1 /*
2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "libadt/vectset.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/block.hpp"
29 #include "opto/c2compiler.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/machnode.hpp"
33 #include "opto/opcodes.hpp"
34 #include "opto/phaseX.hpp"
35 #include "opto/rootnode.hpp"
36 #include "opto/runtime.hpp"
37 #include "runtime/deoptimization.hpp"
38 #if defined AD_MD_HPP
39 # include AD_MD_HPP
40 #elif defined TARGET_ARCH_MODEL_x86_32
41 # include "adfiles/ad_x86_32.hpp"
42 #elif defined TARGET_ARCH_MODEL_x86_64
43 # include "adfiles/ad_x86_64.hpp"
44 #elif defined TARGET_ARCH_MODEL_aarch64
45 # include "adfiles/ad_aarch64.hpp"
46 #elif defined TARGET_ARCH_MODEL_sparc
47 # include "adfiles/ad_sparc.hpp"
48 #elif defined TARGET_ARCH_MODEL_zero
49 # include "adfiles/ad_zero.hpp"
50 #elif defined TARGET_ARCH_MODEL_ppc_64
51 # include "adfiles/ad_ppc_64.hpp"
52 #endif
53
54
55 // Portions of code courtesy of Clifford Click
56
57 // Optimization - Graph Style
58
59 // To avoid float value underflow
60 #define MIN_BLOCK_FREQUENCY 1.e-35f
61
62 //----------------------------schedule_node_into_block-------------------------
63 // Insert node n into block b. Look for projections of n and make sure they
64 // are in b also.
schedule_node_into_block(Node * n,Block * b)65 void PhaseCFG::schedule_node_into_block( Node *n, Block *b ) {
66 // Set basic block of n, Add n to b,
67 map_node_to_block(n, b);
68 b->add_inst(n);
69
70 // After Matching, nearly any old Node may have projections trailing it.
71 // These are usually machine-dependent flags. In any case, they might
72 // float to another block below this one. Move them up.
73 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
74 Node* use = n->fast_out(i);
75 if (use->is_Proj()) {
76 Block* buse = get_block_for_node(use);
77 if (buse != b) { // In wrong block?
78 if (buse != NULL) {
79 buse->find_remove(use); // Remove from wrong block
80 }
81 map_node_to_block(use, b);
82 b->add_inst(use);
83 }
84 }
85 }
86 }
87
88 //----------------------------replace_block_proj_ctrl-------------------------
89 // Nodes that have is_block_proj() nodes as their control need to use
90 // the appropriate Region for their actual block as their control since
91 // the projection will be in a predecessor block.
replace_block_proj_ctrl(Node * n)92 void PhaseCFG::replace_block_proj_ctrl( Node *n ) {
93 const Node *in0 = n->in(0);
94 assert(in0 != NULL, "Only control-dependent");
95 const Node *p = in0->is_block_proj();
96 if (p != NULL && p != n) { // Control from a block projection?
97 assert(!n->pinned() || n->is_MachConstantBase(), "only pinned MachConstantBase node is expected here");
98 // Find trailing Region
99 Block *pb = get_block_for_node(in0); // Block-projection already has basic block
100 uint j = 0;
101 if (pb->_num_succs != 1) { // More then 1 successor?
102 // Search for successor
103 uint max = pb->number_of_nodes();
104 assert( max > 1, "" );
105 uint start = max - pb->_num_succs;
106 // Find which output path belongs to projection
107 for (j = start; j < max; j++) {
108 if( pb->get_node(j) == in0 )
109 break;
110 }
111 assert( j < max, "must find" );
112 // Change control to match head of successor basic block
113 j -= start;
114 }
115 n->set_req(0, pb->_succs[j]->head());
116 }
117 }
118
119
120 //------------------------------schedule_pinned_nodes--------------------------
121 // Set the basic block for Nodes pinned into blocks
schedule_pinned_nodes(VectorSet & visited)122 void PhaseCFG::schedule_pinned_nodes(VectorSet &visited) {
123 // Allocate node stack of size C->live_nodes()+8 to avoid frequent realloc
124 GrowableArray <Node *> spstack(C->live_nodes() + 8);
125 spstack.push(_root);
126 while (spstack.is_nonempty()) {
127 Node* node = spstack.pop();
128 if (!visited.test_set(node->_idx)) { // Test node and flag it as visited
129 if (node->pinned() && !has_block(node)) { // Pinned? Nail it down!
130 assert(node->in(0), "pinned Node must have Control");
131 // Before setting block replace block_proj control edge
132 replace_block_proj_ctrl(node);
133 Node* input = node->in(0);
134 while (!input->is_block_start()) {
135 input = input->in(0);
136 }
137 Block* block = get_block_for_node(input); // Basic block of controlling input
138 schedule_node_into_block(node, block);
139 }
140
141 // process all inputs that are non NULL
142 for (int i = node->req() - 1; i >= 0; --i) {
143 if (node->in(i) != NULL) {
144 spstack.push(node->in(i));
145 }
146 }
147 }
148 }
149 }
150
151 #ifdef ASSERT
152 // Assert that new input b2 is dominated by all previous inputs.
153 // Check this by by seeing that it is dominated by b1, the deepest
154 // input observed until b2.
assert_dom(Block * b1,Block * b2,Node * n,const PhaseCFG * cfg)155 static void assert_dom(Block* b1, Block* b2, Node* n, const PhaseCFG* cfg) {
156 if (b1 == NULL) return;
157 assert(b1->_dom_depth < b2->_dom_depth, "sanity");
158 Block* tmp = b2;
159 while (tmp != b1 && tmp != NULL) {
160 tmp = tmp->_idom;
161 }
162 if (tmp != b1) {
163 // Detected an unschedulable graph. Print some nice stuff and die.
164 tty->print_cr("!!! Unschedulable graph !!!");
165 for (uint j=0; j<n->len(); j++) { // For all inputs
166 Node* inn = n->in(j); // Get input
167 if (inn == NULL) continue; // Ignore NULL, missing inputs
168 Block* inb = cfg->get_block_for_node(inn);
169 tty->print("B%d idom=B%d depth=%2d ",inb->_pre_order,
170 inb->_idom ? inb->_idom->_pre_order : 0, inb->_dom_depth);
171 inn->dump();
172 }
173 tty->print("Failing node: ");
174 n->dump();
175 assert(false, "unscheduable graph");
176 }
177 }
178 #endif
179
find_deepest_input(Node * n,const PhaseCFG * cfg)180 static Block* find_deepest_input(Node* n, const PhaseCFG* cfg) {
181 // Find the last input dominated by all other inputs.
182 Block* deepb = NULL; // Deepest block so far
183 int deepb_dom_depth = 0;
184 for (uint k = 0; k < n->len(); k++) { // For all inputs
185 Node* inn = n->in(k); // Get input
186 if (inn == NULL) continue; // Ignore NULL, missing inputs
187 Block* inb = cfg->get_block_for_node(inn);
188 assert(inb != NULL, "must already have scheduled this input");
189 if (deepb_dom_depth < (int) inb->_dom_depth) {
190 // The new inb must be dominated by the previous deepb.
191 // The various inputs must be linearly ordered in the dom
192 // tree, or else there will not be a unique deepest block.
193 DEBUG_ONLY(assert_dom(deepb, inb, n, cfg));
194 deepb = inb; // Save deepest block
195 deepb_dom_depth = deepb->_dom_depth;
196 }
197 }
198 assert(deepb != NULL, "must be at least one input to n");
199 return deepb;
200 }
201
202
203 //------------------------------schedule_early---------------------------------
204 // Find the earliest Block any instruction can be placed in. Some instructions
205 // are pinned into Blocks. Unpinned instructions can appear in last block in
206 // which all their inputs occur.
schedule_early(VectorSet & visited,Node_List & roots)207 bool PhaseCFG::schedule_early(VectorSet &visited, Node_List &roots) {
208 // Allocate stack with enough space to avoid frequent realloc
209 Node_Stack nstack(roots.Size() + 8);
210 // _root will be processed among C->top() inputs
211 roots.push(C->top());
212 visited.set(C->top()->_idx);
213
214 while (roots.size() != 0) {
215 // Use local variables nstack_top_n & nstack_top_i to cache values
216 // on stack's top.
217 Node* parent_node = roots.pop();
218 uint input_index = 0;
219
220 while (true) {
221 if (input_index == 0) {
222 // Fixup some control. Constants without control get attached
223 // to root and nodes that use is_block_proj() nodes should be attached
224 // to the region that starts their block.
225 const Node* control_input = parent_node->in(0);
226 if (control_input != NULL) {
227 replace_block_proj_ctrl(parent_node);
228 } else {
229 // Is a constant with NO inputs?
230 if (parent_node->req() == 1) {
231 parent_node->set_req(0, _root);
232 }
233 }
234 }
235
236 // First, visit all inputs and force them to get a block. If an
237 // input is already in a block we quit following inputs (to avoid
238 // cycles). Instead we put that Node on a worklist to be handled
239 // later (since IT'S inputs may not have a block yet).
240
241 // Assume all n's inputs will be processed
242 bool done = true;
243
244 while (input_index < parent_node->len()) {
245 Node* in = parent_node->in(input_index++);
246 if (in == NULL) {
247 continue;
248 }
249
250 int is_visited = visited.test_set(in->_idx);
251 if (!has_block(in)) {
252 if (is_visited) {
253 assert(false, "graph should be schedulable");
254 return false;
255 }
256 // Save parent node and next input's index.
257 nstack.push(parent_node, input_index);
258 // Process current input now.
259 parent_node = in;
260 input_index = 0;
261 // Not all n's inputs processed.
262 done = false;
263 break;
264 } else if (!is_visited) {
265 // Visit this guy later, using worklist
266 roots.push(in);
267 }
268 }
269
270 if (done) {
271 // All of n's inputs have been processed, complete post-processing.
