1 /*
2 * Copyright (c) 1997, 2025, 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 "libadt/vectset.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "memory/resourceArea.hpp"
28 #include "opto/block.hpp"
29 #include "opto/c2compiler.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/machnode.hpp"
33 #include "opto/opcodes.hpp"
34 #include "opto/phaseX.hpp"
35 #include "opto/rootnode.hpp"
36 #include "opto/runtime.hpp"
37 #include "opto/chaitin.hpp"
38 #include "runtime/deoptimization.hpp"
39
40 // Portions of code courtesy of Clifford Click
41
42 // Optimization - Graph Style
43
44 // To avoid float value underflow
45 #define MIN_BLOCK_FREQUENCY 1.e-35f
46
47 //----------------------------schedule_node_into_block-------------------------
48 // Insert node n into block b. Look for projections of n and make sure they
49 // are in b also.
50 void PhaseCFG::schedule_node_into_block( Node *n, Block *b ) {
51 // Set basic block of n, Add n to b,
52 map_node_to_block(n, b);
53 b->add_inst(n);
54
55 // After Matching, nearly any old Node may have projections trailing it.
56 // These are usually machine-dependent flags. In any case, they might
57 // float to another block below this one. Move them up.
58 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
59 Node* use = n->fast_out(i);
60 if (use->is_Proj()) {
61 Block* buse = get_block_for_node(use);
62 if (buse != b) { // In wrong block?
63 if (buse != nullptr) {
64 buse->find_remove(use); // Remove from wrong block
65 }
66 map_node_to_block(use, b);
67 b->add_inst(use);
68 }
69 }
70 }
71 }
72
73 //----------------------------replace_block_proj_ctrl-------------------------
74 // Nodes that have is_block_proj() nodes as their control need to use
75 // the appropriate Region for their actual block as their control since
76 // the projection will be in a predecessor block.
77 void PhaseCFG::replace_block_proj_ctrl( Node *n ) {
78 const Node *in0 = n->in(0);
79 assert(in0 != nullptr, "Only control-dependent");
80 const Node *p = in0->is_block_proj();
81 if (p != nullptr && p != n) { // Control from a block projection?
82 assert(!n->pinned() || n->is_MachConstantBase(), "only pinned MachConstantBase node is expected here");
83 // Find trailing Region
84 Block *pb = get_block_for_node(in0); // Block-projection already has basic block
85 uint j = 0;
86 if (pb->_num_succs != 1) { // More then 1 successor?
87 // Search for successor
88 uint max = pb->number_of_nodes();
89 assert( max > 1, "" );
90 uint start = max - pb->_num_succs;
91 // Find which output path belongs to projection
92 for (j = start; j < max; j++) {
93 if( pb->get_node(j) == in0 )
94 break;
95 }
96 assert( j < max, "must find" );
97 // Change control to match head of successor basic block
98 j -= start;
99 }
100 n->set_req(0, pb->_succs[j]->head());
101 }
102 }
103
104 bool PhaseCFG::is_dominator(Node* dom_node, Node* node) {
105 assert(is_CFG(node) && is_CFG(dom_node), "node and dom_node must be CFG nodes");
106 if (dom_node == node) {
107 return true;
108 }
109 Block* d = find_block_for_node(dom_node);
110 Block* n = find_block_for_node(node);
111 assert(n != nullptr && d != nullptr, "blocks must exist");
112
113 if (d == n) {
114 if (dom_node->is_block_start()) {
115 return true;
116 }
117 if (node->is_block_start()) {
118 return false;
119 }
120 if (dom_node->is_block_proj()) {
121 return false;
122 }
123 if (node->is_block_proj()) {
124 return true;
125 }
126
127 assert(is_control_proj_or_safepoint(node), "node must be control projection or safepoint");
128 assert(is_control_proj_or_safepoint(dom_node), "dom_node must be control projection or safepoint");
129
130 // Neither 'node' nor 'dom_node' is a block start or block projection.
131 // Check if 'dom_node' is above 'node' in the control graph.
132 if (is_dominating_control(dom_node, node)) {
133 return true;
134 }
135
136 #ifdef ASSERT
137 // If 'dom_node' does not dominate 'node' then 'node' has to dominate 'dom_node'
138 if (!is_dominating_control(node, dom_node)) {
139 node->dump();
140 dom_node->dump();
141 assert(false, "neither dom_node nor node dominates the other");
142 }
143 #endif
144
145 return false;
146 }
147 return d->dom_lca(n) == d;
148 }
149
150 bool PhaseCFG::is_CFG(Node* n) {
151 return n->is_block_proj() || n->is_block_start() || is_control_proj_or_safepoint(n);
152 }
153
154 bool PhaseCFG::is_control_proj_or_safepoint(Node* n) const {
155 bool result = (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_SafePoint) || (n->is_Proj() && n->as_Proj()->bottom_type() == Type::CONTROL);
156 assert(!result || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_SafePoint)
157 || (n->is_Proj() && n->as_Proj()->_con == 0), "If control projection, it must be projection 0");
158 return result;
159 }
160
161 Block* PhaseCFG::find_block_for_node(Node* n) const {
162 if (n->is_block_start() || n->is_block_proj()) {
163 return get_block_for_node(n);
164 } else {
165 // Walk the control graph up if 'n' is not a block start nor a block projection. In this case 'n' must be
166 // an unmatched control projection or a not yet matched safepoint precedence edge in the middle of a block.
167 assert(is_control_proj_or_safepoint(n), "must be control projection or safepoint");
168 Node* ctrl = n->in(0);
169 while (!ctrl->is_block_start()) {
170 ctrl = ctrl->in(0);
171 }
172 return get_block_for_node(ctrl);
173 }
174 }
175
176 // Walk up the control graph from 'n' and check if 'dom_ctrl' is found.
177 bool PhaseCFG::is_dominating_control(Node* dom_ctrl, Node* n) {
178 Node* ctrl = n->in(0);
179 while (!ctrl->is_block_start()) {
180 if (ctrl == dom_ctrl) {
181 return true;
182 }
183 ctrl = ctrl->in(0);
184 }
185 return false;
186 }
187
188
189 //------------------------------schedule_pinned_nodes--------------------------
190 // Set the basic block for Nodes pinned into blocks
191 void PhaseCFG::schedule_pinned_nodes(VectorSet &visited) {
192 // Allocate node stack of size C->live_nodes()+8 to avoid frequent realloc
193 GrowableArray <Node*> spstack(C->live_nodes() + 8);
194 spstack.push(_root);
195 while (spstack.is_nonempty()) {
196 Node* node = spstack.pop();
197 if (!visited.test_set(node->_idx)) { // Test node and flag it as visited
198 if (node->pinned() && !has_block(node)) { // Pinned? Nail it down!
199 assert(node->in(0), "pinned Node must have Control");
200 // Before setting block replace block_proj control edge
201 replace_block_proj_ctrl(node);
202 Node* input = node->in(0);
203 while (!input->is_block_start()) {
204 input = input->in(0);
205 }
206 Block* block = get_block_for_node(input); // Basic block of controlling input
207 schedule_node_into_block(node, block);
208 }
209
210 // If the node has precedence edges (added when CastPP nodes are
211 // removed in final_graph_reshaping), fix the control of the
212 // node to cover the precedence edges and remove the
213 // dependencies.
214 Node* n = nullptr;
215 for (uint i = node->len()-1; i >= node->req(); i--) {
216 Node* m = node->in(i);
217 if (m == nullptr) continue;
218 assert(is_CFG(m), "must be a CFG node");
219 node->rm_prec(i);
220 if (n == nullptr) {
221 n = m;
222 } else {
223 assert(is_dominator(n, m) || is_dominator(m, n), "one must dominate the other");
224 n = is_dominator(n, m) ? m : n;
225 }
226 }
227 if (n != nullptr) {
228 assert(node->in(0), "control should have been set");
229 assert(is_dominator(n, node->in(0)) || is_dominator(node->in(0), n), "one must dominate the other");
230 if (!is_dominator(n, node->in(0))) {
231 node->set_req(0, n);
232 }
233 }
234
235 // process all inputs that are non null
236 for (int i = node->len()-1; i >= 0; --i) {
237 if (node->in(i) != nullptr) {
238 spstack.push(node->in(i));
239 }
240 }
241 }
242 }
243 }
244
245 // Assert that new input b2 is dominated by all previous inputs.
246 // Check this by by seeing that it is dominated by b1, the deepest
247 // input observed until b2.
248 static void assert_dom(Block* b1, Block* b2, Node* n, const PhaseCFG* cfg) {
249 if (b1 == nullptr) return;
250 assert(b1->_dom_depth < b2->_dom_depth, "sanity");
251 Block* tmp = b2;
252 while (tmp != b1 && tmp != nullptr) {
253 tmp = tmp->_idom;
254 }
255 if (tmp != b1) {
256 #ifdef ASSERT
257 // Detected an unschedulable graph. Print some nice stuff and die.
258 tty->print_cr("!!! Unschedulable graph !!!");
259 for (uint j=0; j<n->len(); j++) { // For all inputs
260 Node* inn = n->in(j); // Get input
261 if (inn == nullptr) continue; // Ignore null, missing inputs
262 Block* inb = cfg->get_block_for_node(inn);
263 tty->print("B%d idom=B%d depth=%2d ",inb->_pre_order,
264 inb->_idom ? inb->_idom->_pre_order : 0, inb->_dom_depth);
265 inn->dump();
266 }
267 tty->print("Failing node: ");
268 n->dump();
269 assert(false, "unschedulable graph");
270 #endif
271 cfg->C->record_failure("unschedulable graph");
272 }
273 }
274
275 static Block* find_deepest_input(Node* n, const PhaseCFG* cfg) {
276 // Find the last input dominated by all other inputs.
277 Block* deepb = nullptr; // Deepest block so far
278 int deepb_dom_depth = 0;
279 for (uint k = 0; k < n->len(); k++) { // For all inputs
280 Node* inn = n->in(k); // Get input
281 if (inn == nullptr) continue; // Ignore null, missing inputs
282 Block* inb = cfg->get_block_for_node(inn);
283 assert(inb != nullptr, "must already have scheduled this input");
284 if (deepb_dom_depth < (int) inb->_dom_depth) {
285 // The new inb must be dominated by the previous deepb.
286 // The various inputs must be linearly ordered in the dom
287 // tree, or else there will not be a unique deepest block.
288 assert_dom(deepb, inb, n, cfg);
289 if (cfg->C->failing()) {
290 return nullptr;
291 }
292 deepb = inb; // Save deepest block
293 deepb_dom_depth = deepb->_dom_depth;
294 }
295 }
296 assert(deepb != nullptr, "must be at least one input to n");
297 return deepb;
298 }
299
300
301 //------------------------------schedule_early---------------------------------
302 // Find the earliest Block any instruction can be placed in. Some instructions
303 // are pinned into Blocks. Unpinned instructions can appear in last block in
304 // which all their inputs occur.
