads and stores with different alias
546 // indices do not need anti-dependence edges. Wide MemBar's are
547 // anti-dependent on everything (except immutable memories).
548 const TypePtr* adr_type = store->adr_type();
549 if (!C->can_alias(adr_type, load_alias_idx)) continue;
551 // Most slow-path runtime calls do NOT modify Java memory, but
552 // they can block and so write Raw memory.
553 if (store->is_Mach()) {
554 MachNode* mstore = store->as_Mach();
555 if (load_alias_idx != Compile::AliasIdxRaw) {
556 // Check for call into the runtime using the Java calling
557 // convention (and from there into a wrapper); it has no
558 // _method. Can't do this optimization for Native calls because
559 // they CAN write to Java memory.
560 if (mstore->ideal_Opcode() == Op_CallStaticJava) {
561 assert(mstore->is_MachSafePoint(), "");
562 MachSafePointNode* ms = (MachSafePointNode*) mstore;
563 assert(ms->is_MachCallJava(), "");
564 MachCallJavaNode* mcj = (MachCallJavaNode*) ms;
565 if (mcj->_method == NULL) {
566 // These runtime calls do not write to Java visible memory
567 // (other than Raw) and so do not require anti-dependence edges.
568 continue;
569 }
570 }
571 // Same for SafePoints: they read/write Raw but only read otherwise.
572 // This is basically a workaround for SafePoints only defining control
573 // instead of control + memory.
574 if (mstore->ideal_Opcode() == Op_SafePoint)
575 continue;
576 } else {
577 // Some raw memory, such as the load of "top" at an allocation,
578 // can be control dependent on the previous safepoint. See
579 // comments in GraphKit::allocate_heap() about control input.
580 // Inserting an anti-dep between such a safepoint and a use
581 // creates a cycle, and will cause a subsequent failure in
582 // local scheduling. (BugId 4919904)
583 // (%%% How can a control input be a safepoint and not a projection??)
584 if (mstore->ideal_Opcode() == Op_SafePoint && load->in(0) == mstore)
585 continue;
586 }
587 }
589 // Identify a block that the current load must be above,
590 // or else observe that 'store' is all the way up in the
591 // earliest legal block for 'load'. In the latter case,
592 // immediately insert an anti-dependence edge.
593 Block* store_block = _bbs[store->_idx];
594 assert(store_block != NULL, "unused killing projections skipped above");
596 if (store->is_Phi()) {
597 // 'load' uses memory which is one (or more) of the Phi's inputs.
598 // It must be scheduled not before the Phi, but rather before
599 // each of the relevant Phi inputs.
600 //
601 // Instead of finding the LCA of all inputs to a Phi that match 'mem',
602 // we mark each corresponding predecessor block and do a combined
603 // hoisting operation later (raise_LCA_above_marks).
604 //
605 // Do not assert(store_block != early, "Phi merging memory after access")
606 // PhiNode may be at start of block 'early' with backedge to 'early'
607 DEBUG_ONLY(bool found_match = false);
608 for (uint j = PhiNode::Input, jmax = store->req(); j < jmax; j++) {
609 if (store->in(j) == mem) { // Found matching input?
610 DEBUG_ONLY(found_match = true);
611 Block* pred_block = _bbs[store_block->pred(j)->_idx];
612 if (pred_block != early) {
613 // If any predecessor of the Phi matches the load's "early block",
614 // we do not need a precedence edge between the Phi and 'load'
615 // since the load will be forced into a block preceding the Phi.
616 pred_block->set_raise_LCA_mark(load_index);
617 assert(!LCA_orig->dominates(pred_block) ||
618 early->dominates(pred_block), "early is high enough");
619 must_raise_LCA = true;
620 } else {
621 // anti-dependent upon PHI pinned below 'early', no edge needed
622 LCA = early; // but can not schedule below 'early'
623 }
624 }
625 }
626 assert(found_match, "no worklist bug");
627 #ifdef TRACK_PHI_INPUTS
628 #ifdef ASSERT
629 // This assert asks about correct handling of PhiNodes, which may not
630 // have all input edges directly from 'mem'. See BugId 4621264
631 int num_mem_inputs = phi_inputs.at_grow(store->_idx,0) + 1;
632 // Increment by exactly one even if there are multiple copies of 'mem'
633 // coming into the phi, because we will run this block several times
634 // if there are several copies of 'mem'. (That's how DU iterators work.)
635 phi_inputs.at_put(store->_idx, num_mem_inputs);
636 assert(PhiNode::Input + num_mem_inputs < store->req(),
637 "Expect at least one phi input will not be from original memory state");
638 #endif //ASSERT
639 #endif //TRACK_PHI_INPUTS
640 } else if (store_block != early) {
641 // 'store' is between the current LCA and earliest possible block.
642 // Label its block, and decide later on how to raise the LCA
643 // to include the effect on LCA of this store.
644 // If this store's block gets chosen as the raised LCA, we
645 // will find him on the non_early_stores list and stick him
646 // with a precedence edge.
647 // (But, don't bother if LCA is already raised all the way.)
648 if (LCA != early) {
649 store_block->set_raise_LCA_mark(load_index);
650 must_raise_LCA = true;
651 non_early_stores.push(store);
652 }
653 } else {
654 // Found a possibly-interfering store in the load's 'early' block.
655 // This means 'load' cannot sink at all in the dominator tree.
656 // Add an anti-dep edge, and squeeze 'load' into the highest block.
657 assert(store != load->in(0), "dependence cycle found");
658 if (verify) {
659 assert(store->find_edge(load) != -1, "missing precedence edge");
660 } else {
661 store->add_prec(load);
662 }
663 LCA = early;
664 // This turns off the process of gathering non_early_stores.
665 }
666 }
667 // (Worklist is now empty; all nearby stores have been visited.)
669 // Finished if 'load' must be scheduled in its 'early' block.
670 // If we found any stores there, they have already been given
671 // precedence edges.
672 if (LCA == early) return LCA;
674 // We get here only if there are no possibly-interfering stores
675 // in the load's 'early' block. Move LCA up above all predecessors
676 // which contain stores we have noted.
677 //
678 // The raised LCA block can be a home to such interfering stores,
679 // but its predecessors must not contain any such stores.
680 //
681 // The raised LCA will be a lower bound for placing the load,
682 // preventing the load from sinking past any block containing
683 // a store that may invalidate the memory state required by 'load'.
684 if (must_raise_LCA)
685 LCA = raise_LCA_above_marks(LCA, load->_idx, early, _bbs);
686 if (LCA == early) return LCA;
688 // Insert anti-dependence edges from 'load' to each store
689 // in the non-early LCA block.
690 // Mine the non_early_stores list for such stores.
