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 "memory/resourceArea.hpp"
29 #include "opto/block.hpp"
30 #include "opto/c2compiler.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/machnode.hpp"
34 #include "opto/opcodes.hpp"
35 #include "opto/phaseX.hpp"
36 #include "opto/rootnode.hpp"
37 #include "opto/runtime.hpp"
38 #include "opto/chaitin.hpp"
39 #include "runtime/deoptimization.hpp"
40
41 // Portions of code courtesy of Clifford Click
42
43 // Optimization - Graph Style
44
45 // To avoid float value underflow
46 #define MIN_BLOCK_FREQUENCY 1.e-35f
47
48 //----------------------------schedule_node_into_block-------------------------
49 // Insert node n into block b. Look for projections of n and make sure they
50 // are in b also.
schedule_node_into_block(Node * n,Block * b)51 void PhaseCFG::schedule_node_into_block( Node *n, Block *b ) {
52 // Set basic block of n, Add n to b,
53 map_node_to_block(n, b);
54 b->add_inst(n);
55
56 // After Matching, nearly any old Node may have projections trailing it.
57 // These are usually machine-dependent flags. In any case, they might
58 // float to another block below this one. Move them up.
59 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
60 Node* use = n->fast_out(i);
61 if (use->is_Proj()) {
62 Block* buse = get_block_for_node(use);
63 if (buse != b) { // In wrong block?
64 if (buse != NULL) {
65 buse->find_remove(use); // Remove from wrong block
66 }
67 map_node_to_block(use, b);
68 b->add_inst(use);
69 }
70 }
71 }
72 }
73
74 //----------------------------replace_block_proj_ctrl-------------------------
75 // Nodes that have is_block_proj() nodes as their control need to use
76 // the appropriate Region for their actual block as their control since
77 // the projection will be in a predecessor block.
replace_block_proj_ctrl(Node * n)78 void PhaseCFG::replace_block_proj_ctrl( Node *n ) {
79 const Node *in0 = n->in(0);
80 assert(in0 != NULL, "Only control-dependent");
81 const Node *p = in0->is_block_proj();
82 if (p != NULL && p != n) { // Control from a block projection?
83 assert(!n->pinned() || n->is_MachConstantBase(), "only pinned MachConstantBase node is expected here");
84 // Find trailing Region
85 Block *pb = get_block_for_node(in0); // Block-projection already has basic block
86 uint j = 0;
87 if (pb->_num_succs != 1) { // More then 1 successor?
88 // Search for successor
89 uint max = pb->number_of_nodes();
90 assert( max > 1, "" );
91 uint start = max - pb->_num_succs;
92 // Find which output path belongs to projection
93 for (j = start; j < max; j++) {
94 if( pb->get_node(j) == in0 )
95 break;
96 }
97 assert( j < max, "must find" );
98 // Change control to match head of successor basic block
99 j -= start;
100 }
101 n->set_req(0, pb->_succs[j]->head());
102 }
103 }
104
is_dominator(Node * dom_node,Node * node)105 bool PhaseCFG::is_dominator(Node* dom_node, Node* node) {
106 assert(is_CFG(node) && is_CFG(dom_node), "node and dom_node must be CFG nodes");
107 if (dom_node == node) {
108 return true;
109 }
110 Block* d = find_block_for_node(dom_node);
111 Block* n = find_block_for_node(node);
112 assert(n != NULL && d != NULL, "blocks must exist");
113
114 if (d == n) {
115 if (dom_node->is_block_start()) {
116 return true;
117 }
118 if (node->is_block_start()) {
119 return false;
120 }
121 if (dom_node->is_block_proj()) {
122 return false;
123 }
124 if (node->is_block_proj()) {
125 return true;
126 }
127
128 assert(is_control_proj_or_safepoint(node), "node must be control projection or safepoint");
129 assert(is_control_proj_or_safepoint(dom_node), "dom_node must be control projection or safepoint");
130
131 // Neither 'node' nor 'dom_node' is a block start or block projection.
132 // Check if 'dom_node' is above 'node' in the control graph.
133 if (is_dominating_control(dom_node, node)) {
134 return true;
135 }
136
137 #ifdef ASSERT
138 // If 'dom_node' does not dominate 'node' then 'node' has to dominate 'dom_node'
139 if (!is_dominating_control(node, dom_node)) {
140 node->dump();
141 dom_node->dump();
142 assert(false, "neither dom_node nor node dominates the other");
143 }
144 #endif
145
146 return false;
147 }
148 return d->dom_lca(n) == d;
149 }
150
is_CFG(Node * n)151 bool PhaseCFG::is_CFG(Node* n) {
152 return n->is_block_proj() || n->is_block_start() || is_control_proj_or_safepoint(n);
153 }
154
is_control_proj_or_safepoint(Node * n)155 bool PhaseCFG::is_control_proj_or_safepoint(Node* n) {
156 bool result = (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_SafePoint) || (n->is_Proj() && n->as_Proj()->bottom_type() == Type::CONTROL);
157 assert(!result || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_SafePoint)
158 || (n->is_Proj() && n->as_Proj()->_con == 0), "If control projection, it must be projection 0");
159 return result;
160 }
161
find_block_for_node(Node * n)162 Block* PhaseCFG::find_block_for_node(Node* n) {
163 if (n->is_block_start() || n->is_block_proj()) {
164 return get_block_for_node(n);
165 } else {
166 // Walk the control graph up if 'n' is not a block start nor a block projection. In this case 'n' must be
167 // an unmatched control projection or a not yet matched safepoint precedence edge in the middle of a block.
168 assert(is_control_proj_or_safepoint(n), "must be control projection or safepoint");
169 Node* ctrl = n->in(0);
170 while (!ctrl->is_block_start()) {
171 ctrl = ctrl->in(0);
172 }
173 return get_block_for_node(ctrl);
174 }
175 }
176
177 // Walk up the control graph from 'n' and check if 'dom_ctrl' is found.
is_dominating_control(Node * dom_ctrl,Node * n)178 bool PhaseCFG::is_dominating_control(Node* dom_ctrl, Node* n) {
179 Node* ctrl = n->in(0);
180 while (!ctrl->is_block_start()) {
181 if (ctrl == dom_ctrl) {
182 return true;
183 }
184 ctrl = ctrl->in(0);
185 }
186 return false;
187 }
188
189
190 //------------------------------schedule_pinned_nodes--------------------------
191 // Set the basic block for Nodes pinned into blocks
schedule_pinned_nodes(VectorSet & visited)192 void PhaseCFG::schedule_pinned_nodes(VectorSet &visited) {
193 // Allocate node stack of size C->live_nodes()+8 to avoid frequent realloc
194 GrowableArray <Node*> spstack(C->live_nodes() + 8);
195 spstack.push(_root);
196 while (spstack.is_nonempty()) {
197 Node* node = spstack.pop();
198 if (!visited.test_set(node->_idx)) { // Test node and flag it as visited
199 if (node->pinned() && !has_block(node)) { // Pinned? Nail it down!
200 assert(node->in(0), "pinned Node must have Control");
201 // Before setting block replace block_proj control edge
202 replace_block_proj_ctrl(node);
203 Node* input = node->in(0);
204 while (!input->is_block_start()) {
205 input = input->in(0);
206 }
207 Block* block = get_block_for_node(input); // Basic block of controlling input
208 schedule_node_into_block(node, block);
209 }
210
211 // If the node has precedence edges (added when CastPP nodes are
212 // removed in final_graph_reshaping), fix the control of the
213 // node to cover the precedence edges and remove the
214 // dependencies.
215 Node* n = NULL;
216 for (uint i = node->len()-1; i >= node->req(); i--) {
217 Node* m = node->in(i);
218 if (m == NULL) continue;
219
220 // Only process precedence edges that are CFG nodes. Safepoints and control projections can be in the middle of a block
221 if (is_CFG(m)) {
222 node->rm_prec(i);
223 if (n == NULL) {
224 n = m;
225 } else {
226 assert(is_dominator(n, m) || is_dominator(m, n), "one must dominate the other");
227 n = is_dominator(n, m) ? m : n;
228 }
229 } else {
230 assert(node->is_Mach(), "sanity");
231 assert(node->as_Mach()->ideal_Opcode() == Op_StoreCM, "must be StoreCM node");
232 }
233 }
234 if (n != NULL) {
235 assert(node->in(0), "control should have been set");
236 assert(is_dominator(n, node->in(0)) || is_dominator(node->in(0), n), "one must dominate the other");
237 if (!is_dominator(n, node->in(0))) {
238 node->set_req(0, n);
239 }
240 }
241
242 // process all inputs that are non NULL
243 for (int i = node->req()-1; i >= 0; --i) {
244 if (node->in(i) != NULL) {
245 spstack.push(node->in(i));
246 }
247 }
248 }
249 }
250 }
251
252 #ifdef ASSERT
253 // Assert that new input b2 is dominated by all previous inputs.
254 // Check this by by seeing that it is dominated by b1, the deepest
255 // input observed until b2.
assert_dom(Block * b1,Block * b2,Node * n,const PhaseCFG * cfg)256 static void assert_dom(Block* b1, Block* b2, Node* n, const PhaseCFG* cfg) {
257 if (b1 == NULL) return;
258 assert(b1->_dom_depth < b2->_dom_depth, "sanity");
259 Block* tmp = b2;
260 while (tmp != b1 && tmp != NULL) {
261 tmp = tmp->_idom;
262 }
263 if (tmp != b1) {
264 // Detected an unschedulable graph. Print some nice stuff and die.
265 tty->print_cr("!!! Unschedulable graph !!!");
266 for (uint j=0; j<n->len(); j++) { // For all inputs
267 Node* inn = n->in(j); // Get input
268 if (inn == NULL) continue; // Ignore NULL, missing inputs
269 Block* inb = cfg->get_block_for_node(inn);
270 tty->print("B%d idom=B%d depth=%2d ",inb->_pre_order,
271 inb->_idom ? inb->_idom->_pre_order : 0, inb->_dom_depth);
272 inn->dump();
273 }
274 tty->print("Failing node: ");
275 n->dump();
276 assert(false, "unscheduable graph");
277 }
278 }
279 #endif
280
find_deepest_input(Node * n,const PhaseCFG * cfg)281 static Block* find_deepest_input(Node* n, const PhaseCFG* cfg) {
282 // Find the last input dominated by all other inputs.
283 Block* deepb = NULL; // Deepest block so far
284 int deepb_dom_depth = 0;
285 for (uint k = 0; k < n->len(); k++) { // For all inputs
286 Node* inn = n->in(k); // Get input
287 if (inn == NULL) continue; // Ignore NULL, missing inputs
288 Block* inb = cfg->get_block_for_node(inn);
289 assert(inb != NULL, "must already have scheduled this input");
290 if (deepb_dom_depth < (int) inb->_dom_depth) {
291 // The new inb must be dominated by the previous deepb.
