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) const155 bool PhaseCFG::is_control_proj_or_safepoint(Node* n) const {
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) const162 Block* PhaseCFG::find_block_for_node(Node* n) const {
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 assert(Compile::AliasIdxTop <= load_alias_idx && load_alias_idx < C->num_alias_types(), "Invalid alias index");
569 if (load_alias_idx == Compile::AliasIdxBot && C->AliasLevel() > 0 &&
570 (PrintOpto || VerifyAliases ||
571 (PrintMiscellaneous && (WizardMode || Verbose)))) {
572 // Load nodes should not consume all of memory.
573 // Reporting a bottom type indicates a bug in adlc.
574 // If some particular type of node validly consumes all of memory,
575 // sharpen the preceding "if" to exclude it, so we can catch bugs here.
576 tty->print_cr("*** Possible Anti-Dependence Bug: Load consumes all of memory.");
577 load->dump(2);
578 if (VerifyAliases) assert(load_alias_idx != Compile::AliasIdxBot, "");
579 }
580 #endif
581
582 if (!C->alias_type(load_alias_idx)->is_rewritable()) {
583 // It is impossible to spoil this load by putting stores before it,
584 // because we know that the stores will never update the value
585 // which 'load' must witness.
586 return LCA;
587 }
588
589 node_idx_t load_index = load->_idx;
590
591 // Note the earliest legal placement of 'load', as determined by
592 // by the unique point in the dom tree where all memory effects
593 // and other inputs are first available. (Computed by schedule_early.)
594 // For normal loads, 'early' is the shallowest place (dom graph wise)
595 // to look for anti-deps between this load and any store.
596 Block* early = get_block_for_node(load);
597
598 // If we are subsuming loads, compute an "early" block that only considers
599 // memory or address inputs. This block may be different than the
600 // schedule_early block in that it could be at an even shallower depth in the
601 // dominator tree, and allow for a broader discovery of anti-dependences.
602 if (C->subsume_loads()) {
603 early = memory_early_block(load, early, this);
604 }
605
606 ResourceArea *area = Thread::current()->resource_area();
607 Node_List worklist_mem(area); // prior memory state to store
608 Node_List worklist_store(area); // possible-def to explore
609 Node_List worklist_visited(area); // visited mergemem nodes
610 Node_List non_early_stores(area); // all relevant stores outside of early
611 bool must_raise_LCA = false;
612
613 // 'load' uses some memory state; look for users of the same state.
614 // Recurse through MergeMem nodes to the stores that use them.
615
616 // Each of these stores is a possible definition of memory
617 // that 'load' needs to use. We need to force 'load'
618 // to occur before each such store. When the store is in
619 // the same block as 'load', we insert an anti-dependence
620 // edge load->store.
621
622 // The relevant stores "nearby" the load consist of a tree rooted
623 // at initial_mem, with internal nodes of type MergeMem.
624 // Therefore, the branches visited by the worklist are of this form:
625 // initial_mem -> (MergeMem ->)* store
626 // The anti-dependence constraints apply only to the fringe of this tree.
627
628 Node* initial_mem = load->in(MemNode::Memory);
629 worklist_store.push(initial_mem);
630 worklist_visited.push(initial_mem);
631 worklist_mem.push(NULL);
632 while (worklist_store.size() > 0) {
633 // Examine a nearby store to see if it might interfere with our load.
634 Node* mem = worklist_mem.pop();
635 Node* store = worklist_store.pop();
636 uint op = store->Opcode();
637
638 // MergeMems do not directly have anti-deps.
639 // Treat them as internal nodes in a forward tree of memory states,
640 // the leaves of which are each a 'possible-def'.
641 if (store == initial_mem // root (exclusive) of tree we are searching
642 || op == Op_MergeMem // internal node of tree we are searching
643 ) {
644 mem = store; // It's not a possibly interfering store.
645 if (store == initial_mem)
646 initial_mem = NULL; // only process initial memory once
647
648 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
649 store = mem->fast_out(i);
650 if (store->is_MergeMem()) {
651 // Be sure we don't get into combinatorial problems.
652 // (Allow phis to be repeated; they can merge two relevant states.)
653 uint j = worklist_visited.size();
654 for (; j > 0; j--) {
655 if (worklist_visited.at(j-1) == store) break;
656 }
657 if (j > 0) continue; // already on work list; do not repeat
658 worklist_visited.push(store);
659 }
660 worklist_mem.push(mem);
661 worklist_store.push(store);
662 }
663 continue;
664 }
665
666 if (op == Op_MachProj || op == Op_Catch) continue;
667 if (store->needs_anti_dependence_check()) continue; // not really a store
668
669 // Compute the alias index. Loads and stores with different alias
670 // indices do not need anti-dependence edges. Wide MemBar's are
671 // anti-dependent on everything (except immutable memories).
672 const TypePtr* adr_type = store->adr_type();
673 if (!C->can_alias(adr_type, load_alias_idx)) continue;
674
675 // Most slow-path runtime calls do NOT modify Java memory, but
676 // they can block and so write Raw memory.
677 if (store->is_Mach()) {
678 MachNode* mstore = store->as_Mach();
679 if (load_alias_idx != Compile::AliasIdxRaw) {
680 // Check for call into the runtime using the Java calling
681 // convention (and from there into a wrapper); it has no
682 // _method. Can't do this optimization for Native calls because
683 // they CAN write to Java memory.
684 if (mstore->ideal_Opcode() == Op_CallStaticJava) {
685 assert(mstore->is_MachSafePoint(), "");
686 MachSafePointNode* ms = (MachSafePointNode*) mstore;
687 assert(ms->is_MachCallJava(), "");
688 MachCallJavaNode* mcj = (MachCallJavaNode*) ms;
689 if (mcj->_method == NULL) {
690 // These runtime calls do not write to Java visible memory
691 // (other than Raw) and so do not require anti-dependence edges.
692 continue;
693 }
694 }
695 // Same for SafePoints: they read/write Raw but only read otherwise.
696 // This is basically a workaround for SafePoints only defining control
697 // instead of control + memory.
698 if (mstore->ideal_Opcode() == Op_SafePoint)
699 continue;
700 } else {
701 // Some raw memory, such as the load of "top" at an allocation,
702 // can be control dependent on the previous safepoint. See
703 // comments in GraphKit::allocate_heap() about control input.
704 // Inserting an anti-dep between such a safepoint and a use
705 // creates a cycle, and will cause a subsequent failure in
706 // local scheduling. (BugId 4919904)
707 // (%%% How can a control input be a safepoint and not a projection??)
708 if (mstore->ideal_Opcode() == Op_SafePoint && load->in(0) == mstore)
709 continue;
710 }
711 }
712
713 // Identify a block that the current load must be above,
714 // or else observe that 'store' is all the way up in the
715 // earliest legal block for 'load'. In the latter case,
716 // immediately insert an anti-dependence edge.
717 Block* store_block = get_block_for_node(store);
718 assert(store_block != NULL, "unused killing projections skipped above");
719
720 if (store->is_Phi()) {
721 // Loop-phis need to raise load before input. (Other phis are treated
722 // as store below.)
723 //
724 // 'load' uses memory which is one (or more) of the Phi's inputs.
725 // It must be scheduled not before the Phi, but rather before
726 // each of the relevant Phi inputs.
727 //
728 // Instead of finding the LCA of all inputs to a Phi that match 'mem',
729 // we mark each corresponding predecessor block and do a combined
730 // hoisting operation later (raise_LCA_above_marks).
731 //
732 // Do not assert(store_block != early, "Phi merging memory after access")
733 // PhiNode may be at start of block 'early' with backedge to 'early'
734 DEBUG_ONLY(bool found_match = false);
735 for (uint j = PhiNode::Input, jmax = store->req(); j < jmax; j++) {
736 if (store->in(j) == mem) { // Found matching input?