272
273 // Some instructions are pinned into a block. These include Region,
274 // Phi, Start, Return, and other control-dependent instructions and
275 // any projections which depend on them.
276 if (!parent_node->pinned()) {
277 // Set earliest legal block.
278 Block* earliest_block = find_deepest_input(parent_node, this);
279 map_node_to_block(parent_node, earliest_block);
280 } else {
281 assert(get_block_for_node(parent_node) == get_block_for_node(parent_node->in(0)), "Pinned Node should be at the same block as its control edge");
282 }
283
284 if (nstack.is_empty()) {
285 // Finished all nodes on stack.
286 // Process next node on the worklist 'roots'.
287 break;
288 }
289 // Get saved parent node and next input's index.
290 parent_node = nstack.node();
291 input_index = nstack.index();
292 nstack.pop();
293 }
294 }
295 }
296 return true;
297 }
298
299 //------------------------------dom_lca----------------------------------------
300 // Find least common ancestor in dominator tree
301 // LCA is a current notion of LCA, to be raised above 'this'.
302 // As a convenient boundary condition, return 'this' if LCA is NULL.
303 // Find the LCA of those two nodes.
dom_lca(Block * LCA)304 Block* Block::dom_lca(Block* LCA) {
305 if (LCA == NULL || LCA == this) return this;
306
307 Block* anc = this;
308 while (anc->_dom_depth > LCA->_dom_depth)
309 anc = anc->_idom; // Walk up till anc is as high as LCA
310
311 while (LCA->_dom_depth > anc->_dom_depth)
312 LCA = LCA->_idom; // Walk up till LCA is as high as anc
313
314 while (LCA != anc) { // Walk both up till they are the same
315 LCA = LCA->_idom;
316 anc = anc->_idom;
317 }
318
319 return LCA;
320 }
321
322 //--------------------------raise_LCA_above_use--------------------------------
323 // We are placing a definition, and have been given a def->use edge.
324 // The definition must dominate the use, so move the LCA upward in the
325 // dominator tree to dominate the use. If the use is a phi, adjust
326 // the LCA only with the phi input paths which actually use this def.
raise_LCA_above_use(Block * LCA,Node * use,Node * def,const PhaseCFG * cfg)327 static Block* raise_LCA_above_use(Block* LCA, Node* use, Node* def, const PhaseCFG* cfg) {
328 Block* buse = cfg->get_block_for_node(use);
329 if (buse == NULL) return LCA; // Unused killing Projs have no use block
330 if (!use->is_Phi()) return buse->dom_lca(LCA);
331 uint pmax = use->req(); // Number of Phi inputs
332 // Why does not this loop just break after finding the matching input to
333 // the Phi? Well...it's like this. I do not have true def-use/use-def
334 // chains. Means I cannot distinguish, from the def-use direction, which
335 // of many use-defs lead from the same use to the same def. That is, this
336 // Phi might have several uses of the same def. Each use appears in a
337 // different predecessor block. But when I enter here, I cannot distinguish
338 // which use-def edge I should find the predecessor block for. So I find
339 // them all. Means I do a little extra work if a Phi uses the same value
340 // more than once.
341 for (uint j=1; j<pmax; j++) { // For all inputs
342 if (use->in(j) == def) { // Found matching input?
343 Block* pred = cfg->get_block_for_node(buse->pred(j));
344 LCA = pred->dom_lca(LCA);
345 }
346 }
347 return LCA;
348 }
349
350 //----------------------------raise_LCA_above_marks----------------------------
351 // Return a new LCA that dominates LCA and any of its marked predecessors.
352 // Search all my parents up to 'early' (exclusive), looking for predecessors
353 // which are marked with the given index. Return the LCA (in the dom tree)
354 // of all marked blocks. If there are none marked, return the original
355 // LCA.
raise_LCA_above_marks(Block * LCA,node_idx_t mark,Block * early,const PhaseCFG * cfg)356 static Block* raise_LCA_above_marks(Block* LCA, node_idx_t mark, Block* early, const PhaseCFG* cfg) {
357 Block_List worklist;
358 worklist.push(LCA);
359 while (worklist.size() > 0) {
360 Block* mid = worklist.pop();
361 if (mid == early) continue; // stop searching here
362
363 // Test and set the visited bit.
364 if (mid->raise_LCA_visited() == mark) continue; // already visited
365
366 // Don't process the current LCA, otherwise the search may terminate early
367 if (mid != LCA && mid->raise_LCA_mark() == mark) {
368 // Raise the LCA.
369 LCA = mid->dom_lca(LCA);
370 if (LCA == early) break; // stop searching everywhere
371 assert(early->dominates(LCA), "early is high enough");
372 // Resume searching at that point, skipping intermediate levels.
373 worklist.push(LCA);
374 if (LCA == mid)
375 continue; // Don't mark as visited to avoid early termination.
376 } else {
377 // Keep searching through this block's predecessors.
378 for (uint j = 1, jmax = mid->num_preds(); j < jmax; j++) {
379 Block* mid_parent = cfg->get_block_for_node(mid->pred(j));
380 worklist.push(mid_parent);
381 }
382 }
383 mid->set_raise_LCA_visited(mark);
384 }
385 return LCA;
386 }
387
388 //--------------------------memory_early_block--------------------------------
389 // This is a variation of find_deepest_input, the heart of schedule_early.
390 // Find the "early" block for a load, if we considered only memory and
391 // address inputs, that is, if other data inputs were ignored.
392 //
393 // Because a subset of edges are considered, the resulting block will
394 // be earlier (at a shallower dom_depth) than the true schedule_early
395 // point of the node. We compute this earlier block as a more permissive
396 // site for anti-dependency insertion, but only if subsume_loads is enabled.
memory_early_block(Node * load,Block * early,const PhaseCFG * cfg)397 static Block* memory_early_block(Node* load, Block* early, const PhaseCFG* cfg) {
398 Node* base;
399 Node* index;
400 Node* store = load->in(MemNode::Memory);
401 load->as_Mach()->memory_inputs(base, index);
402
403 assert(base != NodeSentinel && index != NodeSentinel,
404 "unexpected base/index inputs");
405
406 Node* mem_inputs[4];
407 int mem_inputs_length = 0;
408 if (base != NULL) mem_inputs[mem_inputs_length++] = base;
409 if (index != NULL) mem_inputs[mem_inputs_length++] = index;
410 if (store != NULL) mem_inputs[mem_inputs_length++] = store;
411
412 // In the comparision below, add one to account for the control input,
413 // which may be null, but always takes up a spot in the in array.
414 if (mem_inputs_length + 1 < (int) load->req()) {
415 // This "load" has more inputs than just the memory, base and index inputs.
416 // For purposes of checking anti-dependences, we need to start
417 // from the early block of only the address portion of the instruction,
418 // and ignore other blocks that may have factored into the wider
419 // schedule_early calculation.
420 if (load->in(0) != NULL) mem_inputs[mem_inputs_length++] = load->in(0);
421
422 Block* deepb = NULL; // Deepest block so far
423 int deepb_dom_depth = 0;
424 for (int i = 0; i < mem_inputs_length; i++) {
425 Block* inb = cfg->get_block_for_node(mem_inputs[i]);
426 if (deepb_dom_depth < (int) inb->_dom_depth) {
427 // The new inb must be dominated by the previous deepb.
428 // The various inputs must be linearly ordered in the dom
429 // tree, or else there will not be a unique deepest block.
430 DEBUG_ONLY(assert_dom(deepb, inb, load, cfg));
431 deepb = inb; // Save deepest block
432 deepb_dom_depth = deepb->_dom_depth;
433 }
434 }
435 early = deepb;
436 }
437
438 return early;
439 }
440
441 //--------------------------insert_anti_dependences---------------------------
442 // A load may need to witness memory that nearby stores can overwrite.
443 // For each nearby store, either insert an "anti-dependence" edge
444 // from the load to the store, or else move LCA upward to force the
445 // load to (eventually) be scheduled in a block above the store.
446 //
447 // Do not add edges to stores on distinct control-flow paths;
448 // only add edges to stores which might interfere.
449 //
450 // Return the (updated) LCA. There will not be any possibly interfering
451 // store between the load's "early block" and the updated LCA.
452 // Any stores in the updated LCA will have new precedence edges
453 // back to the load. The caller is expected to schedule the load
454 // in the LCA, in which case the precedence edges will make LCM
455 // preserve anti-dependences. The caller may also hoist the load
456 // above the LCA, if it is not the early block.
insert_anti_dependences(Block * LCA,Node * load,bool verify)457 Block* PhaseCFG::insert_anti_dependences(Block* LCA, Node* load, bool verify) {
458 assert(load->needs_anti_dependence_check(), "must be a load of some sort");
459 assert(LCA != NULL, "");
460 DEBUG_ONLY(Block* LCA_orig = LCA);
461
462 // Compute the alias index. Loads and stores with different alias indices
463 // do not need anti-dependence edges.
464 uint load_alias_idx = C->get_alias_index(load->adr_type());
465 #ifdef ASSERT
466 if (load_alias_idx == Compile::AliasIdxBot && C->AliasLevel() > 0 &&
467 (PrintOpto || VerifyAliases ||
468 PrintMiscellaneous && (WizardMode || Verbose))) {
469 // Load nodes should not consume all of memory.
470 // Reporting a bottom type indicates a bug in adlc.
471 // If some particular type of node validly consumes all of memory,
472 // sharpen the preceding "if" to exclude it, so we can catch bugs here.
473 tty->print_cr("*** Possible Anti-Dependence Bug: Load consumes all of memory.");
474 load->dump(2);
475 if (VerifyAliases) assert(load_alias_idx != Compile::AliasIdxBot, "");
476 }
477 #endif
478 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrComp),
479 "String compare is only known 'load' that does not conflict with any stores");
480 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrEquals),
481 "String equals is a 'load' that does not conflict with any stores");
482 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrIndexOf),
483 "String indexOf is a 'load' that does not conflict with any stores");
484 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_AryEq),
485 "Arrays equals is a 'load' that do not conflict with any stores");
486
487 if (!C->alias_type(load_alias_idx)->is_rewritable()) {
488 // It is impossible to spoil this load by putting stores before it,
489 // because we know that the stores will never update the value
490 // which 'load' must witness.