292 // The various inputs must be linearly ordered in the dom
293 // tree, or else there will not be a unique deepest block.
294 DEBUG_ONLY(assert_dom(deepb, inb, n, cfg));
295 deepb = inb; // Save deepest block
296 deepb_dom_depth = deepb->_dom_depth;
297 }
298 }
299 assert(deepb != NULL, "must be at least one input to n");
300 return deepb;
301 }
302
303
304 //------------------------------schedule_early---------------------------------
305 // Find the earliest Block any instruction can be placed in. Some instructions
306 // are pinned into Blocks. Unpinned instructions can appear in last block in
307 // which all their inputs occur.
schedule_early(VectorSet & visited,Node_Stack & roots)308 bool PhaseCFG::schedule_early(VectorSet &visited, Node_Stack &roots) {
309 // Allocate stack with enough space to avoid frequent realloc
310 Node_Stack nstack(roots.size() + 8);
311 // _root will be processed among C->top() inputs
312 roots.push(C->top(), 0);
313 visited.set(C->top()->_idx);
314
315 while (roots.size() != 0) {
316 // Use local variables nstack_top_n & nstack_top_i to cache values
317 // on stack's top.
318 Node* parent_node = roots.node();
319 uint input_index = 0;
320 roots.pop();
321
322 while (true) {
323 if (input_index == 0) {
324 // Fixup some control. Constants without control get attached
325 // to root and nodes that use is_block_proj() nodes should be attached
326 // to the region that starts their block.
327 const Node* control_input = parent_node->in(0);
328 if (control_input != NULL) {
329 replace_block_proj_ctrl(parent_node);
330 } else {
331 // Is a constant with NO inputs?
332 if (parent_node->req() == 1) {
333 parent_node->set_req(0, _root);
334 }
335 }
336 }
337
338 // First, visit all inputs and force them to get a block. If an
339 // input is already in a block we quit following inputs (to avoid
340 // cycles). Instead we put that Node on a worklist to be handled
341 // later (since IT'S inputs may not have a block yet).
342
343 // Assume all n's inputs will be processed
344 bool done = true;
345
346 while (input_index < parent_node->len()) {
347 Node* in = parent_node->in(input_index++);
348 if (in == NULL) {
349 continue;
350 }
351
352 int is_visited = visited.test_set(in->_idx);
353 if (!has_block(in)) {
354 if (is_visited) {
355 assert(false, "graph should be schedulable");
356 return false;
357 }
358 // Save parent node and next input's index.
359 nstack.push(parent_node, input_index);
360 // Process current input now.
361 parent_node = in;
362 input_index = 0;
363 // Not all n's inputs processed.
364 done = false;
365 break;
366 } else if (!is_visited) {
367 // Visit this guy later, using worklist
368 roots.push(in, 0);
369 }
370 }
371
372 if (done) {
373 // All of n's inputs have been processed, complete post-processing.
374
375 // Some instructions are pinned into a block. These include Region,
376 // Phi, Start, Return, and other control-dependent instructions and
377 // any projections which depend on them.
378 if (!parent_node->pinned()) {
379 // Set earliest legal block.
380 Block* earliest_block = find_deepest_input(parent_node, this);
381 map_node_to_block(parent_node, earliest_block);
382 } else {
383 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");
384 }
385
386 if (nstack.is_empty()) {
387 // Finished all nodes on stack.
388 // Process next node on the worklist 'roots'.
389 break;
390 }
391 // Get saved parent node and next input's index.
392 parent_node = nstack.node();
393 input_index = nstack.index();
394 nstack.pop();
395 }
396 }
397 }
398 return true;
399 }
400
401 //------------------------------dom_lca----------------------------------------
402 // Find least common ancestor in dominator tree
403 // LCA is a current notion of LCA, to be raised above 'this'.
404 // As a convenient boundary condition, return 'this' if LCA is NULL.
405 // Find the LCA of those two nodes.
dom_lca(Block * LCA)406 Block* Block::dom_lca(Block* LCA) {
407 if (LCA == NULL || LCA == this) return this;
408
409 Block* anc = this;
410 while (anc->_dom_depth > LCA->_dom_depth)
411 anc = anc->_idom; // Walk up till anc is as high as LCA
412
413 while (LCA->_dom_depth > anc->_dom_depth)
414 LCA = LCA->_idom; // Walk up till LCA is as high as anc
415
416 while (LCA != anc) { // Walk both up till they are the same
417 LCA = LCA->_idom;
418 anc = anc->_idom;
419 }
420
421 return LCA;
422 }
423
424 //--------------------------raise_LCA_above_use--------------------------------
425 // We are placing a definition, and have been given a def->use edge.
426 // The definition must dominate the use, so move the LCA upward in the
427 // dominator tree to dominate the use. If the use is a phi, adjust
428 // 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)429 static Block* raise_LCA_above_use(Block* LCA, Node* use, Node* def, const PhaseCFG* cfg) {
430 Block* buse = cfg->get_block_for_node(use);
431 if (buse == NULL) return LCA; // Unused killing Projs have no use block
432 if (!use->is_Phi()) return buse->dom_lca(LCA);
433 uint pmax = use->req(); // Number of Phi inputs
434 // Why does not this loop just break after finding the matching input to
435 // the Phi? Well...it's like this. I do not have true def-use/use-def
436 // chains. Means I cannot distinguish, from the def-use direction, which
437 // of many use-defs lead from the same use to the same def. That is, this
438 // Phi might have several uses of the same def. Each use appears in a
439 // different predecessor block. But when I enter here, I cannot distinguish
440 // which use-def edge I should find the predecessor block for. So I find
441 // them all. Means I do a little extra work if a Phi uses the same value
442 // more than once.
443 for (uint j=1; j<pmax; j++) { // For all inputs
444 if (use->in(j) == def) { // Found matching input?
445 Block* pred = cfg->get_block_for_node(buse->pred(j));
446 LCA = pred->dom_lca(LCA);
447 }
448 }
449 return LCA;
450 }
451
452 //----------------------------raise_LCA_above_marks----------------------------
453 // Return a new LCA that dominates LCA and any of its marked predecessors.
454 // Search all my parents up to 'early' (exclusive), looking for predecessors
455 // which are marked with the given index. Return the LCA (in the dom tree)
456 // of all marked blocks. If there are none marked, return the original
457 // LCA.
raise_LCA_above_marks(Block * LCA,node_idx_t mark,Block * early,const PhaseCFG * cfg)458 static Block* raise_LCA_above_marks(Block* LCA, node_idx_t mark, Block* early, const PhaseCFG* cfg) {
459 Block_List worklist;
460 worklist.push(LCA);
461 while (worklist.size() > 0) {
462 Block* mid = worklist.pop();
463 if (mid == early) continue; // stop searching here
464
465 // Test and set the visited bit.
466 if (mid->raise_LCA_visited() == mark) continue; // already visited
467
468 // Don't process the current LCA, otherwise the search may terminate early
469 if (mid != LCA && mid->raise_LCA_mark() == mark) {
470 // Raise the LCA.
471 LCA = mid->dom_lca(LCA);
472 if (LCA == early) break; // stop searching everywhere
473 assert(early->dominates(LCA), "early is high enough");
474 // Resume searching at that point, skipping intermediate levels.
475 worklist.push(LCA);
476 if (LCA == mid)
477 continue; // Don't mark as visited to avoid early termination.
478 } else {
479 // Keep searching through this block's predecessors.
480 for (uint j = 1, jmax = mid->num_preds(); j < jmax; j++) {
481 Block* mid_parent = cfg->get_block_for_node(mid->pred(j));
482 worklist.push(mid_parent);
483 }
484 }
485 mid->set_raise_LCA_visited(mark);
486 }
487 return LCA;
488 }
489
490 //--------------------------memory_early_block--------------------------------
491 // This is a variation of find_deepest_input, the heart of schedule_early.
492 // Find the "early" block for a load, if we considered only memory and
493 // address inputs, that is, if other data inputs were ignored.
494 //
495 // Because a subset of edges are considered, the resulting block will
496 // be earlier (at a shallower dom_depth) than the true schedule_early
497 // point of the node. We compute this earlier block as a more permissive
498 // site for anti-dependency insertion, but only if subsume_loads is enabled.
memory_early_block(Node * load,Block * early,const PhaseCFG * cfg)499 static Block* memory_early_block(Node* load, Block* early, const PhaseCFG* cfg) {
500 Node* base;
501 Node* index;
502 Node* store = load->in(MemNode::Memory);
503 load->as_Mach()->memory_inputs(base, index);
504
505 assert(base != NodeSentinel && index != NodeSentinel,
506 "unexpected base/index inputs");
507
508 Node* mem_inputs[4];
509 int mem_inputs_length = 0;
510 if (base != NULL) mem_inputs[mem_inputs_length++] = base;
511 if (index != NULL) mem_inputs[mem_inputs_length++] = index;
512 if (store != NULL) mem_inputs[mem_inputs_length++] = store;
513
514 // In the comparision below, add one to account for the control input,
515 // which may be null, but always takes up a spot in the in array.
516 if (mem_inputs_length + 1 < (int) load->req()) {
517 // This "load" has more inputs than just the memory, base and index inputs.
518 // For purposes of checking anti-dependences, we need to start
519 // from the early block of only the address portion of the instruction,
520 // and ignore other blocks that may have factored into the wider
521 // schedule_early calculation.
522 if (load->in(0) != NULL) mem_inputs[mem_inputs_length++] = load->in(0);
523
524 Block* deepb = NULL; // Deepest block so far
525 int deepb_dom_depth = 0;
526 for (int i = 0; i < mem_inputs_length; i++) {
527 Block* inb = cfg->get_block_for_node(mem_inputs[i]);
528 if (deepb_dom_depth < (int) inb->_dom_depth) {
529 // The new inb must be dominated by the previous deepb.
530 // The various inputs must be linearly ordered in the dom
531 // tree, or else there will not be a unique deepest block.
532 DEBUG_ONLY(assert_dom(deepb, inb, load, cfg));
533 deepb = inb; // Save deepest block
534 deepb_dom_depth = deepb->_dom_depth;
535 }
536 }
537 early = deepb;
538 }
539
540 return early;
541 }
542
543 //--------------------------insert_anti_dependences---------------------------
544 // A load may need to witness memory that nearby stores can overwrite.
545 // For each nearby store, either insert an "anti-dependence" edge
546 // from the load to the store, or else move LCA upward to force the
547 // load to (eventually) be scheduled in a block above the store.
548 //
549 // Do not add edges to stores on distinct control-flow paths;
550 // only add edges to stores which might interfere.
551 //
552 // Return the (updated) LCA. There will not be any possibly interfering
553 // store between the load's "early block" and the updated LCA.