691 if (LCA->raise_LCA_mark() == load_index) {
692 while (non_early_stores.size() > 0) {
693 Node* store = non_early_stores.pop();
694 Block* store_block = _bbs[store->_idx];
695 if (store_block == LCA) {
696 // add anti_dependence from store to load in its own block
697 assert(store != load->in(0), "dependence cycle found");
698 if (verify) {
699 assert(store->find_edge(load) != -1, "missing precedence edge");
700 } else {
701 store->add_prec(load);
702 }
703 } else {
704 assert(store_block->raise_LCA_mark() == load_index, "block was marked");
705 // Any other stores we found must be either inside the new LCA
706 // or else outside the original LCA. In the latter case, they
707 // did not interfere with any use of 'load'.
708 assert(LCA->dominates(store_block)
709 || !LCA_orig->dominates(store_block), "no stray stores");
710 }
711 }
712 }
714 // Return the highest block containing stores; any stores
715 // within that block have been given anti-dependence edges.
716 return LCA;
717 }
719 // This class is used to iterate backwards over the nodes in the graph.
721 class Node_Backward_Iterator {
723 private:
724 Node_Backward_Iterator();
726 public:
727 // Constructor for the iterator
728 Node_Backward_Iterator(Node *root, VectorSet &visited, Node_List &stack, Block_Array &bbs);
730 // Postincrement operator to iterate over the nodes
731 Node *next();
733 private:
734 VectorSet &_visited;
735 Node_List &_stack;
736 Block_Array &_bbs;
737 };
739 // Constructor for the Node_Backward_Iterator
740 Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_List &stack, Block_Array &bbs )
741 : _visited(visited), _stack(stack), _bbs(bbs) {
742 // The stack should contain exactly the root
743 stack.clear();
744 stack.push(root);
746 // Clear the visited bits
747 visited.Clear();
748 }
750 // Iterator for the Node_Backward_Iterator
751 Node *Node_Backward_Iterator::next() {
753 // If the _stack is empty, then just return NULL: finished.
754 if ( !_stack.size() )
755 return NULL;
757 // '_stack' is emulating a real _stack. The 'visit-all-users' loop has been
758 // made stateless, so I do not need to record the index 'i' on my _stack.
759 // Instead I visit all users each time, scanning for unvisited users.
760 // I visit unvisited not-anti-dependence users first, then anti-dependent
761 // children next.
762 Node *self = _stack.pop();
764 // I cycle here when I am entering a deeper level of recursion.
765 // The key variable 'self' was set prior to jumping here.
766 while( 1 ) {
768 _visited.set(self->_idx);
770 // Now schedule all uses as late as possible.
771 uint src = self->is_Proj() ? self->in(0)->_idx : self->_idx;
772 uint src_rpo = _bbs[src]->_rpo;
774 // Schedule all nodes in a post-order visit
775 Node *unvisited = NULL; // Unvisited anti-dependent Node, if any
777 // Scan for unvisited nodes
778 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
779 // For all uses, schedule late
780 Node* n = self->fast_out(i); // Use
782 // Skip already visited children
783 if ( _visited.test(n->_idx) )
784 continue;
786 // do not traverse backward control edges
787 Node *use = n->is_Proj() ? n->in(0) : n;
788 uint use_rpo = _bbs[use->_idx]->_rpo;
790 if ( use_rpo < src_rpo )
791 continue;
793 // Phi nodes always precede uses in a basic block
794 if ( use_rpo == src_rpo && use->is_Phi() )
795 continue;
797 unvisited = n; // Found unvisited
799 // Check for possible-anti-dependent
800 if( !n->needs_anti_dependence_check() )
801 break; // Not visited, not anti-dep; schedule it NOW
802 }
804 // Did I find an unvisited not-anti-dependent Node?
805 if ( !unvisited )
806 break; // All done with children; post-visit 'self'
808 // Visit the unvisited Node. Contains the obvious push to
809 // indicate I'm entering a deeper level of recursion. I push the
810 // old state onto the _stack and set a new state and loop (recurse).
811 _stack.push(self);
812 self = unvisited;
813 } // End recursion loop
815 return self;
816 }
818 //------------------------------ComputeLatenciesBackwards----------------------
819 // Compute the latency of all the instructions.
820 void PhaseCFG::ComputeLatenciesBackwards(VectorSet &visited, Node_List &stack) {
821 #ifndef PRODUCT
822 if (trace_opto_pipelining())
823 tty->print("\n#---- ComputeLatenciesBackwards ----\n");
824 #endif
826 Node_Backward_Iterator iter((Node *)_root, visited, stack, _bbs);
827 Node *n;
829 // Walk over all the nodes from last to first
830 while (n = iter.next()) {
831 // Set the latency for the definitions of this instruction
832 partial_latency_of_defs(n);
833 }
834 } // end ComputeLatenciesBackwards
836 //------------------------------partial_latency_of_defs------------------------
837 // Compute the latency impact of this node on all defs. This computes
838 // a number that increases as we approach the beginning of the routine.
839 void PhaseCFG::partial_latency_of_defs(Node *n) {
840 // Set the latency for this instruction
841 #ifndef PRODUCT
842 if (trace_opto_pipelining()) {
843 tty->print("# latency_to_inputs: node_latency[%d] = %d for node",
844 n->_idx, _node_latency.at_grow(n->_idx));
845 dump();
846 }
847 #endif
849 if (n->is_Proj())
850 n = n->in(0);
852 if (n->is_Root())
853 return;
855 uint nlen = n->len();
856 uint use_latency = _node_latency.at_grow(n->_idx);
857 uint use_pre_order = _bbs[n->_idx]->_pre_order;
859 for ( uint j=0; j<nlen; j++ ) {
860 Node *def = n->in(j);
862 if (!def || def == n)
863 continue;
865 // Walk backwards thru projections
866 if (def->is_Proj())
867 def = def->in(0);
869 #ifndef PRODUCT
870 if (trace_opto_pipelining()) {
871 tty->print("# in(%2d): ", j);
872 def->dump();
873 }
874 #endif
876 // If the defining block is not known, assume it is ok
877 Block *def_block = _bbs[def->_idx];
878 uint def_pre_order = def_block ? def_block->_pre_order : 0;
880 if ( (use_pre_order < def_pre_order) ||
881 (use_pre_order == def_pre_order && n->is_Phi()) )
882 continue;
884 uint delta_latency = n->latency(j);
885 uint current_latency = delta_latency + use_latency;
887 if (_node_latency.at_grow(def->_idx) < current_latency) {
888 _node_latency.at_put_grow(def->_idx, current_latency);
889 }
891 #ifndef PRODUCT
892 if (trace_opto_pipelining()) {
893 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, node_latency[%d] = %d",
894 use_latency, j, delta_latency, current_latency, def->_idx,
895 _node_latency.at_grow(def->_idx));
896 }
897 #endif
898 }
899 }
901 //------------------------------latency_from_use-------------------------------
902 // Compute the latency of a specific use
903 int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) {
904 // If self-reference, return no latency
905 if (use == n || use->is_Root())
906 return 0;
908 uint def_pre_order = _bbs[def->_idx]->_pre_order;
909 uint latency = 0;
911 // If the use is not a projection, then it is simple...