737 DEBUG_ONLY(found_match = true);
738 Block* pred_block = get_block_for_node(store_block->pred(j));
739 if (pred_block != early) {
740 // If any predecessor of the Phi matches the load's "early block",
741 // we do not need a precedence edge between the Phi and 'load'
742 // since the load will be forced into a block preceding the Phi.
743 pred_block->set_raise_LCA_mark(load_index);
744 assert(!LCA_orig->dominates(pred_block) ||
745 early->dominates(pred_block), "early is high enough");
746 must_raise_LCA = true;
747 } else {
748 // anti-dependent upon PHI pinned below 'early', no edge needed
749 LCA = early; // but can not schedule below 'early'
750 }
751 }
752 }
753 assert(found_match, "no worklist bug");
754 } else if (store_block != early) {
755 // 'store' is between the current LCA and earliest possible block.
756 // Label its block, and decide later on how to raise the LCA
757 // to include the effect on LCA of this store.
758 // If this store's block gets chosen as the raised LCA, we
759 // will find him on the non_early_stores list and stick him
760 // with a precedence edge.
761 // (But, don't bother if LCA is already raised all the way.)
762 if (LCA != early) {
763 store_block->set_raise_LCA_mark(load_index);
764 must_raise_LCA = true;
765 non_early_stores.push(store);
766 }
767 } else {
768 // Found a possibly-interfering store in the load's 'early' block.
769 // This means 'load' cannot sink at all in the dominator tree.
770 // Add an anti-dep edge, and squeeze 'load' into the highest block.
771 assert(store != load->find_exact_control(load->in(0)), "dependence cycle found");
772 if (verify) {
773 assert(store->find_edge(load) != -1, "missing precedence edge");
774 } else {
775 store->add_prec(load);
776 }
777 LCA = early;
778 // This turns off the process of gathering non_early_stores.
779 }
780 }
781 // (Worklist is now empty; all nearby stores have been visited.)
782
783 // Finished if 'load' must be scheduled in its 'early' block.
784 // If we found any stores there, they have already been given
785 // precedence edges.
786 if (LCA == early) return LCA;
787
788 // We get here only if there are no possibly-interfering stores
789 // in the load's 'early' block. Move LCA up above all predecessors
790 // which contain stores we have noted.
791 //
792 // The raised LCA block can be a home to such interfering stores,
793 // but its predecessors must not contain any such stores.
794 //
795 // The raised LCA will be a lower bound for placing the load,
796 // preventing the load from sinking past any block containing
797 // a store that may invalidate the memory state required by 'load'.
798 if (must_raise_LCA)
799 LCA = raise_LCA_above_marks(LCA, load->_idx, early, this);
800 if (LCA == early) return LCA;
801
802 // Insert anti-dependence edges from 'load' to each store
803 // in the non-early LCA block.
804 // Mine the non_early_stores list for such stores.
805 if (LCA->raise_LCA_mark() == load_index) {
806 while (non_early_stores.size() > 0) {
807 Node* store = non_early_stores.pop();
808 Block* store_block = get_block_for_node(store);
809 if (store_block == LCA) {
810 // add anti_dependence from store to load in its own block
811 assert(store != load->find_exact_control(load->in(0)), "dependence cycle found");
812 if (verify) {
813 assert(store->find_edge(load) != -1, "missing precedence edge");
814 } else {
815 store->add_prec(load);
816 }
817 } else {
818 assert(store_block->raise_LCA_mark() == load_index, "block was marked");
819 // Any other stores we found must be either inside the new LCA
820 // or else outside the original LCA. In the latter case, they
821 // did not interfere with any use of 'load'.
822 assert(LCA->dominates(store_block)
823 || !LCA_orig->dominates(store_block), "no stray stores");
824 }
825 }
826 }
827
828 // Return the highest block containing stores; any stores
829 // within that block have been given anti-dependence edges.
830 return LCA;
831 }
832
833 // This class is used to iterate backwards over the nodes in the graph.
834
835 class Node_Backward_Iterator {
836
837 private:
838 Node_Backward_Iterator();
839
840 public:
841 // Constructor for the iterator
842 Node_Backward_Iterator(Node *root, VectorSet &visited, Node_Stack &stack, PhaseCFG &cfg);
843
844 // Postincrement operator to iterate over the nodes
845 Node *next();
846
847 private:
848 VectorSet &_visited;
849 Node_Stack &_stack;
850 PhaseCFG &_cfg;
851 };
852
853 // Constructor for the Node_Backward_Iterator
Node_Backward_Iterator(Node * root,VectorSet & visited,Node_Stack & stack,PhaseCFG & cfg)854 Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_Stack &stack, PhaseCFG &cfg)
855 : _visited(visited), _stack(stack), _cfg(cfg) {
856 // The stack should contain exactly the root
857 stack.clear();
858 stack.push(root, root->outcnt());
859
860 // Clear the visited bits
861 visited.clear();
862 }
863
864 // Iterator for the Node_Backward_Iterator
next()865 Node *Node_Backward_Iterator::next() {
866
867 // If the _stack is empty, then just return NULL: finished.
868 if ( !_stack.size() )
869 return NULL;
870
871 // I visit unvisited not-anti-dependence users first, then anti-dependent
872 // children next. I iterate backwards to support removal of nodes.
873 // The stack holds states consisting of 3 values:
874 // current Def node, flag which indicates 1st/2nd pass, index of current out edge
875 Node *self = (Node*)(((uintptr_t)_stack.node()) & ~1);
876 bool iterate_anti_dep = (((uintptr_t)_stack.node()) & 1);
877 uint idx = MIN2(_stack.index(), self->outcnt()); // Support removal of nodes.
878 _stack.pop();
879
880 // I cycle here when I am entering a deeper level of recursion.
881 // The key variable 'self' was set prior to jumping here.
882 while( 1 ) {
883
884 _visited.set(self->_idx);
885
886 // Now schedule all uses as late as possible.
887 const Node* src = self->is_Proj() ? self->in(0) : self;
888 uint src_rpo = _cfg.get_block_for_node(src)->_rpo;
889
890 // Schedule all nodes in a post-order visit
891 Node *unvisited = NULL; // Unvisited anti-dependent Node, if any
892
893 // Scan for unvisited nodes
894 while (idx > 0) {
895 // For all uses, schedule late
896 Node* n = self->raw_out(--idx); // Use
897
898 // Skip already visited children
899 if ( _visited.test(n->_idx) )
900 continue;
901
902 // do not traverse backward control edges
903 Node *use = n->is_Proj() ? n->in(0) : n;
904 uint use_rpo = _cfg.get_block_for_node(use)->_rpo;
905
906 if ( use_rpo < src_rpo )
907 continue;
908
909 // Phi nodes always precede uses in a basic block
910 if ( use_rpo == src_rpo && use->is_Phi() )
911 continue;
912
913 unvisited = n; // Found unvisited
914
915 // Check for possible-anti-dependent
916 // 1st pass: No such nodes, 2nd pass: Only such nodes.
917 if (n->needs_anti_dependence_check() == iterate_anti_dep) {
918 unvisited = n; // Found unvisited
919 break;
920 }
921 }
922
923 // Did I find an unvisited not-anti-dependent Node?
924 if (!unvisited) {
925 if (!iterate_anti_dep) {
926 // 2nd pass: Iterate over nodes which needs_anti_dependence_check.
927 iterate_anti_dep = true;
928 idx = self->outcnt();
929 continue;
930 }
931 break; // All done with children; post-visit 'self'
932 }
933
934 // Visit the unvisited Node. Contains the obvious push to
935 // indicate I'm entering a deeper level of recursion. I push the
936 // old state onto the _stack and set a new state and loop (recurse).