554 // Any stores in the updated LCA will have new precedence edges
555 // back to the load. The caller is expected to schedule the load
556 // in the LCA, in which case the precedence edges will make LCM
557 // preserve anti-dependences. The caller may also hoist the load
558 // above the LCA, if it is not the early block.
insert_anti_dependences(Block * LCA,Node * load,bool verify)559 Block* PhaseCFG::insert_anti_dependences(Block* LCA, Node* load, bool verify) {
560 assert(load->needs_anti_dependence_check(), "must be a load of some sort");
561 assert(LCA != NULL, "");
562 DEBUG_ONLY(Block* LCA_orig = LCA);
563
564 // Compute the alias index. Loads and stores with different alias indices
565 // do not need anti-dependence edges.
566 int load_alias_idx = C->get_alias_index(load->adr_type());
567 #ifdef ASSERT
568 if (load_alias_idx == Compile::AliasIdxBot && C->AliasLevel() > 0 &&
569 (PrintOpto || VerifyAliases ||
570 (PrintMiscellaneous && (WizardMode || Verbose)))) {
571 // Load nodes should not consume all of memory.
572 // Reporting a bottom type indicates a bug in adlc.
573 // If some particular type of node validly consumes all of memory,
574 // sharpen the preceding "if" to exclude it, so we can catch bugs here.
575 tty->print_cr("*** Possible Anti-Dependence Bug: Load consumes all of memory.");
576 load->dump(2);
577 if (VerifyAliases) assert(load_alias_idx != Compile::AliasIdxBot, "");
578 }
579 #endif
580 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrComp),
581 "String compare is only known 'load' that does not conflict with any stores");
582 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrEquals),
583 "String equals is a 'load' that does not conflict with any stores");
584 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrIndexOf),
585 "String indexOf is a 'load' that does not conflict with any stores");
586 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrIndexOfChar),
587 "String indexOfChar is a 'load' that does not conflict with any stores");
588 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_AryEq),
589 "Arrays equals is a 'load' that does not conflict with any stores");
590 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_HasNegatives),
591 "HasNegatives is a 'load' that does not conflict with any stores");
592
593 if (!C->alias_type(load_alias_idx)->is_rewritable()) {
594 // It is impossible to spoil this load by putting stores before it,
595 // because we know that the stores will never update the value
596 // which 'load' must witness.
597 return LCA;
598 }
599
600 node_idx_t load_index = load->_idx;
601
602 // Note the earliest legal placement of 'load', as determined by
603 // by the unique point in the dom tree where all memory effects
604 // and other inputs are first available. (Computed by schedule_early.)
605 // For normal loads, 'early' is the shallowest place (dom graph wise)
606 // to look for anti-deps between this load and any store.
607 Block* early = get_block_for_node(load);
608
609 // If we are subsuming loads, compute an "early" block that only considers
610 // memory or address inputs. This block may be different than the
611 // schedule_early block in that it could be at an even shallower depth in the
612 // dominator tree, and allow for a broader discovery of anti-dependences.
613 if (C->subsume_loads()) {
614 early = memory_early_block(load, early, this);
615 }
616
617 ResourceArea *area = Thread::current()->resource_area();
618 Node_List worklist_mem(area); // prior memory state to store
619 Node_List worklist_store(area); // possible-def to explore
620 Node_List worklist_visited(area); // visited mergemem nodes
621 Node_List non_early_stores(area); // all relevant stores outside of early
622 bool must_raise_LCA = false;
623
624 #ifdef TRACK_PHI_INPUTS
625 // %%% This extra checking fails because MergeMem nodes are not GVNed.
626 // Provide "phi_inputs" to check if every input to a PhiNode is from the
627 // original memory state. This indicates a PhiNode for which should not
628 // prevent the load from sinking. For such a block, set_raise_LCA_mark
629 // may be overly conservative.
630 // Mechanism: count inputs seen for each Phi encountered in worklist_store.
631 DEBUG_ONLY(GrowableArray<uint> phi_inputs(area, C->unique(),0,0));
632 #endif
633
634 // 'load' uses some memory state; look for users of the same state.
635 // Recurse through MergeMem nodes to the stores that use them.
636
637 // Each of these stores is a possible definition of memory
638 // that 'load' needs to use. We need to force 'load'
639 // to occur before each such store. When the store is in
640 // the same block as 'load', we insert an anti-dependence
641 // edge load->store.
642
643 // The relevant stores "nearby" the load consist of a tree rooted
644 // at initial_mem, with internal nodes of type MergeMem.
645 // Therefore, the branches visited by the worklist are of this form:
646 // initial_mem -> (MergeMem ->)* store
647 // The anti-dependence constraints apply only to the fringe of this tree.
648
649 Node* initial_mem = load->in(MemNode::Memory);
650 worklist_store.push(initial_mem);
651 worklist_visited.push(initial_mem);
652 worklist_mem.push(NULL);
653 while (worklist_store.size() > 0) {
654 // Examine a nearby store to see if it might interfere with our load.
655 Node* mem = worklist_mem.pop();
656 Node* store = worklist_store.pop();
657 uint op = store->Opcode();
658
659 // MergeMems do not directly have anti-deps.
660 // Treat them as internal nodes in a forward tree of memory states,
661 // the leaves of which are each a 'possible-def'.
662 if (store == initial_mem // root (exclusive) of tree we are searching
663 || op == Op_MergeMem // internal node of tree we are searching
664 ) {
665 mem = store; // It's not a possibly interfering store.
666 if (store == initial_mem)
667 initial_mem = NULL; // only process initial memory once
668
669 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
670 store = mem->fast_out(i);
671 if (store->is_MergeMem()) {
672 // Be sure we don't get into combinatorial problems.
673 // (Allow phis to be repeated; they can merge two relevant states.)
674 uint j = worklist_visited.size();
675 for (; j > 0; j--) {
676 if (worklist_visited.at(j-1) == store) break;
677 }
678 if (j > 0) continue; // already on work list; do not repeat
679 worklist_visited.push(store);
680 }
681 worklist_mem.push(mem);
682 worklist_store.push(store);
683 }
684 continue;
685 }
686
687 if (op == Op_MachProj || op == Op_Catch) continue;
688 if (store->needs_anti_dependence_check()) continue; // not really a store
689
690 // Compute the alias index. Loads and stores with different alias
691 // indices do not need anti-dependence edges. Wide MemBar's are
692 // anti-dependent on everything (except immutable memories).
693 const TypePtr* adr_type = store->adr_type();
694 if (!C->can_alias(adr_type, load_alias_idx)) continue;
695
696 // Most slow-path runtime calls do NOT modify Java memory, but
697 // they can block and so write Raw memory.
698 if (store->is_Mach()) {
699 MachNode* mstore = store->as_Mach();
700 if (load_alias_idx != Compile::AliasIdxRaw) {
701 // Check for call into the runtime using the Java calling
702 // convention (and from there into a wrapper); it has no
703 // _method. Can't do this optimization for Native calls because
704 // they CAN write to Java memory.
705 if (mstore->ideal_Opcode() == Op_CallStaticJava) {
706 assert(mstore->is_MachSafePoint(), "");
707 MachSafePointNode* ms = (MachSafePointNode*) mstore;
708 assert(ms->is_MachCallJava(), "");
709 MachCallJavaNode* mcj = (MachCallJavaNode*) ms;
710 if (mcj->_method == NULL) {
711 // These runtime calls do not write to Java visible memory
712 // (other than Raw) and so do not require anti-dependence edges.
713 continue;
714 }
715 }
716 // Same for SafePoints: they read/write Raw but only read otherwise.
717 // This is basically a workaround for SafePoints only defining control
718 // instead of control + memory.
719 if (mstore->ideal_Opcode() == Op_SafePoint)
720 continue;
721
722 // Check if the store is a membar on which the load is control dependent.
723 // Inserting an anti-dependency between that membar and the load would
724 // create a cycle that causes local scheduling to fail.
725 if (mstore->isa_MachMemBar()) {
726 Node* dom = load->find_exact_control(load->in(0));
727 while (dom != NULL && dom != dom->in(0) && dom != mstore) {
728 dom = dom->in(0);
729 }
730 if (dom == mstore) {
731 continue;
732 }
733 }
734 } else {
735 // Some raw memory, such as the load of "top" at an allocation,
736 // can be control dependent on the previous safepoint. See
737 // comments in GraphKit::allocate_heap() about control input.
738 // Inserting an anti-dep between such a safepoint and a use
739 // creates a cycle, and will cause a subsequent failure in
740 // local scheduling. (BugId 4919904)
741 // (%%% How can a control input be a safepoint and not a projection??)
742 if (mstore->ideal_Opcode() == Op_SafePoint && load->in(0) == mstore)
743 continue;
744 }
745 }
746
747 // Identify a block that the current load must be above,
748 // or else observe that 'store' is all the way up in the
749 // earliest legal block for 'load'. In the latter case,
750 // immediately insert an anti-dependence edge.
751 Block* store_block = get_block_for_node(store);
752 assert(store_block != NULL, "unused killing projections skipped above");
753
754 if (store->is_Phi()) {
755 // Loop-phis need to raise load before input. (Other phis are treated
756 // as store below.)
757 //
758 // 'load' uses memory which is one (or more) of the Phi's inputs.
759 // It must be scheduled not before the Phi, but rather before
760 // each of the relevant Phi inputs.
761 //
762 // Instead of finding the LCA of all inputs to a Phi that match 'mem',
763 // we mark each corresponding predecessor block and do a combined
764 // hoisting operation later (raise_LCA_above_marks).
765 //
766 // Do not assert(store_block != early, "Phi merging memory after access")
767 // PhiNode may be at start of block 'early' with backedge to 'early'
768 DEBUG_ONLY(bool found_match = false);
769 for (uint j = PhiNode::Input, jmax = store->req(); j < jmax; j++) {
770 if (store->in(j) == mem) { // Found matching input?
771 DEBUG_ONLY(found_match = true);
772 Block* pred_block = get_block_for_node(store_block->pred(j));
773 if (pred_block != early) {
774 // If any predecessor of the Phi matches the load's "early block",
775 // we do not need a precedence edge between the Phi and 'load'
776 // since the load will be forced into a block preceding the Phi.
777 pred_block->set_raise_LCA_mark(load_index);
778 assert(!LCA_orig->dominates(pred_block) ||
779 early->dominates(pred_block), "early is high enough");
780 must_raise_LCA = true;
781 } else {
782 // anti-dependent upon PHI pinned below 'early', no edge needed
783 LCA = early; // but can not schedule below 'early'
784 }
785 }
786 }
787 assert(found_match, "no worklist bug");
788 #ifdef TRACK_PHI_INPUTS
789 #ifdef ASSERT
790 // This assert asks about correct handling of PhiNodes, which may not
791 // have all input edges directly from 'mem'. See BugId 4621264
792 int num_mem_inputs = phi_inputs.at_grow(store->_idx,0) + 1;
793 // Increment by exactly one even if there are multiple copies of 'mem'
794 // coming into the phi, because we will run this block several times
795 // if there are several copies of 'mem'. (That's how DU iterators work.)