912 if (!use->is_Proj()) {
913 #ifndef PRODUCT
914 if (trace_opto_pipelining()) {
915 tty->print("# out(): ");
916 use->dump();
917 }
918 #endif
920 uint use_pre_order = _bbs[use->_idx]->_pre_order;
922 if (use_pre_order < def_pre_order)
923 return 0;
925 if (use_pre_order == def_pre_order && use->is_Phi())
926 return 0;
928 uint nlen = use->len();
929 uint nl = _node_latency.at_grow(use->_idx);
931 for ( uint j=0; j<nlen; j++ ) {
932 if (use->in(j) == n) {
933 // Change this if we want local latencies
934 uint ul = use->latency(j);
935 uint l = ul + nl;
936 if (latency < l) latency = l;
937 #ifndef PRODUCT
938 if (trace_opto_pipelining()) {
939 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, latency = %d",
940 nl, j, ul, l, latency);
941 }
942 #endif
943 }
944 }
945 } else {
946 // This is a projection, just grab the latency of the use(s)
947 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
948 uint l = latency_from_use(use, def, use->fast_out(j));
949 if (latency < l) latency = l;
950 }
951 }
953 return latency;
954 }
956 //------------------------------latency_from_uses------------------------------
957 // Compute the latency of this instruction relative to all of it's uses.
958 // This computes a number that increases as we approach the beginning of the
959 // routine.
960 void PhaseCFG::latency_from_uses(Node *n) {
961 // Set the latency for this instruction
962 #ifndef PRODUCT
963 if (trace_opto_pipelining()) {
964 tty->print("# latency_from_outputs: node_latency[%d] = %d for node",
965 n->_idx, _node_latency.at_grow(n->_idx));
966 dump();
967 }
968 #endif
969 uint latency=0;
970 const Node *def = n->is_Proj() ? n->in(0): n;
972 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
973 uint l = latency_from_use(n, def, n->fast_out(i));
975 if (latency < l) latency = l;
976 }
978 _node_latency.at_put_grow(n->_idx, latency);
979 }
981 //------------------------------hoist_to_cheaper_block-------------------------
982 // Pick a block for node self, between early and LCA, that is a cheaper
983 // alternative to LCA.
984 Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) {
985 const double delta = 1+PROB_UNLIKELY_MAG(4);
986 Block* least = LCA;
987 double least_freq = least->_freq;
988 uint target = _node_latency.at_grow(self->_idx);
989 uint start_latency = _node_latency.at_grow(LCA->_nodes[0]->_idx);
990 uint end_latency = _node_latency.at_grow(LCA->_nodes[LCA->end_idx()]->_idx);
991 bool in_latency = (target <= start_latency);
992 const Block* root_block = _bbs[_root->_idx];
994 // Turn off latency scheduling if scheduling is just plain off
995 if (!C->do_scheduling())
996 in_latency = true;
998 // Do not hoist (to cover latency) instructions which target a
999 // single register. Hoisting stretches the live range of the
1000 // single register and may force spilling.
1001 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1002 if (mach && mach->out_RegMask().is_bound1() && mach->out_RegMask().is_NotEmpty())
1003 in_latency = true;
1005 #ifndef PRODUCT
1006 if (trace_opto_pipelining()) {
1007 tty->print("# Find cheaper block for latency %d: ",
1008 _node_latency.at_grow(self->_idx));
1009 self->dump();
1010 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1011 LCA->_pre_order,
1012 LCA->_nodes[0]->_idx,
1013 start_latency,
1014 LCA->_nodes[LCA->end_idx()]->_idx,
1015 end_latency,
1016 least_freq);
1017 }
1018 #endif
1020 // Walk up the dominator tree from LCA (Lowest common ancestor) to
1021 // the earliest legal location. Capture the least execution frequency.
1022 while (LCA != early) {
1023 LCA = LCA->_idom; // Follow up the dominator tree
1025 if (LCA == NULL) {
1026 // Bailout without retry
1027 C->record_method_not_compilable("late schedule failed: LCA == NULL");
1028 return least;
1029 }
1031 // Don't hoist machine instructions to the root basic block
1032 if (mach && LCA == root_block)
1033 break;
1035 uint start_lat = _node_latency.at_grow(LCA->_nodes[0]->_idx);
1036 uint end_idx = LCA->end_idx();
1037 uint end_lat = _node_latency.at_grow(LCA->_nodes[end_idx]->_idx);
1038 double LCA_freq = LCA->_freq;
1039 #ifndef PRODUCT
1040 if (trace_opto_pipelining()) {
1041 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1042 LCA->_pre_order, LCA->_nodes[0]->_idx, start_lat, end_idx, end_lat, LCA_freq);
1043 }
1044 #endif
1045 if (LCA_freq < least_freq || // Better Frequency
1046 ( !in_latency && // No block containing latency
1047 LCA_freq < least_freq * delta && // No worse frequency
1048 target >= end_lat && // within latency range
1049 !self->is_iteratively_computed() ) // But don't hoist IV increments
1050 // because they may end up above other uses of their phi forcing
1051 // their result register to be different from their input.
1052 ) {
1053 least = LCA; // Found cheaper block
1054 least_freq = LCA_freq;
1055 start_latency = start_lat;
1056 end_latency = end_lat;
1057 if (target <= start_lat)
1058 in_latency = true;
1059 }
1060 }
1062 #ifndef PRODUCT
1063 if (trace_opto_pipelining()) {
1064 tty->print_cr("# Choose block B%d with start latency=%d and freq=%g",
1065 least->_pre_order, start_latency, least_freq);
1066 }
1067 #endif
1069 // See if the latency needs to be updated
1070 if (target < end_latency) {
1071 #ifndef PRODUCT
1072 if (trace_opto_pipelining()) {
1073 tty->print_cr("# Change latency for [%4d] from %d to %d", self->_idx, target, end_latency);
1074 }
1075 #endif
1076 _node_latency.at_put_grow(self->_idx, end_latency);
1077 partial_latency_of_defs(self);
1078 }
1080 return least;
1081 }
1084 //------------------------------schedule_late-----------------------------------
1085 // Now schedule all codes as LATE as possible. This is the LCA in the
1086 // dominator tree of all USES of a value. Pick the block with the least
1087 // loop nesting depth that is lowest in the dominator tree.