305 bool PhaseCFG::schedule_early(VectorSet &visited, Node_Stack &roots) {
306 // Allocate stack with enough space to avoid frequent realloc
307 Node_Stack nstack(roots.size() + 8);
308 // _root will be processed among C->top() inputs
309 roots.push(C->top(), 0);
310 visited.set(C->top()->_idx);
311
312 while (roots.size() != 0) {
313 // Use local variables nstack_top_n & nstack_top_i to cache values
314 // on stack's top.
315 Node* parent_node = roots.node();
316 uint input_index = 0;
317 roots.pop();
318
319 while (true) {
320 if (input_index == 0) {
321 // Fixup some control. Constants without control get attached
322 // to root and nodes that use is_block_proj() nodes should be attached
323 // to the region that starts their block.
324 const Node* control_input = parent_node->in(0);
325 if (control_input != nullptr) {
326 replace_block_proj_ctrl(parent_node);
327 } else {
328 // Is a constant with NO inputs?
329 if (parent_node->req() == 1) {
330 parent_node->set_req(0, _root);
331 }
332 }
333 }
334
335 // First, visit all inputs and force them to get a block. If an
336 // input is already in a block we quit following inputs (to avoid
337 // cycles). Instead we put that Node on a worklist to be handled
338 // later (since IT'S inputs may not have a block yet).
339
340 // Assume all n's inputs will be processed
341 bool done = true;
342
343 while (input_index < parent_node->len()) {
344 Node* in = parent_node->in(input_index++);
345 if (in == nullptr) {
346 continue;
347 }
348
349 int is_visited = visited.test_set(in->_idx);
350 if (!has_block(in)) {
351 if (is_visited) {
352 assert(false, "graph should be schedulable");
353 return false;
354 }
355 // Save parent node and next input's index.
356 nstack.push(parent_node, input_index);
357 // Process current input now.
358 parent_node = in;
359 input_index = 0;
360 // Not all n's inputs processed.
361 done = false;
362 break;
363 } else if (!is_visited) {
364 // Visit this guy later, using worklist
365 roots.push(in, 0);
366 }
367 }
368
369 if (done) {
370 // All of n's inputs have been processed, complete post-processing.
371
372 // Some instructions are pinned into a block. These include Region,
373 // Phi, Start, Return, and other control-dependent instructions and
374 // any projections which depend on them.
375 if (!parent_node->pinned()) {
376 // Set earliest legal block.
377 Block* earliest_block = find_deepest_input(parent_node, this);
378 if (C->failing()) {
379 return false;
380 }
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.
406 Block* Block::dom_lca(Block* LCA) {
407 if (LCA == nullptr || 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.
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 == nullptr) 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.
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.
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 != nullptr) mem_inputs[mem_inputs_length++] = base;
511 if (index != nullptr) mem_inputs[mem_inputs_length++] = index;
512 if (store != nullptr) mem_inputs[mem_inputs_length++] = store;
513
514 // In the comparison 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) != nullptr) mem_inputs[mem_inputs_length++] = load->in(0);
523
524 Block* deepb = nullptr; // 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 assert_dom(deepb, inb, load, cfg);
533 if (cfg->C->failing()) {
534 return nullptr;
535 }
536 deepb = inb; // Save deepest block
537 deepb_dom_depth = deepb->_dom_depth;
538 }
539 }
540 early = deepb;
541 }
542
543 return early;
544 }
545
546 // This function is used by insert_anti_dependences to find unrelated loads for stores in implicit null checks.
547 bool PhaseCFG::unrelated_load_in_store_null_block(Node* store, Node* load) {
548 // We expect an anti-dependence edge from 'load' to 'store', except when
549 // implicit_null_check() has hoisted 'store' above its early block to
550 // perform an implicit null check, and 'load' is placed in the null
551 // block. In this case it is safe to ignore the anti-dependence, as the
552 // null block is only reached if 'store' tries to write to null object and
553 // 'load' read from non-null object (there is preceding check for that)
554 // These objects can't be the same.
555 Block* store_block = get_block_for_node(store);
556 Block* load_block = get_block_for_node(load);
557 Node* end = store_block->end();
558 if (end->is_MachNullCheck() && (end->in(1) == store) && store_block->dominates(load_block)) {
559 Node* if_true = end->find_out_with(Op_IfTrue);
560 assert(if_true != nullptr, "null check without null projection");
561 Node* null_block_region = if_true->find_out_with(Op_Region);
562 assert(null_block_region != nullptr, "null check without null region");
563 return get_block_for_node(null_block_region) == load_block;
564 }
565 return false;
566 }
567
568 class DefUseMemStatesQueue : public StackObj {
569 private:
570 class DefUsePair : public StackObj {
571 private:
572 Node* _def; // memory state
573 Node* _use; // use of the memory state that also modifies the memory state
574
575 public:
576 DefUsePair(Node* def, Node* use) :
577 _def(def), _use(use) {
578 }
579
580 DefUsePair() :
581 _def(nullptr), _use(nullptr) {
582 }
583
584 Node* def() const {
585 return _def;
586 }
587
588 Node* use() const {
589 return _use;
590 }
591 };
592
593 GrowableArray<DefUsePair> _queue;
594 GrowableArray<MergeMemNode*> _worklist_visited; // visited mergemem nodes
595
596 bool already_enqueued(Node* def_mem, PhiNode* use_phi) const {
597 // def_mem is one of the inputs of use_phi and at least one input of use_phi is
598 // not def_mem. It's however possible that use_phi has def_mem as input multiple
599 // times. If that happens, use_phi is recorded as a use of def_mem multiple
600 // times as well. When PhaseCFG::insert_anti_dependences() goes over
601 // uses of def_mem and enqueues them for processing, use_phi would then be
602 // enqueued for processing multiple times when it only needs to be
603 // processed once. The code below checks if use_phi as a use of def_mem was
604 // already enqueued to avoid redundant processing of use_phi.
605 int j = _queue.length()-1;
606 // If there are any use of def_mem already enqueued, they were enqueued
607 // last (all use of def_mem are processed in one go).
608 for (; j >= 0; j--) {
609 const DefUsePair& def_use_pair = _queue.at(j);
610 if (def_use_pair.def() != def_mem) {
611 // We're done with the uses of def_mem
612 break;
613 }
614 if (def_use_pair.use() == use_phi) {
615 return true;
616 }
617 }
618 #ifdef ASSERT
619 for (; j >= 0; j--) {
620 const DefUsePair& def_use_pair = _queue.at(j);
621 assert(def_use_pair.def() != def_mem, "Should be done with the uses of def_mem");
622 }
623 #endif
624 return false;
625 }
626
627 public:
628 DefUseMemStatesQueue(ResourceArea* area) {
629 }
630
631 void push(Node* def_mem_state, Node* use_mem_state) {
632 if (use_mem_state->is_MergeMem()) {
633 // Be sure we don't get into combinatorial problems.
634 if (!_worklist_visited.append_if_missing(use_mem_state->as_MergeMem())) {
635 return; // already on work list; do not repeat
636 }
637 } else if (use_mem_state->is_Phi()) {
638 // A Phi could have the same mem as input multiple times. If that's the case, we don't need to enqueue it
639 // more than once. We otherwise allow phis to be repeated; they can merge two relevant states.
640 if (already_enqueued(def_mem_state, use_mem_state->as_Phi())) {
641 return;
642 }
643 }
644
645 _queue.push(DefUsePair(def_mem_state, use_mem_state));
646 }
647
648 bool is_nonempty() const {
649 return _queue.is_nonempty();
650 }
651
652 Node* top_def() const {
653 return _queue.top().def();
654 }
655
656 Node* top_use() const {
657 return _queue.top().use();
658 }
659
660 void pop() {
661 _queue.pop();
662 }
663 };
664
665 //--------------------------insert_anti_dependences---------------------------
666 // A load may need to witness memory that nearby stores can overwrite.
667 // For each nearby store, either insert an "anti-dependence" edge
668 // from the load to the store, or else move LCA upward to force the
669 // load to (eventually) be scheduled in a block above the store.
670 //
671 // Do not add edges to stores on distinct control-flow paths;
672 // only add edges to stores which might interfere.
673 //
674 // Return the (updated) LCA. There will not be any possibly interfering
675 // store between the load's "early block" and the updated LCA.
676 // Any stores in the updated LCA will have new precedence edges
677 // back to the load. The caller is expected to schedule the load
678 // in the LCA, in which case the precedence edges will make LCM
679 // preserve anti-dependences. The caller may also hoist the load
680 // above the LCA, if it is not the early block.
681 Block* PhaseCFG::insert_anti_dependences(Block* LCA, Node* load, bool verify) {
682 ResourceMark rm;
683 assert(load->needs_anti_dependence_check(), "must be a load of some sort");
684 assert(LCA != nullptr, "");
685 DEBUG_ONLY(Block* LCA_orig = LCA);
686
687 // Compute the alias index. Loads and stores with different alias indices
688 // do not need anti-dependence edges.
689 int load_alias_idx = C->get_alias_index(load->adr_type());
690 #ifdef ASSERT
691 assert(Compile::AliasIdxTop <= load_alias_idx && load_alias_idx < C->num_alias_types(), "Invalid alias index");
692 if (load_alias_idx == Compile::AliasIdxBot && C->do_aliasing() &&
693 (PrintOpto || VerifyAliases ||
694 (PrintMiscellaneous && (WizardMode || Verbose)))) {
695 // Load nodes should not consume all of memory.
696 // Reporting a bottom type indicates a bug in adlc.
697 // If some particular type of node validly consumes all of memory,
698 // sharpen the preceding "if" to exclude it, so we can catch bugs here.
699 tty->print_cr("*** Possible Anti-Dependence Bug: Load consumes all of memory.");
700 load->dump(2);
701 if (VerifyAliases) assert(load_alias_idx != Compile::AliasIdxBot, "");
702 }
703 #endif
704
705 if (!C->alias_type(load_alias_idx)->is_rewritable()) {
706 // It is impossible to spoil this load by putting stores before it,
707 // because we know that the stores will never update the value
708 // which 'load' must witness.
709 return LCA;
710 }
711
712 node_idx_t load_index = load->_idx;
713
714 // Note the earliest legal placement of 'load', as determined by
715 // by the unique point in the dom tree where all memory effects
716 // and other inputs are first available. (Computed by schedule_early.)
717 // For normal loads, 'early' is the shallowest place (dom graph wise)
718 // to look for anti-deps between this load and any store.
719 Block* early = get_block_for_node(load);
720
721 // If we are subsuming loads, compute an "early" block that only considers
722 // memory or address inputs. This block may be different than the
723 // schedule_early block in that it could be at an even shallower depth in the
724 // dominator tree, and allow for a broader discovery of anti-dependences.