796 phi_inputs.at_put(store->_idx, num_mem_inputs);
797 assert(PhiNode::Input + num_mem_inputs < store->req(),
798 "Expect at least one phi input will not be from original memory state");
799 #endif //ASSERT
800 #endif //TRACK_PHI_INPUTS
801 } else if (store_block != early) {
802 // 'store' is between the current LCA and earliest possible block.
803 // Label its block, and decide later on how to raise the LCA
804 // to include the effect on LCA of this store.
805 // If this store's block gets chosen as the raised LCA, we
806 // will find him on the non_early_stores list and stick him
807 // with a precedence edge.
808 // (But, don't bother if LCA is already raised all the way.)
809 if (LCA != early) {
810 store_block->set_raise_LCA_mark(load_index);
811 must_raise_LCA = true;
812 non_early_stores.push(store);
813 }
814 } else {
815 // Found a possibly-interfering store in the load's 'early' block.
816 // This means 'load' cannot sink at all in the dominator tree.
817 // Add an anti-dep edge, and squeeze 'load' into the highest block.
818 assert(store != load->find_exact_control(load->in(0)), "dependence cycle found");
819 if (verify) {
820 assert(store->find_edge(load) != -1, "missing precedence edge");
821 } else {
822 store->add_prec(load);
823 }
824 LCA = early;
825 // This turns off the process of gathering non_early_stores.
826 }
827 }
828 // (Worklist is now empty; all nearby stores have been visited.)
829
830 // Finished if 'load' must be scheduled in its 'early' block.
831 // If we found any stores there, they have already been given
832 // precedence edges.
833 if (LCA == early) return LCA;
834
835 // We get here only if there are no possibly-interfering stores
836 // in the load's 'early' block. Move LCA up above all predecessors
837 // which contain stores we have noted.
838 //
839 // The raised LCA block can be a home to such interfering stores,
840 // but its predecessors must not contain any such stores.
841 //
842 // The raised LCA will be a lower bound for placing the load,
843 // preventing the load from sinking past any block containing
844 // a store that may invalidate the memory state required by 'load'.
845 if (must_raise_LCA)
846 LCA = raise_LCA_above_marks(LCA, load->_idx, early, this);
847 if (LCA == early) return LCA;
848
849 // Insert anti-dependence edges from 'load' to each store
850 // in the non-early LCA block.
851 // Mine the non_early_stores list for such stores.
852 if (LCA->raise_LCA_mark() == load_index) {
853 while (non_early_stores.size() > 0) {
854 Node* store = non_early_stores.pop();
855 Block* store_block = get_block_for_node(store);
856 if (store_block == LCA) {
857 // add anti_dependence from store to load in its own block
858 assert(store != load->find_exact_control(load->in(0)), "dependence cycle found");
859 if (verify) {
860 assert(store->find_edge(load) != -1, "missing precedence edge");
861 } else {
862 store->add_prec(load);
863 }
864 } else {
865 assert(store_block->raise_LCA_mark() == load_index, "block was marked");
866 // Any other stores we found must be either inside the new LCA
867 // or else outside the original LCA. In the latter case, they
868 // did not interfere with any use of 'load'.
869 assert(LCA->dominates(store_block)
870 || !LCA_orig->dominates(store_block), "no stray stores");
871 }
872 }
873 }
874
875 // Return the highest block containing stores; any stores
876 // within that block have been given anti-dependence edges.
877 return LCA;
878 }
879
880 // This class is used to iterate backwards over the nodes in the graph.
881
882 class Node_Backward_Iterator {
883
884 private:
885 Node_Backward_Iterator();
886
887 public:
888 // Constructor for the iterator
889 Node_Backward_Iterator(Node *root, VectorSet &visited, Node_Stack &stack, PhaseCFG &cfg);
890
891 // Postincrement operator to iterate over the nodes
892 Node *next();
893
894 private:
895 VectorSet &_visited;
896 Node_Stack &_stack;
897 PhaseCFG &_cfg;
898 };
899
900 // Constructor for the Node_Backward_Iterator
Node_Backward_Iterator(Node * root,VectorSet & visited,Node_Stack & stack,PhaseCFG & cfg)901 Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_Stack &stack, PhaseCFG &cfg)
902 : _visited(visited), _stack(stack), _cfg(cfg) {
903 // The stack should contain exactly the root
904 stack.clear();
905 stack.push(root, root->outcnt());
906
907 // Clear the visited bits
908 visited.Clear();
909 }
910
911 // Iterator for the Node_Backward_Iterator
next()912 Node *Node_Backward_Iterator::next() {
913
914 // If the _stack is empty, then just return NULL: finished.
915 if ( !_stack.size() )
916 return NULL;
917
918 // I visit unvisited not-anti-dependence users first, then anti-dependent
919 // children next. I iterate backwards to support removal of nodes.
920 // The stack holds states consisting of 3 values:
921 // current Def node, flag which indicates 1st/2nd pass, index of current out edge
922 Node *self = (Node*)(((uintptr_t)_stack.node()) & ~1);
923 bool iterate_anti_dep = (((uintptr_t)_stack.node()) & 1);
924 uint idx = MIN2(_stack.index(), self->outcnt()); // Support removal of nodes.
925 _stack.pop();
926
927 // I cycle here when I am entering a deeper level of recursion.
928 // The key variable 'self' was set prior to jumping here.
929 while( 1 ) {
930
931 _visited.set(self->_idx);
932
933 // Now schedule all uses as late as possible.
934 const Node* src = self->is_Proj() ? self->in(0) : self;
935 uint src_rpo = _cfg.get_block_for_node(src)->_rpo;
936
937 // Schedule all nodes in a post-order visit
938 Node *unvisited = NULL; // Unvisited anti-dependent Node, if any
939
940 // Scan for unvisited nodes
941 while (idx > 0) {
942 // For all uses, schedule late
943 Node* n = self->raw_out(--idx); // Use
944
945 // Skip already visited children
946 if ( _visited.test(n->_idx) )
947 continue;
948
949 // do not traverse backward control edges
950 Node *use = n->is_Proj() ? n->in(0) : n;
951 uint use_rpo = _cfg.get_block_for_node(use)->_rpo;
952
953 if ( use_rpo < src_rpo )
954 continue;
955
956 // Phi nodes always precede uses in a basic block
957 if ( use_rpo == src_rpo && use->is_Phi() )
958 continue;
959
960 unvisited = n; // Found unvisited
961
962 // Check for possible-anti-dependent
963 // 1st pass: No such nodes, 2nd pass: Only such nodes.
964 if (n->needs_anti_dependence_check() == iterate_anti_dep) {
965 unvisited = n; // Found unvisited
966 break;
967 }
968 }
969
970 // Did I find an unvisited not-anti-dependent Node?
971 if (!unvisited) {
972 if (!iterate_anti_dep) {
973 // 2nd pass: Iterate over nodes which needs_anti_dependence_check.
974 iterate_anti_dep = true;
975 idx = self->outcnt();
976 continue;
977 }
978 break; // All done with children; post-visit 'self'
979 }
980
981 // Visit the unvisited Node. Contains the obvious push to
982 // indicate I'm entering a deeper level of recursion. I push the
983 // old state onto the _stack and set a new state and loop (recurse).
984 _stack.push((Node*)((uintptr_t)self | (uintptr_t)iterate_anti_dep), idx);
985 self = unvisited;
986 iterate_anti_dep = false;
987 idx = self->outcnt();
988 } // End recursion loop
989
990 return self;
991 }
992
993 //------------------------------ComputeLatenciesBackwards----------------------
994 // Compute the latency of all the instructions.
compute_latencies_backwards(VectorSet & visited,Node_Stack & stack)995 void PhaseCFG::compute_latencies_backwards(VectorSet &visited, Node_Stack &stack) {
996 #ifndef PRODUCT
997 if (trace_opto_pipelining())
998 tty->print("\n#---- ComputeLatenciesBackwards ----\n");
999 #endif
1000
1001 Node_Backward_Iterator iter((Node *)_root, visited, stack, *this);
1002 Node *n;
1003
1004 // Walk over all the nodes from last to first
1005 while ((n = iter.next())) {
1006 // Set the latency for the definitions of this instruction
1007 partial_latency_of_defs(n);
1008 }
1009 } // end ComputeLatenciesBackwards
1010
1011 //------------------------------partial_latency_of_defs------------------------
1012 // Compute the latency impact of this node on all defs. This computes
1013 // a number that increases as we approach the beginning of the routine.
partial_latency_of_defs(Node * n)1014 void PhaseCFG::partial_latency_of_defs(Node *n) {
1015 // Set the latency for this instruction
1016 #ifndef PRODUCT
1017 if (trace_opto_pipelining()) {
1018 tty->print("# latency_to_inputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n));
1019 dump();
1020 }
1021 #endif
1022
1023 if (n->is_Proj()) {
1024 n = n->in(0);
1025 }
1026
1027 if (n->is_Root()) {
1028 return;
1029 }
1030
1031 uint nlen = n->len();
1032 uint use_latency = get_latency_for_node(n);
1033 uint use_pre_order = get_block_for_node(n)->_pre_order;
1034
1035 for (uint j = 0; j < nlen; j++) {
1036 Node *def = n->in(j);
1037
1038 if (!def || def == n) {
1039 continue;
1040 }
1041
1042 // Walk backwards thru projections
1043 if (def->is_Proj()) {
1044 def = def->in(0);
1045 }
1046
1047 #ifndef PRODUCT
1048 if (trace_opto_pipelining()) {
1049 tty->print("# in(%2d): ", j);
1050 def->dump();
1051 }
1052 #endif
1053
1054 // If the defining block is not known, assume it is ok
1055 Block *def_block = get_block_for_node(def);
1056 uint def_pre_order = def_block ? def_block->_pre_order : 0;
1057
1058 if ((use_pre_order < def_pre_order) || (use_pre_order == def_pre_order && n->is_Phi())) {
1059 continue;
1060 }
1061
1062 uint delta_latency = n->latency(j);
1063 uint current_latency = delta_latency + use_latency;
1064
1065 if (get_latency_for_node(def) < current_latency) {
1066 set_latency_for_node(def, current_latency);
1067 }
1068
1069 #ifndef PRODUCT
1070 if (trace_opto_pipelining()) {
1071 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));
1072 }
1073 #endif
1074 }
1075 }
1076
1077 //------------------------------latency_from_use-------------------------------
1078 // Compute the latency of a specific use
latency_from_use(Node * n,const Node * def,Node * use)1079 int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) {
1080 // If self-reference, return no latency
1081 if (use == n || use->is_Root()) {
1082 return 0;
1083 }
1084
1085 uint def_pre_order = get_block_for_node(def)->_pre_order;
1086 uint latency = 0;
1087
1088 // If the use is not a projection, then it is simple...