1088 extern const char must_clone[];
1089 void PhaseCFG::schedule_late(VectorSet &visited, Node_List &stack) {
1090 #ifndef PRODUCT
1091 if (trace_opto_pipelining())
1092 tty->print("\n#---- schedule_late ----\n");
1093 #endif
1095 Node_Backward_Iterator iter((Node *)_root, visited, stack, _bbs);
1096 Node *self;
1098 // Walk over all the nodes from last to first
1099 while (self = iter.next()) {
1100 Block* early = _bbs[self->_idx]; // Earliest legal placement
1102 if (self->is_top()) {
1103 // Top node goes in bb #2 with other constants.
1104 // It must be special-cased, because it has no out edges.
1105 early->add_inst(self);
1106 continue;
1107 }
1109 // No uses, just terminate
1110 if (self->outcnt() == 0) {
1111 assert(self->Opcode() == Op_MachProj, "sanity");
1112 continue; // Must be a dead machine projection
1113 }
1115 // If node is pinned in the block, then no scheduling can be done.
1116 if( self->pinned() ) // Pinned in block?
1117 continue;
1119 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1120 if (mach) {
1121 switch (mach->ideal_Opcode()) {
1122 case Op_CreateEx:
1123 // Don't move exception creation
1124 early->add_inst(self);
1125 continue;
1126 break;
1127 case Op_CheckCastPP:
1128 // Don't move CheckCastPP nodes away from their input, if the input
1129 // is a rawptr (5071820).
1130 Node *def = self->in(1);
1131 if (def != NULL && def->bottom_type()->base() == Type::RawPtr) {
1132 early->add_inst(self);
1133 continue;
1134 }
1135 break;
1136 }
1137 }
1139 // Gather LCA of all uses
1140 Block *LCA = NULL;
1141 {
1142 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
1143 // For all uses, find LCA
1144 Node* use = self->fast_out(i);
1145 LCA = raise_LCA_above_use(LCA, use, self, _bbs);
1146 }
1147 } // (Hide defs of imax, i from rest of block.)
1149 // Place temps in the block of their use. This isn't a
1150 // requirement for correctness but it reduces useless
1151 // interference between temps and other nodes.
1152 if (mach != NULL && mach->is_MachTemp()) {
1153 _bbs.map(self->_idx, LCA);
1154 LCA->add_inst(self);
1155 continue;
1156 }
1158 // Check if 'self' could be anti-dependent on memory
1159 if (self->needs_anti_dependence_check()) {
1160 // Hoist LCA above possible-defs and insert anti-dependences to
1161 // defs in new LCA block.
1162 LCA = insert_anti_dependences(LCA, self);
1163 }
1165 if (early->_dom_depth > LCA->_dom_depth) {
1166 // Somehow the LCA has moved above the earliest legal point.
1167 // (One way this can happen is via memory_early_block.)
1168 if (C->subsume_loads() == true && !C->failing()) {
1169 // Retry with subsume_loads == false
1170 // If this is the first failure, the sentinel string will "stick"
1171 // to the Compile object, and the C2Compiler will see it and retry.
1172 C->record_failure(C2Compiler::retry_no_subsuming_loads());
1173 } else {
1174 // Bailout without retry when (early->_dom_depth > LCA->_dom_depth)
1175 C->record_method_not_compilable("late schedule failed: incorrect graph");
1176 }
1177 return;
1178 }
1180 // If there is no opportunity to hoist, then we're done.
1181 bool try_to_hoist = (LCA != early);
1183 // Must clone guys stay next to use; no hoisting allowed.
1184 // Also cannot hoist guys that alter memory or are otherwise not
1185 // allocatable (hoisting can make a value live longer, leading to
1186 // anti and output dependency problems which are normally resolved
1187 // by the register allocator giving everyone a different register).
1188 if (mach != NULL && must_clone[mach->ideal_Opcode()])
1189 try_to_hoist = false;
1191 Block* late = NULL;
1192 if (try_to_hoist) {
1193 // Now find the block with the least execution frequency.
1194 // Start at the latest schedule and work up to the earliest schedule
1195 // in the dominator tree. Thus the Node will dominate all its uses.
1196 late = hoist_to_cheaper_block(LCA, early, self);
1197 } else {
1198 // Just use the LCA of the uses.
1199 late = LCA;
1200 }
1202 // Put the node into target block
1203 schedule_node_into_block(self, late);
1205 #ifdef ASSERT
1206 if (self->needs_anti_dependence_check()) {
1207 // since precedence edges are only inserted when we're sure they
1208 // are needed make sure that after placement in a block we don't
1209 // need any new precedence edges.
1210 verify_anti_dependences(late, self);
1211 }
1212 #endif
1213 } // Loop until all nodes have been visited
1215 } // end ScheduleLate
1217 //------------------------------GlobalCodeMotion-------------------------------
1218 void PhaseCFG::GlobalCodeMotion( Matcher &matcher, uint unique, Node_List &proj_list ) {
1219 ResourceMark rm;
1221 #ifndef PRODUCT
1222 if (trace_opto_pipelining()) {
1223 tty->print("\n---- Start GlobalCodeMotion ----\n");
1224 }
1225 #endif
1227 // Initialize the bbs.map for things on the proj_list
1228 uint i;
1229 for( i=0; i < proj_list.size(); i++ )
1230 _bbs.map(proj_list[i]->_idx, NULL);
1232 // Set the basic block for Nodes pinned into blocks
1233 Arena *a = Thread::current()->resource_area();
1234 VectorSet visited(a);
1235 schedule_pinned_nodes( visited );
1237 // Find the earliest Block any instruction can be placed in. Some
1238 // instructions are pinned into Blocks. Unpinned instructions can
1239 // appear in last block in which all their inputs occur.
1240 visited.Clear();
1241 Node_List stack(a);
1242 stack.map( (unique >> 1) + 16, NULL); // Pre-grow the list
1243 if (!schedule_early(visited, stack)) {
1244 // Bailout without retry
1245 C->record_method_not_compilable("early schedule failed");
1246 return;
1247 }
1249 // Build Def-Use edges.
1250 proj_list.push(_root); // Add real root as another root
1251 proj_list.pop();
1253 // Compute the latency information (via backwards walk) for all the
1254 // instructions in the graph
1255 GrowableArray<uint> node_latency;
1256 _node_latency = node_latency;
1258 if( C->do_scheduling() )
1259 ComputeLatenciesBackwards(visited, stack);
1261 // Now schedule all codes as LATE as possible. This is the LCA in the
1262 // dominator tree of all USES of a value. Pick the block with the least
1263 // loop nesting depth that is lowest in the dominator tree.