937 _stack.push((Node*)((uintptr_t)self | (uintptr_t)iterate_anti_dep), idx);
938 self = unvisited;
939 iterate_anti_dep = false;
940 idx = self->outcnt();
941 } // End recursion loop
942
943 return self;
944 }
945
946 //------------------------------ComputeLatenciesBackwards----------------------
947 // Compute the latency of all the instructions.
compute_latencies_backwards(VectorSet & visited,Node_Stack & stack)948 void PhaseCFG::compute_latencies_backwards(VectorSet &visited, Node_Stack &stack) {
949 #ifndef PRODUCT
950 if (trace_opto_pipelining())
951 tty->print("\n#---- ComputeLatenciesBackwards ----\n");
952 #endif
953
954 Node_Backward_Iterator iter((Node *)_root, visited, stack, *this);
955 Node *n;
956
957 // Walk over all the nodes from last to first
958 while ((n = iter.next())) {
959 // Set the latency for the definitions of this instruction
960 partial_latency_of_defs(n);
961 }
962 } // end ComputeLatenciesBackwards
963
964 //------------------------------partial_latency_of_defs------------------------
965 // Compute the latency impact of this node on all defs. This computes
966 // a number that increases as we approach the beginning of the routine.
partial_latency_of_defs(Node * n)967 void PhaseCFG::partial_latency_of_defs(Node *n) {
968 // Set the latency for this instruction
969 #ifndef PRODUCT
970 if (trace_opto_pipelining()) {
971 tty->print("# latency_to_inputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n));
972 dump();
973 }
974 #endif
975
976 if (n->is_Proj()) {
977 n = n->in(0);
978 }
979
980 if (n->is_Root()) {
981 return;
982 }
983
984 uint nlen = n->len();
985 uint use_latency = get_latency_for_node(n);
986 uint use_pre_order = get_block_for_node(n)->_pre_order;
987
988 for (uint j = 0; j < nlen; j++) {
989 Node *def = n->in(j);
990
991 if (!def || def == n) {
992 continue;
993 }
994
995 // Walk backwards thru projections
996 if (def->is_Proj()) {
997 def = def->in(0);
998 }
999
1000 #ifndef PRODUCT
1001 if (trace_opto_pipelining()) {
1002 tty->print("# in(%2d): ", j);
1003 def->dump();
1004 }
1005 #endif
1006
1007 // If the defining block is not known, assume it is ok
1008 Block *def_block = get_block_for_node(def);
1009 uint def_pre_order = def_block ? def_block->_pre_order : 0;
1010
1011 if ((use_pre_order < def_pre_order) || (use_pre_order == def_pre_order && n->is_Phi())) {
1012 continue;
1013 }
1014
1015 uint delta_latency = n->latency(j);
1016 uint current_latency = delta_latency + use_latency;
1017
1018 if (get_latency_for_node(def) < current_latency) {
1019 set_latency_for_node(def, current_latency);
1020 }
1021
1022 #ifndef PRODUCT
1023 if (trace_opto_pipelining()) {
1024 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));
1025 }
1026 #endif
1027 }
1028 }
1029
1030 //------------------------------latency_from_use-------------------------------
1031 // Compute the latency of a specific use
latency_from_use(Node * n,const Node * def,Node * use)1032 int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) {
1033 // If self-reference, return no latency
1034 if (use == n || use->is_Root()) {
1035 return 0;
1036 }
1037
1038 uint def_pre_order = get_block_for_node(def)->_pre_order;
1039 uint latency = 0;
1040
1041 // If the use is not a projection, then it is simple...
1042 if (!use->is_Proj()) {
1043 #ifndef PRODUCT
1044 if (trace_opto_pipelining()) {
1045 tty->print("# out(): ");
1046 use->dump();
1047 }
1048 #endif
1049
1050 uint use_pre_order = get_block_for_node(use)->_pre_order;
1051
1052 if (use_pre_order < def_pre_order)
1053 return 0;
1054
1055 if (use_pre_order == def_pre_order && use->is_Phi())
1056 return 0;
1057
1058 uint nlen = use->len();
1059 uint nl = get_latency_for_node(use);
1060
1061 for ( uint j=0; j<nlen; j++ ) {
1062 if (use->in(j) == n) {
1063 // Change this if we want local latencies
1064 uint ul = use->latency(j);
1065 uint l = ul + nl;
1066 if (latency < l) latency = l;
1067 #ifndef PRODUCT
1068 if (trace_opto_pipelining()) {
1069 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, latency = %d",
1070 nl, j, ul, l, latency);
1071 }
1072 #endif
1073 }
1074 }
1075 } else {
1076 // This is a projection, just grab the latency of the use(s)
1077 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
1078 uint l = latency_from_use(use, def, use->fast_out(j));
1079 if (latency < l) latency = l;
1080 }
1081 }
1082
1083 return latency;
1084 }
1085
1086 //------------------------------latency_from_uses------------------------------
1087 // Compute the latency of this instruction relative to all of it's uses.
1088 // This computes a number that increases as we approach the beginning of the
1089 // routine.
latency_from_uses(Node * n)1090 void PhaseCFG::latency_from_uses(Node *n) {
1091 // Set the latency for this instruction
1092 #ifndef PRODUCT
1093 if (trace_opto_pipelining()) {
1094 tty->print("# latency_from_outputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n));
1095 dump();
1096 }
1097 #endif
1098 uint latency=0;
1099 const Node *def = n->is_Proj() ? n->in(0): n;
1100
1101 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1102 uint l = latency_from_use(n, def, n->fast_out(i));
1103
1104 if (latency < l) latency = l;
1105 }
1106
1107 set_latency_for_node(n, latency);
1108 }
1109
1110 //------------------------------hoist_to_cheaper_block-------------------------
1111 // Pick a block for node self, between early and LCA, that is a cheaper
1112 // alternative to LCA.
hoist_to_cheaper_block(Block * LCA,Block * early,Node * self)1113 Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) {
1114 const double delta = 1+PROB_UNLIKELY_MAG(4);
1115 Block* least = LCA;
1116 double least_freq = least->_freq;
1117 uint target = get_latency_for_node(self);
1118 uint start_latency = get_latency_for_node(LCA->head());
1119 uint end_latency = get_latency_for_node(LCA->get_node(LCA->end_idx()));
1120 bool in_latency = (target <= start_latency);
1121 const Block* root_block = get_block_for_node(_root);
1122
1123 // Turn off latency scheduling if scheduling is just plain off
1124 if (!C->do_scheduling())
1125 in_latency = true;
1126
1127 // Do not hoist (to cover latency) instructions which target a
1128 // single register. Hoisting stretches the live range of the
1129 // single register and may force spilling.
1130 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1131 if (mach && mach->out_RegMask().is_bound1() && mach->out_RegMask().is_NotEmpty())
1132 in_latency = true;
1133
1134 #ifndef PRODUCT
1135 if (trace_opto_pipelining()) {
1136 tty->print("# Find cheaper block for latency %d: ", get_latency_for_node(self));
1137 self->dump();
1138 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1139 LCA->_pre_order,
1140 LCA->head()->_idx,
1141 start_latency,
1142 LCA->get_node(LCA->end_idx())->_idx,
1143 end_latency,
1144 least_freq);
1145 }
1146 #endif
1147
1148 int cand_cnt = 0; // number of candidates tried
1149
1150 // Walk up the dominator tree from LCA (Lowest common ancestor) to
1151 // the earliest legal location. Capture the least execution frequency.
1152 while (LCA != early) {
1153 LCA = LCA->_idom; // Follow up the dominator tree
1154
1155 if (LCA == NULL) {
1156 // Bailout without retry
1157 assert(false, "graph should be schedulable");
1158 C->record_method_not_compilable("late schedule failed: LCA == NULL");
1159 return least;
1160 }
1161
1162 // Don't hoist machine instructions to the root basic block
1163 if (mach && LCA == root_block)
1164 break;
1165
1166 uint start_lat = get_latency_for_node(LCA->head());
1167 uint end_idx = LCA->end_idx();
1168 uint end_lat = get_latency_for_node(LCA->get_node(end_idx));
1169 double LCA_freq = LCA->_freq;
1170 #ifndef PRODUCT
1171 if (trace_opto_pipelining()) {
1172 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1173 LCA->_pre_order, LCA->head()->_idx, start_lat, end_idx, end_lat, LCA_freq);
1174 }
1175 #endif
1176 cand_cnt++;
1177 if (LCA_freq < least_freq || // Better Frequency
1178 (StressGCM && C->randomized_select(cand_cnt)) || // Should be randomly accepted in stress mode
1179 (!StressGCM && // Otherwise, choose with latency
1180 !in_latency && // No block containing latency
1181 LCA_freq < least_freq * delta && // No worse frequency
1182 target >= end_lat && // within latency range
1183 !self->is_iteratively_computed() ) // But don't hoist IV increments
1184 // because they may end up above other uses of their phi forcing
1185 // their result register to be different from their input.
