1 /*
2 * Copyright (c) 1997, 2024, 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.
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 != nullptr) {
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.
78 void PhaseCFG::replace_block_proj_ctrl( Node *n ) {
79 const Node *in0 = n->in(0);
80 assert(in0 != nullptr, "Only control-dependent");
81 const Node *p = in0->is_block_proj();
82 if (p != nullptr && 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
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 != nullptr && d != nullptr, "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
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
155 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
162 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.
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
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 = nullptr;
216 for (uint i = node->len()-1; i >= node->req(); i--) {
217 Node* m = node->in(i);
218 if (m == nullptr) continue;
219 assert(is_CFG(m), "must be a CFG node");
220 node->rm_prec(i);
221 if (n == nullptr) {
222 n = m;
223 } else {
224 assert(is_dominator(n, m) || is_dominator(m, n), "one must dominate the other");
225 n = is_dominator(n, m) ? m : n;
226 }
227 }
228 if (n != nullptr) {
229 assert(node->in(0), "control should have been set");
230 assert(is_dominator(n, node->in(0)) || is_dominator(node->in(0), n), "one must dominate the other");
231 if (!is_dominator(n, node->in(0))) {
232 node->set_req(0, n);
233 }
234 }
235
236 // process all inputs that are non null
237 for (int i = node->req()-1; i >= 0; --i) {
238 if (node->in(i) != nullptr) {
239 spstack.push(node->in(i));
240 }
241 }
242 }
243 }
244 }
245
246 #ifdef ASSERT
247 // Assert that new input b2 is dominated by all previous inputs.
248 // Check this by by seeing that it is dominated by b1, the deepest
249 // input observed until b2.
250 static void assert_dom(Block* b1, Block* b2, Node* n, const PhaseCFG* cfg) {
251 if (b1 == nullptr) return;
252 assert(b1->_dom_depth < b2->_dom_depth, "sanity");
253 Block* tmp = b2;
254 while (tmp != b1 && tmp != nullptr) {
255 tmp = tmp->_idom;
256 }
257 if (tmp != b1) {
258 // Detected an unschedulable graph. Print some nice stuff and die.
259 tty->print_cr("!!! Unschedulable graph !!!");
260 for (uint j=0; j<n->len(); j++) { // For all inputs
261 Node* inn = n->in(j); // Get input
262 if (inn == nullptr) continue; // Ignore null, missing inputs
263 Block* inb = cfg->get_block_for_node(inn);
264 tty->print("B%d idom=B%d depth=%2d ",inb->_pre_order,
265 inb->_idom ? inb->_idom->_pre_order : 0, inb->_dom_depth);
266 inn->dump();
267 }
268 tty->print("Failing node: ");
269 n->dump();
270 assert(false, "unscheduable graph");
271 }
272 }
273 #endif
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 DEBUG_ONLY(assert_dom(deepb, inb, n, cfg));
289 deepb = inb; // Save deepest block
290 deepb_dom_depth = deepb->_dom_depth;
291 }
292 }
293 assert(deepb != nullptr, "must be at least one input to n");
294 return deepb;
295 }
296
297
298 //------------------------------schedule_early---------------------------------
299 // Find the earliest Block any instruction can be placed in. Some instructions
300 // are pinned into Blocks. Unpinned instructions can appear in last block in
301 // which all their inputs occur.
302 bool PhaseCFG::schedule_early(VectorSet &visited, Node_Stack &roots) {
303 // Allocate stack with enough space to avoid frequent realloc
304 Node_Stack nstack(roots.size() + 8);
305 // _root will be processed among C->top() inputs
306 roots.push(C->top(), 0);
307 visited.set(C->top()->_idx);
308
309 while (roots.size() != 0) {
310 // Use local variables nstack_top_n & nstack_top_i to cache values
311 // on stack's top.
312 Node* parent_node = roots.node();
313 uint input_index = 0;
314 roots.pop();
315
316 while (true) {
317 if (input_index == 0) {
318 // Fixup some control. Constants without control get attached
319 // to root and nodes that use is_block_proj() nodes should be attached
320 // to the region that starts their block.