1264 // ( visited.Clear() called in schedule_late()->Node_Backward_Iterator() )
1265 schedule_late(visited, stack);
1266 if( C->failing() ) {
1267 // schedule_late fails only when graph is incorrect.
1268 assert(!VerifyGraphEdges, "verification should have failed");
1269 return;
1270 }
1272 unique = C->unique();
1274 #ifndef PRODUCT
1275 if (trace_opto_pipelining()) {
1276 tty->print("\n---- Detect implicit null checks ----\n");
1277 }
1278 #endif
1280 // Detect implicit-null-check opportunities. Basically, find NULL checks
1281 // with suitable memory ops nearby. Use the memory op to do the NULL check.
1282 // I can generate a memory op if there is not one nearby.
1283 if (C->is_method_compilation()) {
1284 // Don't do it for natives, adapters, or runtime stubs
1285 int allowed_reasons = 0;
1286 // ...and don't do it when there have been too many traps, globally.
1287 for (int reason = (int)Deoptimization::Reason_none+1;
1288 reason < Compile::trapHistLength; reason++) {
1289 assert(reason < BitsPerInt, "recode bit map");
1290 if (!C->too_many_traps((Deoptimization::DeoptReason) reason))
1291 allowed_reasons |= nth_bit(reason);
1292 }
1293 // By reversing the loop direction we get a very minor gain on mpegaudio.
1294 // Feel free to revert to a forward loop for clarity.
1295 // for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) {
1296 for( int i= matcher._null_check_tests.size()-2; i>=0; i-=2 ) {
1297 Node *proj = matcher._null_check_tests[i ];
1298 Node *val = matcher._null_check_tests[i+1];
1299 _bbs[proj->_idx]->implicit_null_check(this, proj, val, allowed_reasons);
1300 // The implicit_null_check will only perform the transformation
1301 // if the null branch is truly uncommon, *and* it leads to an
1302 // uncommon trap. Combined with the too_many_traps guards
1303 // above, this prevents SEGV storms reported in 6366351,
1304 // by recompiling offending methods without this optimization.
1305 }
1306 }
1308 #ifndef PRODUCT
1309 if (trace_opto_pipelining()) {
1310 tty->print("\n---- Start Local Scheduling ----\n");
1311 }
1312 #endif
1314 // Schedule locally. Right now a simple topological sort.
1315 // Later, do a real latency aware scheduler.
1316 int *ready_cnt = NEW_RESOURCE_ARRAY(int,C->unique());
1317 memset( ready_cnt, -1, C->unique() * sizeof(int) );
1318 visited.Clear();
1319 for (i = 0; i < _num_blocks; i++) {
1320 if (!_blocks[i]->schedule_local(this, matcher, ready_cnt, visited)) {
1321 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
1322 C->record_method_not_compilable("local schedule failed");
1323 }
1324 return;
1325 }
1326 }
1328 // If we inserted any instructions between a Call and his CatchNode,
1329 // clone the instructions on all paths below the Catch.
1330 for( i=0; i < _num_blocks; i++ )
1331 _blocks[i]->call_catch_cleanup(_bbs);
1333 #ifndef PRODUCT
1334 if (trace_opto_pipelining()) {
1335 tty->print("\n---- After GlobalCodeMotion ----\n");
1336 for (uint i = 0; i < _num_blocks; i++) {
1337 _blocks[i]->dump();
1338 }
1339 }
1340 #endif
1341 }
1344 //------------------------------Estimate_Block_Frequency-----------------------
1345 // Estimate block frequencies based on IfNode probabilities.
1346 void PhaseCFG::Estimate_Block_Frequency() {
1348 // Force conditional branches leading to uncommon traps to be unlikely,
1349 // not because we get to the uncommon_trap with less relative frequency,
1350 // but because an uncommon_trap typically causes a deopt, so we only get
1351 // there once.
1352 if (C->do_freq_based_layout()) {
1353 Block_List worklist;
1354 Block* root_blk = _blocks[0];
1355 for (uint i = 1; i < root_blk->num_preds(); i++) {
1356 Block *pb = _bbs[root_blk->pred(i)->_idx];
1357 if (pb->has_uncommon_code()) {
1358 worklist.push(pb);
1359 }
1360 }
1361 while (worklist.size() > 0) {
1362 Block* uct = worklist.pop();
1363 if (uct == _broot) continue;
1364 for (uint i = 1; i < uct->num_preds(); i++) {
1365 Block *pb = _bbs[uct->pred(i)->_idx];
1366 if (pb->_num_succs == 1) {
1367 worklist.push(pb);
1368 } else if (pb->num_fall_throughs() == 2) {
1369 pb->update_uncommon_branch(uct);
1370 }
1371 }
1372 }
1373 }
1375 // Create the loop tree and calculate loop depth.
1376 _root_loop = create_loop_tree();
1377 _root_loop->compute_loop_depth(0);
1379 // Compute block frequency of each block, relative to a single loop entry.
1380 _root_loop->compute_freq();
1382 // Adjust all frequencies to be relative to a single method entry
1383 _root_loop->_freq = 1.0;
1384 _root_loop->scale_freq();
1386 // Save outmost loop frequency for LRG frequency threshold
1387 _outer_loop_freq = _root_loop->outer_loop_freq();
1389 // force paths ending at uncommon traps to be infrequent
1390 if (!C->do_freq_based_layout()) {
1391 Block_List worklist;
1392 Block* root_blk = _blocks[0];
1393 for (uint i = 1; i < root_blk->num_preds(); i++) {
1394 Block *pb = _bbs[root_blk->pred(i)->_idx];
1395 if (pb->has_uncommon_code()) {
1396 worklist.push(pb);
1397 }
1398 }
1399 while (worklist.size() > 0) {
1400 Block* uct = worklist.pop();
1401 uct->_freq = PROB_MIN;
1402 for (uint i = 1; i < uct->num_preds(); i++) {
1403 Block *pb = _bbs[uct->pred(i)->_idx];
1404 if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) {
1405 worklist.push(pb);
1406 }
1407 }
1408 }
1409 }
1411 #ifdef ASSERT
1412 for (uint i = 0; i < _num_blocks; i++ ) {
1413 Block *b = _blocks[i];
1414 assert(b->_freq >= MIN_BLOCK_FREQUENCY, "Register Allocator requires meaningful block frequency");
1415 }
1416 #endif
1418 #ifndef PRODUCT
1419 if (PrintCFGBlockFreq) {
1420 tty->print_cr("CFG Block Frequencies");
1421 _root_loop->dump_tree();
1422 if (Verbose) {
1423 tty->print_cr("PhaseCFG dump");
1424 dump();
1425 tty->print_cr("Node dump");
1426 _root->dump(99999);
1427 }
1428 }
1429 #endif
1430 }
1432 //----------------------------create_loop_tree--------------------------------
1433 // Create a loop tree from the CFG
1434 CFGLoop* PhaseCFG::create_loop_tree() {
1436 #ifdef ASSERT
1437 assert( _blocks[0] == _broot, "" );
1438 for (uint i = 0; i < _num_blocks; i++ ) {
1439 Block *b = _blocks[i];
1440 // Check that _loop field are clear...we could clear them if not.