725 if (C->subsume_loads()) {
726 early = memory_early_block(load, early, this);
727 if (C->failing()) {
728 return nullptr;
729 }
730 }
731
732 ResourceArea* area = Thread::current()->resource_area();
733 DefUseMemStatesQueue worklist_def_use_mem_states(area); // prior memory state to store and possible-def to explore
734 Node_List non_early_stores(area); // all relevant stores outside of early
735 bool must_raise_LCA = false;
736
737 // 'load' uses some memory state; look for users of the same state.
738 // Recurse through MergeMem nodes to the stores that use them.
739
740 // Each of these stores is a possible definition of memory
741 // that 'load' needs to use. We need to force 'load'
742 // to occur before each such store. When the store is in
743 // the same block as 'load', we insert an anti-dependence
744 // edge load->store.
745
746 // The relevant stores "nearby" the load consist of a tree rooted
747 // at initial_mem, with internal nodes of type MergeMem.
748 // Therefore, the branches visited by the worklist are of this form:
749 // initial_mem -> (MergeMem ->)* Memory state modifying node
750 // Memory state modifying nodes include Store and Phi nodes and any node for which needs_anti_dependence_check()
751 // returns false.
752 // The anti-dependence constraints apply only to the fringe of this tree.
753
754 Node* initial_mem = load->in(MemNode::Memory);
755
756 // We don't optimize the memory graph for pinned loads, so we may need to raise the
757 // root of our search tree through the corresponding slices of MergeMem nodes to
758 // get to the node that really creates the memory state for this slice.
759 if (load_alias_idx >= Compile::AliasIdxRaw) {
760 while (initial_mem->is_MergeMem()) {
761 MergeMemNode* mm = initial_mem->as_MergeMem();
762 Node* p = mm->memory_at(load_alias_idx);
763 if (p != mm->base_memory()) {
764 initial_mem = p;
765 } else {
766 break;
767 }
768 }
769 }
770 worklist_def_use_mem_states.push(nullptr, initial_mem);
771 while (worklist_def_use_mem_states.is_nonempty()) {
772 // Examine a nearby store to see if it might interfere with our load.
773 Node* def_mem_state = worklist_def_use_mem_states.top_def();
774 Node* use_mem_state = worklist_def_use_mem_states.top_use();
775 worklist_def_use_mem_states.pop();
776
777 uint op = use_mem_state->Opcode();
778
779 #ifdef ASSERT
780 // CacheWB nodes are peculiar in a sense that they both are anti-dependent and produce memory.
781 // Allow them to be treated as a store.
782 bool is_cache_wb = false;
783 if (use_mem_state->is_Mach()) {
784 int ideal_op = use_mem_state->as_Mach()->ideal_Opcode();
785 is_cache_wb = (ideal_op == Op_CacheWB);
786 }
787 assert(!use_mem_state->needs_anti_dependence_check() || is_cache_wb, "no loads");
788 #endif
789
790 // MergeMems do not directly have anti-deps.
791 // Treat them as internal nodes in a forward tree of memory states,
792 // the leaves of which are each a 'possible-def'.
793 if (use_mem_state == initial_mem // root (exclusive) of tree we are searching
794 || op == Op_MergeMem // internal node of tree we are searching
795 ) {
796 def_mem_state = use_mem_state; // It's not a possibly interfering store.
797 if (use_mem_state == initial_mem)
798 initial_mem = nullptr; // only process initial memory once
799
800 for (DUIterator_Fast imax, i = def_mem_state->fast_outs(imax); i < imax; i++) {
801 use_mem_state = def_mem_state->fast_out(i);
802 if (use_mem_state->needs_anti_dependence_check()) {
803 // use_mem_state is also a kind of load (i.e. needs_anti_dependence_check), and it is not a memory state
804 // modifying node (store, Phi or MergeMem). Hence, load can't be anti dependent on this node.
805 continue;
806 }
807 worklist_def_use_mem_states.push(def_mem_state, use_mem_state);
808 }
809 continue;
810 }
811
812 if (op == Op_MachProj || op == Op_Catch) continue;
813
814 // Compute the alias index. Loads and stores with different alias
815 // indices do not need anti-dependence edges. Wide MemBar's are
816 // anti-dependent on everything (except immutable memories).
817 const TypePtr* adr_type = use_mem_state->adr_type();
818 if (!C->can_alias(adr_type, load_alias_idx)) continue;
819
820 // Most slow-path runtime calls do NOT modify Java memory, but
821 // they can block and so write Raw memory.
822 if (use_mem_state->is_Mach()) {
823 MachNode* mstore = use_mem_state->as_Mach();
824 if (load_alias_idx != Compile::AliasIdxRaw) {
825 // Check for call into the runtime using the Java calling
826 // convention (and from there into a wrapper); it has no
827 // _method. Can't do this optimization for Native calls because
828 // they CAN write to Java memory.
829 if (mstore->ideal_Opcode() == Op_CallStaticJava) {
830 assert(mstore->is_MachSafePoint(), "");
831 MachSafePointNode* ms = (MachSafePointNode*) mstore;
832 assert(ms->is_MachCallJava(), "");
833 MachCallJavaNode* mcj = (MachCallJavaNode*) ms;
834 if (mcj->_method == nullptr) {
835 // These runtime calls do not write to Java visible memory
836 // (other than Raw) and so do not require anti-dependence edges.
837 continue;
838 }
839 }
840 // Same for SafePoints: they read/write Raw but only read otherwise.
841 // This is basically a workaround for SafePoints only defining control
842 // instead of control + memory.
843 if (mstore->ideal_Opcode() == Op_SafePoint)
844 continue;
845 } else {
846 // Some raw memory, such as the load of "top" at an allocation,
847 // can be control dependent on the previous safepoint. See
848 // comments in GraphKit::allocate_heap() about control input.
849 // Inserting an anti-dep between such a safepoint and a use
850 // creates a cycle, and will cause a subsequent failure in
851 // local scheduling. (BugId 4919904)
852 // (%%% How can a control input be a safepoint and not a projection??)
853 if (mstore->ideal_Opcode() == Op_SafePoint && load->in(0) == mstore)
854 continue;
855 }
856 }
857
858 // Identify a block that the current load must be above,
859 // or else observe that 'store' is all the way up in the
860 // earliest legal block for 'load'. In the latter case,
861 // immediately insert an anti-dependence edge.
862 Block* store_block = get_block_for_node(use_mem_state);
863 assert(store_block != nullptr, "unused killing projections skipped above");
864
865 if (use_mem_state->is_Phi()) {
866 // Loop-phis need to raise load before input. (Other phis are treated
867 // as store below.)
868 //
869 // 'load' uses memory which is one (or more) of the Phi's inputs.
870 // It must be scheduled not before the Phi, but rather before
871 // each of the relevant Phi inputs.
872 //
873 // Instead of finding the LCA of all inputs to a Phi that match 'mem',
874 // we mark each corresponding predecessor block and do a combined
875 // hoisting operation later (raise_LCA_above_marks).
876 //
877 // Do not assert(store_block != early, "Phi merging memory after access")
878 // PhiNode may be at start of block 'early' with backedge to 'early'
879 DEBUG_ONLY(bool found_match = false);
880 for (uint j = PhiNode::Input, jmax = use_mem_state->req(); j < jmax; j++) {
881 if (use_mem_state->in(j) == def_mem_state) { // Found matching input?
882 DEBUG_ONLY(found_match = true);
883 Block* pred_block = get_block_for_node(store_block->pred(j));
884 if (pred_block != early) {
885 // If any predecessor of the Phi matches the load's "early block",
886 // we do not need a precedence edge between the Phi and 'load'
887 // since the load will be forced into a block preceding the Phi.
888 pred_block->set_raise_LCA_mark(load_index);
889 assert(!LCA_orig->dominates(pred_block) ||
890 early->dominates(pred_block), "early is high enough");
891 must_raise_LCA = true;
892 } else {
893 // anti-dependent upon PHI pinned below 'early', no edge needed
894 LCA = early; // but can not schedule below 'early'
895 }
896 }
897 }
898 assert(found_match, "no worklist bug");
899 } else if (store_block != early) {
900 // 'store' is between the current LCA and earliest possible block.
901 // Label its block, and decide later on how to raise the LCA
902 // to include the effect on LCA of this store.
903 // If this store's block gets chosen as the raised LCA, we
904 // will find him on the non_early_stores list and stick him
905 // with a precedence edge.
906 // (But, don't bother if LCA is already raised all the way.)
907 if (LCA != early && !unrelated_load_in_store_null_block(use_mem_state, load)) {
908 store_block->set_raise_LCA_mark(load_index);
909 must_raise_LCA = true;
910 non_early_stores.push(use_mem_state);
911 }
912 } else {
913 // Found a possibly-interfering store in the load's 'early' block.
914 // This means 'load' cannot sink at all in the dominator tree.
915 // Add an anti-dep edge, and squeeze 'load' into the highest block.
916 assert(use_mem_state != load->find_exact_control(load->in(0)), "dependence cycle found");
917 if (verify) {
918 assert(use_mem_state->find_edge(load) != -1 || unrelated_load_in_store_null_block(use_mem_state, load),
919 "missing precedence edge");
920 } else {
921 use_mem_state->add_prec(load);
922 }
923 LCA = early;
924 // This turns off the process of gathering non_early_stores.
925 }
926 }
927 // (Worklist is now empty; all nearby stores have been visited.)
928
929 // Finished if 'load' must be scheduled in its 'early' block.
930 // If we found any stores there, they have already been given
931 // precedence edges.
932 if (LCA == early) return LCA;
933
934 // We get here only if there are no possibly-interfering stores
935 // in the load's 'early' block. Move LCA up above all predecessors
936 // which contain stores we have noted.
937 //
938 // The raised LCA block can be a home to such interfering stores,
939 // but its predecessors must not contain any such stores.
940 //
941 // The raised LCA will be a lower bound for placing the load,
942 // preventing the load from sinking past any block containing
943 // a store that may invalidate the memory state required by 'load'.
944 if (must_raise_LCA)
945 LCA = raise_LCA_above_marks(LCA, load->_idx, early, this);
946 if (LCA == early) return LCA;
947
948 // Insert anti-dependence edges from 'load' to each store
949 // in the non-early LCA block.
950 // Mine the non_early_stores list for such stores.
951 if (LCA->raise_LCA_mark() == load_index) {
952 while (non_early_stores.size() > 0) {
953 Node* store = non_early_stores.pop();
954 Block* store_block = get_block_for_node(store);
955 if (store_block == LCA) {
956 // add anti_dependence from store to load in its own block
957 assert(store != load->find_exact_control(load->in(0)), "dependence cycle found");
958 if (verify) {
959 assert(store->find_edge(load) != -1, "missing precedence edge");
960 } else {
961 store->add_prec(load);
962 }
963 } else {
964 assert(store_block->raise_LCA_mark() == load_index, "block was marked");
965 // Any other stores we found must be either inside the new LCA
966 // or else outside the original LCA. In the latter case, they
967 // did not interfere with any use of 'load'.