321 const Node* control_input = parent_node->in(0);
322 if (control_input != nullptr) {
323 replace_block_proj_ctrl(parent_node);
324 } else {
325 // Is a constant with NO inputs?
326 if (parent_node->req() == 1) {
327 parent_node->set_req(0, _root);
328 }
329 }
330 }
331
332 // First, visit all inputs and force them to get a block. If an
333 // input is already in a block we quit following inputs (to avoid
334 // cycles). Instead we put that Node on a worklist to be handled
335 // later (since IT'S inputs may not have a block yet).
336
337 // Assume all n's inputs will be processed
338 bool done = true;
339
340 while (input_index < parent_node->len()) {
341 Node* in = parent_node->in(input_index++);
342 if (in == nullptr) {
343 continue;
344 }
345
346 int is_visited = visited.test_set(in->_idx);
347 if (!has_block(in)) {
348 if (is_visited) {
349 assert(false, "graph should be schedulable");
350 return false;
351 }
352 // Save parent node and next input's index.
353 nstack.push(parent_node, input_index);
354 // Process current input now.
355 parent_node = in;
356 input_index = 0;
357 // Not all n's inputs processed.
358 done = false;
359 break;
360 } else if (!is_visited) {
361 // Visit this guy later, using worklist
362 roots.push(in, 0);
363 }
364 }
365
366 if (done) {
367 // All of n's inputs have been processed, complete post-processing.
368
369 // Some instructions are pinned into a block. These include Region,
370 // Phi, Start, Return, and other control-dependent instructions and
371 // any projections which depend on them.
372 if (!parent_node->pinned()) {
373 // Set earliest legal block.
374 Block* earliest_block = find_deepest_input(parent_node, this);
375 map_node_to_block(parent_node, earliest_block);
376 } else {
377 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");
378 }
379
380 if (nstack.is_empty()) {
381 // Finished all nodes on stack.
382 // Process next node on the worklist 'roots'.
383 break;
384 }
385 // Get saved parent node and next input's index.
386 parent_node = nstack.node();
387 input_index = nstack.index();
388 nstack.pop();
389 }
390 }
391 }
392 return true;
393 }
394
395 //------------------------------dom_lca----------------------------------------
396 // Find least common ancestor in dominator tree
397 // LCA is a current notion of LCA, to be raised above 'this'.
398 // As a convenient boundary condition, return 'this' if LCA is null.
399 // Find the LCA of those two nodes.
400 Block* Block::dom_lca(Block* LCA) {
401 if (LCA == nullptr || LCA == this) return this;
402
403 Block* anc = this;
404 while (anc->_dom_depth > LCA->_dom_depth)
405 anc = anc->_idom; // Walk up till anc is as high as LCA
406
407 while (LCA->_dom_depth > anc->_dom_depth)
408 LCA = LCA->_idom; // Walk up till LCA is as high as anc
409
410 while (LCA != anc) { // Walk both up till they are the same
411 LCA = LCA->_idom;
412 anc = anc->_idom;
413 }
414
415 return LCA;
416 }
417
418 //--------------------------raise_LCA_above_use--------------------------------
419 // We are placing a definition, and have been given a def->use edge.
420 // The definition must dominate the use, so move the LCA upward in the
421 // dominator tree to dominate the use. If the use is a phi, adjust
422 // the LCA only with the phi input paths which actually use this def.
423 static Block* raise_LCA_above_use(Block* LCA, Node* use, Node* def, const PhaseCFG* cfg) {
424 Block* buse = cfg->get_block_for_node(use);
425 if (buse == nullptr) return LCA; // Unused killing Projs have no use block
426 if (!use->is_Phi()) return buse->dom_lca(LCA);
427 uint pmax = use->req(); // Number of Phi inputs
428 // Why does not this loop just break after finding the matching input to
429 // the Phi? Well...it's like this. I do not have true def-use/use-def
430 // chains. Means I cannot distinguish, from the def-use direction, which
431 // of many use-defs lead from the same use to the same def. That is, this
432 // Phi might have several uses of the same def. Each use appears in a
433 // different predecessor block. But when I enter here, I cannot distinguish
434 // which use-def edge I should find the predecessor block for. So I find
435 // them all. Means I do a little extra work if a Phi uses the same value
436 // more than once.