491 return LCA;
492 }
493
494 node_idx_t load_index = load->_idx;
495
496 // Note the earliest legal placement of 'load', as determined by
497 // by the unique point in the dom tree where all memory effects
498 // and other inputs are first available. (Computed by schedule_early.)
499 // For normal loads, 'early' is the shallowest place (dom graph wise)
500 // to look for anti-deps between this load and any store.
501 Block* early = get_block_for_node(load);
502
503 // If we are subsuming loads, compute an "early" block that only considers
504 // memory or address inputs. This block may be different than the
505 // schedule_early block in that it could be at an even shallower depth in the
506 // dominator tree, and allow for a broader discovery of anti-dependences.
507 if (C->subsume_loads()) {
508 early = memory_early_block(load, early, this);
509 }
510
511 ResourceArea *area = Thread::current()->resource_area();
512 Node_List worklist_mem(area); // prior memory state to store
513 Node_List worklist_store(area); // possible-def to explore
514 Node_List worklist_visited(area); // visited mergemem nodes
515 Node_List non_early_stores(area); // all relevant stores outside of early
516 bool must_raise_LCA = false;
517
518 #ifdef TRACK_PHI_INPUTS
519 // %%% This extra checking fails because MergeMem nodes are not GVNed.
520 // Provide "phi_inputs" to check if every input to a PhiNode is from the
521 // original memory state. This indicates a PhiNode for which should not
522 // prevent the load from sinking. For such a block, set_raise_LCA_mark
523 // may be overly conservative.
524 // Mechanism: count inputs seen for each Phi encountered in worklist_store.
525 DEBUG_ONLY(GrowableArray<uint> phi_inputs(area, C->unique(),0,0));
526 #endif
527
528 // 'load' uses some memory state; look for users of the same state.
529 // Recurse through MergeMem nodes to the stores that use them.
530
531 // Each of these stores is a possible definition of memory
532 // that 'load' needs to use. We need to force 'load'
533 // to occur before each such store. When the store is in
534 // the same block as 'load', we insert an anti-dependence
535 // edge load->store.
536
537 // The relevant stores "nearby" the load consist of a tree rooted
538 // at initial_mem, with internal nodes of type MergeMem.
539 // Therefore, the branches visited by the worklist are of this form:
540 // initial_mem -> (MergeMem ->)* store
541 // The anti-dependence constraints apply only to the fringe of this tree.
542
543 Node* initial_mem = load->in(MemNode::Memory);
544 worklist_store.push(initial_mem);
545 worklist_visited.push(initial_mem);
546 worklist_mem.push(NULL);
547 while (worklist_store.size() > 0) {
548 // Examine a nearby store to see if it might interfere with our load.
549 Node* mem = worklist_mem.pop();
550 Node* store = worklist_store.pop();
551 uint op = store->Opcode();
552
553 // MergeMems do not directly have anti-deps.
554 // Treat them as internal nodes in a forward tree of memory states,
555 // the leaves of which are each a 'possible-def'.
556 if (store == initial_mem // root (exclusive) of tree we are searching
557 || op == Op_MergeMem // internal node of tree we are searching
558 ) {
559 mem = store; // It's not a possibly interfering store.
560 if (store == initial_mem)
561 initial_mem = NULL; // only process initial memory once
562
563 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
564 store = mem->fast_out(i);
565 if (store->is_MergeMem()) {
566 // Be sure we don't get into combinatorial problems.
567 // (Allow phis to be repeated; they can merge two relevant states.)
568 uint j = worklist_visited.size();
569 for (; j > 0; j--) {
570 if (worklist_visited.at(j-1) == store) break;
571 }
572 if (j > 0) continue; // already on work list; do not repeat
573 worklist_visited.push(store);
574 }
575 worklist_mem.push(mem);
576 worklist_store.push(store);
577 }
578 continue;
579 }
580
581 if (op == Op_MachProj || op == Op_Catch) continue;
582 if (store->needs_anti_dependence_check()) continue; // not really a store
583
584 // Compute the alias index. Loads and stores with different alias
585 // indices do not need anti-dependence edges. Wide MemBar's are
586 // anti-dependent on everything (except immutable memories).
587 const TypePtr* adr_type = store->adr_type();
588 if (!C->can_alias(adr_type, load_alias_idx)) continue;
589
590 // Most slow-path runtime calls do NOT modify Java memory, but
591 // they can block and so write Raw memory.
592 if (store->is_Mach()) {
593 MachNode* mstore = store->as_Mach();
594 if (load_alias_idx != Compile::AliasIdxRaw) {
595 // Check for call into the runtime using the Java calling
596 // convention (and from there into a wrapper); it has no
597 // _method. Can't do this optimization for Native calls because
598 // they CAN write to Java memory.
599 if (mstore->ideal_Opcode() == Op_CallStaticJava) {
600 assert(mstore->is_MachSafePoint(), "");
601 MachSafePointNode* ms = (MachSafePointNode*) mstore;
602 assert(ms->is_MachCallJava(), "");
603 MachCallJavaNode* mcj = (MachCallJavaNode*) ms;
604 if (mcj->_method == NULL) {
605 // These runtime calls do not write to Java visible memory
606 // (other than Raw) and so do not require anti-dependence edges.
607 continue;
608 }
609 }
610 // Same for SafePoints: they read/write Raw but only read otherwise.
611 // This is basically a workaround for SafePoints only defining control
612 // instead of control + memory.
613 if (mstore->ideal_Opcode() == Op_SafePoint)
614 continue;
615 } else {
616 // Some raw memory, such as the load of "top" at an allocation,
617 // can be control dependent on the previous safepoint. See
618 // comments in GraphKit::allocate_heap() about control input.
619 // Inserting an anti-dep between such a safepoint and a use
620 // creates a cycle, and will cause a subsequent failure in
621 // local scheduling. (BugId 4919904)
622 // (%%% How can a control input be a safepoint and not a projection??)
623 if (mstore->ideal_Opcode() == Op_SafePoint && load->in(0) == mstore)
624 continue;
625 }
626 }
627
628 // Identify a block that the current load must be above,
629 // or else observe that 'store' is all the way up in the
630 // earliest legal block for 'load'. In the latter case,
631 // immediately insert an anti-dependence edge.
632 Block* store_block = get_block_for_node(store);
633 assert(store_block != NULL, "unused killing projections skipped above");
634
635 if (store->is_Phi()) {
636 // 'load' uses memory which is one (or more) of the Phi's inputs.
637 // It must be scheduled not before the Phi, but rather before
638 // each of the relevant Phi inputs.
639 //
640 // Instead of finding the LCA of all inputs to a Phi that match 'mem',
641 // we mark each corresponding predecessor block and do a combined
642 // hoisting operation later (raise_LCA_above_marks).
643 //
644 // Do not assert(store_block != early, "Phi merging memory after access")
645 // PhiNode may be at start of block 'early' with backedge to 'early'
646 DEBUG_ONLY(bool found_match = false);
647 for (uint j = PhiNode::Input, jmax = store->req(); j < jmax; j++) {
648 if (store->in(j) == mem) { // Found matching input?
649 DEBUG_ONLY(found_match = true);
650 Block* pred_block = get_block_for_node(store_block->pred(j));
651 if (pred_block != early) {
652 // If any predecessor of the Phi matches the load's "early block",
653 // we do not need a precedence edge between the Phi and 'load'
654 // since the load will be forced into a block preceding the Phi.
655 pred_block->set_raise_LCA_mark(load_index);
656 assert(!LCA_orig->dominates(pred_block) ||
657 early->dominates(pred_block), "early is high enough");
658 must_raise_LCA = true;
659 } else {
660 // anti-dependent upon PHI pinned below 'early', no edge needed
661 LCA = early; // but can not schedule below 'early'
662 }
663 }
664 }
665 assert(found_match, "no worklist bug");
666 #ifdef TRACK_PHI_INPUTS
667 #ifdef ASSERT
668 // This assert asks about correct handling of PhiNodes, which may not
669 // have all input edges directly from 'mem'. See BugId 4621264
670 int num_mem_inputs = phi_inputs.at_grow(store->_idx,0) + 1;
671 // Increment by exactly one even if there are multiple copies of 'mem'
672 // coming into the phi, because we will run this block several times
673 // if there are several copies of 'mem'. (That's how DU iterators work.)
674 phi_inputs.at_put(store->_idx, num_mem_inputs);
675 assert(PhiNode::Input + num_mem_inputs < store->req(),
676 "Expect at least one phi input will not be from original memory state");
677 #endif //ASSERT
678 #endif //TRACK_PHI_INPUTS
679 } else if (store_block != early) {
680 // 'store' is between the current LCA and earliest possible block.
681 // Label its block, and decide later on how to raise the LCA
682 // to include the effect on LCA of this store.
683 // If this store's block gets chosen as the raised LCA, we
684 // will find him on the non_early_stores list and stick him
685 // with a precedence edge.
686 // (But, don't bother if LCA is already raised all the way.)
687 if (LCA != early) {
688 store_block->set_raise_LCA_mark(load_index);
689 must_raise_LCA = true;
690 non_early_stores.push(store);
691 }
692 } else {
693 // Found a possibly-interfering store in the load's 'early' block.
694 // This means 'load' cannot sink at all in the dominator tree.
695 // Add an anti-dep edge, and squeeze 'load' into the highest block.
696 assert(store != load->in(0), "dependence cycle found");
697 if (verify) {
698 assert(store->find_edge(load) != -1, "missing precedence edge");
699 } else {
700 store->add_prec(load);
701 }
702 LCA = early;
703 // This turns off the process of gathering non_early_stores.