968 assert(LCA->dominates(store_block)
969 || !LCA_orig->dominates(store_block), "no stray stores");
970 }
971 }
972 }
973
974 // Return the highest block containing stores; any stores
975 // within that block have been given anti-dependence edges.
976 return LCA;
977 }
978
979 // This class is used to iterate backwards over the nodes in the graph.
980
981 class Node_Backward_Iterator {
982
983 private:
984 Node_Backward_Iterator();
985
986 public:
987 // Constructor for the iterator
988 Node_Backward_Iterator(Node *root, VectorSet &visited, Node_Stack &stack, PhaseCFG &cfg);
989
990 // Postincrement operator to iterate over the nodes
991 Node *next();
992
993 private:
994 VectorSet &_visited;
995 Node_Stack &_stack;
996 PhaseCFG &_cfg;
997 };
998
999 // Constructor for the Node_Backward_Iterator
1000 Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_Stack &stack, PhaseCFG &cfg)
1001 : _visited(visited), _stack(stack), _cfg(cfg) {
1002 // The stack should contain exactly the root
1003 stack.clear();
1004 stack.push(root, root->outcnt());
1005
1006 // Clear the visited bits
1007 visited.clear();
1008 }
1009
1010 // Iterator for the Node_Backward_Iterator
1011 Node *Node_Backward_Iterator::next() {
1012
1013 // If the _stack is empty, then just return null: finished.
1014 if ( !_stack.size() )
1015 return nullptr;
1016
1017 // I visit unvisited not-anti-dependence users first, then anti-dependent
1018 // children next. I iterate backwards to support removal of nodes.
1019 // The stack holds states consisting of 3 values:
1020 // current Def node, flag which indicates 1st/2nd pass, index of current out edge
1021 Node *self = (Node*)(((uintptr_t)_stack.node()) & ~1);
1022 bool iterate_anti_dep = (((uintptr_t)_stack.node()) & 1);
1023 uint idx = MIN2(_stack.index(), self->outcnt()); // Support removal of nodes.
1024 _stack.pop();
1025
1026 // I cycle here when I am entering a deeper level of recursion.
1027 // The key variable 'self' was set prior to jumping here.
1028 while( 1 ) {
1029
1030 _visited.set(self->_idx);
1031
1032 // Now schedule all uses as late as possible.
1033 const Node* src = self->is_Proj() ? self->in(0) : self;
1034 uint src_rpo = _cfg.get_block_for_node(src)->_rpo;
1035
1036 // Schedule all nodes in a post-order visit
1037 Node *unvisited = nullptr; // Unvisited anti-dependent Node, if any
1038
1039 // Scan for unvisited nodes
1040 while (idx > 0) {
1041 // For all uses, schedule late
1042 Node* n = self->raw_out(--idx); // Use
1043
1044 // Skip already visited children
1045 if ( _visited.test(n->_idx) )
1046 continue;
1047
1048 // do not traverse backward control edges
1049 Node *use = n->is_Proj() ? n->in(0) : n;
1050 uint use_rpo = _cfg.get_block_for_node(use)->_rpo;
1051
1052 if ( use_rpo < src_rpo )
1053 continue;
1054
1055 // Phi nodes always precede uses in a basic block
1056 if ( use_rpo == src_rpo && use->is_Phi() )
1057 continue;
1058
1059 unvisited = n; // Found unvisited
1060
1061 // Check for possible-anti-dependent
1062 // 1st pass: No such nodes, 2nd pass: Only such nodes.
1063 if (n->needs_anti_dependence_check() == iterate_anti_dep) {
1064 unvisited = n; // Found unvisited
1065 break;
1066 }
1067 }
1068
1069 // Did I find an unvisited not-anti-dependent Node?
1070 if (!unvisited) {
1071 if (!iterate_anti_dep) {
1072 // 2nd pass: Iterate over nodes which needs_anti_dependence_check.
1073 iterate_anti_dep = true;
1074 idx = self->outcnt();
1075 continue;
1076 }
1077 break; // All done with children; post-visit 'self'
1078 }
1079
1080 // Visit the unvisited Node. Contains the obvious push to
1081 // indicate I'm entering a deeper level of recursion. I push the
1082 // old state onto the _stack and set a new state and loop (recurse).
1083 _stack.push((Node*)((uintptr_t)self | (uintptr_t)iterate_anti_dep), idx);
1084 self = unvisited;
1085 iterate_anti_dep = false;
1086 idx = self->outcnt();
1087 } // End recursion loop
1088
1089 return self;
1090 }
1091
1092 //------------------------------ComputeLatenciesBackwards----------------------
1093 // Compute the latency of all the instructions.
1094 void PhaseCFG::compute_latencies_backwards(VectorSet &visited, Node_Stack &stack) {
1095 #ifndef PRODUCT
1096 if (trace_opto_pipelining())
1097 tty->print("\n#---- ComputeLatenciesBackwards ----\n");
1098 #endif
1099
1100 Node_Backward_Iterator iter((Node *)_root, visited, stack, *this);
1101 Node *n;
1102
1103 // Walk over all the nodes from last to first
1104 while ((n = iter.next())) {
1105 // Set the latency for the definitions of this instruction
1106 partial_latency_of_defs(n);
1107 }
1108 } // end ComputeLatenciesBackwards
1109
1110 //------------------------------partial_latency_of_defs------------------------
1111 // Compute the latency impact of this node on all defs. This computes
1112 // a number that increases as we approach the beginning of the routine.
1113 void PhaseCFG::partial_latency_of_defs(Node *n) {
1114 // Set the latency for this instruction
1115 #ifndef PRODUCT
1116 if (trace_opto_pipelining()) {
1117 tty->print("# latency_to_inputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n));
1118 dump();
1119 }
1120 #endif
1121
1122 if (n->is_Proj()) {
1123 n = n->in(0);
1124 }
1125
1126 if (n->is_Root()) {
1127 return;
1128 }
1129
1130 uint nlen = n->len();
1131 uint use_latency = get_latency_for_node(n);
1132 uint use_pre_order = get_block_for_node(n)->_pre_order;
1133
1134 for (uint j = 0; j < nlen; j++) {
1135 Node *def = n->in(j);
1136
1137 if (!def || def == n) {
1138 continue;
1139 }
1140
1141 // Walk backwards thru projections
1142 if (def->is_Proj()) {
1143 def = def->in(0);
1144 }
1145
1146 #ifndef PRODUCT
1147 if (trace_opto_pipelining()) {
1148 tty->print("# in(%2d): ", j);
1149 def->dump();
1150 }
1151 #endif
1152
1153 // If the defining block is not known, assume it is ok
1154 Block *def_block = get_block_for_node(def);
1155 uint def_pre_order = def_block ? def_block->_pre_order : 0;
1156
1157 if ((use_pre_order < def_pre_order) || (use_pre_order == def_pre_order && n->is_Phi())) {
1158 continue;
1159 }
1160
1161 uint delta_latency = n->latency(j);
1162 uint current_latency = delta_latency + use_latency;
1163
1164 if (get_latency_for_node(def) < current_latency) {
1165 set_latency_for_node(def, current_latency);
1166 }
1167
1168 #ifndef PRODUCT
1169 if (trace_opto_pipelining()) {
1170 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));
1171 }
1172 #endif
1173 }
1174 }
1175
1176 //------------------------------latency_from_use-------------------------------
1177 // Compute the latency of a specific use
1178 int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) {
1179 // If self-reference, return no latency
1180 if (use == n || use->is_Root()) {
1181 return 0;
1182 }
1183
1184 uint def_pre_order = get_block_for_node(def)->_pre_order;
1185 uint latency = 0;
1186
1187 // If the use is not a projection, then it is simple...
1188 if (!use->is_Proj()) {
1189 #ifndef PRODUCT
1190 if (trace_opto_pipelining()) {
1191 tty->print("# out(): ");
1192 use->dump();
1193 }
1194 #endif
1195
1196 uint use_pre_order = get_block_for_node(use)->_pre_order;
1197
1198 if (use_pre_order < def_pre_order)
1199 return 0;
1200
1201 if (use_pre_order == def_pre_order && use->is_Phi())
1202 return 0;
1203
1204 uint nlen = use->len();
1205 uint nl = get_latency_for_node(use);
1206
1207 for ( uint j=0; j<nlen; j++ ) {
1208 if (use->in(j) == n) {
1209 // Change this if we want local latencies
1210 uint ul = use->latency(j);
1211 uint l = ul + nl;
1212 if (latency < l) latency = l;
1213 #ifndef PRODUCT
1214 if (trace_opto_pipelining()) {
1215 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, latency = %d",
1216 nl, j, ul, l, latency);
1217 }
1218 #endif
1219 }
1220 }
1221 } else {
1222 // This is a projection, just grab the latency of the use(s)
1223 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
1224 uint l = latency_from_use(use, def, use->fast_out(j));
1225 if (latency < l) latency = l;
1226 }
1227 }
1228
1229 return latency;
1230 }
1231
1232 //------------------------------latency_from_uses------------------------------
1233 // Compute the latency of this instruction relative to all of it's uses.
1234 // This computes a number that increases as we approach the beginning of the
1235 // routine.
1236 void PhaseCFG::latency_from_uses(Node *n) {
1237 // Set the latency for this instruction
1238 #ifndef PRODUCT
1239 if (trace_opto_pipelining()) {
1240 tty->print("# latency_from_outputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n));
1241 dump();
1242 }
1243 #endif
1244 uint latency=0;
1245 const Node *def = n->is_Proj() ? n->in(0): n;
1246
1247 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1248 uint l = latency_from_use(n, def, n->fast_out(i));
1249
1250 if (latency < l) latency = l;
1251 }
1252
1253 set_latency_for_node(n, latency);
1254 }
1255
1256 //------------------------------is_cheaper_block-------------------------
1257 // Check if a block between early and LCA block of uses is cheaper by
1258 // frequency-based policy, latency-based policy and random-based policy
1259 bool PhaseCFG::is_cheaper_block(Block* LCA, Node* self, uint target_latency,
1260 uint end_latency, double least_freq,
1261 int cand_cnt, bool in_latency) {
1262 if (StressGCM) {
1263 // Should be randomly accepted in stress mode
1264 return C->randomized_select(cand_cnt);
1265 }
1266
1267 const double delta = 1 + PROB_UNLIKELY_MAG(4);
1268
1269 // Better Frequency. Add a small delta to the comparison to not needlessly
1270 // hoist because of, e.g., small numerical inaccuracies.
1271 if (LCA->_freq * delta < least_freq) {
1272 return true;
1273 }
1274
1275 // Otherwise, choose with latency
1276 if (!in_latency && // No block containing latency
1277 LCA->_freq < least_freq * delta && // No worse frequency
1278 target_latency >= end_latency && // within latency range
1279 !self->is_iteratively_computed() // But don't hoist IV increments
1280 // because they may end up above other uses of their phi forcing
1281 // their result register to be different from their input.