437 for (uint j=1; j<pmax; j++) { // For all inputs
438 if (use->in(j) == def) { // Found matching input?
439 Block* pred = cfg->get_block_for_node(buse->pred(j));
440 LCA = pred->dom_lca(LCA);
441 }
442 }
443 return LCA;
444 }
445
446 //----------------------------raise_LCA_above_marks----------------------------
447 // Return a new LCA that dominates LCA and any of its marked predecessors.
448 // Search all my parents up to 'early' (exclusive), looking for predecessors
449 // which are marked with the given index. Return the LCA (in the dom tree)
450 // of all marked blocks. If there are none marked, return the original
451 // LCA.
452 static Block* raise_LCA_above_marks(Block* LCA, node_idx_t mark, Block* early, const PhaseCFG* cfg) {
453 Block_List worklist;
454 worklist.push(LCA);
455 while (worklist.size() > 0) {
456 Block* mid = worklist.pop();
457 if (mid == early) continue; // stop searching here
458
459 // Test and set the visited bit.
460 if (mid->raise_LCA_visited() == mark) continue; // already visited
461
462 // Don't process the current LCA, otherwise the search may terminate early
463 if (mid != LCA && mid->raise_LCA_mark() == mark) {
464 // Raise the LCA.
465 LCA = mid->dom_lca(LCA);
466 if (LCA == early) break; // stop searching everywhere
467 assert(early->dominates(LCA), "early is high enough");
468 // Resume searching at that point, skipping intermediate levels.
469 worklist.push(LCA);
470 if (LCA == mid)
471 continue; // Don't mark as visited to avoid early termination.
472 } else {
473 // Keep searching through this block's predecessors.
474 for (uint j = 1, jmax = mid->num_preds(); j < jmax; j++) {
475 Block* mid_parent = cfg->get_block_for_node(mid->pred(j));
476 worklist.push(mid_parent);
477 }
478 }
479 mid->set_raise_LCA_visited(mark);
480 }
481 return LCA;
482 }
483
484 //--------------------------memory_early_block--------------------------------
485 // This is a variation of find_deepest_input, the heart of schedule_early.
486 // Find the "early" block for a load, if we considered only memory and
487 // address inputs, that is, if other data inputs were ignored.
488 //
489 // Because a subset of edges are considered, the resulting block will
490 // be earlier (at a shallower dom_depth) than the true schedule_early
491 // point of the node. We compute this earlier block as a more permissive
492 // site for anti-dependency insertion, but only if subsume_loads is enabled.
493 static Block* memory_early_block(Node* load, Block* early, const PhaseCFG* cfg) {
494 Node* base;
495 Node* index;
496 Node* store = load->in(MemNode::Memory);
497 load->as_Mach()->memory_inputs(base, index);
498
499 assert(base != NodeSentinel && index != NodeSentinel,
500 "unexpected base/index inputs");
501
502 Node* mem_inputs[4];
503 int mem_inputs_length = 0;
504 if (base != nullptr) mem_inputs[mem_inputs_length++] = base;
505 if (index != nullptr) mem_inputs[mem_inputs_length++] = index;
506 if (store != nullptr) mem_inputs[mem_inputs_length++] = store;
507
508 // In the comparison below, add one to account for the control input,
509 // which may be null, but always takes up a spot in the in array.
510 if (mem_inputs_length + 1 < (int) load->req()) {
511 // This "load" has more inputs than just the memory, base and index inputs.
512 // For purposes of checking anti-dependences, we need to start
513 // from the early block of only the address portion of the instruction,
514 // and ignore other blocks that may have factored into the wider
515 // schedule_early calculation.
516 if (load->in(0) != nullptr) mem_inputs[mem_inputs_length++] = load->in(0);
517
518 Block* deepb = nullptr; // Deepest block so far
519 int deepb_dom_depth = 0;
520 for (int i = 0; i < mem_inputs_length; i++) {
521 Block* inb = cfg->get_block_for_node(mem_inputs[i]);
522 if (deepb_dom_depth < (int) inb->_dom_depth) {
523 // The new inb must be dominated by the previous deepb.