1186 ) {
1187 least = LCA; // Found cheaper block
1188 least_freq = LCA_freq;
1189 start_latency = start_lat;
1190 end_latency = end_lat;
1191 if (target <= start_lat)
1192 in_latency = true;
1193 }
1194 }
1195
1196 #ifndef PRODUCT
1197 if (trace_opto_pipelining()) {
1198 tty->print_cr("# Choose block B%d with start latency=%d and freq=%g",
1199 least->_pre_order, start_latency, least_freq);
1200 }
1201 #endif
1202
1203 // See if the latency needs to be updated
1204 if (target < end_latency) {
1205 #ifndef PRODUCT
1206 if (trace_opto_pipelining()) {
1207 tty->print_cr("# Change latency for [%4d] from %d to %d", self->_idx, target, end_latency);
1208 }
1209 #endif
1210 set_latency_for_node(self, end_latency);
1211 partial_latency_of_defs(self);
1212 }
1213
1214 return least;
1215 }
1216
1217
1218 //------------------------------schedule_late-----------------------------------
1219 // Now schedule all codes as LATE as possible. This is the LCA in the
1220 // dominator tree of all USES of a value. Pick the block with the least
1221 // loop nesting depth that is lowest in the dominator tree.
1222 extern const char must_clone[];
schedule_late(VectorSet & visited,Node_Stack & stack)1223 void PhaseCFG::schedule_late(VectorSet &visited, Node_Stack &stack) {
1224 #ifndef PRODUCT
1225 if (trace_opto_pipelining())
1226 tty->print("\n#---- schedule_late ----\n");
1227 #endif
1228
1229 Node_Backward_Iterator iter((Node *)_root, visited, stack, *this);
1230 Node *self;
1231
1232 // Walk over all the nodes from last to first
1233 while ((self = iter.next())) {
1234 Block* early = get_block_for_node(self); // Earliest legal placement
1235
1236 if (self->is_top()) {
1237 // Top node goes in bb #2 with other constants.
1238 // It must be special-cased, because it has no out edges.
1239 early->add_inst(self);
1240 continue;
1241 }
1242
1243 // No uses, just terminate
1244 if (self->outcnt() == 0) {
1245 assert(self->is_MachProj(), "sanity");
1246 continue; // Must be a dead machine projection
1247 }
1248
1249 // If node is pinned in the block, then no scheduling can be done.
1250 if( self->pinned() ) // Pinned in block?
1251 continue;
1252
1253 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1254 if (mach) {
1255 switch (mach->ideal_Opcode()) {
1256 case Op_CreateEx:
1257 // Don't move exception creation
1258 early->add_inst(self);
1259 continue;
1260 break;
1261 case Op_CheckCastPP: {
1262 // Don't move CheckCastPP nodes away from their input, if the input
1263 // is a rawptr (5071820).
1264 Node *def = self->in(1);
1265 if (def != NULL && def->bottom_type()->base() == Type::RawPtr) {
1266 early->add_inst(self);
1267 #ifdef ASSERT
1268 _raw_oops.push(def);
1269 #endif
1270 continue;
1271 }
1272 break;
1273 }
1274 default:
1275 break;
1276 }
1277 if (C->has_irreducible_loop() && self->bottom_type()->has_memory()) {
1278 // If the CFG is irreducible, keep memory-writing nodes as close as
1279 // possible to their original block (given by the control input). This
1280 // prevents PhaseCFG::hoist_to_cheaper_block() from placing such nodes
1281 // into descendants of their original loop, as in the following example:
1282 //
1283 // Original placement of store in B1 (loop L1):
1284 //
1285 // B1 (L1):
1286 // m1 <- ..
1287 // m2 <- store m1, ..
1288 // B2 (L2):
1289 // jump B2
1290 // B3 (L1):
1291 // .. <- .. m2, ..
1292 //
1293 // Wrong "hoisting" of store to B2 (in loop L2, child of L1):
1294 //
1295 // B1 (L1):
1296 // m1 <- ..
1297 // B2 (L2):
1298 // m2 <- store m1, ..
1299 // # Wrong: m1 and m2 interfere at this point.
1300 // jump B2
1301 // B3 (L1):
1302 // .. <- .. m2, ..
1303 //
1304 // This "hoist inversion" can happen due to CFGLoop::compute_freq()'s
1305 // inaccurate estimation of frequencies for irreducible CFGs, which can
1306 // lead to for example assigning B1 and B3 a higher frequency than B2.
1307 #ifndef PRODUCT
1308 if (trace_opto_pipelining()) {
1309 tty->print_cr("# Irreducible loops: schedule in earliest block B%d:",
1310 early->_pre_order);
1311 self->dump();
1312 }
1313 #endif
1314 schedule_node_into_block(self, early);
1315 continue;
1316 }
1317 }
1318
1319 // Gather LCA of all uses
1320 Block *LCA = NULL;
1321 {
1322 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
1323 // For all uses, find LCA
1324 Node* use = self->fast_out(i);
1325 LCA = raise_LCA_above_use(LCA, use, self, this);
1326 }
1327 guarantee(LCA != NULL, "There must be a LCA");
1328 } // (Hide defs of imax, i from rest of block.)
1329
1330 // Place temps in the block of their use. This isn't a
1331 // requirement for correctness but it reduces useless
1332 // interference between temps and other nodes.
1333 if (mach != NULL && mach->is_MachTemp()) {
1334 map_node_to_block(self, LCA);
1335 LCA->add_inst(self);
1336 continue;
1337 }
1338
1339 // Check if 'self' could be anti-dependent on memory
1340 if (self->needs_anti_dependence_check()) {
1341 // Hoist LCA above possible-defs and insert anti-dependences to
1342 // defs in new LCA block.
1343 LCA = insert_anti_dependences(LCA, self);
1344 }
1345
1346 if (early->_dom_depth > LCA->_dom_depth) {
1347 // Somehow the LCA has moved above the earliest legal point.
1348 // (One way this can happen is via memory_early_block.)
1349 if (C->subsume_loads() == true && !C->failing()) {
1350 // Retry with subsume_loads == false
1351 // If this is the first failure, the sentinel string will "stick"
1352 // to the Compile object, and the C2Compiler will see it and retry.
1353 C->record_failure(C2Compiler::retry_no_subsuming_loads());
1354 } else {
1355 // Bailout without retry when (early->_dom_depth > LCA->_dom_depth)
1356 assert(false, "graph should be schedulable");
1357 C->record_method_not_compilable("late schedule failed: incorrect graph");
1358 }
1359 return;
1360 }
1361
1362 // If there is no opportunity to hoist, then we're done.
1363 // In stress mode, try to hoist even the single operations.
1364 bool try_to_hoist = StressGCM || (LCA != early);
1365
1366 // Must clone guys stay next to use; no hoisting allowed.
1367 // Also cannot hoist guys that alter memory or are otherwise not
1368 // allocatable (hoisting can make a value live longer, leading to
1369 // anti and output dependency problems which are normally resolved
1370 // by the register allocator giving everyone a different register).
1371 if (mach != NULL && must_clone[mach->ideal_Opcode()])
1372 try_to_hoist = false;
1373
1374 Block* late = NULL;
1375 if (try_to_hoist) {
1376 // Now find the block with the least execution frequency.
1377 // Start at the latest schedule and work up to the earliest schedule
1378 // in the dominator tree. Thus the Node will dominate all its uses.