1441 assert(b->_loop == NULL, "clear _loop expected");
1442 // Sanity check that the RPO numbering is reflected in the _blocks array.
1443 // It doesn't have to be for the loop tree to be built, but if it is not,
1444 // then the blocks have been reordered since dom graph building...which
1445 // may question the RPO numbering
1446 assert(b->_rpo == i, "unexpected reverse post order number");
1447 }
1448 #endif
1450 int idct = 0;
1451 CFGLoop* root_loop = new CFGLoop(idct++);
1453 Block_List worklist;
1455 // Assign blocks to loops
1456 for(uint i = _num_blocks - 1; i > 0; i-- ) { // skip Root block
1457 Block *b = _blocks[i];
1459 if (b->head()->is_Loop()) {
1460 Block* loop_head = b;
1461 assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1462 Node* tail_n = loop_head->pred(LoopNode::LoopBackControl);
1463 Block* tail = _bbs[tail_n->_idx];
1465 // Defensively filter out Loop nodes for non-single-entry loops.
1466 // For all reasonable loops, the head occurs before the tail in RPO.
1467 if (i <= tail->_rpo) {
1469 // The tail and (recursive) predecessors of the tail
1470 // are made members of a new loop.
1472 assert(worklist.size() == 0, "nonempty worklist");
1473 CFGLoop* nloop = new CFGLoop(idct++);
1474 assert(loop_head->_loop == NULL, "just checking");
1475 loop_head->_loop = nloop;
1476 // Add to nloop so push_pred() will skip over inner loops
1477 nloop->add_member(loop_head);
1478 nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, _bbs);
1480 while (worklist.size() > 0) {
1481 Block* member = worklist.pop();
1482 if (member != loop_head) {
1483 for (uint j = 1; j < member->num_preds(); j++) {
1484 nloop->push_pred(member, j, worklist, _bbs);
1485 }
1486 }
1487 }
1488 }
1489 }
1490 }
1492 // Create a member list for each loop consisting
1493 // of both blocks and (immediate child) loops.
1494 for (uint i = 0; i < _num_blocks; i++) {
1495 Block *b = _blocks[i];
1496 CFGLoop* lp = b->_loop;
1497 if (lp == NULL) {
1498 // Not assigned to a loop. Add it to the method's pseudo loop.
1499 b->_loop = root_loop;
1500 lp = root_loop;
1501 }
1502 if (lp == root_loop || b != lp->head()) { // loop heads are already members
1503 lp->add_member(b);
1504 }
1505 if (lp != root_loop) {
1506 if (lp->parent() == NULL) {
1507 // Not a nested loop. Make it a child of the method's pseudo loop.
1508 root_loop->add_nested_loop(lp);
1509 }
1510 if (b == lp->head()) {
1511 // Add nested loop to member list of parent loop.
1512 lp->parent()->add_member(lp);
1513 }
1514 }
1515 }
1517 return root_loop;
1518 }
1520 //------------------------------push_pred--------------------------------------
1521 void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, Block_Array& node_to_blk) {
1522 Node* pred_n = blk->pred(i);
1523 Block* pred = node_to_blk[pred_n->_idx];
1524 CFGLoop *pred_loop = pred->_loop;
1525 if (pred_loop == NULL) {
1526 // Filter out blocks for non-single-entry loops.
1527 // For all reasonable loops, the head occurs before the tail in RPO.
1528 if (pred->_rpo > head()->_rpo) {
1529 pred->_loop = this;
1530 worklist.push(pred);
1531 }
1532 } else if (pred_loop != this) {
1533 // Nested loop.
1534 while (pred_loop->_parent != NULL && pred_loop->_parent != this) {
1535 pred_loop = pred_loop->_parent;
1536 }
1537 // Make pred's loop be a child
1538 if (pred_loop->_parent == NULL) {
1539 add_nested_loop(pred_loop);
1540 // Continue with loop entry predecessor.
1541 Block* pred_head = pred_loop->head();
1542 assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1543 assert(pred_head != head(), "loop head in only one loop");
1544 push_pred(pred_head, LoopNode::EntryControl, worklist, node_to_blk);
1545 } else {
1546 assert(pred_loop->_parent == this && _parent == NULL, "just checking");
1547 }
1548 }
1549 }
1551 //------------------------------add_nested_loop--------------------------------
1552 // Make cl a child of the current loop in the loop tree.
1553 void CFGLoop::add_nested_loop(CFGLoop* cl) {
1554 assert(_parent == NULL, "no parent yet");
1555 assert(cl != this, "not my own parent");
1556 cl->_parent = this;
1557 CFGLoop* ch = _child;
1558 if (ch == NULL) {
1559 _child = cl;
1560 } else {
1561 while (ch->_sibling != NULL) { ch = ch->_sibling; }
1562 ch->_sibling = cl;
1563 }
1564 }
1566 //------------------------------compute_loop_depth-----------------------------
1567 // Store the loop depth in each CFGLoop object.
1568 // Recursively walk the children to do the same for them.
1569 void CFGLoop::compute_loop_depth(int depth) {
1570 _depth = depth;
1571 CFGLoop* ch = _child;
1572 while (ch != NULL) {
1573 ch->compute_loop_depth(depth + 1);
1574 ch = ch->_sibling;
1575 }
1576 }
1578 //------------------------------compute_freq-----------------------------------
1579 // Compute the frequency of each block and loop, relative to a single entry
1580 // into the dominating loop head.
1581 void CFGLoop::compute_freq() {
1582 // Bottom up traversal of loop tree (visit inner loops first.)
1583 // Set loop head frequency to 1.0, then transitively
1584 // compute frequency for all successors in the loop,
1585 // as well as for each exit edge. Inner loops are
1586 // treated as single blocks with loop exit targets
1587 // as the successor blocks.