1282 ) {
1283 return true;
1284 }
1285
1286 return false;
1287 }
1288
1289 //------------------------------hoist_to_cheaper_block-------------------------
1290 // Pick a block for node self, between early and LCA block of uses, that is a
1291 // cheaper alternative to LCA.
1292 Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) {
1293 Block* least = LCA;
1294 double least_freq = least->_freq;
1295 uint target = get_latency_for_node(self);
1296 uint start_latency = get_latency_for_node(LCA->head());
1297 uint end_latency = get_latency_for_node(LCA->get_node(LCA->end_idx()));
1298 bool in_latency = (target <= start_latency);
1299 const Block* root_block = get_block_for_node(_root);
1300
1301 // Turn off latency scheduling if scheduling is just plain off
1302 if (!C->do_scheduling())
1303 in_latency = true;
1304
1305 // Do not hoist (to cover latency) instructions which target a
1306 // single register. Hoisting stretches the live range of the
1307 // single register and may force spilling.
1308 MachNode* mach = self->is_Mach() ? self->as_Mach() : nullptr;
1309 if (mach && mach->out_RegMask().is_bound1() && mach->out_RegMask().is_NotEmpty())
1310 in_latency = true;
1311
1312 #ifndef PRODUCT
1313 if (trace_opto_pipelining()) {
1314 tty->print("# Find cheaper block for latency %d: ", get_latency_for_node(self));
1315 self->dump();
1316 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1317 LCA->_pre_order,
1318 LCA->head()->_idx,
1319 start_latency,
1320 LCA->get_node(LCA->end_idx())->_idx,
1321 end_latency,
1322 least_freq);
1323 }
1324 #endif
1325
1326 int cand_cnt = 0; // number of candidates tried
1327
1328 // Walk up the dominator tree from LCA (Lowest common ancestor) to
1329 // the earliest legal location. Capture the least execution frequency,
1330 // or choose a random block if -XX:+StressGCM, or using latency-based policy
1331 while (LCA != early) {
1332 LCA = LCA->_idom; // Follow up the dominator tree
1333
1334 if (LCA == nullptr) {
1335 // Bailout without retry
1336 assert(false, "graph should be schedulable");
1337 C->record_method_not_compilable("late schedule failed: LCA is null");
1338 return least;
1339 }
1340
1341 // Don't hoist machine instructions to the root basic block
1342 if (mach && LCA == root_block)
1343 break;
1344
1345 if (self->is_memory_writer() &&
1346 (LCA->_loop->depth() > early->_loop->depth())) {
1347 // LCA is an invalid placement for a memory writer: choosing it would
1348 // cause memory interference, as illustrated in schedule_late().
1349 continue;
1350 }
1351 verify_memory_writer_placement(LCA, self);
1352
1353 uint start_lat = get_latency_for_node(LCA->head());
1354 uint end_idx = LCA->end_idx();
1355 uint end_lat = get_latency_for_node(LCA->get_node(end_idx));
1356 double LCA_freq = LCA->_freq;
1357 #ifndef PRODUCT
1358 if (trace_opto_pipelining()) {
1359 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1360 LCA->_pre_order, LCA->head()->_idx, start_lat, end_idx, end_lat, LCA_freq);
1361 }
1362 #endif
1363 cand_cnt++;
1364 if (is_cheaper_block(LCA, self, target, end_lat, least_freq, cand_cnt, in_latency)) {
1365 least = LCA; // Found cheaper block
1366 least_freq = LCA_freq;
1367 start_latency = start_lat;
1368 end_latency = end_lat;
1369 if (target <= start_lat)
1370 in_latency = true;
1371 }
1372 }
1373
1374 #ifndef PRODUCT
1375 if (trace_opto_pipelining()) {
1376 tty->print_cr("# Choose block B%d with start latency=%d and freq=%g",
1377 least->_pre_order, start_latency, least_freq);
1378 }
1379 #endif
1380
1381 // See if the latency needs to be updated
1382 if (target < end_latency) {
1383 #ifndef PRODUCT
1384 if (trace_opto_pipelining()) {
1385 tty->print_cr("# Change latency for [%4d] from %d to %d", self->_idx, target, end_latency);
1386 }
1387 #endif
1388 set_latency_for_node(self, end_latency);
1389 partial_latency_of_defs(self);
1390 }
1391
1392 return least;
1393 }
1394
1395
1396 //------------------------------schedule_late-----------------------------------
1397 // Now schedule all codes as LATE as possible. This is the LCA in the
1398 // dominator tree of all USES of a value. Pick the block with the least
1399 // loop nesting depth that is lowest in the dominator tree.
1400 extern const char must_clone[];
1401 void PhaseCFG::schedule_late(VectorSet &visited, Node_Stack &stack) {
1402 #ifndef PRODUCT
1403 if (trace_opto_pipelining())
1404 tty->print("\n#---- schedule_late ----\n");
1405 #endif
1406
1407 Node_Backward_Iterator iter((Node *)_root, visited, stack, *this);
1408 Node *self;
1409
1410 // Walk over all the nodes from last to first
1411 while ((self = iter.next())) {
1412 Block* early = get_block_for_node(self); // Earliest legal placement
1413
1414 if (self->is_top()) {
1415 // Top node goes in bb #2 with other constants.
1416 // It must be special-cased, because it has no out edges.
1417 early->add_inst(self);
1418 continue;
1419 }
1420
1421 // No uses, just terminate
1422 if (self->outcnt() == 0) {
1423 assert(self->is_MachProj(), "sanity");
1424 continue; // Must be a dead machine projection
1425 }
1426
1427 // If node is pinned in the block, then no scheduling can be done.
1428 if( self->pinned() ) // Pinned in block?
1429 continue;
1430
1431 #ifdef ASSERT
1432 // Assert that memory writers (e.g. stores) have a "home" block (the block
1433 // given by their control input), and that this block corresponds to their
1434 // earliest possible placement. This guarantees that
1435 // hoist_to_cheaper_block() will always have at least one valid choice.
1436 if (self->is_memory_writer()) {
1437 assert(find_block_for_node(self->in(0)) == early,
1438 "The home of a memory writer must also be its earliest placement");
1439 }
1440 #endif
1441
1442 MachNode* mach = self->is_Mach() ? self->as_Mach() : nullptr;
1443 if (mach) {
1444 switch (mach->ideal_Opcode()) {
1445 case Op_CreateEx:
1446 // Don't move exception creation
1447 early->add_inst(self);
1448 continue;
1449 break;
1450 case Op_CastI2N:
1451 early->add_inst(self);
1452 continue;
1453 case Op_CheckCastPP: {
1454 // Don't move CheckCastPP nodes away from their input, if the input
1455 // is a rawptr (5071820).
1456 Node *def = self->in(1);
1457 if (def != nullptr && def->bottom_type()->base() == Type::RawPtr) {
1458 early->add_inst(self);
1459 #ifdef ASSERT
1460 _raw_oops.push(def);
1461 #endif
1462 continue;
1463 }
1464 break;
1465 }
1466 default:
1467 break;
1468 }
1469 if (C->has_irreducible_loop() && self->is_memory_writer()) {
1470 // If the CFG is irreducible, place memory writers in their home block.
1471 // This prevents hoist_to_cheaper_block() from accidentally placing such
1472 // nodes into deeper loops, as in the following example:
1473 //
1474 // Home placement of store in B1 (loop L1):
1475 //
1476 // B1 (L1):
1477 // m1 <- ..
1478 // m2 <- store m1, ..
1479 // B2 (L2):
1480 // jump B2
1481 // B3 (L1):
1482 // .. <- .. m2, ..
1483 //
1484 // Wrong "hoisting" of store to B2 (in loop L2, child of L1):
1485 //
1486 // B1 (L1):
1487 // m1 <- ..
1488 // B2 (L2):
1489 // m2 <- store m1, ..
1490 // # Wrong: m1 and m2 interfere at this point.
1491 // jump B2
1492 // B3 (L1):
1493 // .. <- .. m2, ..
1494 //
1495 // This "hoist inversion" can happen due to different factors such as
1496 // inaccurate estimation of frequencies for irreducible CFGs, and loops
1497 // with always-taken exits in reducible CFGs. In the reducible case,
1498 // hoist inversion is prevented by discarding invalid blocks (those in
1499 // deeper loops than the home block). In the irreducible case, the
1500 // invalid blocks cannot be identified due to incomplete loop nesting
1501 // information, hence a conservative solution is taken.
1502 #ifndef PRODUCT
1503 if (trace_opto_pipelining()) {
1504 tty->print_cr("# Irreducible loops: schedule in home block B%d:",
1505 early->_pre_order);
1506 self->dump();
1507 }
1508 #endif
1509 schedule_node_into_block(self, early);
1510 continue;
1511 }
1512 }
1513
1514 // Gather LCA of all uses
1515 Block *LCA = nullptr;
1516 {
1517 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
1518 // For all uses, find LCA
1519 Node* use = self->fast_out(i);
1520 LCA = raise_LCA_above_use(LCA, use, self, this);
1521 }
1522 guarantee(LCA != nullptr, "There must be a LCA");
1523 } // (Hide defs of imax, i from rest of block.)
1524
1525 // Place temps in the block of their use. This isn't a
1526 // requirement for correctness but it reduces useless
1527 // interference between temps and other nodes.
1528 if (mach != nullptr && mach->is_MachTemp()) {
1529 map_node_to_block(self, LCA);
1530 LCA->add_inst(self);
1531 continue;
1532 }
1533
1534 // Check if 'self' could be anti-dependent on memory
1535 if (self->needs_anti_dependence_check()) {
1536 // Hoist LCA above possible-defs and insert anti-dependences to
1537 // defs in new LCA block.
1538 LCA = insert_anti_dependences(LCA, self);
1539 if (C->failing()) {
1540 return;
1541 }
1542 }
1543
1544 if (early->_dom_depth > LCA->_dom_depth) {
1545 // Somehow the LCA has moved above the earliest legal point.
1546 // (One way this can happen is via memory_early_block.)
1547 if (C->subsume_loads() == true && !C->failing()) {
1548 // Retry with subsume_loads == false
1549 // If this is the first failure, the sentinel string will "stick"
1550 // to the Compile object, and the C2Compiler will see it and retry.
1551 C->record_failure(C2Compiler::retry_no_subsuming_loads());
1552 } else {
1553 // Bailout without retry when (early->_dom_depth > LCA->_dom_depth)
1554 assert(C->failure_is_artificial(), "graph should be schedulable");
1555 C->record_method_not_compilable("late schedule failed: incorrect graph" DEBUG_ONLY(COMMA true));
1556 }
1557 return;
1558 }
1559
1560 if (self->is_memory_writer()) {
1561 // If the LCA of a memory writer is a descendant of its home loop, hoist
1562 // it into a valid placement.