704 }
705 }
706 // (Worklist is now empty; all nearby stores have been visited.)
707
708 // Finished if 'load' must be scheduled in its 'early' block.
709 // If we found any stores there, they have already been given
710 // precedence edges.
711 if (LCA == early) return LCA;
712
713 // We get here only if there are no possibly-interfering stores
714 // in the load's 'early' block. Move LCA up above all predecessors
715 // which contain stores we have noted.
716 //
717 // The raised LCA block can be a home to such interfering stores,
718 // but its predecessors must not contain any such stores.
719 //
720 // The raised LCA will be a lower bound for placing the load,
721 // preventing the load from sinking past any block containing
722 // a store that may invalidate the memory state required by 'load'.
723 if (must_raise_LCA)
724 LCA = raise_LCA_above_marks(LCA, load->_idx, early, this);
725 if (LCA == early) return LCA;
726
727 // Insert anti-dependence edges from 'load' to each store
728 // in the non-early LCA block.
729 // Mine the non_early_stores list for such stores.
730 if (LCA->raise_LCA_mark() == load_index) {
731 while (non_early_stores.size() > 0) {
732 Node* store = non_early_stores.pop();
733 Block* store_block = get_block_for_node(store);
734 if (store_block == LCA) {
735 // add anti_dependence from store to load in its own block
736 assert(store != load->in(0), "dependence cycle found");
737 if (verify) {
738 assert(store->find_edge(load) != -1, "missing precedence edge");
739 } else {
740 store->add_prec(load);
741 }
742 } else {
743 assert(store_block->raise_LCA_mark() == load_index, "block was marked");
744 // Any other stores we found must be either inside the new LCA
745 // or else outside the original LCA. In the latter case, they
746 // did not interfere with any use of 'load'.
747 assert(LCA->dominates(store_block)
748 || !LCA_orig->dominates(store_block), "no stray stores");
749 }
750 }
751 }
752
753 // Return the highest block containing stores; any stores
754 // within that block have been given anti-dependence edges.
755 return LCA;
756 }
757
758 // This class is used to iterate backwards over the nodes in the graph.
759
760 class Node_Backward_Iterator {
761
762 private:
763 Node_Backward_Iterator();
764
765 public:
766 // Constructor for the iterator
767 Node_Backward_Iterator(Node *root, VectorSet &visited, Node_List &stack, PhaseCFG &cfg);
768
769 // Postincrement operator to iterate over the nodes
770 Node *next();
771
772 private:
773 VectorSet &_visited;
774 Node_List &_stack;
775 PhaseCFG &_cfg;
776 };
777
778 // Constructor for the Node_Backward_Iterator
Node_Backward_Iterator(Node * root,VectorSet & visited,Node_List & stack,PhaseCFG & cfg)779 Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_List &stack, PhaseCFG &cfg)
780 : _visited(visited), _stack(stack), _cfg(cfg) {
781 // The stack should contain exactly the root
782 stack.clear();
783 stack.push(root);
784
785 // Clear the visited bits
786 visited.Clear();
787 }
788
789 // Iterator for the Node_Backward_Iterator
next()790 Node *Node_Backward_Iterator::next() {
791
792 // If the _stack is empty, then just return NULL: finished.
793 if ( !_stack.size() )
794 return NULL;
795
796 // '_stack' is emulating a real _stack. The 'visit-all-users' loop has been
797 // made stateless, so I do not need to record the index 'i' on my _stack.
798 // Instead I visit all users each time, scanning for unvisited users.
799 // I visit unvisited not-anti-dependence users first, then anti-dependent
800 // children next.
801 Node *self = _stack.pop();
802
803 // I cycle here when I am entering a deeper level of recursion.
804 // The key variable 'self' was set prior to jumping here.
805 while( 1 ) {
806
807 _visited.set(self->_idx);
808
809 // Now schedule all uses as late as possible.
810 const Node* src = self->is_Proj() ? self->in(0) : self;
811 uint src_rpo = _cfg.get_block_for_node(src)->_rpo;
812
813 // Schedule all nodes in a post-order visit
814 Node *unvisited = NULL; // Unvisited anti-dependent Node, if any
815
816 // Scan for unvisited nodes
817 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
818 // For all uses, schedule late
819 Node* n = self->fast_out(i); // Use
820
821 // Skip already visited children
822 if ( _visited.test(n->_idx) )
823 continue;
824
825 // do not traverse backward control edges
826 Node *use = n->is_Proj() ? n->in(0) : n;
827 uint use_rpo = _cfg.get_block_for_node(use)->_rpo;
828
829 if ( use_rpo < src_rpo )
830 continue;
831
832 // Phi nodes always precede uses in a basic block
833 if ( use_rpo == src_rpo && use->is_Phi() )
834 continue;
835
836 unvisited = n; // Found unvisited
837
838 // Check for possible-anti-dependent
839 if( !n->needs_anti_dependence_check() )
840 break; // Not visited, not anti-dep; schedule it NOW
841 }
842
843 // Did I find an unvisited not-anti-dependent Node?
844 if ( !unvisited )
845 break; // All done with children; post-visit 'self'
846
847 // Visit the unvisited Node. Contains the obvious push to
848 // indicate I'm entering a deeper level of recursion. I push the
849 // old state onto the _stack and set a new state and loop (recurse).
850 _stack.push(self);
851 self = unvisited;
852 } // End recursion loop
853
854 return self;
855 }
856
857 //------------------------------ComputeLatenciesBackwards----------------------
858 // Compute the latency of all the instructions.
compute_latencies_backwards(VectorSet & visited,Node_List & stack)859 void PhaseCFG::compute_latencies_backwards(VectorSet &visited, Node_List &stack) {
860 #ifndef PRODUCT
861 if (trace_opto_pipelining())
862 tty->print("\n#---- ComputeLatenciesBackwards ----\n");
863 #endif
864
865 Node_Backward_Iterator iter((Node *)_root, visited, stack, *this);
866 Node *n;
867
868 // Walk over all the nodes from last to first
869 while (n = iter.next()) {
870 // Set the latency for the definitions of this instruction
871 partial_latency_of_defs(n);
872 }
873 } // end ComputeLatenciesBackwards
874
875 //------------------------------partial_latency_of_defs------------------------
876 // Compute the latency impact of this node on all defs. This computes
877 // a number that increases as we approach the beginning of the routine.
partial_latency_of_defs(Node * n)878 void PhaseCFG::partial_latency_of_defs(Node *n) {
879 // Set the latency for this instruction
880 #ifndef PRODUCT
881 if (trace_opto_pipelining()) {
882 tty->print("# latency_to_inputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n));
883 dump();
884 }
885 #endif
886
887 if (n->is_Proj()) {
888 n = n->in(0);
889 }
890
891 if (n->is_Root()) {
892 return;
893 }
894
895 uint nlen = n->len();
896 uint use_latency = get_latency_for_node(n);
897 uint use_pre_order = get_block_for_node(n)->_pre_order;
898
899 for (uint j = 0; j < nlen; j++) {
900 Node *def = n->in(j);
901
902 if (!def || def == n) {
903 continue;
904 }
905
906 // Walk backwards thru projections
907 if (def->is_Proj()) {
908 def = def->in(0);
909 }
910
911 #ifndef PRODUCT
912 if (trace_opto_pipelining()) {
913 tty->print("# in(%2d): ", j);
914 def->dump();
915 }
916 #endif
917
918 // If the defining block is not known, assume it is ok
919 Block *def_block = get_block_for_node(def);
920 uint def_pre_order = def_block ? def_block->_pre_order : 0;
921
922 if ((use_pre_order < def_pre_order) || (use_pre_order == def_pre_order && n->is_Phi())) {
923 continue;
924 }
925
926 uint delta_latency = n->latency(j);
927 uint current_latency = delta_latency + use_latency;
928
929 if (get_latency_for_node(def) < current_latency) {
930 set_latency_for_node(def, current_latency);
931 }
932
933 #ifndef PRODUCT
934 if (trace_opto_pipelining()) {
935 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, node_latency[%d] = %d", use_latency, j, delta_latency, current_latency, def->_idx, get_latency_for_node(def));
936 }
937 #endif
938 }
939 }
940
941 //------------------------------latency_from_use-------------------------------
942 // Compute the latency of a specific use
latency_from_use(Node * n,const Node * def,Node * use)943 int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) {
944 // If self-reference, return no latency
945 if (use == n || use->is_Root()) {
946 return 0;
947 }
948
949 uint def_pre_order = get_block_for_node(def)->_pre_order;
950 uint latency = 0;
951
952 // If the use is not a projection, then it is simple...
953 if (!use->is_Proj()) {
954 #ifndef PRODUCT
955 if (trace_opto_pipelining()) {
956 tty->print("# out(): ");
957 use->dump();
958 }
959 #endif
960
961 uint use_pre_order = get_block_for_node(use)->_pre_order;
962
963 if (use_pre_order < def_pre_order)
964 return 0;
965
966 if (use_pre_order == def_pre_order && use->is_Phi())
967 return 0;
968
969 uint nlen = use->len();
970 uint nl = get_latency_for_node(use);
971
972 for ( uint j=0; j<nlen; j++ ) {
973 if (use->in(j) == n) {
974 // Change this if we want local latencies
975 uint ul = use->latency(j);
976 uint l = ul + nl;
977 if (latency < l) latency = l;
978 #ifndef PRODUCT
979 if (trace_opto_pipelining()) {
980 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, latency = %d",
981 nl, j, ul, l, latency);
982 }
983 #endif
984 }
985 }
986 } else {
987 // This is a projection, just grab the latency of the use(s)
988 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
989 uint l = latency_from_use(use, def, use->fast_out(j));
990 if (latency < l) latency = l;
991 }
992 }
993
994 return latency;
995 }
996
997 //------------------------------latency_from_uses------------------------------
998 // Compute the latency of this instruction relative to all of it's uses.