1089 if (!use->is_Proj()) {
1090 #ifndef PRODUCT
1091 if (trace_opto_pipelining()) {
1092 tty->print("# out(): ");
1093 use->dump();
1094 }
1095 #endif
1096
1097 uint use_pre_order = get_block_for_node(use)->_pre_order;
1098
1099 if (use_pre_order < def_pre_order)
1100 return 0;
1101
1102 if (use_pre_order == def_pre_order && use->is_Phi())
1103 return 0;
1104
1105 uint nlen = use->len();
1106 uint nl = get_latency_for_node(use);
1107
1108 for ( uint j=0; j<nlen; j++ ) {
1109 if (use->in(j) == n) {
1110 // Change this if we want local latencies
1111 uint ul = use->latency(j);
1112 uint l = ul + nl;
1113 if (latency < l) latency = l;
1114 #ifndef PRODUCT
1115 if (trace_opto_pipelining()) {
1116 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, latency = %d",
1117 nl, j, ul, l, latency);
1118 }
1119 #endif
1120 }
1121 }
1122 } else {
1123 // This is a projection, just grab the latency of the use(s)
1124 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
1125 uint l = latency_from_use(use, def, use->fast_out(j));
1126 if (latency < l) latency = l;
1127 }
1128 }
1129
1130 return latency;
1131 }
1132
1133 //------------------------------latency_from_uses------------------------------
1134 // Compute the latency of this instruction relative to all of it's uses.
1135 // This computes a number that increases as we approach the beginning of the
1136 // routine.
latency_from_uses(Node * n)1137 void PhaseCFG::latency_from_uses(Node *n) {
1138 // Set the latency for this instruction
1139 #ifndef PRODUCT
1140 if (trace_opto_pipelining()) {
1141 tty->print("# latency_from_outputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n));
1142 dump();
1143 }
1144 #endif
1145 uint latency=0;
1146 const Node *def = n->is_Proj() ? n->in(0): n;
1147
1148 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1149 uint l = latency_from_use(n, def, n->fast_out(i));
1150
1151 if (latency < l) latency = l;
1152 }
1153
1154 set_latency_for_node(n, latency);
1155 }
1156
1157 //------------------------------hoist_to_cheaper_block-------------------------
1158 // Pick a block for node self, between early and LCA, that is a cheaper
1159 // alternative to LCA.
hoist_to_cheaper_block(Block * LCA,Block * early,Node * self)1160 Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) {
1161 const double delta = 1+PROB_UNLIKELY_MAG(4);
1162 Block* least = LCA;
1163 double least_freq = least->_freq;
1164 uint target = get_latency_for_node(self);
1165 uint start_latency = get_latency_for_node(LCA->head());
1166 uint end_latency = get_latency_for_node(LCA->get_node(LCA->end_idx()));
1167 bool in_latency = (target <= start_latency);
1168 const Block* root_block = get_block_for_node(_root);
1169
1170 // Turn off latency scheduling if scheduling is just plain off
1171 if (!C->do_scheduling())
1172 in_latency = true;
1173
1174 // Do not hoist (to cover latency) instructions which target a
1175 // single register. Hoisting stretches the live range of the
1176 // single register and may force spilling.
1177 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1178 if (mach && mach->out_RegMask().is_bound1() && mach->out_RegMask().is_NotEmpty())
1179 in_latency = true;
1180
1181 #ifndef PRODUCT
1182 if (trace_opto_pipelining()) {
1183 tty->print("# Find cheaper block for latency %d: ", get_latency_for_node(self));
1184 self->dump();
1185 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1186 LCA->_pre_order,
1187 LCA->head()->_idx,
1188 start_latency,
1189 LCA->get_node(LCA->end_idx())->_idx,
1190 end_latency,
1191 least_freq);
1192 }
1193 #endif
1194
1195 int cand_cnt = 0; // number of candidates tried
1196
1197 // Walk up the dominator tree from LCA (Lowest common ancestor) to
1198 // the earliest legal location. Capture the least execution frequency.
1199 while (LCA != early) {
1200 LCA = LCA->_idom; // Follow up the dominator tree
1201
1202 if (LCA == NULL) {
1203 // Bailout without retry
1204 assert(false, "graph should be schedulable");
1205 C->record_method_not_compilable("late schedule failed: LCA == NULL");
1206 return least;
1207 }
1208
1209 // Don't hoist machine instructions to the root basic block
1210 if (mach && LCA == root_block)
1211 break;
1212
1213 uint start_lat = get_latency_for_node(LCA->head());
1214 uint end_idx = LCA->end_idx();
1215 uint end_lat = get_latency_for_node(LCA->get_node(end_idx));
1216 double LCA_freq = LCA->_freq;
1217 #ifndef PRODUCT
1218 if (trace_opto_pipelining()) {
1219 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1220 LCA->_pre_order, LCA->head()->_idx, start_lat, end_idx, end_lat, LCA_freq);
1221 }
1222 #endif
1223 cand_cnt++;
1224 if (LCA_freq < least_freq || // Better Frequency
1225 (StressGCM && Compile::randomized_select(cand_cnt)) || // Should be randomly accepted in stress mode
1226 (!StressGCM && // Otherwise, choose with latency
1227 !in_latency && // No block containing latency
1228 LCA_freq < least_freq * delta && // No worse frequency
1229 target >= end_lat && // within latency range
1230 !self->is_iteratively_computed() ) // But don't hoist IV increments
1231 // because they may end up above other uses of their phi forcing
1232 // their result register to be different from their input.
1233 ) {
1234 least = LCA; // Found cheaper block
1235 least_freq = LCA_freq;
1236 start_latency = start_lat;
1237 end_latency = end_lat;
1238 if (target <= start_lat)
1239 in_latency = true;
1240 }
1241 }
1242
1243 #ifndef PRODUCT
1244 if (trace_opto_pipelining()) {
1245 tty->print_cr("# Choose block B%d with start latency=%d and freq=%g",
1246 least->_pre_order, start_latency, least_freq);
1247 }
1248 #endif
1249
1250 // See if the latency needs to be updated
1251 if (target < end_latency) {
1252 #ifndef PRODUCT
1253 if (trace_opto_pipelining()) {
1254 tty->print_cr("# Change latency for [%4d] from %d to %d", self->_idx, target, end_latency);
1255 }
1256 #endif
1257 set_latency_for_node(self, end_latency);
1258 partial_latency_of_defs(self);
1259 }
1260
1261 return least;
1262 }
1263
1264
1265 //------------------------------schedule_late-----------------------------------
1266 // Now schedule all codes as LATE as possible. This is the LCA in the
1267 // dominator tree of all USES of a value. Pick the block with the least
1268 // loop nesting depth that is lowest in the dominator tree.
1269 extern const char must_clone[];
schedule_late(VectorSet & visited,Node_Stack & stack)1270 void PhaseCFG::schedule_late(VectorSet &visited, Node_Stack &stack) {
1271 #ifndef PRODUCT
1272 if (trace_opto_pipelining())
1273 tty->print("\n#---- schedule_late ----\n");
1274 #endif
1275
1276 Node_Backward_Iterator iter((Node *)_root, visited, stack, *this);
1277 Node *self;
1278
1279 // Walk over all the nodes from last to first
1280 while ((self = iter.next())) {
1281 Block* early = get_block_for_node(self); // Earliest legal placement
1282
1283 if (self->is_top()) {
1284 // Top node goes in bb #2 with other constants.
1285 // It must be special-cased, because it has no out edges.
1286 early->add_inst(self);
1287 continue;
1288 }
1289
1290 // No uses, just terminate
1291 if (self->outcnt() == 0) {
1292 assert(self->is_MachProj(), "sanity");
1293 continue; // Must be a dead machine projection
1294 }
1295
1296 // If node is pinned in the block, then no scheduling can be done.
1297 if( self->pinned() ) // Pinned in block?
1298 continue;
1299
1300 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1301 if (mach) {
1302 switch (mach->ideal_Opcode()) {
1303 case Op_CreateEx:
1304 // Don't move exception creation
1305 early->add_inst(self);
1306 continue;
1307 break;
1308 case Op_CheckCastPP: {
1309 // Don't move CheckCastPP nodes away from their input, if the input
1310 // is a rawptr (5071820).
1311 Node *def = self->in(1);
1312 if (def != NULL && def->bottom_type()->base() == Type::RawPtr) {
1313 early->add_inst(self);
1314 #ifdef ASSERT
1315 _raw_oops.push(def);
1316 #endif
1317 continue;
1318 }
1319 break;
1320 }
1321 default:
1322 break;
1323 }
1324 }
1325
1326 // Gather LCA of all uses
1327 Block *LCA = NULL;
1328 {
1329 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
1330 // For all uses, find LCA
1331 Node* use = self->fast_out(i);
1332 LCA = raise_LCA_above_use(LCA, use, self, this);
1333 }
1334 guarantee(LCA != NULL, "There must be a LCA");
1335 } // (Hide defs of imax, i from rest of block.)
1336
1337 // Place temps in the block of their use. This isn't a
1338 // requirement for correctness but it reduces useless
1339 // interference between temps and other nodes.
1340 if (mach != NULL && mach->is_MachTemp()) {
1341 map_node_to_block(self, LCA);
1342 LCA->add_inst(self);
1343 continue;
1344 }
1345
1346 // Check if 'self' could be anti-dependent on memory
1347 if (self->needs_anti_dependence_check()) {
1348 // Hoist LCA above possible-defs and insert anti-dependences to
1349 // defs in new LCA block.
1350 LCA = insert_anti_dependences(LCA, self);
1351 }
1352
1353 if (early->_dom_depth > LCA->_dom_depth) {
1354 // Somehow the LCA has moved above the earliest legal point.
1355 // (One way this can happen is via memory_early_block.)
1356 if (C->subsume_loads() == true && !C->failing()) {
1357 // Retry with subsume_loads == false
1358 // If this is the first failure, the sentinel string will "stick"
1359 // to the Compile object, and the C2Compiler will see it and retry.
1360 C->record_failure(C2Compiler::retry_no_subsuming_loads());
1361 } else {
1362 // Bailout without retry when (early->_dom_depth > LCA->_dom_depth)
1363 assert(false, "graph should be schedulable");
1364 C->record_method_not_compilable("late schedule failed: incorrect graph");
1365 }
1366 return;
1367 }
1368
1369 // If there is no opportunity to hoist, then we're done.
1370 // In stress mode, try to hoist even the single operations.
1371 bool try_to_hoist = StressGCM || (LCA != early);
1372
1373 // Must clone guys stay next to use; no hoisting allowed.