1563 while (LCA->_loop->depth() > early->_loop->depth()) {
1564 LCA = LCA->_idom;
1565 }
1566 assert(LCA != nullptr, "a valid LCA must exist");
1567 verify_memory_writer_placement(LCA, self);
1568 }
1569
1570 // If there is no opportunity to hoist, then we're done.
1571 // In stress mode, try to hoist even the single operations.
1572 bool try_to_hoist = StressGCM || (LCA != early);
1573
1574 // Must clone guys stay next to use; no hoisting allowed.
1575 // Also cannot hoist guys that alter memory or are otherwise not
1576 // allocatable (hoisting can make a value live longer, leading to
1577 // anti and output dependency problems which are normally resolved
1578 // by the register allocator giving everyone a different register).
1579 if (mach != nullptr && must_clone[mach->ideal_Opcode()])
1580 try_to_hoist = false;
1581
1582 Block* late = nullptr;
1583 if (try_to_hoist) {
1584 // Now find the block with the least execution frequency.
1585 // Start at the latest schedule and work up to the earliest schedule
1586 // in the dominator tree. Thus the Node will dominate all its uses.
1587 late = hoist_to_cheaper_block(LCA, early, self);
1588 } else {
1589 // Just use the LCA of the uses.
1590 late = LCA;
1591 }
1592
1593 // Put the node into target block
1594 schedule_node_into_block(self, late);
1595
1596 #ifdef ASSERT
1597 if (self->needs_anti_dependence_check()) {
1598 // since precedence edges are only inserted when we're sure they
1599 // are needed make sure that after placement in a block we don't
1600 // need any new precedence edges.
1601 verify_anti_dependences(late, self);
1602 }
1603 #endif
1604 } // Loop until all nodes have been visited
1605
1606 } // end ScheduleLate
1607
1608 //------------------------------GlobalCodeMotion-------------------------------
1609 void PhaseCFG::global_code_motion() {
1610 ResourceMark rm;
1611
1612 #ifndef PRODUCT
1613 if (trace_opto_pipelining()) {
1614 tty->print("\n---- Start GlobalCodeMotion ----\n");
1615 }
1616 #endif
1617
1618 // Initialize the node to block mapping for things on the proj_list
1619 for (uint i = 0; i < _matcher.number_of_projections(); i++) {
1620 unmap_node_from_block(_matcher.get_projection(i));
1621 }
1622
1623 // Set the basic block for Nodes pinned into blocks
1624 VectorSet visited;
1625 schedule_pinned_nodes(visited);
1626
1627 // Find the earliest Block any instruction can be placed in. Some
1628 // instructions are pinned into Blocks. Unpinned instructions can
1629 // appear in last block in which all their inputs occur.
1630 visited.clear();
1631 Node_Stack stack((C->live_nodes() >> 2) + 16); // pre-grow
1632 if (!schedule_early(visited, stack)) {
1633 // Bailout without retry
1634 assert(C->failure_is_artificial(), "early schedule failed");
1635 C->record_method_not_compilable("early schedule failed" DEBUG_ONLY(COMMA true));
1636 return;
1637 }
1638
1639 // Build Def-Use edges.
1640 // Compute the latency information (via backwards walk) for all the
1641 // instructions in the graph
1642 _node_latency = new GrowableArray<uint>(); // resource_area allocation
1643
1644 if (C->do_scheduling()) {
1645 compute_latencies_backwards(visited, stack);
1646 }
1647
1648 // Now schedule all codes as LATE as possible. This is the LCA in the
1649 // dominator tree of all USES of a value. Pick the block with the least
1650 // loop nesting depth that is lowest in the dominator tree.
1651 // ( visited.clear() called in schedule_late()->Node_Backward_Iterator() )
1652 schedule_late(visited, stack);
1653 if (C->failing()) {
1654 return;
1655 }
1656
1657 #ifndef PRODUCT
1658 if (trace_opto_pipelining()) {
1659 tty->print("\n---- Detect implicit null checks ----\n");
1660 }
1661 #endif
1662
1663 // Detect implicit-null-check opportunities. Basically, find null checks
1664 // with suitable memory ops nearby. Use the memory op to do the null check.
1665 // I can generate a memory op if there is not one nearby.
1666 if (C->is_method_compilation()) {
1667 // By reversing the loop direction we get a very minor gain on mpegaudio.
1668 // Feel free to revert to a forward loop for clarity.
1669 // for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) {
1670 for (int i = _matcher._null_check_tests.size() - 2; i >= 0; i -= 2) {
1671 Node* proj = _matcher._null_check_tests[i];
1672 Node* val = _matcher._null_check_tests[i + 1];
1673 Block* block = get_block_for_node(proj);
1674 implicit_null_check(block, proj, val, C->allowed_deopt_reasons());
1675 // The implicit_null_check will only perform the transformation
1676 // if the null branch is truly uncommon, *and* it leads to an
1677 // uncommon trap. Combined with the too_many_traps guards
1678 // above, this prevents SEGV storms reported in 6366351,
1679 // by recompiling offending methods without this optimization.
1680 if (C->failing()) {
1681 return;
1682 }
1683 }
1684 }
1685
1686 bool block_size_threshold_ok = false;
1687 intptr_t *recalc_pressure_nodes = nullptr;
1688 if (OptoRegScheduling) {
1689 for (uint i = 0; i < number_of_blocks(); i++) {
1690 Block* block = get_block(i);
1691 if (block->number_of_nodes() > 10) {
1692 block_size_threshold_ok = true;
1693 break;
1694 }
1695 }
1696 }
1697
1698 // Enabling the scheduler for register pressure plus finding blocks of size to schedule for it
1699 // is key to enabling this feature.
1700 PhaseChaitin regalloc(C->unique(), *this, _matcher, true);
1701 ResourceArea live_arena(mtCompiler, Arena::Tag::tag_reglive); // Arena for liveness
1702 ResourceMark rm_live(&live_arena);
1703 PhaseLive live(*this, regalloc._lrg_map.names(), &live_arena, true);
1704 PhaseIFG ifg(&live_arena);
1705 if (OptoRegScheduling && block_size_threshold_ok) {
1706 regalloc.mark_ssa();
1707 Compile::TracePhase tp(_t_computeLive);
1708 rm_live.reset_to_mark(); // Reclaim working storage
1709 IndexSet::reset_memory(C, &live_arena);
1710 uint node_size = regalloc._lrg_map.max_lrg_id();
1711 ifg.init(node_size); // Empty IFG
1712 regalloc.set_ifg(ifg);
1713 regalloc.set_live(live);
1714 regalloc.gather_lrg_masks(false); // Collect LRG masks
1715 live.compute(node_size); // Compute liveness
1716
1717 recalc_pressure_nodes = NEW_RESOURCE_ARRAY(intptr_t, node_size);
1718 for (uint i = 0; i < node_size; i++) {
1719 recalc_pressure_nodes[i] = 0;
1720 }
1721 }
1722 _regalloc = ®alloc;
1723
1724 #ifndef PRODUCT
1725 if (trace_opto_pipelining()) {
1726 tty->print("\n---- Start Local Scheduling ----\n");
1727 }
1728 #endif
1729
1730 // Schedule locally. Right now a simple topological sort.
1731 // Later, do a real latency aware scheduler.
1732 GrowableArray<int> ready_cnt(C->unique(), C->unique(), -1);
1733 visited.reset();
1734 for (uint i = 0; i < number_of_blocks(); i++) {
1735 Block* block = get_block(i);
1736 if (!schedule_local(block, ready_cnt, visited, recalc_pressure_nodes)) {
1737 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
1738 assert(C->failure_is_artificial(), "local schedule failed");
1739 C->record_method_not_compilable("local schedule failed" DEBUG_ONLY(COMMA true));
1740 }
1741 _regalloc = nullptr;
1742 return;
1743 }
1744 }
1745 _regalloc = nullptr;
1746
1747 // If we inserted any instructions between a Call and his CatchNode,
1748 // clone the instructions on all paths below the Catch.
1749 for (uint i = 0; i < number_of_blocks(); i++) {
1750 Block* block = get_block(i);
1751 call_catch_cleanup(block);
1752 if (C->failing()) {
1753 return;
1754 }
1755 }
1756
1757 #ifndef PRODUCT
1758 if (trace_opto_pipelining()) {
1759 tty->print("\n---- After GlobalCodeMotion ----\n");
1760 for (uint i = 0; i < number_of_blocks(); i++) {
1761 Block* block = get_block(i);
1762 block->dump();
1763 }
1764 }
1765 #endif
1766 // Dead.
1767 _node_latency = (GrowableArray<uint> *)((intptr_t)0xdeadbeef);
1768 }
1769
1770 bool PhaseCFG::do_global_code_motion() {
1771
1772 build_dominator_tree();
1773 if (C->failing()) {
1774 return false;
1775 }
1776
1777 NOT_PRODUCT( C->verify_graph_edges(); )
1778
1779 estimate_block_frequency();
1780
1781 global_code_motion();
1782
1783 if (C->failing()) {
1784 return false;
1785 }
1786
1787 return true;
1788 }
1789
1790 //------------------------------Estimate_Block_Frequency-----------------------
1791 // Estimate block frequencies based on IfNode probabilities.
1792 void PhaseCFG::estimate_block_frequency() {
1793
1794 // Force conditional branches leading to uncommon traps to be unlikely,
1795 // not because we get to the uncommon_trap with less relative frequency,
1796 // but because an uncommon_trap typically causes a deopt, so we only get
1797 // there once.
1798 if (C->do_freq_based_layout()) {
1799 Block_List worklist;
1800 Block* root_blk = get_block(0);
1801 for (uint i = 1; i < root_blk->num_preds(); i++) {
1802 Block *pb = get_block_for_node(root_blk->pred(i));
1803 if (pb->has_uncommon_code()) {
1804 worklist.push(pb);
1805 }
1806 }
1807 while (worklist.size() > 0) {
1808 Block* uct = worklist.pop();
1809 if (uct == get_root_block()) {
1810 continue;
1811 }
1812 for (uint i = 1; i < uct->num_preds(); i++) {
1813 Block *pb = get_block_for_node(uct->pred(i));
1814 if (pb->_num_succs == 1) {
1815 worklist.push(pb);
1816 } else if (pb->num_fall_throughs() == 2) {
1817 pb->update_uncommon_branch(uct);
1818 }
1819 }
1820 }
1821 }
1822
1823 // Create the loop tree and calculate loop depth.
1824 _root_loop = create_loop_tree();
1825 _root_loop->compute_loop_depth(0);
1826
1827 // Compute block frequency of each block, relative to a single loop entry.