999 // This computes a number that increases as we approach the beginning of the
1000 // routine.
latency_from_uses(Node * n)1001 void PhaseCFG::latency_from_uses(Node *n) {
1002 // Set the latency for this instruction
1003 #ifndef PRODUCT
1004 if (trace_opto_pipelining()) {
1005 tty->print("# latency_from_outputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n));
1006 dump();
1007 }
1008 #endif
1009 uint latency=0;
1010 const Node *def = n->is_Proj() ? n->in(0): n;
1011
1012 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1013 uint l = latency_from_use(n, def, n->fast_out(i));
1014
1015 if (latency < l) latency = l;
1016 }
1017
1018 set_latency_for_node(n, latency);
1019 }
1020
1021 //------------------------------hoist_to_cheaper_block-------------------------
1022 // Pick a block for node self, between early and LCA, that is a cheaper
1023 // alternative to LCA.
hoist_to_cheaper_block(Block * LCA,Block * early,Node * self)1024 Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) {
1025 const double delta = 1+PROB_UNLIKELY_MAG(4);
1026 Block* least = LCA;
1027 double least_freq = least->_freq;
1028 uint target = get_latency_for_node(self);
1029 uint start_latency = get_latency_for_node(LCA->head());
1030 uint end_latency = get_latency_for_node(LCA->get_node(LCA->end_idx()));
1031 bool in_latency = (target <= start_latency);
1032 const Block* root_block = get_block_for_node(_root);
1033
1034 // Turn off latency scheduling if scheduling is just plain off
1035 if (!C->do_scheduling())
1036 in_latency = true;
1037
1038 // Do not hoist (to cover latency) instructions which target a
1039 // single register. Hoisting stretches the live range of the
1040 // single register and may force spilling.
1041 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1042 if (mach && mach->out_RegMask().is_bound1() && mach->out_RegMask().is_NotEmpty())
1043 in_latency = true;
1044
1045 #ifndef PRODUCT
1046 if (trace_opto_pipelining()) {
1047 tty->print("# Find cheaper block for latency %d: ", get_latency_for_node(self));
1048 self->dump();
1049 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1050 LCA->_pre_order,
1051 LCA->head()->_idx,
1052 start_latency,
1053 LCA->get_node(LCA->end_idx())->_idx,
1054 end_latency,
1055 least_freq);
1056 }
1057 #endif
1058
1059 int cand_cnt = 0; // number of candidates tried
1060
1061 // Walk up the dominator tree from LCA (Lowest common ancestor) to
1062 // the earliest legal location. Capture the least execution frequency.
1063 while (LCA != early) {
1064 LCA = LCA->_idom; // Follow up the dominator tree
1065
1066 if (LCA == NULL) {
1067 // Bailout without retry
1068 assert(false, "graph should be schedulable");
1069 C->record_method_not_compilable("late schedule failed: LCA == NULL");
1070 return least;
1071 }
1072
1073 // Don't hoist machine instructions to the root basic block
1074 if (mach && LCA == root_block)
1075 break;
1076
1077 uint start_lat = get_latency_for_node(LCA->head());
1078 uint end_idx = LCA->end_idx();
1079 uint end_lat = get_latency_for_node(LCA->get_node(end_idx));
1080 double LCA_freq = LCA->_freq;
1081 #ifndef PRODUCT
1082 if (trace_opto_pipelining()) {
1083 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1084 LCA->_pre_order, LCA->head()->_idx, start_lat, end_idx, end_lat, LCA_freq);
1085 }
1086 #endif
1087 cand_cnt++;
1088 if (LCA_freq < least_freq || // Better Frequency
1089 (StressGCM && Compile::randomized_select(cand_cnt)) || // Should be randomly accepted in stress mode
1090 (!StressGCM && // Otherwise, choose with latency
1091 !in_latency && // No block containing latency
1092 LCA_freq < least_freq * delta && // No worse frequency
1093 target >= end_lat && // within latency range
1094 !self->is_iteratively_computed() ) // But don't hoist IV increments
1095 // because they may end up above other uses of their phi forcing
1096 // their result register to be different from their input.
1097 ) {
1098 least = LCA; // Found cheaper block
1099 least_freq = LCA_freq;
1100 start_latency = start_lat;
1101 end_latency = end_lat;
1102 if (target <= start_lat)
1103 in_latency = true;
1104 }
1105 }
1106
1107 #ifndef PRODUCT
1108 if (trace_opto_pipelining()) {
1109 tty->print_cr("# Choose block B%d with start latency=%d and freq=%g",
1110 least->_pre_order, start_latency, least_freq);
1111 }
1112 #endif
1113
1114 // See if the latency needs to be updated
1115 if (target < end_latency) {
1116 #ifndef PRODUCT
1117 if (trace_opto_pipelining()) {
1118 tty->print_cr("# Change latency for [%4d] from %d to %d", self->_idx, target, end_latency);
1119 }
1120 #endif
1121 set_latency_for_node(self, end_latency);
1122 partial_latency_of_defs(self);
1123 }
1124
1125 return least;
1126 }
1127
1128
1129 //------------------------------schedule_late-----------------------------------
1130 // Now schedule all codes as LATE as possible. This is the LCA in the
1131 // dominator tree of all USES of a value. Pick the block with the least
1132 // loop nesting depth that is lowest in the dominator tree.
1133 extern const char must_clone[];
schedule_late(VectorSet & visited,Node_List & stack)1134 void PhaseCFG::schedule_late(VectorSet &visited, Node_List &stack) {
1135 #ifndef PRODUCT
1136 if (trace_opto_pipelining())
1137 tty->print("\n#---- schedule_late ----\n");
1138 #endif
1139
1140 Node_Backward_Iterator iter((Node *)_root, visited, stack, *this);
1141 Node *self;
1142
1143 // Walk over all the nodes from last to first
1144 while (self = iter.next()) {
1145 Block* early = get_block_for_node(self); // Earliest legal placement
1146
1147 if (self->is_top()) {
1148 // Top node goes in bb #2 with other constants.
1149 // It must be special-cased, because it has no out edges.
1150 early->add_inst(self);
1151 continue;
1152 }
1153
1154 // No uses, just terminate
1155 if (self->outcnt() == 0) {
1156 assert(self->is_MachProj(), "sanity");
1157 continue; // Must be a dead machine projection
1158 }
1159
1160 // If node is pinned in the block, then no scheduling can be done.
1161 if( self->pinned() ) // Pinned in block?
1162 continue;
1163
1164 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1165 if (mach) {
1166 switch (mach->ideal_Opcode()) {
1167 case Op_CreateEx:
1168 // Don't move exception creation
1169 early->add_inst(self);
1170 continue;
1171 break;
1172 case Op_CheckCastPP:
1173 // Don't move CheckCastPP nodes away from their input, if the input
1174 // is a rawptr (5071820).
1175 Node *def = self->in(1);
1176 if (def != NULL && def->bottom_type()->base() == Type::RawPtr) {
1177 early->add_inst(self);
1178 #ifdef ASSERT
1179 _raw_oops.push(def);
1180 #endif
1181 continue;
1182 }
1183 break;
1184 }
1185 }
1186
1187 // Gather LCA of all uses
1188 Block *LCA = NULL;
1189 {
1190 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
1191 // For all uses, find LCA
1192 Node* use = self->fast_out(i);
1193 LCA = raise_LCA_above_use(LCA, use, self, this);
1194 }
1195 } // (Hide defs of imax, i from rest of block.)
1196
1197 // Place temps in the block of their use. This isn't a
1198 // requirement for correctness but it reduces useless
1199 // interference between temps and other nodes.
1200 if (mach != NULL && mach->is_MachTemp()) {
1201 map_node_to_block(self, LCA);
1202 LCA->add_inst(self);
1203 continue;
1204 }
1205
1206 // Check if 'self' could be anti-dependent on memory
1207 if (self->needs_anti_dependence_check()) {
1208 // Hoist LCA above possible-defs and insert anti-dependences to
1209 // defs in new LCA block.
1210 LCA = insert_anti_dependences(LCA, self);
1211 }
1212
1213 if (early->_dom_depth > LCA->_dom_depth) {
1214 // Somehow the LCA has moved above the earliest legal point.
1215 // (One way this can happen is via memory_early_block.)
1216 if (C->subsume_loads() == true && !C->failing()) {
1217 // Retry with subsume_loads == false
1218 // If this is the first failure, the sentinel string will "stick"
1219 // to the Compile object, and the C2Compiler will see it and retry.
1220 C->record_failure(C2Compiler::retry_no_subsuming_loads());
1221 } else {
1222 // Bailout without retry when (early->_dom_depth > LCA->_dom_depth)
1223 assert(false, "graph should be schedulable");
1224 C->record_method_not_compilable("late schedule failed: incorrect graph");
1225 }
1226 return;
1227 }
1228
1229 // If there is no opportunity to hoist, then we're done.
1230 // In stress mode, try to hoist even the single operations.
1231 bool try_to_hoist = StressGCM || (LCA != early);
1232
1233 // Must clone guys stay next to use; no hoisting allowed.
1234 // Also cannot hoist guys that alter memory or are otherwise not
1235 // allocatable (hoisting can make a value live longer, leading to
1236 // anti and output dependency problems which are normally resolved
1237 // by the register allocator giving everyone a different register).
1238 if (mach != NULL && must_clone[mach->ideal_Opcode()])
1239 try_to_hoist = false;
1240
1241 Block* late = NULL;
1242 if (try_to_hoist) {
1243 // Now find the block with the least execution frequency.
1244 // Start at the latest schedule and work up to the earliest schedule
1245 // in the dominator tree. Thus the Node will dominate all its uses.
1246 late = hoist_to_cheaper_block(LCA, early, self);
1247 } else {
1248 // Just use the LCA of the uses.