1374 // Also cannot hoist guys that alter memory or are otherwise not
1375 // allocatable (hoisting can make a value live longer, leading to
1376 // anti and output dependency problems which are normally resolved
1377 // by the register allocator giving everyone a different register).
1378 if (mach != NULL && must_clone[mach->ideal_Opcode()])
1379 try_to_hoist = false;
1380
1381 Block* late = NULL;
1382 if (try_to_hoist) {
1383 // Now find the block with the least execution frequency.
1384 // Start at the latest schedule and work up to the earliest schedule
1385 // in the dominator tree. Thus the Node will dominate all its uses.
1386 late = hoist_to_cheaper_block(LCA, early, self);
1387 } else {
1388 // Just use the LCA of the uses.
1389 late = LCA;
1390 }
1391
1392 // Put the node into target block
1393 schedule_node_into_block(self, late);
1394
1395 #ifdef ASSERT
1396 if (self->needs_anti_dependence_check()) {
1397 // since precedence edges are only inserted when we're sure they
1398 // are needed make sure that after placement in a block we don't
1399 // need any new precedence edges.
1400 verify_anti_dependences(late, self);
1401 }
1402 #endif
1403 } // Loop until all nodes have been visited
1404
1405 } // end ScheduleLate
1406
1407 //------------------------------GlobalCodeMotion-------------------------------
global_code_motion()1408 void PhaseCFG::global_code_motion() {
1409 ResourceMark rm;
1410
1411 #ifndef PRODUCT
1412 if (trace_opto_pipelining()) {
1413 tty->print("\n---- Start GlobalCodeMotion ----\n");
1414 }
1415 #endif
1416
1417 // Initialize the node to block mapping for things on the proj_list
1418 for (uint i = 0; i < _matcher.number_of_projections(); i++) {
1419 unmap_node_from_block(_matcher.get_projection(i));
1420 }
1421
1422 // Set the basic block for Nodes pinned into blocks
1423 Arena* arena = Thread::current()->resource_area();
1424 VectorSet visited(arena);
1425 schedule_pinned_nodes(visited);
1426
1427 // Find the earliest Block any instruction can be placed in. Some
1428 // instructions are pinned into Blocks. Unpinned instructions can
1429 // appear in last block in which all their inputs occur.
1430 visited.Clear();
1431 Node_Stack stack(arena, (C->live_nodes() >> 2) + 16); // pre-grow
1432 if (!schedule_early(visited, stack)) {
1433 // Bailout without retry
1434 C->record_method_not_compilable("early schedule failed");
1435 return;
1436 }
1437
1438 // Build Def-Use edges.
1439 // Compute the latency information (via backwards walk) for all the
1440 // instructions in the graph
1441 _node_latency = new GrowableArray<uint>(); // resource_area allocation
1442
1443 if (C->do_scheduling()) {
1444 compute_latencies_backwards(visited, stack);
1445 }
1446
1447 // Now schedule all codes as LATE as possible. This is the LCA in the
1448 // dominator tree of all USES of a value. Pick the block with the least
1449 // loop nesting depth that is lowest in the dominator tree.
1450 // ( visited.Clear() called in schedule_late()->Node_Backward_Iterator() )
1451 schedule_late(visited, stack);
1452 if (C->failing()) {
1453 return;
1454 }
1455
1456 #ifndef PRODUCT
1457 if (trace_opto_pipelining()) {
1458 tty->print("\n---- Detect implicit null checks ----\n");
1459 }
1460 #endif
1461
1462 // Detect implicit-null-check opportunities. Basically, find NULL checks
1463 // with suitable memory ops nearby. Use the memory op to do the NULL check.
1464 // I can generate a memory op if there is not one nearby.
1465 if (C->is_method_compilation()) {
1466 // By reversing the loop direction we get a very minor gain on mpegaudio.
1467 // Feel free to revert to a forward loop for clarity.
1468 // for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) {
1469 for (int i = _matcher._null_check_tests.size() - 2; i >= 0; i -= 2) {
1470 Node* proj = _matcher._null_check_tests[i];
1471 Node* val = _matcher._null_check_tests[i + 1];
1472 Block* block = get_block_for_node(proj);
1473 implicit_null_check(block, proj, val, C->allowed_deopt_reasons());
1474 // The implicit_null_check will only perform the transformation
1475 // if the null branch is truly uncommon, *and* it leads to an
1476 // uncommon trap. Combined with the too_many_traps guards
1477 // above, this prevents SEGV storms reported in 6366351,
1478 // by recompiling offending methods without this optimization.
1479 }
1480 }
1481
1482 bool block_size_threshold_ok = false;
1483 intptr_t *recalc_pressure_nodes = NULL;
1484 if (OptoRegScheduling) {
1485 for (uint i = 0; i < number_of_blocks(); i++) {
1486 Block* block = get_block(i);
1487 if (block->number_of_nodes() > 10) {
1488 block_size_threshold_ok = true;
1489 break;
1490 }
1491 }
1492 }
1493
1494 // Enabling the scheduler for register pressure plus finding blocks of size to schedule for it
1495 // is key to enabling this feature.
1496 PhaseChaitin regalloc(C->unique(), *this, _matcher, true);
1497 ResourceArea live_arena(mtCompiler); // Arena for liveness
1498 ResourceMark rm_live(&live_arena);
1499 PhaseLive live(*this, regalloc._lrg_map.names(), &live_arena, true);
1500 PhaseIFG ifg(&live_arena);
1501 if (OptoRegScheduling && block_size_threshold_ok) {
1502 regalloc.mark_ssa();
1503 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]);
1504 rm_live.reset_to_mark(); // Reclaim working storage
1505 IndexSet::reset_memory(C, &live_arena);
1506 uint node_size = regalloc._lrg_map.max_lrg_id();
1507 ifg.init(node_size); // Empty IFG
1508 regalloc.set_ifg(ifg);
1509 regalloc.set_live(live);
1510 regalloc.gather_lrg_masks(false); // Collect LRG masks
1511 live.compute(node_size); // Compute liveness
1512
1513 recalc_pressure_nodes = NEW_RESOURCE_ARRAY(intptr_t, node_size);
1514 for (uint i = 0; i < node_size; i++) {
1515 recalc_pressure_nodes[i] = 0;
1516 }
1517 }
1518 _regalloc = ®alloc;
1519
1520 #ifndef PRODUCT
1521 if (trace_opto_pipelining()) {
1522 tty->print("\n---- Start Local Scheduling ----\n");
1523 }
1524 #endif
1525
1526 // Schedule locally. Right now a simple topological sort.
1527 // Later, do a real latency aware scheduler.
1528 GrowableArray<int> ready_cnt(C->unique(), C->unique(), -1);
1529 visited.Clear();
1530 for (uint i = 0; i < number_of_blocks(); i++) {
1531 Block* block = get_block(i);
1532 if (!schedule_local(block, ready_cnt, visited, recalc_pressure_nodes)) {
1533 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
1534 C->record_method_not_compilable("local schedule failed");
1535 }
1536 _regalloc = NULL;
1537 return;
1538 }
1539 }
1540 _regalloc = NULL;
1541
1542 // If we inserted any instructions between a Call and his CatchNode,
1543 // clone the instructions on all paths below the Catch.
1544 for (uint i = 0; i < number_of_blocks(); i++) {
1545 Block* block = get_block(i);
1546 call_catch_cleanup(block);
1547 }
1548
1549 #ifndef PRODUCT
1550 if (trace_opto_pipelining()) {
1551 tty->print("\n---- After GlobalCodeMotion ----\n");
1552 for (uint i = 0; i < number_of_blocks(); i++) {
1553 Block* block = get_block(i);
1554 block->dump();
1555 }
1556 }
1557 #endif
1558 // Dead.
1559 _node_latency = (GrowableArray<uint> *)((intptr_t)0xdeadbeef);
1560 }
1561
do_global_code_motion()1562 bool PhaseCFG::do_global_code_motion() {
1563
1564 build_dominator_tree();
1565 if (C->failing()) {
1566 return false;
1567 }
1568
1569 NOT_PRODUCT( C->verify_graph_edges(); )
1570
1571 estimate_block_frequency();
1572
1573 global_code_motion();
1574
1575 if (C->failing()) {
1576 return false;
1577 }
1578
1579 return true;
1580 }
1581
1582 //------------------------------Estimate_Block_Frequency-----------------------
1583 // Estimate block frequencies based on IfNode probabilities.
estimate_block_frequency()1584 void PhaseCFG::estimate_block_frequency() {
1585
1586 // Force conditional branches leading to uncommon traps to be unlikely,
1587 // not because we get to the uncommon_trap with less relative frequency,
1588 // but because an uncommon_trap typically causes a deopt, so we only get
1589 // there once.
1590 if (C->do_freq_based_layout()) {
1591 Block_List worklist;
1592 Block* root_blk = get_block(0);
1593 for (uint i = 1; i < root_blk->num_preds(); i++) {
1594 Block *pb = get_block_for_node(root_blk->pred(i));
1595 if (pb->has_uncommon_code()) {
1596 worklist.push(pb);
1597 }
1598 }
1599 while (worklist.size() > 0) {
1600 Block* uct = worklist.pop();
1601 if (uct == get_root_block()) {
1602 continue;
1603 }
1604 for (uint i = 1; i < uct->num_preds(); i++) {
1605 Block *pb = get_block_for_node(uct->pred(i));
1606 if (pb->_num_succs == 1) {
1607 worklist.push(pb);
1608 } else if (pb->num_fall_throughs() == 2) {
1609 pb->update_uncommon_branch(uct);
1610 }
1611 }
1612 }
1613 }
1614
1615 // Create the loop tree and calculate loop depth.
1616 _root_loop = create_loop_tree();
1617 _root_loop->compute_loop_depth(0);
1618
1619 // Compute block frequency of each block, relative to a single loop entry.