1379 late = hoist_to_cheaper_block(LCA, early, self);
1380 } else {
1381 // Just use the LCA of the uses.
1382 late = LCA;
1383 }
1384
1385 // Put the node into target block
1386 schedule_node_into_block(self, late);
1387
1388 #ifdef ASSERT
1389 if (self->needs_anti_dependence_check()) {
1390 // since precedence edges are only inserted when we're sure they
1391 // are needed make sure that after placement in a block we don't
1392 // need any new precedence edges.
1393 verify_anti_dependences(late, self);
1394 }
1395 #endif
1396 } // Loop until all nodes have been visited
1397
1398 } // end ScheduleLate
1399
1400 //------------------------------GlobalCodeMotion-------------------------------
global_code_motion()1401 void PhaseCFG::global_code_motion() {
1402 ResourceMark rm;
1403
1404 #ifndef PRODUCT
1405 if (trace_opto_pipelining()) {
1406 tty->print("\n---- Start GlobalCodeMotion ----\n");
1407 }
1408 #endif
1409
1410 // Initialize the node to block mapping for things on the proj_list
1411 for (uint i = 0; i < _matcher.number_of_projections(); i++) {
1412 unmap_node_from_block(_matcher.get_projection(i));
1413 }
1414
1415 // Set the basic block for Nodes pinned into blocks
1416 VectorSet visited;
1417 schedule_pinned_nodes(visited);
1418
1419 // Find the earliest Block any instruction can be placed in. Some
1420 // instructions are pinned into Blocks. Unpinned instructions can
1421 // appear in last block in which all their inputs occur.
1422 visited.clear();
1423 Node_Stack stack((C->live_nodes() >> 2) + 16); // pre-grow
1424 if (!schedule_early(visited, stack)) {
1425 // Bailout without retry
1426 C->record_method_not_compilable("early schedule failed");
1427 return;
1428 }
1429
1430 // Build Def-Use edges.
1431 // Compute the latency information (via backwards walk) for all the
1432 // instructions in the graph
1433 _node_latency = new GrowableArray<uint>(); // resource_area allocation
1434
1435 if (C->do_scheduling()) {
1436 compute_latencies_backwards(visited, stack);
1437 }
1438
1439 // Now schedule all codes as LATE as possible. This is the LCA in the
1440 // dominator tree of all USES of a value. Pick the block with the least
1441 // loop nesting depth that is lowest in the dominator tree.
1442 // ( visited.clear() called in schedule_late()->Node_Backward_Iterator() )
1443 schedule_late(visited, stack);
1444 if (C->failing()) {
1445 return;
1446 }
1447
1448 #ifndef PRODUCT
1449 if (trace_opto_pipelining()) {
1450 tty->print("\n---- Detect implicit null checks ----\n");
1451 }
1452 #endif
1453
1454 // Detect implicit-null-check opportunities. Basically, find NULL checks
1455 // with suitable memory ops nearby. Use the memory op to do the NULL check.
1456 // I can generate a memory op if there is not one nearby.
1457 if (C->is_method_compilation()) {
1458 // By reversing the loop direction we get a very minor gain on mpegaudio.
1459 // Feel free to revert to a forward loop for clarity.
1460 // for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) {
1461 for (int i = _matcher._null_check_tests.size() - 2; i >= 0; i -= 2) {
1462 Node* proj = _matcher._null_check_tests[i];
1463 Node* val = _matcher._null_check_tests[i + 1];
1464 Block* block = get_block_for_node(proj);
1465 implicit_null_check(block, proj, val, C->allowed_deopt_reasons());
1466 // The implicit_null_check will only perform the transformation
1467 // if the null branch is truly uncommon, *and* it leads to an
1468 // uncommon trap. Combined with the too_many_traps guards
1469 // above, this prevents SEGV storms reported in 6366351,
1470 // by recompiling offending methods without this optimization.
1471 }
1472 }
1473
1474 bool block_size_threshold_ok = false;
1475 intptr_t *recalc_pressure_nodes = NULL;
1476 if (OptoRegScheduling) {
1477 for (uint i = 0; i < number_of_blocks(); i++) {
1478 Block* block = get_block(i);
1479 if (block->number_of_nodes() > 10) {
1480 block_size_threshold_ok = true;
1481 break;
1482 }
1483 }
1484 }
1485
1486 // Enabling the scheduler for register pressure plus finding blocks of size to schedule for it
1487 // is key to enabling this feature.
1488 PhaseChaitin regalloc(C->unique(), *this, _matcher, true);
1489 ResourceArea live_arena(mtCompiler); // Arena for liveness
1490 ResourceMark rm_live(&live_arena);
1491 PhaseLive live(*this, regalloc._lrg_map.names(), &live_arena, true);
1492 PhaseIFG ifg(&live_arena);
1493 if (OptoRegScheduling && block_size_threshold_ok) {
1494 regalloc.mark_ssa();
1495 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]);
1496 rm_live.reset_to_mark(); // Reclaim working storage
1497 IndexSet::reset_memory(C, &live_arena);
1498 uint node_size = regalloc._lrg_map.max_lrg_id();
1499 ifg.init(node_size); // Empty IFG
1500 regalloc.set_ifg(ifg);
1501 regalloc.set_live(live);
1502 regalloc.gather_lrg_masks(false); // Collect LRG masks
1503 live.compute(node_size); // Compute liveness
1504
1505 recalc_pressure_nodes = NEW_RESOURCE_ARRAY(intptr_t, node_size);
1506 for (uint i = 0; i < node_size; i++) {
1507 recalc_pressure_nodes[i] = 0;
1508 }
1509 }
1510 _regalloc = ®alloc;
1511
1512 #ifndef PRODUCT
1513 if (trace_opto_pipelining()) {
1514 tty->print("\n---- Start Local Scheduling ----\n");
1515 }
1516 #endif
1517
1518 // Schedule locally. Right now a simple topological sort.
1519 // Later, do a real latency aware scheduler.
1520 GrowableArray<int> ready_cnt(C->unique(), C->unique(), -1);
1521 visited.reset();
1522 for (uint i = 0; i < number_of_blocks(); i++) {
1523 Block* block = get_block(i);
1524 if (!schedule_local(block, ready_cnt, visited, recalc_pressure_nodes)) {
1525 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
1526 C->record_method_not_compilable("local schedule failed");
1527 }
1528 _regalloc = NULL;
1529 return;
1530 }
1531 }
1532 _regalloc = NULL;
1533
1534 // If we inserted any instructions between a Call and his CatchNode,
1535 // clone the instructions on all paths below the Catch.
1536 for (uint i = 0; i < number_of_blocks(); i++) {
1537 Block* block = get_block(i);
1538 call_catch_cleanup(block);
1539 }
1540
1541 #ifndef PRODUCT
1542 if (trace_opto_pipelining()) {
1543 tty->print("\n---- After GlobalCodeMotion ----\n");
1544 for (uint i = 0; i < number_of_blocks(); i++) {
1545 Block* block = get_block(i);
1546 block->dump();
1547 }
1548 }
1549 #endif
1550 // Dead.
1551 _node_latency = (GrowableArray<uint> *)((intptr_t)0xdeadbeef);
1552 }
1553
do_global_code_motion()1554 bool PhaseCFG::do_global_code_motion() {
1555
1556 build_dominator_tree();
1557 if (C->failing()) {
1558 return false;
1559 }
1560
1561 NOT_PRODUCT( C->verify_graph_edges(); )
1562
1563 estimate_block_frequency();
1564
1565 global_code_motion();
1566
1567 if (C->failing()) {
1568 return false;
1569 }
1570
1571 return true;
1572 }
1573
1574 //------------------------------Estimate_Block_Frequency-----------------------
1575 // Estimate block frequencies based on IfNode probabilities.
estimate_block_frequency()1576 void PhaseCFG::estimate_block_frequency() {
1577
1578 // Force conditional branches leading to uncommon traps to be unlikely,
1579 // not because we get to the uncommon_trap with less relative frequency,
1580 // but because an uncommon_trap typically causes a deopt, so we only get
1581 // there once.