1249 late = LCA;
1250 }
1251
1252 // Put the node into target block
1253 schedule_node_into_block(self, late);
1254
1255 #ifdef ASSERT
1256 if (self->needs_anti_dependence_check()) {
1257 // since precedence edges are only inserted when we're sure they
1258 // are needed make sure that after placement in a block we don't
1259 // need any new precedence edges.
1260 verify_anti_dependences(late, self);
1261 }
1262 #endif
1263 } // Loop until all nodes have been visited
1264
1265 } // end ScheduleLate
1266
1267 //------------------------------GlobalCodeMotion-------------------------------
global_code_motion()1268 void PhaseCFG::global_code_motion() {
1269 ResourceMark rm;
1270
1271 #ifndef PRODUCT
1272 if (trace_opto_pipelining()) {
1273 tty->print("\n---- Start GlobalCodeMotion ----\n");
1274 }
1275 #endif
1276
1277 // Initialize the node to block mapping for things on the proj_list
1278 for (uint i = 0; i < _matcher.number_of_projections(); i++) {
1279 unmap_node_from_block(_matcher.get_projection(i));
1280 }
1281
1282 // Set the basic block for Nodes pinned into blocks
1283 Arena* arena = Thread::current()->resource_area();
1284 VectorSet visited(arena);
1285 schedule_pinned_nodes(visited);
1286
1287 // Find the earliest Block any instruction can be placed in. Some
1288 // instructions are pinned into Blocks. Unpinned instructions can
1289 // appear in last block in which all their inputs occur.
1290 visited.Clear();
1291 Node_List stack(arena);
1292 // Pre-grow the list
1293 stack.map((C->live_nodes() >> 1) + 16, NULL);
1294 if (!schedule_early(visited, stack)) {
1295 // Bailout without retry
1296 C->record_method_not_compilable("early schedule failed");
1297 return;
1298 }
1299
1300 // Build Def-Use edges.
1301 // Compute the latency information (via backwards walk) for all the
1302 // instructions in the graph
1303 _node_latency = new GrowableArray<uint>(); // resource_area allocation
1304
1305 if (C->do_scheduling()) {
1306 compute_latencies_backwards(visited, stack);
1307 }
1308
1309 // Now schedule all codes as LATE as possible. This is the LCA in the
1310 // dominator tree of all USES of a value. Pick the block with the least
1311 // loop nesting depth that is lowest in the dominator tree.
1312 // ( visited.Clear() called in schedule_late()->Node_Backward_Iterator() )
1313 schedule_late(visited, stack);
1314 if (C->failing()) {
1315 return;
1316 }
1317
1318 #ifndef PRODUCT
1319 if (trace_opto_pipelining()) {
1320 tty->print("\n---- Detect implicit null checks ----\n");
1321 }
1322 #endif
1323
1324 // Detect implicit-null-check opportunities. Basically, find NULL checks
1325 // with suitable memory ops nearby. Use the memory op to do the NULL check.
1326 // I can generate a memory op if there is not one nearby.
1327 if (C->is_method_compilation()) {
1328 // By reversing the loop direction we get a very minor gain on mpegaudio.
1329 // Feel free to revert to a forward loop for clarity.
1330 // for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) {
1331 for (int i = _matcher._null_check_tests.size() - 2; i >= 0; i -= 2) {
1332 Node* proj = _matcher._null_check_tests[i];
1333 Node* val = _matcher._null_check_tests[i + 1];
1334 Block* block = get_block_for_node(proj);
1335 implicit_null_check(block, proj, val, C->allowed_deopt_reasons());
1336 // The implicit_null_check will only perform the transformation
1337 // if the null branch is truly uncommon, *and* it leads to an
1338 // uncommon trap. Combined with the too_many_traps guards
1339 // above, this prevents SEGV storms reported in 6366351,
1340 // by recompiling offending methods without this optimization.
1341 }
1342 }
1343
1344 #ifndef PRODUCT
1345 if (trace_opto_pipelining()) {
1346 tty->print("\n---- Start Local Scheduling ----\n");
1347 }
1348 #endif
1349
1350 // Schedule locally. Right now a simple topological sort.
1351 // Later, do a real latency aware scheduler.
1352 GrowableArray<int> ready_cnt(C->unique(), C->unique(), -1);
1353 visited.Clear();
1354 for (uint i = 0; i < number_of_blocks(); i++) {
1355 Block* block = get_block(i);
1356 if (!schedule_local(block, ready_cnt, visited)) {
1357 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
1358 C->record_method_not_compilable("local schedule failed");
1359 }
1360 return;
1361 }
1362 }
1363
1364 // If we inserted any instructions between a Call and his CatchNode,
1365 // clone the instructions on all paths below the Catch.
1366 for (uint i = 0; i < number_of_blocks(); i++) {
1367 Block* block = get_block(i);
1368 call_catch_cleanup(block);
1369 }
1370
1371 #ifndef PRODUCT
1372 if (trace_opto_pipelining()) {
1373 tty->print("\n---- After GlobalCodeMotion ----\n");
1374 for (uint i = 0; i < number_of_blocks(); i++) {
1375 Block* block = get_block(i);
1376 block->dump();
1377 }
1378 }
1379 #endif
1380 // Dead.
1381 _node_latency = (GrowableArray<uint> *)((intptr_t)0xdeadbeef);
1382 }
1383
do_global_code_motion()1384 bool PhaseCFG::do_global_code_motion() {
1385
1386 build_dominator_tree();
1387 if (C->failing()) {
1388 return false;
1389 }
1390
1391 NOT_PRODUCT( C->verify_graph_edges(); )
1392
1393 estimate_block_frequency();
1394
1395 global_code_motion();
1396
1397 if (C->failing()) {
1398 return false;
1399 }
1400
1401 return true;
1402 }
1403
1404 //------------------------------Estimate_Block_Frequency-----------------------
1405 // Estimate block frequencies based on IfNode probabilities.
estimate_block_frequency()1406 void PhaseCFG::estimate_block_frequency() {
1407
1408 // Force conditional branches leading to uncommon traps to be unlikely,
1409 // not because we get to the uncommon_trap with less relative frequency,
1410 // but because an uncommon_trap typically causes a deopt, so we only get
1411 // there once.
1412 if (C->do_freq_based_layout()) {
1413 Block_List worklist;
1414 Block* root_blk = get_block(0);
1415 for (uint i = 1; i < root_blk->num_preds(); i++) {
1416 Block *pb = get_block_for_node(root_blk->pred(i));
1417 if (pb->has_uncommon_code()) {
1418 worklist.push(pb);
1419 }
1420 }
1421 while (worklist.size() > 0) {
1422 Block* uct = worklist.pop();
1423 if (uct == get_root_block()) {
1424 continue;
1425 }
1426 for (uint i = 1; i < uct->num_preds(); i++) {
1427 Block *pb = get_block_for_node(uct->pred(i));
1428 if (pb->_num_succs == 1) {
1429 worklist.push(pb);
1430 } else if (pb->num_fall_throughs() == 2) {
1431 pb->update_uncommon_branch(uct);
1432 }
1433 }
1434 }
1435 }
1436
1437 // Create the loop tree and calculate loop depth.
1438 _root_loop = create_loop_tree();
1439 _root_loop->compute_loop_depth(0);
1440
1441 // Compute block frequency of each block, relative to a single loop entry.
1442 _root_loop->compute_freq();
1443
1444 // Adjust all frequencies to be relative to a single method entry
1445 _root_loop->_freq = 1.0;
1446 _root_loop->scale_freq();
1447
1448 // Save outmost loop frequency for LRG frequency threshold
1449 _outer_loop_frequency = _root_loop->outer_loop_freq();
1450
1451 // force paths ending at uncommon traps to be infrequent
1452 if (!C->do_freq_based_layout()) {
1453 Block_List worklist;
1454 Block* root_blk = get_block(0);
1455 for (uint i = 1; i < root_blk->num_preds(); i++) {
1456 Block *pb = get_block_for_node(root_blk->pred(i));
1457 if (pb->has_uncommon_code()) {
1458 worklist.push(pb);
1459 }
1460 }
1461 while (worklist.size() > 0) {
1462 Block* uct = worklist.pop();
1463 uct->_freq = PROB_MIN;
1464 for (uint i = 1; i < uct->num_preds(); i++) {
1465 Block *pb = get_block_for_node(uct->pred(i));
1466 if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) {
1467 worklist.push(pb);
1468 }
1469 }
1470 }
1471 }
1472
1473 #ifdef ASSERT
1474 for (uint i = 0; i < number_of_blocks(); i++) {
1475 Block* b = get_block(i);
1476 assert(b->_freq >= MIN_BLOCK_FREQUENCY, "Register Allocator requires meaningful block frequency");
1477 }
1478 #endif
1479
1480 #ifndef PRODUCT
1481 if (PrintCFGBlockFreq) {
1482 tty->print_cr("CFG Block Frequencies");
1483 _root_loop->dump_tree();
1484 if (Verbose) {
1485 tty->print_cr("PhaseCFG dump");
1486 dump();
1487 tty->print_cr("Node dump");
1488 _root->dump(99999);
1489 }
1490 }
1491 #endif
1492 }
1493
1494 //----------------------------create_loop_tree--------------------------------
1495 // Create a loop tree from the CFG
create_loop_tree()1496 CFGLoop* PhaseCFG::create_loop_tree() {
1497
1498 #ifdef ASSERT
1499 assert(get_block(0) == get_root_block(), "first block should be root block");
1500 for (uint i = 0; i < number_of_blocks(); i++) {
1501 Block* block = get_block(i);
1502 // Check that _loop field are clear...we could clear them if not.
1503 assert(block->_loop == NULL, "clear _loop expected");
1504 // Sanity check that the RPO numbering is reflected in the _blocks array.