1620 _root_loop->compute_freq();
1621
1622 // Adjust all frequencies to be relative to a single method entry
1623 _root_loop->_freq = 1.0;
1624 _root_loop->scale_freq();
1625
1626 // Save outmost loop frequency for LRG frequency threshold
1627 _outer_loop_frequency = _root_loop->outer_loop_freq();
1628
1629 // force paths ending at uncommon traps to be infrequent
1630 if (!C->do_freq_based_layout()) {
1631 Block_List worklist;
1632 Block* root_blk = get_block(0);
1633 for (uint i = 1; i < root_blk->num_preds(); i++) {
1634 Block *pb = get_block_for_node(root_blk->pred(i));
1635 if (pb->has_uncommon_code()) {
1636 worklist.push(pb);
1637 }
1638 }
1639 while (worklist.size() > 0) {
1640 Block* uct = worklist.pop();
1641 uct->_freq = PROB_MIN;
1642 for (uint i = 1; i < uct->num_preds(); i++) {
1643 Block *pb = get_block_for_node(uct->pred(i));
1644 if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) {
1645 worklist.push(pb);
1646 }
1647 }
1648 }
1649 }
1650
1651 #ifdef ASSERT
1652 for (uint i = 0; i < number_of_blocks(); i++) {
1653 Block* b = get_block(i);
1654 assert(b->_freq >= MIN_BLOCK_FREQUENCY, "Register Allocator requires meaningful block frequency");
1655 }
1656 #endif
1657
1658 #ifndef PRODUCT
1659 if (PrintCFGBlockFreq) {
1660 tty->print_cr("CFG Block Frequencies");
1661 _root_loop->dump_tree();
1662 if (Verbose) {
1663 tty->print_cr("PhaseCFG dump");
1664 dump();
1665 tty->print_cr("Node dump");
1666 _root->dump(99999);
1667 }
1668 }
1669 #endif
1670 }
1671
1672 //----------------------------create_loop_tree--------------------------------
1673 // Create a loop tree from the CFG
create_loop_tree()1674 CFGLoop* PhaseCFG::create_loop_tree() {
1675
1676 #ifdef ASSERT
1677 assert(get_block(0) == get_root_block(), "first block should be root block");
1678 for (uint i = 0; i < number_of_blocks(); i++) {
1679 Block* block = get_block(i);
1680 // Check that _loop field are clear...we could clear them if not.
1681 assert(block->_loop == NULL, "clear _loop expected");
1682 // Sanity check that the RPO numbering is reflected in the _blocks array.
1683 // It doesn't have to be for the loop tree to be built, but if it is not,
1684 // then the blocks have been reordered since dom graph building...which
1685 // may question the RPO numbering
1686 assert(block->_rpo == i, "unexpected reverse post order number");
1687 }
1688 #endif
1689
1690 int idct = 0;
1691 CFGLoop* root_loop = new CFGLoop(idct++);
1692
1693 Block_List worklist;
1694
1695 // Assign blocks to loops
1696 for(uint i = number_of_blocks() - 1; i > 0; i-- ) { // skip Root block
1697 Block* block = get_block(i);
1698
1699 if (block->head()->is_Loop()) {
1700 Block* loop_head = block;
1701 assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1702 Node* tail_n = loop_head->pred(LoopNode::LoopBackControl);
1703 Block* tail = get_block_for_node(tail_n);
1704
1705 // Defensively filter out Loop nodes for non-single-entry loops.
1706 // For all reasonable loops, the head occurs before the tail in RPO.
1707 if (i <= tail->_rpo) {
1708
1709 // The tail and (recursive) predecessors of the tail
1710 // are made members of a new loop.
1711
1712 assert(worklist.size() == 0, "nonempty worklist");
1713 CFGLoop* nloop = new CFGLoop(idct++);
1714 assert(loop_head->_loop == NULL, "just checking");
1715 loop_head->_loop = nloop;
1716 // Add to nloop so push_pred() will skip over inner loops
1717 nloop->add_member(loop_head);
1718 nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, this);
1719
1720 while (worklist.size() > 0) {
1721 Block* member = worklist.pop();
1722 if (member != loop_head) {
1723 for (uint j = 1; j < member->num_preds(); j++) {
1724 nloop->push_pred(member, j, worklist, this);
1725 }
1726 }
1727 }
1728 }
1729 }
1730 }
1731
1732 // Create a member list for each loop consisting
1733 // of both blocks and (immediate child) loops.
1734 for (uint i = 0; i < number_of_blocks(); i++) {
1735 Block* block = get_block(i);
1736 CFGLoop* lp = block->_loop;
1737 if (lp == NULL) {
1738 // Not assigned to a loop. Add it to the method's pseudo loop.
1739 block->_loop = root_loop;
1740 lp = root_loop;
1741 }
1742 if (lp == root_loop || block != lp->head()) { // loop heads are already members
1743 lp->add_member(block);
1744 }
1745 if (lp != root_loop) {
1746 if (lp->parent() == NULL) {
1747 // Not a nested loop. Make it a child of the method's pseudo loop.
1748 root_loop->add_nested_loop(lp);
1749 }
1750 if (block == lp->head()) {
1751 // Add nested loop to member list of parent loop.
1752 lp->parent()->add_member(lp);
1753 }
1754 }
1755 }
1756
1757 return root_loop;
1758 }
1759
1760 //------------------------------push_pred--------------------------------------
push_pred(Block * blk,int i,Block_List & worklist,PhaseCFG * cfg)1761 void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg) {
1762 Node* pred_n = blk->pred(i);
1763 Block* pred = cfg->get_block_for_node(pred_n);
1764 CFGLoop *pred_loop = pred->_loop;
1765 if (pred_loop == NULL) {
1766 // Filter out blocks for non-single-entry loops.
1767 // For all reasonable loops, the head occurs before the tail in RPO.
1768 if (pred->_rpo > head()->_rpo) {
1769 pred->_loop = this;
1770 worklist.push(pred);
1771 }
1772 } else if (pred_loop != this) {
1773 // Nested loop.
1774 while (pred_loop->_parent != NULL && pred_loop->_parent != this) {
1775 pred_loop = pred_loop->_parent;
1776 }
1777 // Make pred's loop be a child
1778 if (pred_loop->_parent == NULL) {
1779 add_nested_loop(pred_loop);
1780 // Continue with loop entry predecessor.
1781 Block* pred_head = pred_loop->head();
1782 assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1783 assert(pred_head != head(), "loop head in only one loop");
1784 push_pred(pred_head, LoopNode::EntryControl, worklist, cfg);
1785 } else {
1786 assert(pred_loop->_parent == this && _parent == NULL, "just checking");
1787 }
1788 }
1789 }
1790
1791 //------------------------------add_nested_loop--------------------------------
1792 // Make cl a child of the current loop in the loop tree.
add_nested_loop(CFGLoop * cl)1793 void CFGLoop::add_nested_loop(CFGLoop* cl) {
1794 assert(_parent == NULL, "no parent yet");
1795 assert(cl != this, "not my own parent");
1796 cl->_parent = this;
1797 CFGLoop* ch = _child;
1798 if (ch == NULL) {
1799 _child = cl;
1800 } else {
1801 while (ch->_sibling != NULL) { ch = ch->_sibling; }
1802 ch->_sibling = cl;
1803 }
1804 }
1805
1806 //------------------------------compute_loop_depth-----------------------------
1807 // Store the loop depth in each CFGLoop object.
1808 // Recursively walk the children to do the same for them.
compute_loop_depth(int depth)1809 void CFGLoop::compute_loop_depth(int depth) {
1810 _depth = depth;
1811 CFGLoop* ch = _child;
1812 while (ch != NULL) {
1813 ch->compute_loop_depth(depth + 1);
1814 ch = ch->_sibling;
1815 }
1816 }
1817
1818 //------------------------------compute_freq-----------------------------------
1819 // Compute the frequency of each block and loop, relative to a single entry
1820 // into the dominating loop head.
compute_freq()1821 void CFGLoop::compute_freq() {
1822 // Bottom up traversal of loop tree (visit inner loops first.)
1823 // Set loop head frequency to 1.0, then transitively
1824 // compute frequency for all successors in the loop,
1825 // as well as for each exit edge. Inner loops are
1826 // treated as single blocks with loop exit targets
1827 // as the successor blocks.
1828
1829 // Nested loops first
1830 CFGLoop* ch = _child;
1831 while (ch != NULL) {
1832 ch->compute_freq();
1833 ch = ch->_sibling;
1834 }
1835 assert (_members.length() > 0, "no empty loops");
1836 Block* hd = head();
1837 hd->_freq = 1.0;
1838 for (int i = 0; i < _members.length(); i++) {
1839 CFGElement* s = _members.at(i);
1840 double freq = s->_freq;
1841 if (s->is_block()) {
1842 Block* b = s->as_Block();
1843 for (uint j = 0; j < b->_num_succs; j++) {
1844 Block* sb = b->_succs[j];
1845 update_succ_freq(sb, freq * b->succ_prob(j));
1846 }
1847 } else {
1848 CFGLoop* lp = s->as_CFGLoop();
1849 assert(lp->_parent == this, "immediate child");
1850 for (int k = 0; k < lp->_exits.length(); k++) {
1851 Block* eb = lp->_exits.at(k).get_target();
1852 double prob = lp->_exits.at(k).get_prob();
1853 update_succ_freq(eb, freq * prob);
1854 }
1855 }
1856 }
1857
1858 // For all loops other than the outer, "method" loop,
1859 // sum and normalize the exit probability. The "method" loop
1860 // should keep the initial exit probability of 1, so that
1861 // inner blocks do not get erroneously scaled.
1862 if (_depth != 0) {
1863 // Total the exit probabilities for this loop.
1864 double exits_sum = 0.0f;
1865 for (int i = 0; i < _exits.length(); i++) {
1866 exits_sum += _exits.at(i).get_prob();
1867 }
1868
1869 // Normalize the exit probabilities. Until now, the
1870 // probabilities estimate the possibility of exit per
1871 // a single loop iteration; afterward, they estimate
1872 // the probability of exit per loop entry.
1873 for (int i = 0; i < _exits.length(); i++) {
1874 Block* et = _exits.at(i).get_target();
1875 float new_prob = 0.0f;
1876 if (_exits.at(i).get_prob() > 0.0f) {
1877 new_prob = _exits.at(i).get_prob() / exits_sum;
1878 }
1879 BlockProbPair bpp(et, new_prob);
1880 _exits.at_put(i, bpp);
1881 }
1882
1883 // Save the total, but guard against unreasonable probability,
1884 // as the value is used to estimate the loop trip count.
1885 // An infinite trip count would blur relative block
1886 // frequencies.
1887 if (exits_sum > 1.0f) exits_sum = 1.0;
1888 if (exits_sum < PROB_MIN) exits_sum = PROB_MIN;
1889 _exit_prob = exits_sum;
1890 }
1891 }
1892
1893 //------------------------------succ_prob-------------------------------------
1894 // Determine the probability of reaching successor 'i' from the receiver block.
succ_prob(uint i)1895 float Block::succ_prob(uint i) {
1896 int eidx = end_idx();
1897 Node *n = get_node(eidx); // Get ending Node
1898
1899 int op = n->Opcode();
1900 if (n->is_Mach()) {
1901 if (n->is_MachNullCheck()) {
1902 // Can only reach here if called after lcm. The original Op_If is gone,
1903 // so we attempt to infer the probability from one or both of the
1904 // successor blocks.
1905 assert(_num_succs == 2, "expecting 2 successors of a null check");
1906 // If either successor has only one predecessor, then the
1907 // probability estimate can be derived using the
1908 // relative frequency of the successor and this block.