1582 if (C->do_freq_based_layout()) {
1583 Block_List worklist;
1584 Block* root_blk = get_block(0);
1585 for (uint i = 1; i < root_blk->num_preds(); i++) {
1586 Block *pb = get_block_for_node(root_blk->pred(i));
1587 if (pb->has_uncommon_code()) {
1588 worklist.push(pb);
1589 }
1590 }
1591 while (worklist.size() > 0) {
1592 Block* uct = worklist.pop();
1593 if (uct == get_root_block()) {
1594 continue;
1595 }
1596 for (uint i = 1; i < uct->num_preds(); i++) {
1597 Block *pb = get_block_for_node(uct->pred(i));
1598 if (pb->_num_succs == 1) {
1599 worklist.push(pb);
1600 } else if (pb->num_fall_throughs() == 2) {
1601 pb->update_uncommon_branch(uct);
1602 }
1603 }
1604 }
1605 }
1606
1607 // Create the loop tree and calculate loop depth.
1608 _root_loop = create_loop_tree();
1609 _root_loop->compute_loop_depth(0);
1610
1611 // Compute block frequency of each block, relative to a single loop entry.
1612 _root_loop->compute_freq();
1613
1614 // Adjust all frequencies to be relative to a single method entry
1615 _root_loop->_freq = 1.0;
1616 _root_loop->scale_freq();
1617
1618 // Save outmost loop frequency for LRG frequency threshold
1619 _outer_loop_frequency = _root_loop->outer_loop_freq();
1620
1621 // force paths ending at uncommon traps to be infrequent
1622 if (!C->do_freq_based_layout()) {
1623 Block_List worklist;
1624 Block* root_blk = get_block(0);
1625 for (uint i = 1; i < root_blk->num_preds(); i++) {
1626 Block *pb = get_block_for_node(root_blk->pred(i));
1627 if (pb->has_uncommon_code()) {
1628 worklist.push(pb);
1629 }
1630 }
1631 while (worklist.size() > 0) {
1632 Block* uct = worklist.pop();
1633 uct->_freq = PROB_MIN;
1634 for (uint i = 1; i < uct->num_preds(); i++) {
1635 Block *pb = get_block_for_node(uct->pred(i));
1636 if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) {
1637 worklist.push(pb);
1638 }
1639 }
1640 }
1641 }
1642
1643 #ifdef ASSERT
1644 for (uint i = 0; i < number_of_blocks(); i++) {
1645 Block* b = get_block(i);
1646 assert(b->_freq >= MIN_BLOCK_FREQUENCY, "Register Allocator requires meaningful block frequency");
1647 }
1648 #endif
1649
1650 #ifndef PRODUCT
1651 if (PrintCFGBlockFreq) {
1652 tty->print_cr("CFG Block Frequencies");
1653 _root_loop->dump_tree();
1654 if (Verbose) {
1655 tty->print_cr("PhaseCFG dump");
1656 dump();
1657 tty->print_cr("Node dump");
1658 _root->dump(99999);
1659 }
1660 }
1661 #endif
1662 }
1663
1664 //----------------------------create_loop_tree--------------------------------
1665 // Create a loop tree from the CFG
create_loop_tree()1666 CFGLoop* PhaseCFG::create_loop_tree() {
1667
1668 #ifdef ASSERT
1669 assert(get_block(0) == get_root_block(), "first block should be root block");
1670 for (uint i = 0; i < number_of_blocks(); i++) {
1671 Block* block = get_block(i);
1672 // Check that _loop field are clear...we could clear them if not.
1673 assert(block->_loop == NULL, "clear _loop expected");
1674 // Sanity check that the RPO numbering is reflected in the _blocks array.
1675 // It doesn't have to be for the loop tree to be built, but if it is not,
1676 // then the blocks have been reordered since dom graph building...which
1677 // may question the RPO numbering
1678 assert(block->_rpo == i, "unexpected reverse post order number");
1679 }
1680 #endif
1681
1682 int idct = 0;
1683 CFGLoop* root_loop = new CFGLoop(idct++);
1684
1685 Block_List worklist;
1686
1687 // Assign blocks to loops
1688 for(uint i = number_of_blocks() - 1; i > 0; i-- ) { // skip Root block
1689 Block* block = get_block(i);
1690
1691 if (block->head()->is_Loop()) {
1692 Block* loop_head = block;
1693 assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1694 Node* tail_n = loop_head->pred(LoopNode::LoopBackControl);
1695 Block* tail = get_block_for_node(tail_n);
1696
1697 // Defensively filter out Loop nodes for non-single-entry loops.
1698 // For all reasonable loops, the head occurs before the tail in RPO.
1699 if (i <= tail->_rpo) {
1700
1701 // The tail and (recursive) predecessors of the tail
1702 // are made members of a new loop.
1703
1704 assert(worklist.size() == 0, "nonempty worklist");
1705 CFGLoop* nloop = new CFGLoop(idct++);
1706 assert(loop_head->_loop == NULL, "just checking");
1707 loop_head->_loop = nloop;
1708 // Add to nloop so push_pred() will skip over inner loops
1709 nloop->add_member(loop_head);
1710 nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, this);
1711
1712 while (worklist.size() > 0) {
1713 Block* member = worklist.pop();
1714 if (member != loop_head) {
1715 for (uint j = 1; j < member->num_preds(); j++) {
1716 nloop->push_pred(member, j, worklist, this);
1717 }
1718 }
1719 }
1720 }
1721 }
1722 }
1723
1724 // Create a member list for each loop consisting
1725 // of both blocks and (immediate child) loops.
1726 for (uint i = 0; i < number_of_blocks(); i++) {
1727 Block* block = get_block(i);
1728 CFGLoop* lp = block->_loop;
1729 if (lp == NULL) {
1730 // Not assigned to a loop. Add it to the method's pseudo loop.
1731 block->_loop = root_loop;
1732 lp = root_loop;
1733 }
1734 if (lp == root_loop || block != lp->head()) { // loop heads are already members
1735 lp->add_member(block);
1736 }
1737 if (lp != root_loop) {
1738 if (lp->parent() == NULL) {
1739 // Not a nested loop. Make it a child of the method's pseudo loop.
1740 root_loop->add_nested_loop(lp);
1741 }
1742 if (block == lp->head()) {
1743 // Add nested loop to member list of parent loop.
1744 lp->parent()->add_member(lp);
1745 }
1746 }
1747 }
1748
1749 return root_loop;
1750 }
1751
1752 //------------------------------push_pred--------------------------------------
push_pred(Block * blk,int i,Block_List & worklist,PhaseCFG * cfg)1753 void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg) {
1754 Node* pred_n = blk->pred(i);
1755 Block* pred = cfg->get_block_for_node(pred_n);
1756 CFGLoop *pred_loop = pred->_loop;
1757 if (pred_loop == NULL) {
1758 // Filter out blocks for non-single-entry loops.
1759 // For all reasonable loops, the head occurs before the tail in RPO.
1760 if (pred->_rpo > head()->_rpo) {
1761 pred->_loop = this;
1762 worklist.push(pred);
1763 }
1764 } else if (pred_loop != this) {
1765 // Nested loop.
1766 while (pred_loop->_parent != NULL && pred_loop->_parent != this) {
1767 pred_loop = pred_loop->_parent;
1768 }
1769 // Make pred's loop be a child
1770 if (pred_loop->_parent == NULL) {
1771 add_nested_loop(pred_loop);
1772 // Continue with loop entry predecessor.