1505 // It doesn't have to be for the loop tree to be built, but if it is not,
1506 // then the blocks have been reordered since dom graph building...which
1507 // may question the RPO numbering
1508 assert(block->_rpo == i, "unexpected reverse post order number");
1509 }
1510 #endif
1511
1512 int idct = 0;
1513 CFGLoop* root_loop = new CFGLoop(idct++);
1514
1515 Block_List worklist;
1516
1517 // Assign blocks to loops
1518 for(uint i = number_of_blocks() - 1; i > 0; i-- ) { // skip Root block
1519 Block* block = get_block(i);
1520
1521 if (block->head()->is_Loop()) {
1522 Block* loop_head = block;
1523 assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1524 Node* tail_n = loop_head->pred(LoopNode::LoopBackControl);
1525 Block* tail = get_block_for_node(tail_n);
1526
1527 // Defensively filter out Loop nodes for non-single-entry loops.
1528 // For all reasonable loops, the head occurs before the tail in RPO.
1529 if (i <= tail->_rpo) {
1530
1531 // The tail and (recursive) predecessors of the tail
1532 // are made members of a new loop.
1533
1534 assert(worklist.size() == 0, "nonempty worklist");
1535 CFGLoop* nloop = new CFGLoop(idct++);
1536 assert(loop_head->_loop == NULL, "just checking");
1537 loop_head->_loop = nloop;
1538 // Add to nloop so push_pred() will skip over inner loops
1539 nloop->add_member(loop_head);
1540 nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, this);
1541
1542 while (worklist.size() > 0) {
1543 Block* member = worklist.pop();
1544 if (member != loop_head) {
1545 for (uint j = 1; j < member->num_preds(); j++) {
1546 nloop->push_pred(member, j, worklist, this);
1547 }
1548 }
1549 }
1550 }
1551 }
1552 }
1553
1554 // Create a member list for each loop consisting
1555 // of both blocks and (immediate child) loops.
1556 for (uint i = 0; i < number_of_blocks(); i++) {
1557 Block* block = get_block(i);
1558 CFGLoop* lp = block->_loop;
1559 if (lp == NULL) {
1560 // Not assigned to a loop. Add it to the method's pseudo loop.
1561 block->_loop = root_loop;
1562 lp = root_loop;
1563 }
1564 if (lp == root_loop || block != lp->head()) { // loop heads are already members
1565 lp->add_member(block);
1566 }
1567 if (lp != root_loop) {
1568 if (lp->parent() == NULL) {
1569 // Not a nested loop. Make it a child of the method's pseudo loop.
1570 root_loop->add_nested_loop(lp);
1571 }
1572 if (block == lp->head()) {
1573 // Add nested loop to member list of parent loop.
1574 lp->parent()->add_member(lp);
1575 }
1576 }
1577 }
1578
1579 return root_loop;
1580 }
1581
1582 //------------------------------push_pred--------------------------------------
push_pred(Block * blk,int i,Block_List & worklist,PhaseCFG * cfg)1583 void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg) {
1584 Node* pred_n = blk->pred(i);
1585 Block* pred = cfg->get_block_for_node(pred_n);
1586 CFGLoop *pred_loop = pred->_loop;
1587 if (pred_loop == NULL) {
1588 // Filter out blocks for non-single-entry loops.
1589 // For all reasonable loops, the head occurs before the tail in RPO.
1590 if (pred->_rpo > head()->_rpo) {
1591 pred->_loop = this;
1592 worklist.push(pred);
1593 }
1594 } else if (pred_loop != this) {
1595 // Nested loop.
1596 while (pred_loop->_parent != NULL && pred_loop->_parent != this) {
1597 pred_loop = pred_loop->_parent;
1598 }
1599 // Make pred's loop be a child
1600 if (pred_loop->_parent == NULL) {
1601 add_nested_loop(pred_loop);
1602 // Continue with loop entry predecessor.
1603 Block* pred_head = pred_loop->head();
1604 assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1605 assert(pred_head != head(), "loop head in only one loop");
1606 push_pred(pred_head, LoopNode::EntryControl, worklist, cfg);
1607 } else {
1608 assert(pred_loop->_parent == this && _parent == NULL, "just checking");
1609 }
1610 }
1611 }
1612
1613 //------------------------------add_nested_loop--------------------------------
1614 // Make cl a child of the current loop in the loop tree.
add_nested_loop(CFGLoop * cl)1615 void CFGLoop::add_nested_loop(CFGLoop* cl) {
1616 assert(_parent == NULL, "no parent yet");
1617 assert(cl != this, "not my own parent");
1618 cl->_parent = this;
1619 CFGLoop* ch = _child;
1620 if (ch == NULL) {
1621 _child = cl;
1622 } else {
1623 while (ch->_sibling != NULL) { ch = ch->_sibling; }
1624 ch->_sibling = cl;
1625 }
1626 }
1627
1628 //------------------------------compute_loop_depth-----------------------------
1629 // Store the loop depth in each CFGLoop object.
1630 // Recursively walk the children to do the same for them.
compute_loop_depth(int depth)1631 void CFGLoop::compute_loop_depth(int depth) {
1632 _depth = depth;
1633 CFGLoop* ch = _child;
1634 while (ch != NULL) {
1635 ch->compute_loop_depth(depth + 1);
1636 ch = ch->_sibling;
1637 }
1638 }
1639
1640 //------------------------------compute_freq-----------------------------------
1641 // Compute the frequency of each block and loop, relative to a single entry
1642 // into the dominating loop head.
compute_freq()1643 void CFGLoop::compute_freq() {
1644 // Bottom up traversal of loop tree (visit inner loops first.)
1645 // Set loop head frequency to 1.0, then transitively
1646 // compute frequency for all successors in the loop,
1647 // as well as for each exit edge. Inner loops are
1648 // treated as single blocks with loop exit targets
1649 // as the successor blocks.
1650
1651 // Nested loops first
1652 CFGLoop* ch = _child;
1653 while (ch != NULL) {
1654 ch->compute_freq();
1655 ch = ch->_sibling;
1656 }
1657 assert (_members.length() > 0, "no empty loops");
1658 Block* hd = head();
1659 hd->_freq = 1.0f;
1660 for (int i = 0; i < _members.length(); i++) {
1661 CFGElement* s = _members.at(i);
1662 float freq = s->_freq;
1663 if (s->is_block()) {
1664 Block* b = s->as_Block();
1665 for (uint j = 0; j < b->_num_succs; j++) {
1666 Block* sb = b->_succs[j];
1667 update_succ_freq(sb, freq * b->succ_prob(j));
1668 }
1669 } else {
1670 CFGLoop* lp = s->as_CFGLoop();
1671 assert(lp->_parent == this, "immediate child");
1672 for (int k = 0; k < lp->_exits.length(); k++) {
1673 Block* eb = lp->_exits.at(k).get_target();
1674 float prob = lp->_exits.at(k).get_prob();
1675 update_succ_freq(eb, freq * prob);
1676 }
1677 }
1678 }
1679
1680 // For all loops other than the outer, "method" loop,
1681 // sum and normalize the exit probability. The "method" loop
1682 // should keep the initial exit probability of 1, so that
1683 // inner blocks do not get erroneously scaled.
1684 if (_depth != 0) {
1685 // Total the exit probabilities for this loop.
1686 float exits_sum = 0.0f;
1687 for (int i = 0; i < _exits.length(); i++) {
1688 exits_sum += _exits.at(i).get_prob();
1689 }
1690
1691 // Normalize the exit probabilities. Until now, the
1692 // probabilities estimate the possibility of exit per
1693 // a single loop iteration; afterward, they estimate
1694 // the probability of exit per loop entry.
1695 for (int i = 0; i < _exits.length(); i++) {
1696 Block* et = _exits.at(i).get_target();
1697 float new_prob = 0.0f;
1698 if (_exits.at(i).get_prob() > 0.0f) {
1699 new_prob = _exits.at(i).get_prob() / exits_sum;
1700 }
1701 BlockProbPair bpp(et, new_prob);
1702 _exits.at_put(i, bpp);
1703 }
1704
1705 // Save the total, but guard against unreasonable probability,
1706 // as the value is used to estimate the loop trip count.
1707 // An infinite trip count would blur relative block
1708 // frequencies.
1709 if (exits_sum > 1.0f) exits_sum = 1.0;
1710 if (exits_sum < PROB_MIN) exits_sum = PROB_MIN;
1711 _exit_prob = exits_sum;
1712 }
1713 }
1714
1715 //------------------------------succ_prob-------------------------------------
1716 // Determine the probability of reaching successor 'i' from the receiver block.
succ_prob(uint i)1717 float Block::succ_prob(uint i) {
1718 int eidx = end_idx();
1719 Node *n = get_node(eidx); // Get ending Node
1720
1721 int op = n->Opcode();
1722 if (n->is_Mach()) {
1723 if (n->is_MachNullCheck()) {
1724 // Can only reach here if called after lcm. The original Op_If is gone,
1725 // so we attempt to infer the probability from one or both of the
1726 // successor blocks.
1727 assert(_num_succs == 2, "expecting 2 successors of a null check");
1728 // If either successor has only one predecessor, then the
1729 // probability estimate can be derived using the
1730 // relative frequency of the successor and this block.
1731 if (_succs[i]->num_preds() == 2) {
1732 return _succs[i]->_freq / _freq;
1733 } else if (_succs[1-i]->num_preds() == 2) {
1734 return 1 - (_succs[1-i]->_freq / _freq);
1735 } else {
1736 // Estimate using both successor frequencies
1737 float freq = _succs[i]->_freq;
1738 return freq / (freq + _succs[1-i]->_freq);
1739 }
1740 }
1741 op = n->as_Mach()->ideal_Opcode();
1742 }
1743
1744
1745 // Switch on branch type
1746 switch( op ) {
1747 case Op_CountedLoopEnd:
1748 case Op_If: {
1749 assert (i < 2, "just checking");
1750 // Conditionals pass on only part of their frequency
1751 float prob = n->as_MachIf()->_prob;
1752 assert(prob >= 0.0 && prob <= 1.0, "out of range probability");
1753 // If succ[i] is the FALSE branch, invert path info
1754 if( get_node(i + eidx + 1)->Opcode() == Op_IfFalse ) {
1755 return 1.0f - prob; // not taken
1756 } else {
1757 return prob; // taken
1758 }
1759 }
1760
1761 case Op_Jump:
1762 // Divide the frequency between all successors evenly
1763 return 1.0f/_num_succs;
1764
1765 case Op_Catch: {
1766 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
1767 if (ci->_con == CatchProjNode::fall_through_index) {
1768 // Fall-thru path gets the lion's share.