1909 if (_succs[i]->num_preds() == 2) {
1910 return _succs[i]->_freq / _freq;
1911 } else if (_succs[1-i]->num_preds() == 2) {
1912 return 1 - (_succs[1-i]->_freq / _freq);
1913 } else {
1914 // Estimate using both successor frequencies
1915 float freq = _succs[i]->_freq;
1916 return freq / (freq + _succs[1-i]->_freq);
1917 }
1918 }
1919 op = n->as_Mach()->ideal_Opcode();
1920 }
1921
1922
1923 // Switch on branch type
1924 switch( op ) {
1925 case Op_CountedLoopEnd:
1926 case Op_If: {
1927 assert (i < 2, "just checking");
1928 // Conditionals pass on only part of their frequency
1929 float prob = n->as_MachIf()->_prob;
1930 assert(prob >= 0.0 && prob <= 1.0, "out of range probability");
1931 // If succ[i] is the FALSE branch, invert path info
1932 if( get_node(i + eidx + 1)->Opcode() == Op_IfFalse ) {
1933 return 1.0f - prob; // not taken
1934 } else {
1935 return prob; // taken
1936 }
1937 }
1938
1939 case Op_Jump:
1940 return n->as_MachJump()->_probs[get_node(i + eidx + 1)->as_JumpProj()->_con];
1941
1942 case Op_Catch: {
1943 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
1944 if (ci->_con == CatchProjNode::fall_through_index) {
1945 // Fall-thru path gets the lion's share.
1946 return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs;
1947 } else {
1948 // Presume exceptional paths are equally unlikely
1949 return PROB_UNLIKELY_MAG(5);
1950 }
1951 }
1952
1953 case Op_Root:
1954 case Op_Goto:
1955 // Pass frequency straight thru to target
1956 return 1.0f;
1957
1958 case Op_NeverBranch:
1959 return 0.0f;
1960
1961 case Op_TailCall:
1962 case Op_TailJump:
1963 case Op_Return:
1964 case Op_Halt:
1965 case Op_Rethrow:
1966 // Do not push out freq to root block
1967 return 0.0f;
1968
1969 default:
1970 ShouldNotReachHere();
1971 }
1972
1973 return 0.0f;
1974 }
1975
1976 //------------------------------num_fall_throughs-----------------------------
1977 // Return the number of fall-through candidates for a block
num_fall_throughs()1978 int Block::num_fall_throughs() {
1979 int eidx = end_idx();
1980 Node *n = get_node(eidx); // Get ending Node
1981
1982 int op = n->Opcode();
1983 if (n->is_Mach()) {
1984 if (n->is_MachNullCheck()) {
1985 // In theory, either side can fall-thru, for simplicity sake,
1986 // let's say only the false branch can now.
1987 return 1;
1988 }
1989 op = n->as_Mach()->ideal_Opcode();
1990 }
1991
1992 // Switch on branch type
1993 switch( op ) {
1994 case Op_CountedLoopEnd:
1995 case Op_If:
1996 return 2;
1997
1998 case Op_Root:
1999 case Op_Goto:
2000 return 1;
2001
2002 case Op_Catch: {
2003 for (uint i = 0; i < _num_succs; i++) {
2004 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
2005 if (ci->_con == CatchProjNode::fall_through_index) {
2006 return 1;
2007 }
2008 }
2009 return 0;
2010 }
2011
2012 case Op_Jump:
2013 case Op_NeverBranch:
2014 case Op_TailCall:
2015 case Op_TailJump:
2016 case Op_Return:
2017 case Op_Halt:
2018 case Op_Rethrow:
2019 return 0;
2020
2021 default:
2022 ShouldNotReachHere();
2023 }
2024
2025 return 0;
2026 }
2027
2028 //------------------------------succ_fall_through-----------------------------
2029 // Return true if a specific successor could be fall-through target.
succ_fall_through(uint i)2030 bool Block::succ_fall_through(uint i) {
2031 int eidx = end_idx();
2032 Node *n = get_node(eidx); // Get ending Node
2033
2034 int op = n->Opcode();
2035 if (n->is_Mach()) {
2036 if (n->is_MachNullCheck()) {
2037 // In theory, either side can fall-thru, for simplicity sake,
2038 // let's say only the false branch can now.
2039 return get_node(i + eidx + 1)->Opcode() == Op_IfFalse;
2040 }
2041 op = n->as_Mach()->ideal_Opcode();
2042 }
2043
2044 // Switch on branch type
2045 switch( op ) {
2046 case Op_CountedLoopEnd:
2047 case Op_If:
2048 case Op_Root:
2049 case Op_Goto:
2050 return true;
2051
2052 case Op_Catch: {
2053 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
2054 return ci->_con == CatchProjNode::fall_through_index;
2055 }
2056
2057 case Op_Jump:
2058 case Op_NeverBranch:
2059 case Op_TailCall:
2060 case Op_TailJump:
2061 case Op_Return:
2062 case Op_Halt:
2063 case Op_Rethrow:
2064 return false;
2065
2066 default:
2067 ShouldNotReachHere();
2068 }
2069
2070 return false;
2071 }
2072
2073 //------------------------------update_uncommon_branch------------------------
2074 // Update the probability of a two-branch to be uncommon
update_uncommon_branch(Block * ub)2075 void Block::update_uncommon_branch(Block* ub) {
2076 int eidx = end_idx();
2077 Node *n = get_node(eidx); // Get ending Node
2078
2079 int op = n->as_Mach()->ideal_Opcode();
2080
2081 assert(op == Op_CountedLoopEnd || op == Op_If, "must be a If");
2082 assert(num_fall_throughs() == 2, "must be a two way branch block");
2083
2084 // Which successor is ub?
2085 uint s;
2086 for (s = 0; s <_num_succs; s++) {
2087 if (_succs[s] == ub) break;
2088 }
2089 assert(s < 2, "uncommon successor must be found");
2090
2091 // If ub is the true path, make the proability small, else
2092 // ub is the false path, and make the probability large
2093 bool invert = (get_node(s + eidx + 1)->Opcode() == Op_IfFalse);
2094
2095 // Get existing probability
2096 float p = n->as_MachIf()->_prob;
2097
2098 if (invert) p = 1.0 - p;
2099 if (p > PROB_MIN) {
2100 p = PROB_MIN;
2101 }
2102 if (invert) p = 1.0 - p;
2103
2104 n->as_MachIf()->_prob = p;
2105 }
2106
2107 //------------------------------update_succ_freq-------------------------------
2108 // Update the appropriate frequency associated with block 'b', a successor of
2109 // a block in this loop.
update_succ_freq(Block * b,double freq)2110 void CFGLoop::update_succ_freq(Block* b, double freq) {
2111 if (b->_loop == this) {
2112 if (b == head()) {
2113 // back branch within the loop
2114 // Do nothing now, the loop carried frequency will be
2115 // adjust later in scale_freq().
2116 } else {
2117 // simple branch within the loop
2118 b->_freq += freq;
2119 }
2120 } else if (!in_loop_nest(b)) {
2121 // branch is exit from this loop
2122 BlockProbPair bpp(b, freq);
2123 _exits.append(bpp);
2124 } else {
2125 // branch into nested loop
2126 CFGLoop* ch = b->_loop;
2127 ch->_freq += freq;
2128 }
2129 }
2130
2131 //------------------------------in_loop_nest-----------------------------------
2132 // Determine if block b is in the receiver's loop nest.
in_loop_nest(Block * b)2133 bool CFGLoop::in_loop_nest(Block* b) {
2134 int depth = _depth;
2135 CFGLoop* b_loop = b->_loop;
2136 int b_depth = b_loop->_depth;
2137 if (depth == b_depth) {
2138 return true;
2139 }
2140 while (b_depth > depth) {
2141 b_loop = b_loop->_parent;
2142 b_depth = b_loop->_depth;
2143 }
2144 return b_loop == this;
2145 }
2146
2147 //------------------------------scale_freq-------------------------------------
2148 // Scale frequency of loops and blocks by trip counts from outer loops
2149 // Do a top down traversal of loop tree (visit outer loops first.)
scale_freq()2150 void CFGLoop::scale_freq() {
2151 double loop_freq = _freq * trip_count();
2152 _freq = loop_freq;
2153 for (int i = 0; i < _members.length(); i++) {
2154 CFGElement* s = _members.at(i);
2155 double block_freq = s->_freq * loop_freq;
2156 if (g_isnan(block_freq) || block_freq < MIN_BLOCK_FREQUENCY)
2157 block_freq = MIN_BLOCK_FREQUENCY;
2158 s->_freq = block_freq;
2159 }
2160 CFGLoop* ch = _child;
2161 while (ch != NULL) {
2162 ch->scale_freq();
2163 ch = ch->_sibling;
2164 }
2165 }
2166
2167 // Frequency of outer loop
outer_loop_freq() const2168 double CFGLoop::outer_loop_freq() const {
2169 if (_child != NULL) {
2170 return _child->_freq;
2171 }
2172 return _freq;
2173 }
2174
2175 #ifndef PRODUCT
2176 //------------------------------dump_tree--------------------------------------
dump_tree() const2177 void CFGLoop::dump_tree() const {
2178 dump();
2179 if (_child != NULL) _child->dump_tree();
2180 if (_sibling != NULL) _sibling->dump_tree();
2181 }
2182
2183 //------------------------------dump-------------------------------------------
dump() const2184 void CFGLoop::dump() const {
2185 for (int i = 0; i < _depth; i++) tty->print(" ");
2186 tty->print("%s: %d trip_count: %6.0f freq: %6.0f\n",
2187 _depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq);
2188 for (int i = 0; i < _depth; i++) tty->print(" ");
2189 tty->print(" members:");
2190 int k = 0;
2191 for (int i = 0; i < _members.length(); i++) {
2192 if (k++ >= 6) {
2193 tty->print("\n ");
2194 for (int j = 0; j < _depth+1; j++) tty->print(" ");
2195 k = 0;
2196 }
2197 CFGElement *s = _members.at(i);
2198 if (s->is_block()) {
2199 Block *b = s->as_Block();
2200 tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq);
2201 } else {
2202 CFGLoop* lp = s->as_CFGLoop();
2203 tty->print(" L%d(%6.3f)", lp->_id, lp->_freq);
2204 }
2205 }
2206 tty->print("\n");
2207 for (int i = 0; i < _depth; i++) tty->print(" ");
2208 tty->print(" exits: ");
2209 k = 0;
2210 for (int i = 0; i < _exits.length(); i++) {
2211 if (k++ >= 7) {
2212 tty->print("\n ");
2213 for (int j = 0; j < _depth+1; j++) tty->print(" ");
2214 k = 0;
2215 }
2216 Block *blk = _exits.at(i).get_target();
2217 double prob = _exits.at(i).get_prob();
2218 tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100));
2219 }
2220 tty->print("\n");
2221 }
2222 #endif
2223