1773 Block* pred_head = pred_loop->head();
1774 assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1775 assert(pred_head != head(), "loop head in only one loop");
1776 push_pred(pred_head, LoopNode::EntryControl, worklist, cfg);
1777 } else {
1778 assert(pred_loop->_parent == this && _parent == NULL, "just checking");
1779 }
1780 }
1781 }
1782
1783 //------------------------------add_nested_loop--------------------------------
1784 // Make cl a child of the current loop in the loop tree.
add_nested_loop(CFGLoop * cl)1785 void CFGLoop::add_nested_loop(CFGLoop* cl) {
1786 assert(_parent == NULL, "no parent yet");
1787 assert(cl != this, "not my own parent");
1788 cl->_parent = this;
1789 CFGLoop* ch = _child;
1790 if (ch == NULL) {
1791 _child = cl;
1792 } else {
1793 while (ch->_sibling != NULL) { ch = ch->_sibling; }
1794 ch->_sibling = cl;
1795 }
1796 }
1797
1798 //------------------------------compute_loop_depth-----------------------------
1799 // Store the loop depth in each CFGLoop object.
1800 // Recursively walk the children to do the same for them.
compute_loop_depth(int depth)1801 void CFGLoop::compute_loop_depth(int depth) {
1802 _depth = depth;
1803 CFGLoop* ch = _child;
1804 while (ch != NULL) {
1805 ch->compute_loop_depth(depth + 1);
1806 ch = ch->_sibling;
1807 }
1808 }
1809
1810 //------------------------------compute_freq-----------------------------------
1811 // Compute the frequency of each block and loop, relative to a single entry
1812 // into the dominating loop head.
compute_freq()1813 void CFGLoop::compute_freq() {
1814 // Bottom up traversal of loop tree (visit inner loops first.)
1815 // Set loop head frequency to 1.0, then transitively
1816 // compute frequency for all successors in the loop,
1817 // as well as for each exit edge. Inner loops are
1818 // treated as single blocks with loop exit targets
1819 // as the successor blocks.
1820
1821 // Nested loops first
1822 CFGLoop* ch = _child;
1823 while (ch != NULL) {
1824 ch->compute_freq();
1825 ch = ch->_sibling;
1826 }
1827 assert (_members.length() > 0, "no empty loops");
1828 Block* hd = head();
1829 hd->_freq = 1.0;
1830 for (int i = 0; i < _members.length(); i++) {
1831 CFGElement* s = _members.at(i);
1832 double freq = s->_freq;
1833 if (s->is_block()) {
1834 Block* b = s->as_Block();
1835 for (uint j = 0; j < b->_num_succs; j++) {
1836 Block* sb = b->_succs[j];
1837 update_succ_freq(sb, freq * b->succ_prob(j));
1838 }
1839 } else {
1840 CFGLoop* lp = s->as_CFGLoop();
1841 assert(lp->_parent == this, "immediate child");
1842 for (int k = 0; k < lp->_exits.length(); k++) {
1843 Block* eb = lp->_exits.at(k).get_target();
1844 double prob = lp->_exits.at(k).get_prob();
1845 update_succ_freq(eb, freq * prob);
1846 }
1847 }
1848 }
1849
1850 // For all loops other than the outer, "method" loop,
1851 // sum and normalize the exit probability. The "method" loop
1852 // should keep the initial exit probability of 1, so that
1853 // inner blocks do not get erroneously scaled.
1854 if (_depth != 0) {
1855 // Total the exit probabilities for this loop.
1856 double exits_sum = 0.0f;
1857 for (int i = 0; i < _exits.length(); i++) {
1858 exits_sum += _exits.at(i).get_prob();
1859 }
1860
1861 // Normalize the exit probabilities. Until now, the
1862 // probabilities estimate the possibility of exit per
1863 // a single loop iteration; afterward, they estimate
1864 // the probability of exit per loop entry.
1865 for (int i = 0; i < _exits.length(); i++) {
1866 Block* et = _exits.at(i).get_target();
1867 float new_prob = 0.0f;
1868 if (_exits.at(i).get_prob() > 0.0f) {
1869 new_prob = _exits.at(i).get_prob() / exits_sum;
1870 }
1871 BlockProbPair bpp(et, new_prob);
1872 _exits.at_put(i, bpp);
1873 }
1874
1875 // Save the total, but guard against unreasonable probability,
1876 // as the value is used to estimate the loop trip count.
1877 // An infinite trip count would blur relative block
1878 // frequencies.
1879 if (exits_sum > 1.0f) exits_sum = 1.0;
1880 if (exits_sum < PROB_MIN) exits_sum = PROB_MIN;
1881 _exit_prob = exits_sum;
1882 }
1883 }
1884
1885 //------------------------------succ_prob-------------------------------------
1886 // Determine the probability of reaching successor 'i' from the receiver block.
succ_prob(uint i)1887 float Block::succ_prob(uint i) {
1888 int eidx = end_idx();
1889 Node *n = get_node(eidx); // Get ending Node
1890
1891 int op = n->Opcode();
1892 if (n->is_Mach()) {
1893 if (n->is_MachNullCheck()) {
1894 // Can only reach here if called after lcm. The original Op_If is gone,
1895 // so we attempt to infer the probability from one or both of the
1896 // successor blocks.
1897 assert(_num_succs == 2, "expecting 2 successors of a null check");
1898 // If either successor has only one predecessor, then the
1899 // probability estimate can be derived using the
1900 // relative frequency of the successor and this block.
1901 if (_succs[i]->num_preds() == 2) {
1902 return _succs[i]->_freq / _freq;
1903 } else if (_succs[1-i]->num_preds() == 2) {
1904 return 1 - (_succs[1-i]->_freq / _freq);
1905 } else {
1906 // Estimate using both successor frequencies
1907 float freq = _succs[i]->_freq;
1908 return freq / (freq + _succs[1-i]->_freq);
1909 }
1910 }
1911 op = n->as_Mach()->ideal_Opcode();
1912 }
1913
1914
1915 // Switch on branch type
1916 switch( op ) {
1917 case Op_CountedLoopEnd:
1918 case Op_If: {
1919 assert (i < 2, "just checking");
1920 // Conditionals pass on only part of their frequency
1921 float prob = n->as_MachIf()->_prob;
1922 assert(prob >= 0.0 && prob <= 1.0, "out of range probability");
1923 // If succ[i] is the FALSE branch, invert path info
1924 if( get_node(i + eidx + 1)->Opcode() == Op_IfFalse ) {
1925 return 1.0f - prob; // not taken
1926 } else {
1927 return prob; // taken
1928 }
1929 }
1930
1931 case Op_Jump:
1932 return n->as_MachJump()->_probs[get_node(i + eidx + 1)->as_JumpProj()->_con];
1933
1934 case Op_Catch: {
1935 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
1936 if (ci->_con == CatchProjNode::fall_through_index) {
1937 // Fall-thru path gets the lion's share.
1938 return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs;
1939 } else {
1940 // Presume exceptional paths are equally unlikely
1941 return PROB_UNLIKELY_MAG(5);
1942 }
1943 }
1944
1945 case Op_Root:
1946 case Op_Goto:
1947 // Pass frequency straight thru to target
1948 return 1.0f;
1949
1950 case Op_NeverBranch:
1951 return 0.0f;
1952
1953 case Op_TailCall:
1954 case Op_TailJump:
1955 case Op_Return:
1956 case Op_Halt:
1957 case Op_Rethrow:
1958 // Do not push out freq to root block
1959 return 0.0f;
1960
1961 default:
1962 ShouldNotReachHere();
1963 }
1964
1965 return 0.0f;
1966 }
1967
1968 //------------------------------num_fall_throughs-----------------------------
1969 // Return the number of fall-through candidates for a block
num_fall_throughs()1970 int Block::num_fall_throughs() {
1971 int eidx = end_idx();
1972 Node *n = get_node(eidx); // Get ending Node
1973
1974 int op = n->Opcode();
1975 if (n->is_Mach()) {
1976 if (n->is_MachNullCheck()) {
1977 // In theory, either side can fall-thru, for simplicity sake,
1978 // let's say only the false branch can now.