1589 // Nested loops first
1590 CFGLoop* ch = _child;
1591 while (ch != NULL) {
1592 ch->compute_freq();
1593 ch = ch->_sibling;
1594 }
1595 assert (_members.length() > 0, "no empty loops");
1596 Block* hd = head();
1597 hd->_freq = 1.0f;
1598 for (int i = 0; i < _members.length(); i++) {
1599 CFGElement* s = _members.at(i);
1600 float freq = s->_freq;
1601 if (s->is_block()) {
1602 Block* b = s->as_Block();
1603 for (uint j = 0; j < b->_num_succs; j++) {
1604 Block* sb = b->_succs[j];
1605 update_succ_freq(sb, freq * b->succ_prob(j));
1606 }
1607 } else {
1608 CFGLoop* lp = s->as_CFGLoop();
1609 assert(lp->_parent == this, "immediate child");
1610 for (int k = 0; k < lp->_exits.length(); k++) {
1611 Block* eb = lp->_exits.at(k).get_target();
1612 float prob = lp->_exits.at(k).get_prob();
1613 update_succ_freq(eb, freq * prob);
1614 }
1615 }
1616 }
1618 // For all loops other than the outer, "method" loop,
1619 // sum and normalize the exit probability. The "method" loop
1620 // should keep the initial exit probability of 1, so that
1621 // inner blocks do not get erroneously scaled.
1622 if (_depth != 0) {
1623 // Total the exit probabilities for this loop.
1624 float exits_sum = 0.0f;
1625 for (int i = 0; i < _exits.length(); i++) {
1626 exits_sum += _exits.at(i).get_prob();
1627 }
1629 // Normalize the exit probabilities. Until now, the
1630 // probabilities estimate the possibility of exit per
1631 // a single loop iteration; afterward, they estimate
1632 // the probability of exit per loop entry.
1633 for (int i = 0; i < _exits.length(); i++) {
1634 Block* et = _exits.at(i).get_target();
1635 float new_prob = 0.0f;
1636 if (_exits.at(i).get_prob() > 0.0f) {
1637 new_prob = _exits.at(i).get_prob() / exits_sum;
1638 }
1639 BlockProbPair bpp(et, new_prob);
1640 _exits.at_put(i, bpp);
1641 }
1643 // Save the total, but guard against unreasonable probability,
1644 // as the value is used to estimate the loop trip count.
1645 // An infinite trip count would blur relative block
1646 // frequencies.
1647 if (exits_sum > 1.0f) exits_sum = 1.0;
1648 if (exits_sum < PROB_MIN) exits_sum = PROB_MIN;
1649 _exit_prob = exits_sum;
1650 }
1651 }
1653 //------------------------------succ_prob-------------------------------------
1654 // Determine the probability of reaching successor 'i' from the receiver block.
1655 float Block::succ_prob(uint i) {
1656 int eidx = end_idx();
1657 Node *n = _nodes[eidx]; // Get ending Node
1659 int op = n->Opcode();
1660 if (n->is_Mach()) {
1661 if (n->is_MachNullCheck()) {
1662 // Can only reach here if called after lcm. The original Op_If is gone,
1663 // so we attempt to infer the probability from one or both of the
1664 // successor blocks.
1665 assert(_num_succs == 2, "expecting 2 successors of a null check");
1666 // If either successor has only one predecessor, then the
1667 // probability estimate can be derived using the
1668 // relative frequency of the successor and this block.
1669 if (_succs[i]->num_preds() == 2) {
1670 return _succs[i]->_freq / _freq;
1671 } else if (_succs[1-i]->num_preds() == 2) {
1672 return 1 - (_succs[1-i]->_freq / _freq);
1673 } else {
1674 // Estimate using both successor frequencies
1675 float freq = _succs[i]->_freq;
1676 return freq / (freq + _succs[1-i]->_freq);
1677 }
1678 }
1679 op = n->as_Mach()->ideal_Opcode();
1680 }
1683 // Switch on branch type
1684 switch( op ) {
1685 case Op_CountedLoopEnd:
1686 case Op_If: {
1687 assert (i < 2, "just checking");
1688 // Conditionals pass on only part of their frequency
1689 float prob = n->as_MachIf()->_prob;
1690 assert(prob >= 0.0 && prob <= 1.0, "out of range probability");
1691 // If succ[i] is the FALSE branch, invert path info
1692 if( _nodes[i + eidx + 1]->Opcode() == Op_IfFalse ) {
1693 return 1.0f - prob; // not taken
1694 } else {
1695 return prob; // taken
1696 }
1697 }
1699 case Op_Jump:
1700 // Divide the frequency between all successors evenly
1701 return 1.0f/_num_succs;
1703 case Op_Catch: {
1704 const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
1705 if (ci->_con == CatchProjNode::fall_through_index) {
1706 // Fall-thru path gets the lion's share.
1707 return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs;
1708 } else {
1709 // Presume exceptional paths are equally unlikely
1710 return PROB_UNLIKELY_MAG(5);
1711 }
1712 }
1714 case Op_Root:
1715 case Op_Goto:
1716 // Pass frequency straight thru to target
1717 return 1.0f;
1719 case Op_NeverBranch:
1720 return 0.0f;
1722 case Op_TailCall:
1723 case Op_TailJump:
1724 case Op_Return:
1725 case Op_Halt:
1726 case Op_Rethrow:
1727 // Do not push out freq to root block
1728 return 0.0f;
1730 default:
1731 ShouldNotReachHere();
1732 }
1734 return 0.0f;
1735 }
1737 //------------------------------num_fall_throughs-----------------------------
1738 // Return the number of fall-through candidates for a block
1739 int Block::num_fall_throughs() {
1740 int eidx = end_idx();
1741 Node *n = _nodes[eidx]; // Get ending Node
1743 int op = n->Opcode();
1744 if (n->is_Mach()) {
1745 if (n->is_MachNullCheck()) {
1746 // In theory, either side can fall-thru, for simplicity sake,
1747 // let's say only the false branch can now.
1748 return 1;
1749 }
1750 op = n->as_Mach()->ideal_Opcode();
1751 }
1753 // Switch on branch type
1754 switch( op ) {
1755 case Op_CountedLoopEnd:
1756 case Op_If:
1757 return 2;
1759 case Op_Root:
1760 case Op_Goto:
1761 return 1;
1763 case Op_Catch: {
1764 for (uint i = 0; i < _num_succs; i++) {
1765 const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
1766 if (ci->_con == CatchProjNode::fall_through_index) {
1767 return 1;
1768 }
1769 }
1770 return 0;
1771 }
1773 case Op_Jump:
1774 case Op_NeverBranch:
1775 case Op_TailCall:
1776 case Op_TailJump:
1777 case Op_Return:
1778 case Op_Halt:
1779 case Op_Rethrow:
1780 return 0;
1782 default:
1783 ShouldNotReachHere();
1784 }
1786 return 0;
1787 }
1789 //------------------------------succ_fall_through-----------------------------
1790 // Return true if a specific successor could be fall-through target.