1828 _root_loop->compute_freq();
1829
1830 // Adjust all frequencies to be relative to a single method entry
1831 _root_loop->_freq = 1.0;
1832 _root_loop->scale_freq();
1833
1834 // Save outmost loop frequency for LRG frequency threshold
1835 _outer_loop_frequency = _root_loop->outer_loop_freq();
1836
1837 // force paths ending at uncommon traps to be infrequent
1838 if (!C->do_freq_based_layout()) {
1839 Block_List worklist;
1840 Block* root_blk = get_block(0);
1841 for (uint i = 1; i < root_blk->num_preds(); i++) {
1842 Block *pb = get_block_for_node(root_blk->pred(i));
1843 if (pb->has_uncommon_code()) {
1844 worklist.push(pb);
1845 }
1846 }
1847 while (worklist.size() > 0) {
1848 Block* uct = worklist.pop();
1849 uct->_freq = PROB_MIN;
1850 for (uint i = 1; i < uct->num_preds(); i++) {
1851 Block *pb = get_block_for_node(uct->pred(i));
1852 if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) {
1853 worklist.push(pb);
1854 }
1855 }
1856 }
1857 }
1858
1859 #ifdef ASSERT
1860 for (uint i = 0; i < number_of_blocks(); i++) {
1861 Block* b = get_block(i);
1862 assert(b->_freq >= MIN_BLOCK_FREQUENCY, "Register Allocator requires meaningful block frequency");
1863 }
1864 #endif
1865
1866 #ifndef PRODUCT
1867 if (PrintCFGBlockFreq) {
1868 tty->print_cr("CFG Block Frequencies");
1869 _root_loop->dump_tree();
1870 if (Verbose) {
1871 tty->print_cr("PhaseCFG dump");
1872 dump();
1873 tty->print_cr("Node dump");
1874 _root->dump(99999);
1875 }
1876 }
1877 #endif
1878 }
1879
1880 //----------------------------create_loop_tree--------------------------------
1881 // Create a loop tree from the CFG
1882 CFGLoop* PhaseCFG::create_loop_tree() {
1883
1884 #ifdef ASSERT
1885 assert(get_block(0) == get_root_block(), "first block should be root block");
1886 for (uint i = 0; i < number_of_blocks(); i++) {
1887 Block* block = get_block(i);
1888 // Check that _loop field are clear...we could clear them if not.
1889 assert(block->_loop == nullptr, "clear _loop expected");
1890 // Sanity check that the RPO numbering is reflected in the _blocks array.
1891 // It doesn't have to be for the loop tree to be built, but if it is not,
1892 // then the blocks have been reordered since dom graph building...which
1893 // may question the RPO numbering
1894 assert(block->_rpo == i, "unexpected reverse post order number");
1895 }
1896 #endif
1897
1898 int idct = 0;
1899 CFGLoop* root_loop = new CFGLoop(idct++);
1900
1901 Block_List worklist;
1902
1903 // Assign blocks to loops
1904 for(uint i = number_of_blocks() - 1; i > 0; i-- ) { // skip Root block
1905 Block* block = get_block(i);
1906
1907 if (block->head()->is_Loop()) {
1908 Block* loop_head = block;
1909 assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1910 Node* tail_n = loop_head->pred(LoopNode::LoopBackControl);
1911 Block* tail = get_block_for_node(tail_n);
1912
1913 // Defensively filter out Loop nodes for non-single-entry loops.
1914 // For all reasonable loops, the head occurs before the tail in RPO.
1915 if (i <= tail->_rpo) {
1916
1917 // The tail and (recursive) predecessors of the tail
1918 // are made members of a new loop.
1919
1920 assert(worklist.size() == 0, "nonempty worklist");
1921 CFGLoop* nloop = new CFGLoop(idct++);
1922 assert(loop_head->_loop == nullptr, "just checking");
1923 loop_head->_loop = nloop;
1924 // Add to nloop so push_pred() will skip over inner loops
1925 nloop->add_member(loop_head);
1926 nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, this);
1927
1928 while (worklist.size() > 0) {
1929 Block* member = worklist.pop();
1930 if (member != loop_head) {
1931 for (uint j = 1; j < member->num_preds(); j++) {
1932 nloop->push_pred(member, j, worklist, this);
1933 }
1934 }
1935 }
1936 }
1937 }
1938 }
1939
1940 // Create a member list for each loop consisting
1941 // of both blocks and (immediate child) loops.
1942 for (uint i = 0; i < number_of_blocks(); i++) {
1943 Block* block = get_block(i);
1944 CFGLoop* lp = block->_loop;
1945 if (lp == nullptr) {
1946 // Not assigned to a loop. Add it to the method's pseudo loop.
1947 block->_loop = root_loop;
1948 lp = root_loop;
1949 }
1950 if (lp == root_loop || block != lp->head()) { // loop heads are already members
1951 lp->add_member(block);
1952 }
1953 if (lp != root_loop) {
1954 if (lp->parent() == nullptr) {
1955 // Not a nested loop. Make it a child of the method's pseudo loop.
1956 root_loop->add_nested_loop(lp);
1957 }
1958 if (block == lp->head()) {
1959 // Add nested loop to member list of parent loop.
1960 lp->parent()->add_member(lp);
1961 }
1962 }
1963 }
1964
1965 return root_loop;
1966 }
1967
1968 //------------------------------push_pred--------------------------------------
1969 void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg) {
1970 Node* pred_n = blk->pred(i);
1971 Block* pred = cfg->get_block_for_node(pred_n);
1972 CFGLoop *pred_loop = pred->_loop;
1973 if (pred_loop == nullptr) {
1974 // Filter out blocks for non-single-entry loops.
1975 // For all reasonable loops, the head occurs before the tail in RPO.
1976 if (pred->_rpo > head()->_rpo) {
1977 pred->_loop = this;
1978 worklist.push(pred);
1979 }
1980 } else if (pred_loop != this) {
1981 // Nested loop.
1982 while (pred_loop->_parent != nullptr && pred_loop->_parent != this) {
1983 pred_loop = pred_loop->_parent;
1984 }
1985 // Make pred's loop be a child
1986 if (pred_loop->_parent == nullptr) {
1987 add_nested_loop(pred_loop);
1988 // Continue with loop entry predecessor.
1989 Block* pred_head = pred_loop->head();
1990 assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1991 assert(pred_head != head(), "loop head in only one loop");
1992 push_pred(pred_head, LoopNode::EntryControl, worklist, cfg);
1993 } else {
1994 assert(pred_loop->_parent == this && _parent == nullptr, "just checking");
1995 }
1996 }
1997 }
1998
1999 //------------------------------add_nested_loop--------------------------------
2000 // Make cl a child of the current loop in the loop tree.
2001 void CFGLoop::add_nested_loop(CFGLoop* cl) {
2002 assert(_parent == nullptr, "no parent yet");
2003 assert(cl != this, "not my own parent");
2004 cl->_parent = this;
2005 CFGLoop* ch = _child;
2006 if (ch == nullptr) {
2007 _child = cl;
2008 } else {
2009 while (ch->_sibling != nullptr) { ch = ch->_sibling; }
2010 ch->_sibling = cl;
2011 }
2012 }
2013
2014 //------------------------------compute_loop_depth-----------------------------
2015 // Store the loop depth in each CFGLoop object.
2016 // Recursively walk the children to do the same for them.
2017 void CFGLoop::compute_loop_depth(int depth) {
2018 _depth = depth;
2019 CFGLoop* ch = _child;
2020 while (ch != nullptr) {
2021 ch->compute_loop_depth(depth + 1);
2022 ch = ch->_sibling;
2023 }
2024 }
2025
2026 //------------------------------compute_freq-----------------------------------
2027 // Compute the frequency of each block and loop, relative to a single entry
2028 // into the dominating loop head.
2029 void CFGLoop::compute_freq() {
2030 // Bottom up traversal of loop tree (visit inner loops first.)
2031 // Set loop head frequency to 1.0, then transitively
2032 // compute frequency for all successors in the loop,
2033 // as well as for each exit edge. Inner loops are
2034 // treated as single blocks with loop exit targets
2035 // as the successor blocks.
2036
2037 // Nested loops first
2038 CFGLoop* ch = _child;
2039 while (ch != nullptr) {
2040 ch->compute_freq();
2041 ch = ch->_sibling;
2042 }
2043 assert (_members.length() > 0, "no empty loops");
2044 Block* hd = head();
2045 hd->_freq = 1.0;
2046 for (int i = 0; i < _members.length(); i++) {
2047 CFGElement* s = _members.at(i);
2048 double freq = s->_freq;
2049 if (s->is_block()) {
2050 Block* b = s->as_Block();
2051 for (uint j = 0; j < b->_num_succs; j++) {
2052 Block* sb = b->_succs[j];
2053 update_succ_freq(sb, freq * b->succ_prob(j));
2054 }
2055 } else {
2056 CFGLoop* lp = s->as_CFGLoop();
2057 assert(lp->_parent == this, "immediate child");
2058 for (int k = 0; k < lp->_exits.length(); k++) {
2059 Block* eb = lp->_exits.at(k).get_target();
2060 double prob = lp->_exits.at(k).get_prob();
2061 update_succ_freq(eb, freq * prob);
2062 }
2063 }
2064 }
2065
2066 // For all loops other than the outer, "method" loop,
2067 // sum and normalize the exit probability. The "method" loop
2068 // should keep the initial exit probability of 1, so that
2069 // inner blocks do not get erroneously scaled.
2070 if (_depth != 0) {
2071 // Total the exit probabilities for this loop.
2072 double exits_sum = 0.0f;
2073 for (int i = 0; i < _exits.length(); i++) {
2074 exits_sum += _exits.at(i).get_prob();
2075 }
2076
2077 // Normalize the exit probabilities. Until now, the
2078 // probabilities estimate the possibility of exit per
2079 // a single loop iteration; afterward, they estimate
2080 // the probability of exit per loop entry.
2081 for (int i = 0; i < _exits.length(); i++) {
2082 Block* et = _exits.at(i).get_target();
2083 float new_prob = 0.0f;
2084 if (_exits.at(i).get_prob() > 0.0f) {
2085 new_prob = _exits.at(i).get_prob() / exits_sum;
2086 }
2087 BlockProbPair bpp(et, new_prob);
2088 _exits.at_put(i, bpp);
2089 }
2090
2091 // Save the total, but guard against unreasonable probability,
2092 // as the value is used to estimate the loop trip count.
2093 // An infinite trip count would blur relative block
2094 // frequencies.
2095 if (exits_sum > 1.0f) exits_sum = 1.0;
2096 if (exits_sum < PROB_MIN) exits_sum = PROB_MIN;
2097 _exit_prob = exits_sum;
2098 }
2099 }
2100
2101 //------------------------------succ_prob-------------------------------------
2102 // Determine the probability of reaching successor 'i' from the receiver block.
2103 float Block::succ_prob(uint i) {
2104 int eidx = end_idx();
2105 Node *n = get_node(eidx); // Get ending Node
2106
2107 int op = n->Opcode();
2108 if (n->is_Mach()) {
2109 if (n->is_MachNullCheck()) {
2110 // Can only reach here if called after lcm. The original Op_If is gone,
2111 // so we attempt to infer the probability from one or both of the
2112 // successor blocks.