1769 return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs;
1770 } else {
1771 // Presume exceptional paths are equally unlikely
1772 return PROB_UNLIKELY_MAG(5);
1773 }
1774 }
1775
1776 case Op_Root:
1777 case Op_Goto:
1778 // Pass frequency straight thru to target
1779 return 1.0f;
1780
1781 case Op_NeverBranch:
1782 return 0.0f;
1783
1784 case Op_TailCall:
1785 case Op_TailJump:
1786 case Op_Return:
1787 case Op_Halt:
1788 case Op_Rethrow:
1789 // Do not push out freq to root block
1790 return 0.0f;
1791
1792 default:
1793 ShouldNotReachHere();
1794 }
1795
1796 return 0.0f;
1797 }
1798
1799 //------------------------------num_fall_throughs-----------------------------
1800 // Return the number of fall-through candidates for a block
num_fall_throughs()1801 int Block::num_fall_throughs() {
1802 int eidx = end_idx();
1803 Node *n = get_node(eidx); // Get ending Node
1804
1805 int op = n->Opcode();
1806 if (n->is_Mach()) {
1807 if (n->is_MachNullCheck()) {
1808 // In theory, either side can fall-thru, for simplicity sake,
1809 // let's say only the false branch can now.
1810 return 1;
1811 }
1812 op = n->as_Mach()->ideal_Opcode();
1813 }
1814
1815 // Switch on branch type
1816 switch( op ) {
1817 case Op_CountedLoopEnd:
1818 case Op_If:
1819 return 2;
1820
1821 case Op_Root:
1822 case Op_Goto:
1823 return 1;
1824
1825 case Op_Catch: {
1826 for (uint i = 0; i < _num_succs; i++) {
1827 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
1828 if (ci->_con == CatchProjNode::fall_through_index) {
1829 return 1;
1830 }
1831 }
1832 return 0;
1833 }
1834
1835 case Op_Jump:
1836 case Op_NeverBranch:
1837 case Op_TailCall:
1838 case Op_TailJump:
1839 case Op_Return:
1840 case Op_Halt:
1841 case Op_Rethrow:
1842 return 0;
1843
1844 default:
1845 ShouldNotReachHere();
1846 }
1847
1848 return 0;
1849 }
1850
1851 //------------------------------succ_fall_through-----------------------------
1852 // Return true if a specific successor could be fall-through target.
succ_fall_through(uint i)1853 bool Block::succ_fall_through(uint i) {
1854 int eidx = end_idx();
1855 Node *n = get_node(eidx); // Get ending Node
1856
1857 int op = n->Opcode();
1858 if (n->is_Mach()) {
1859 if (n->is_MachNullCheck()) {
1860 // In theory, either side can fall-thru, for simplicity sake,
1861 // let's say only the false branch can now.
1862 return get_node(i + eidx + 1)->Opcode() == Op_IfFalse;
1863 }
1864 op = n->as_Mach()->ideal_Opcode();
1865 }
1866
1867 // Switch on branch type
1868 switch( op ) {
1869 case Op_CountedLoopEnd:
1870 case Op_If:
1871 case Op_Root:
1872 case Op_Goto:
1873 return true;
1874
1875 case Op_Catch: {
1876 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
1877 return ci->_con == CatchProjNode::fall_through_index;
1878 }
1879
1880 case Op_Jump:
1881 case Op_NeverBranch:
1882 case Op_TailCall:
1883 case Op_TailJump:
1884 case Op_Return:
1885 case Op_Halt:
1886 case Op_Rethrow:
1887 return false;
1888
1889 default:
1890 ShouldNotReachHere();
1891 }
1892
1893 return false;
1894 }
1895
1896 //------------------------------update_uncommon_branch------------------------
1897 // Update the probability of a two-branch to be uncommon
update_uncommon_branch(Block * ub)1898 void Block::update_uncommon_branch(Block* ub) {
1899 int eidx = end_idx();
1900 Node *n = get_node(eidx); // Get ending Node
1901
1902 int op = n->as_Mach()->ideal_Opcode();
1903
1904 assert(op == Op_CountedLoopEnd || op == Op_If, "must be a If");
1905 assert(num_fall_throughs() == 2, "must be a two way branch block");
1906
1907 // Which successor is ub?
1908 uint s;
1909 for (s = 0; s <_num_succs; s++) {
1910 if (_succs[s] == ub) break;
1911 }
1912 assert(s < 2, "uncommon successor must be found");
1913
1914 // If ub is the true path, make the proability small, else
1915 // ub is the false path, and make the probability large
1916 bool invert = (get_node(s + eidx + 1)->Opcode() == Op_IfFalse);
1917
1918 // Get existing probability
1919 float p = n->as_MachIf()->_prob;
1920
1921 if (invert) p = 1.0 - p;
1922 if (p > PROB_MIN) {
1923 p = PROB_MIN;
1924 }
1925 if (invert) p = 1.0 - p;
1926
1927 n->as_MachIf()->_prob = p;
1928 }
1929
1930 //------------------------------update_succ_freq-------------------------------
1931 // Update the appropriate frequency associated with block 'b', a successor of
1932 // a block in this loop.
update_succ_freq(Block * b,float freq)1933 void CFGLoop::update_succ_freq(Block* b, float freq) {
1934 if (b->_loop == this) {
1935 if (b == head()) {
1936 // back branch within the loop
1937 // Do nothing now, the loop carried frequency will be
1938 // adjust later in scale_freq().
1939 } else {
1940 // simple branch within the loop
1941 b->_freq += freq;
1942 }
1943 } else if (!in_loop_nest(b)) {
1944 // branch is exit from this loop
1945 BlockProbPair bpp(b, freq);
1946 _exits.append(bpp);
1947 } else {
1948 // branch into nested loop
1949 CFGLoop* ch = b->_loop;
1950 ch->_freq += freq;
1951 }
1952 }
1953
1954 //------------------------------in_loop_nest-----------------------------------
1955 // Determine if block b is in the receiver's loop nest.
in_loop_nest(Block * b)1956 bool CFGLoop::in_loop_nest(Block* b) {
1957 int depth = _depth;
1958 CFGLoop* b_loop = b->_loop;
1959 int b_depth = b_loop->_depth;
1960 if (depth == b_depth) {
1961 return true;
1962 }
1963 while (b_depth > depth) {
1964 b_loop = b_loop->_parent;
1965 b_depth = b_loop->_depth;
1966 }
1967 return b_loop == this;
1968 }
1969
1970 //------------------------------scale_freq-------------------------------------
1971 // Scale frequency of loops and blocks by trip counts from outer loops
1972 // Do a top down traversal of loop tree (visit outer loops first.)
scale_freq()1973 void CFGLoop::scale_freq() {
1974 float loop_freq = _freq * trip_count();
1975 _freq = loop_freq;
1976 for (int i = 0; i < _members.length(); i++) {
1977 CFGElement* s = _members.at(i);
1978 float block_freq = s->_freq * loop_freq;
1979 if (g_isnan(block_freq) || block_freq < MIN_BLOCK_FREQUENCY)
1980 block_freq = MIN_BLOCK_FREQUENCY;
1981 s->_freq = block_freq;
1982 }
1983 CFGLoop* ch = _child;
1984 while (ch != NULL) {
1985 ch->scale_freq();
1986 ch = ch->_sibling;
1987 }
1988 }
1989
1990 // Frequency of outer loop
outer_loop_freq() const1991 float CFGLoop::outer_loop_freq() const {
1992 if (_child != NULL) {
1993 return _child->_freq;
1994 }
1995 return _freq;
1996 }
1997
1998 #ifndef PRODUCT
1999 //------------------------------dump_tree--------------------------------------
dump_tree() const2000 void CFGLoop::dump_tree() const {
2001 dump();
2002 if (_child != NULL) _child->dump_tree();
2003 if (_sibling != NULL) _sibling->dump_tree();
2004 }
2005
2006 //------------------------------dump-------------------------------------------
dump() const2007 void CFGLoop::dump() const {
2008 for (int i = 0; i < _depth; i++) tty->print(" ");
2009 tty->print("%s: %d trip_count: %6.0f freq: %6.0f\n",
2010 _depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq);
2011 for (int i = 0; i < _depth; i++) tty->print(" ");
2012 tty->print(" members:");
2013 int k = 0;
2014 for (int i = 0; i < _members.length(); i++) {
2015 if (k++ >= 6) {
2016 tty->print("\n ");
2017 for (int j = 0; j < _depth+1; j++) tty->print(" ");
2018 k = 0;
2019 }
2020 CFGElement *s = _members.at(i);
2021 if (s->is_block()) {
2022 Block *b = s->as_Block();
2023 tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq);
2024 } else {
2025 CFGLoop* lp = s->as_CFGLoop();
2026 tty->print(" L%d(%6.3f)", lp->_id, lp->_freq);
2027 }
2028 }
2029 tty->print("\n");
2030 for (int i = 0; i < _depth; i++) tty->print(" ");
2031 tty->print(" exits: ");
2032 k = 0;
2033 for (int i = 0; i < _exits.length(); i++) {
2034 if (k++ >= 7) {
2035 tty->print("\n ");
2036 for (int j = 0; j < _depth+1; j++) tty->print(" ");
2037 k = 0;
2038 }
2039 Block *blk = _exits.at(i).get_target();
2040 float prob = _exits.at(i).get_prob();
2041 tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100));
2042 }
2043 tty->print("\n");
2044 }
2045 #endif
2046