1979 return 1;
1980 }
1981 op = n->as_Mach()->ideal_Opcode();
1982 }
1983
1984 // Switch on branch type
1985 switch( op ) {
1986 case Op_CountedLoopEnd:
1987 case Op_If:
1988 return 2;
1989
1990 case Op_Root:
1991 case Op_Goto:
1992 return 1;
1993
1994 case Op_Catch: {
1995 for (uint i = 0; i < _num_succs; i++) {
1996 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
1997 if (ci->_con == CatchProjNode::fall_through_index) {
1998 return 1;
1999 }
2000 }
2001 return 0;
2002 }
2003
2004 case Op_Jump:
2005 case Op_NeverBranch:
2006 case Op_TailCall:
2007 case Op_TailJump:
2008 case Op_Return:
2009 case Op_Halt:
2010 case Op_Rethrow:
2011 return 0;
2012
2013 default:
2014 ShouldNotReachHere();
2015 }
2016
2017 return 0;
2018 }
2019
2020 //------------------------------succ_fall_through-----------------------------
2021 // Return true if a specific successor could be fall-through target.
succ_fall_through(uint i)2022 bool Block::succ_fall_through(uint i) {
2023 int eidx = end_idx();
2024 Node *n = get_node(eidx); // Get ending Node
2025
2026 int op = n->Opcode();
2027 if (n->is_Mach()) {
2028 if (n->is_MachNullCheck()) {
2029 // In theory, either side can fall-thru, for simplicity sake,
2030 // let's say only the false branch can now.
2031 return get_node(i + eidx + 1)->Opcode() == Op_IfFalse;
2032 }
2033 op = n->as_Mach()->ideal_Opcode();
2034 }
2035
2036 // Switch on branch type
2037 switch( op ) {
2038 case Op_CountedLoopEnd:
2039 case Op_If:
2040 case Op_Root:
2041 case Op_Goto:
2042 return true;
2043
2044 case Op_Catch: {
2045 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
2046 return ci->_con == CatchProjNode::fall_through_index;
2047 }
2048
2049 case Op_Jump:
2050 case Op_NeverBranch:
2051 case Op_TailCall:
2052 case Op_TailJump:
2053 case Op_Return:
2054 case Op_Halt:
2055 case Op_Rethrow:
2056 return false;
2057
2058 default:
2059 ShouldNotReachHere();
2060 }
2061
2062 return false;
2063 }
2064
2065 //------------------------------update_uncommon_branch------------------------
2066 // Update the probability of a two-branch to be uncommon
update_uncommon_branch(Block * ub)2067 void Block::update_uncommon_branch(Block* ub) {
2068 int eidx = end_idx();
2069 Node *n = get_node(eidx); // Get ending Node
2070
2071 int op = n->as_Mach()->ideal_Opcode();
2072
2073 assert(op == Op_CountedLoopEnd || op == Op_If, "must be a If");
2074 assert(num_fall_throughs() == 2, "must be a two way branch block");
2075
2076 // Which successor is ub?
2077 uint s;
2078 for (s = 0; s <_num_succs; s++) {
2079 if (_succs[s] == ub) break;
2080 }
2081 assert(s < 2, "uncommon successor must be found");
2082
2083 // If ub is the true path, make the proability small, else
2084 // ub is the false path, and make the probability large
2085 bool invert = (get_node(s + eidx + 1)->Opcode() == Op_IfFalse);
2086
2087 // Get existing probability
2088 float p = n->as_MachIf()->_prob;
2089
2090 if (invert) p = 1.0 - p;
2091 if (p > PROB_MIN) {
2092 p = PROB_MIN;
2093 }
2094 if (invert) p = 1.0 - p;
2095
2096 n->as_MachIf()->_prob = p;
2097 }
2098
2099 //------------------------------update_succ_freq-------------------------------
2100 // Update the appropriate frequency associated with block 'b', a successor of
2101 // a block in this loop.
update_succ_freq(Block * b,double freq)2102 void CFGLoop::update_succ_freq(Block* b, double freq) {
2103 if (b->_loop == this) {
2104 if (b == head()) {
2105 // back branch within the loop
2106 // Do nothing now, the loop carried frequency will be
2107 // adjust later in scale_freq().
2108 } else {
2109 // simple branch within the loop
2110 b->_freq += freq;
2111 }
2112 } else if (!in_loop_nest(b)) {
2113 // branch is exit from this loop
2114 BlockProbPair bpp(b, freq);
2115 _exits.append(bpp);
2116 } else {
2117 // branch into nested loop
2118 CFGLoop* ch = b->_loop;
2119 ch->_freq += freq;
2120 }
2121 }
2122
2123 //------------------------------in_loop_nest-----------------------------------
2124 // Determine if block b is in the receiver's loop nest.
in_loop_nest(Block * b)2125 bool CFGLoop::in_loop_nest(Block* b) {
2126 int depth = _depth;
2127 CFGLoop* b_loop = b->_loop;
2128 int b_depth = b_loop->_depth;
2129 if (depth == b_depth) {
2130 return true;
2131 }
2132 while (b_depth > depth) {
2133 b_loop = b_loop->_parent;
2134 b_depth = b_loop->_depth;
2135 }
2136 return b_loop == this;
2137 }
2138
2139 //------------------------------scale_freq-------------------------------------
2140 // Scale frequency of loops and blocks by trip counts from outer loops
2141 // Do a top down traversal of loop tree (visit outer loops first.)
scale_freq()2142 void CFGLoop::scale_freq() {
2143 double loop_freq = _freq * trip_count();
2144 _freq = loop_freq;
2145 for (int i = 0; i < _members.length(); i++) {
2146 CFGElement* s = _members.at(i);
2147 double block_freq = s->_freq * loop_freq;
2148 if (g_isnan(block_freq) || block_freq < MIN_BLOCK_FREQUENCY)
2149 block_freq = MIN_BLOCK_FREQUENCY;
2150 s->_freq = block_freq;
2151 }
2152 CFGLoop* ch = _child;
2153 while (ch != NULL) {
2154 ch->scale_freq();
2155 ch = ch->_sibling;
2156 }
2157 }
2158
2159 // Frequency of outer loop
outer_loop_freq() const2160 double CFGLoop::outer_loop_freq() const {
2161 if (_child != NULL) {
2162 return _child->_freq;
2163 }
2164 return _freq;
2165 }
2166
2167 #ifndef PRODUCT
2168 //------------------------------dump_tree--------------------------------------
dump_tree() const2169 void CFGLoop::dump_tree() const {
2170 dump();
2171 if (_child != NULL) _child->dump_tree();
2172 if (_sibling != NULL) _sibling->dump_tree();
2173 }
2174
2175 //------------------------------dump-------------------------------------------
dump() const2176 void CFGLoop::dump() const {
2177 for (int i = 0; i < _depth; i++) tty->print(" ");
2178 tty->print("%s: %d trip_count: %6.0f freq: %6.0f\n",
2179 _depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq);
2180 for (int i = 0; i < _depth; i++) tty->print(" ");
2181 tty->print(" members:");
2182 int k = 0;
2183 for (int i = 0; i < _members.length(); i++) {
2184 if (k++ >= 6) {
2185 tty->print("\n ");
2186 for (int j = 0; j < _depth+1; j++) tty->print(" ");
2187 k = 0;
2188 }
2189 CFGElement *s = _members.at(i);
2190 if (s->is_block()) {
2191 Block *b = s->as_Block();
2192 tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq);
2193 } else {
2194 CFGLoop* lp = s->as_CFGLoop();
2195 tty->print(" L%d(%6.3f)", lp->_id, lp->_freq);
2196 }
2197 }
2198 tty->print("\n");
2199 for (int i = 0; i < _depth; i++) tty->print(" ");
2200 tty->print(" exits: ");
2201 k = 0;
2202 for (int i = 0; i < _exits.length(); i++) {
2203 if (k++ >= 7) {
2204 tty->print("\n ");
2205 for (int j = 0; j < _depth+1; j++) tty->print(" ");
2206 k = 0;
2207 }
2208 Block *blk = _exits.at(i).get_target();
2209 double prob = _exits.at(i).get_prob();
2210 tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100));
2211 }
2212 tty->print("\n");
2213 }
2214 #endif
2215