1791 bool Block::succ_fall_through(uint i) {
1792 int eidx = end_idx();
1793 Node *n = _nodes[eidx]; // Get ending Node
1795 int op = n->Opcode();
1796 if (n->is_Mach()) {
1797 if (n->is_MachNullCheck()) {
1798 // In theory, either side can fall-thru, for simplicity sake,
1799 // let's say only the false branch can now.
1800 return _nodes[i + eidx + 1]->Opcode() == Op_IfFalse;
1801 }
1802 op = n->as_Mach()->ideal_Opcode();
1803 }
1805 // Switch on branch type
1806 switch( op ) {
1807 case Op_CountedLoopEnd:
1808 case Op_If:
1809 case Op_Root:
1810 case Op_Goto:
1811 return true;
1813 case Op_Catch: {
1814 const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
1815 return ci->_con == CatchProjNode::fall_through_index;
1816 }
1818 case Op_Jump:
1819 case Op_NeverBranch:
1820 case Op_TailCall:
1821 case Op_TailJump:
1822 case Op_Return:
1823 case Op_Halt:
1824 case Op_Rethrow:
1825 return false;
1827 default:
1828 ShouldNotReachHere();
1829 }
1831 return false;
1832 }
1834 //------------------------------update_uncommon_branch------------------------
1835 // Update the probability of a two-branch to be uncommon
1836 void Block::update_uncommon_branch(Block* ub) {
1837 int eidx = end_idx();
1838 Node *n = _nodes[eidx]; // Get ending Node
1840 int op = n->as_Mach()->ideal_Opcode();
1842 assert(op == Op_CountedLoopEnd || op == Op_If, "must be a If");
1843 assert(num_fall_throughs() == 2, "must be a two way branch block");
1845 // Which successor is ub?
1846 uint s;
1847 for (s = 0; s <_num_succs; s++) {
1848 if (_succs[s] == ub) break;
1849 }
1850 assert(s < 2, "uncommon successor must be found");
1852 // If ub is the true path, make the proability small, else
1853 // ub is the false path, and make the probability large
1854 bool invert = (_nodes[s + eidx + 1]->Opcode() == Op_IfFalse);
1856 // Get existing probability
1857 float p = n->as_MachIf()->_prob;
1859 if (invert) p = 1.0 - p;
1860 if (p > PROB_MIN) {
1861 p = PROB_MIN;
1862 }
1863 if (invert) p = 1.0 - p;
1865 n->as_MachIf()->_prob = p;
1866 }
1868 //------------------------------update_succ_freq-------------------------------
1869 // Update the appropriate frequency associated with block 'b', a successor of
1870 // a block in this loop.
1871 void CFGLoop::update_succ_freq(Block* b, float freq) {
1872 if (b->_loop == this) {
1873 if (b == head()) {
1874 // back branch within the loop
1875 // Do nothing now, the loop carried frequency will be
1876 // adjust later in scale_freq().
1877 } else {
1878 // simple branch within the loop
1879 b->_freq += freq;
1880 }
1881 } else if (!in_loop_nest(b)) {
1882 // branch is exit from this loop
1883 BlockProbPair bpp(b, freq);
1884 _exits.append(bpp);
1885 } else {
1886 // branch into nested loop
1887 CFGLoop* ch = b->_loop;
1888 ch->_freq += freq;
1889 }
1890 }
1892 //------------------------------in_loop_nest-----------------------------------
1893 // Determine if block b is in the receiver's loop nest.
1894 bool CFGLoop::in_loop_nest(Block* b) {
1895 int depth = _depth;
1896 CFGLoop* b_loop = b->_loop;
1897 int b_depth = b_loop->_depth;
1898 if (depth == b_depth) {
1899 return true;
1900 }
1901 while (b_depth > depth) {
1902 b_loop = b_loop->_parent;
1903 b_depth = b_loop->_depth;
1904 }
1905 return b_loop == this;
1906 }
1908 //------------------------------scale_freq-------------------------------------
1909 // Scale frequency of loops and blocks by trip counts from outer loops
1910 // Do a top down traversal of loop tree (visit outer loops first.)
1911 void CFGLoop::scale_freq() {
1912 float loop_freq = _freq * trip_count();
1913 _freq = loop_freq;
1914 for (int i = 0; i < _members.length(); i++) {
1915 CFGElement* s = _members.at(i);
1916 float block_freq = s->_freq * loop_freq;
1917 if (g_isnan(block_freq) || block_freq < MIN_BLOCK_FREQUENCY)
1918 block_freq = MIN_BLOCK_FREQUENCY;
1919 s->_freq = block_freq;
1920 }
1921 CFGLoop* ch = _child;
1922 while (ch != NULL) {
1923 ch->scale_freq();
1924 ch = ch->_sibling;
1925 }
1926 }
1928 // Frequency of outer loop
1929 float CFGLoop::outer_loop_freq() const {
1930 if (_child != NULL) {
1931 return _child->_freq;
1932 }
1933 return _freq;
1934 }
1936 #ifndef PRODUCT
1937 //------------------------------dump_tree--------------------------------------
1938 void CFGLoop::dump_tree() const {
1939 dump();
1940 if (_child != NULL) _child->dump_tree();
1941 if (_sibling != NULL) _sibling->dump_tree();
1942 }
1944 //------------------------------dump-------------------------------------------
1945 void CFGLoop::dump() const {
1946 for (int i = 0; i < _depth; i++) tty->print(" ");
1947 tty->print("%s: %d trip_count: %6.0f freq: %6.0f\n",
1948 _depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq);
1949 for (int i = 0; i < _depth; i++) tty->print(" ");
1950 tty->print(" members:", _id);
1951 int k = 0;
1952 for (int i = 0; i < _members.length(); i++) {
1953 if (k++ >= 6) {
1954 tty->print("\n ");
1955 for (int j = 0; j < _depth+1; j++) tty->print(" ");
1956 k = 0;
1957 }
1958 CFGElement *s = _members.at(i);
1959 if (s->is_block()) {
1960 Block *b = s->as_Block();
1961 tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq);
1962 } else {
1963 CFGLoop* lp = s->as_CFGLoop();
1964 tty->print(" L%d(%6.3f)", lp->_id, lp->_freq);
1965 }
1966 }
1967 tty->print("\n");
1968 for (int i = 0; i < _depth; i++) tty->print(" ");
1969 tty->print(" exits: ");
1970 k = 0;
1971 for (int i = 0; i < _exits.length(); i++) {
1972 if (k++ >= 7) {
1973 tty->print("\n ");
1974 for (int j = 0; j < _depth+1; j++) tty->print(" ");
1975 k = 0;
1976 }
1977 Block *blk = _exits.at(i).get_target();
1978 float prob = _exits.at(i).get_prob();
1979 tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100));
1980 }
1981 tty->print("\n");
1982 }
1983 #endif