2113 assert(_num_succs == 2, "expecting 2 successors of a null check");
2114 // If either successor has only one predecessor, then the
2115 // probability estimate can be derived using the
2116 // relative frequency of the successor and this block.
2117 if (_succs[i]->num_preds() == 2) {
2118 return _succs[i]->_freq / _freq;
2119 } else if (_succs[1-i]->num_preds() == 2) {
2120 return 1 - (_succs[1-i]->_freq / _freq);
2121 } else {
2122 // Estimate using both successor frequencies
2123 float freq = _succs[i]->_freq;
2124 return freq / (freq + _succs[1-i]->_freq);
2125 }
2126 }
2127 op = n->as_Mach()->ideal_Opcode();
2128 }
2129
2130
2131 // Switch on branch type
2132 switch( op ) {
2133 case Op_CountedLoopEnd:
2134 case Op_If: {
2135 assert (i < 2, "just checking");
2136 // Conditionals pass on only part of their frequency
2137 float prob = n->as_MachIf()->_prob;
2138 assert(prob >= 0.0 && prob <= 1.0, "out of range probability");
2139 // If succ[i] is the FALSE branch, invert path info
2140 if( get_node(i + eidx + 1)->Opcode() == Op_IfFalse ) {
2141 return 1.0f - prob; // not taken
2142 } else {
2143 return prob; // taken
2144 }
2145 }
2146
2147 case Op_Jump:
2148 return n->as_MachJump()->_probs[get_node(i + eidx + 1)->as_JumpProj()->_con];
2149
2150 case Op_Catch: {
2151 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
2152 if (ci->_con == CatchProjNode::fall_through_index) {
2153 // Fall-thru path gets the lion's share.
2154 return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs;
2155 } else {
2156 // Presume exceptional paths are equally unlikely
2157 return PROB_UNLIKELY_MAG(5);
2158 }
2159 }
2160
2161 case Op_Root:
2162 case Op_Goto:
2163 // Pass frequency straight thru to target
2164 return 1.0f;
2165
2166 case Op_NeverBranch:
2167 return 0.0f;
2168
2169 case Op_TailCall:
2170 case Op_TailJump:
2171 case Op_ForwardException:
2172 case Op_Return:
2173 case Op_Halt:
2174 case Op_Rethrow:
2175 // Do not push out freq to root block
2176 return 0.0f;
2177
2178 default:
2179 ShouldNotReachHere();
2180 }
2181
2182 return 0.0f;
2183 }
2184
2185 //------------------------------num_fall_throughs-----------------------------
2186 // Return the number of fall-through candidates for a block
2187 int Block::num_fall_throughs() {
2188 int eidx = end_idx();
2189 Node *n = get_node(eidx); // Get ending Node
2190
2191 int op = n->Opcode();
2192 if (n->is_Mach()) {
2193 if (n->is_MachNullCheck()) {
2194 // In theory, either side can fall-thru, for simplicity sake,
2195 // let's say only the false branch can now.
2196 return 1;
2197 }
2198 op = n->as_Mach()->ideal_Opcode();
2199 }
2200
2201 // Switch on branch type
2202 switch( op ) {
2203 case Op_CountedLoopEnd:
2204 case Op_If:
2205 return 2;
2206
2207 case Op_Root:
2208 case Op_Goto:
2209 return 1;
2210
2211 case Op_Catch: {
2212 for (uint i = 0; i < _num_succs; i++) {
2213 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
2214 if (ci->_con == CatchProjNode::fall_through_index) {
2215 return 1;
2216 }
2217 }
2218 return 0;
2219 }
2220
2221 case Op_Jump:
2222 case Op_NeverBranch:
2223 case Op_TailCall:
2224 case Op_TailJump:
2225 case Op_ForwardException:
2226 case Op_Return:
2227 case Op_Halt:
2228 case Op_Rethrow:
2229 return 0;
2230
2231 default:
2232 ShouldNotReachHere();
2233 }
2234
2235 return 0;
2236 }
2237
2238 //------------------------------succ_fall_through-----------------------------
2239 // Return true if a specific successor could be fall-through target.
2240 bool Block::succ_fall_through(uint i) {
2241 int eidx = end_idx();
2242 Node *n = get_node(eidx); // Get ending Node
2243
2244 int op = n->Opcode();
2245 if (n->is_Mach()) {
2246 if (n->is_MachNullCheck()) {
2247 // In theory, either side can fall-thru, for simplicity sake,
2248 // let's say only the false branch can now.
2249 return get_node(i + eidx + 1)->Opcode() == Op_IfFalse;
2250 }
2251 op = n->as_Mach()->ideal_Opcode();
2252 }
2253
2254 // Switch on branch type
2255 switch( op ) {
2256 case Op_CountedLoopEnd:
2257 case Op_If:
2258 case Op_Root:
2259 case Op_Goto:
2260 return true;
2261
2262 case Op_Catch: {
2263 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
2264 return ci->_con == CatchProjNode::fall_through_index;
2265 }
2266
2267 case Op_Jump:
2268 case Op_NeverBranch:
2269 case Op_TailCall:
2270 case Op_TailJump:
2271 case Op_ForwardException:
2272 case Op_Return:
2273 case Op_Halt:
2274 case Op_Rethrow:
2275 return false;
2276
2277 default:
2278 ShouldNotReachHere();
2279 }
2280
2281 return false;
2282 }
2283
2284 //------------------------------update_uncommon_branch------------------------
2285 // Update the probability of a two-branch to be uncommon
2286 void Block::update_uncommon_branch(Block* ub) {
2287 int eidx = end_idx();
2288 Node *n = get_node(eidx); // Get ending Node
2289
2290 int op = n->as_Mach()->ideal_Opcode();
2291
2292 assert(op == Op_CountedLoopEnd || op == Op_If, "must be a If");
2293 assert(num_fall_throughs() == 2, "must be a two way branch block");
2294
2295 // Which successor is ub?
2296 uint s;
2297 for (s = 0; s <_num_succs; s++) {
2298 if (_succs[s] == ub) break;
2299 }
2300 assert(s < 2, "uncommon successor must be found");
2301
2302 // If ub is the true path, make the proability small, else
2303 // ub is the false path, and make the probability large
2304 bool invert = (get_node(s + eidx + 1)->Opcode() == Op_IfFalse);
2305
2306 // Get existing probability
2307 float p = n->as_MachIf()->_prob;
2308
2309 if (invert) p = 1.0 - p;
2310 if (p > PROB_MIN) {
2311 p = PROB_MIN;
2312 }
2313 if (invert) p = 1.0 - p;
2314
2315 n->as_MachIf()->_prob = p;
2316 }
2317
2318 //------------------------------update_succ_freq-------------------------------
2319 // Update the appropriate frequency associated with block 'b', a successor of
2320 // a block in this loop.
2321 void CFGLoop::update_succ_freq(Block* b, double freq) {
2322 if (b->_loop == this) {
2323 if (b == head()) {
2324 // back branch within the loop
2325 // Do nothing now, the loop carried frequency will be
2326 // adjust later in scale_freq().
2327 } else {
2328 // simple branch within the loop
2329 b->_freq += freq;
2330 }
2331 } else if (!in_loop_nest(b)) {
2332 // branch is exit from this loop
2333 BlockProbPair bpp(b, freq);
2334 _exits.append(bpp);
2335 } else {
2336 // branch into nested loop
2337 CFGLoop* ch = b->_loop;
2338 ch->_freq += freq;
2339 }
2340 }
2341
2342 //------------------------------in_loop_nest-----------------------------------
2343 // Determine if block b is in the receiver's loop nest.
2344 bool CFGLoop::in_loop_nest(Block* b) {
2345 int depth = _depth;
2346 CFGLoop* b_loop = b->_loop;
2347 int b_depth = b_loop->_depth;
2348 if (depth == b_depth) {
2349 return true;
2350 }
2351 while (b_depth > depth) {
2352 b_loop = b_loop->_parent;
2353 b_depth = b_loop->_depth;
2354 }
2355 return b_loop == this;
2356 }
2357
2358 //------------------------------scale_freq-------------------------------------
2359 // Scale frequency of loops and blocks by trip counts from outer loops
2360 // Do a top down traversal of loop tree (visit outer loops first.)
2361 void CFGLoop::scale_freq() {
2362 double loop_freq = _freq * trip_count();
2363 _freq = loop_freq;
2364 for (int i = 0; i < _members.length(); i++) {
2365 CFGElement* s = _members.at(i);
2366 double block_freq = s->_freq * loop_freq;
2367 if (g_isnan(block_freq) || block_freq < MIN_BLOCK_FREQUENCY)
2368 block_freq = MIN_BLOCK_FREQUENCY;
2369 s->_freq = block_freq;
2370 }
2371 CFGLoop* ch = _child;
2372 while (ch != nullptr) {
2373 ch->scale_freq();
2374 ch = ch->_sibling;
2375 }
2376 }
2377
2378 // Frequency of outer loop
2379 double CFGLoop::outer_loop_freq() const {
2380 if (_child != nullptr) {
2381 return _child->_freq;
2382 }
2383 return _freq;
2384 }
2385
2386 #ifndef PRODUCT
2387 //------------------------------dump_tree--------------------------------------
2388 void CFGLoop::dump_tree() const {
2389 dump();
2390 if (_child != nullptr) _child->dump_tree();
2391 if (_sibling != nullptr) _sibling->dump_tree();
2392 }
2393
2394 //------------------------------dump-------------------------------------------
2395 void CFGLoop::dump() const {
2396 for (int i = 0; i < _depth; i++) tty->print(" ");
2397 tty->print("%s: %d trip_count: %6.0f freq: %6.0f\n",
2398 _depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq);
2399 for (int i = 0; i < _depth; i++) tty->print(" ");
2400 tty->print(" members:");
2401 int k = 0;
2402 for (int i = 0; i < _members.length(); i++) {
2403 if (k++ >= 6) {
2404 tty->print("\n ");
2405 for (int j = 0; j < _depth+1; j++) tty->print(" ");
2406 k = 0;
2407 }
2408 CFGElement *s = _members.at(i);
2409 if (s->is_block()) {
2410 Block *b = s->as_Block();
2411 tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq);
2412 } else {
2413 CFGLoop* lp = s->as_CFGLoop();
2414 tty->print(" L%d(%6.3f)", lp->_id, lp->_freq);
2415 }
2416 }
2417 tty->print("\n");
2418 for (int i = 0; i < _depth; i++) tty->print(" ");
2419 tty->print(" exits: ");
2420 k = 0;
2421 for (int i = 0; i < _exits.length(); i++) {
2422 if (k++ >= 7) {
2423 tty->print("\n ");
2424 for (int j = 0; j < _depth+1; j++) tty->print(" ");
2425 k = 0;
2426 }
2427 Block *blk = _exits.at(i).get_target();
2428 double prob = _exits.at(i).get_prob();
2429 tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100));
2430 }
2431 tty->print("\n");
2432 }
2433 #endif