524 // The various inputs must be linearly ordered in the dom
525 // tree, or else there will not be a unique deepest block.
526 DEBUG_ONLY(assert_dom(deepb, inb, load, cfg));
527 deepb = inb; // Save deepest block
528 deepb_dom_depth = deepb->_dom_depth;
529 }
530 }
531 early = deepb;
532 }
533
534 return early;
535 }
536
537 // This function is used by insert_anti_dependences to find unrelated loads for stores in implicit null checks.
538 bool PhaseCFG::unrelated_load_in_store_null_block(Node* store, Node* load) {
539 // We expect an anti-dependence edge from 'load' to 'store', except when
540 // implicit_null_check() has hoisted 'store' above its early block to
541 // perform an implicit null check, and 'load' is placed in the null
542 // block. In this case it is safe to ignore the anti-dependence, as the
543 // null block is only reached if 'store' tries to write to null object and
544 // 'load' read from non-null object (there is preceding check for that)
545 // These objects can't be the same.
546 Block* store_block = get_block_for_node(store);
547 Block* load_block = get_block_for_node(load);
548 Node* end = store_block->end();
549 if (end->is_MachNullCheck() && (end->in(1) == store) && store_block->dominates(load_block)) {
550 Node* if_true = end->find_out_with(Op_IfTrue);
551 assert(if_true != nullptr, "null check without null projection");
552 Node* null_block_region = if_true->find_out_with(Op_Region);
553 assert(null_block_region != nullptr, "null check without null region");
554 return get_block_for_node(null_block_region) == load_block;
555 }
556 return false;
557 }
558
559 class DefUseMemStatesQueue : public StackObj {
560 private:
561 class DefUsePair : public StackObj {
562 private:
563 Node* _def; // memory state
564 Node* _use; // use of the memory state that also modifies the memory state
565
566 public:
567 DefUsePair(Node* def, Node* use) :
568 _def(def), _use(use) {
569 }
570
571 DefUsePair() :
572 _def(nullptr), _use(nullptr) {
573 }
574
575 Node* def() const {
576 return _def;
577 }
578
579 Node* use() const {
580 return _use;
581 }
582 };
583
584 GrowableArray<DefUsePair> _queue;
585 GrowableArray<MergeMemNode*> _worklist_visited; // visited mergemem nodes
586
587 bool already_enqueued(Node* def_mem, PhiNode* use_phi) const {
588 // def_mem is one of the inputs of use_phi and at least one input of use_phi is
589 // not def_mem. It's however possible that use_phi has def_mem as input multiple
590 // times. If that happens, use_phi is recorded as a use of def_mem multiple
591 // times as well. When PhaseCFG::insert_anti_dependences() goes over
592 // uses of def_mem and enqueues them for processing, use_phi would then be
593 // enqueued for processing multiple times when it only needs to be
594 // processed once. The code below checks if use_phi as a use of def_mem was
595 // already enqueued to avoid redundant processing of use_phi.
596 int j = _queue.length()-1;
597 // If there are any use of def_mem already enqueued, they were enqueued
598 // last (all use of def_mem are processed in one go).
599 for (; j >= 0; j--) {
600 const DefUsePair& def_use_pair = _queue.at(j);
601 if (def_use_pair.def() != def_mem) {
602 // We're done with the uses of def_mem
603 break;
604 }
605 if (def_use_pair.use() == use_phi) {
606 return true;
607 }
608 }
609 #ifdef ASSERT
610 for (; j >= 0; j--) {
611 const DefUsePair& def_use_pair = _queue.at(j);
612 assert(def_use_pair.def() != def_mem, "Should be done with the uses of def_mem");
613 }
614 #endif
615 return false;
616 }
617
618 public:
619 DefUseMemStatesQueue(ResourceArea* area) {
620 }
621
622 void push(Node* def_mem_state, Node* use_mem_state) {
623 if (use_mem_state->is_MergeMem()) {
624 // Be sure we don't get into combinatorial problems.
625 if (!_worklist_visited.append_if_missing(use_mem_state->as_MergeMem())) {
626 return; // already on work list; do not repeat
627 }
628 } else if (use_mem_state->is_Phi()) {
629 // A Phi could have the same mem as input multiple times. If that's the case, we don't need to enqueue it
630 // more than once. We otherwise allow phis to be repeated; they can merge two relevant states.
631 if (already_enqueued(def_mem_state, use_mem_state->as_Phi())) {
632 return;
633 }
634 }
635
636 _queue.push(DefUsePair(def_mem_state, use_mem_state));
637 }
638
639 bool is_nonempty() const {
640 return _queue.is_nonempty();
641 }
642
643 Node* top_def() const {
644 return _queue.top().def();
645 }
646
647 Node* top_use() const {
648 return _queue.top().use();
649 }
650
651 void pop() {
652 _queue.pop();
653 }
654 };
655
656 //--------------------------insert_anti_dependences---------------------------
657 // A load may need to witness memory that nearby stores can overwrite.
658 // For each nearby store, either insert an "anti-dependence" edge
659 // from the load to the store, or else move LCA upward to force the
660 // load to (eventually) be scheduled in a block above the store.
661 //
662 // Do not add edges to stores on distinct control-flow paths;
663 // only add edges to stores which might interfere.
664 //
665 // Return the (updated) LCA. There will not be any possibly interfering
666 // store between the load's "early block" and the updated LCA.
667 // Any stores in the updated LCA will have new precedence edges
668 // back to the load. The caller is expected to schedule the load
669 // in the LCA, in which case the precedence edges will make LCM
670 // preserve anti-dependences. The caller may also hoist the load
671 // above the LCA, if it is not the early block.
672 Block* PhaseCFG::insert_anti_dependences(Block* LCA, Node* load, bool verify) {
673 ResourceMark rm;
674 assert(load->needs_anti_dependence_check(), "must be a load of some sort");
675 assert(LCA != nullptr, "");
676 DEBUG_ONLY(Block* LCA_orig = LCA);
677
678 // Compute the alias index. Loads and stores with different alias indices
679 // do not need anti-dependence edges.
680 int load_alias_idx = C->get_alias_index(load->adr_type());
681 #ifdef ASSERT
682 assert(Compile::AliasIdxTop <= load_alias_idx && load_alias_idx < C->num_alias_types(), "Invalid alias index");
683 if (load_alias_idx == Compile::AliasIdxBot && C->do_aliasing() &&
684 (PrintOpto || VerifyAliases ||
685 (PrintMiscellaneous && (WizardMode || Verbose)))) {
686 // Load nodes should not consume all of memory.
687 // Reporting a bottom type indicates a bug in adlc.
688 // If some particular type of node validly consumes all of memory,
689 // sharpen the preceding "if" to exclude it, so we can catch bugs here.
690 tty->print_cr("*** Possible Anti-Dependence Bug: Load consumes all of memory.");
691 load->dump(2);
692 if (VerifyAliases) assert(load_alias_idx != Compile::AliasIdxBot, "");
693 }
694 #endif
695
696 if (!C->alias_type(load_alias_idx)->is_rewritable()) {
697 // It is impossible to spoil this load by putting stores before it,
698 // because we know that the stores will never update the value
699 // which 'load' must witness.
700 return LCA;
701 }
702
703 node_idx_t load_index = load->_idx;
704
705 // Note the earliest legal placement of 'load', as determined by
706 // by the unique point in the dom tree where all memory effects
707 // and other inputs are first available. (Computed by schedule_early.)
708 // For normal loads, 'early' is the shallowest place (dom graph wise)
709 // to look for anti-deps between this load and any store.
710 Block* early = get_block_for_node(load);
711
712 // If we are subsuming loads, compute an "early" block that only considers
713 // memory or address inputs. This block may be different than the
714 // schedule_early block in that it could be at an even shallower depth in the
715 // dominator tree, and allow for a broader discovery of anti-dependences.
716 if (C->subsume_loads()) {
717 early = memory_early_block(load, early, this);
718 }
719
720 ResourceArea* area = Thread::current()->resource_area();
721 DefUseMemStatesQueue worklist_def_use_mem_states(area); // prior memory state to store and possible-def to explore
722 Node_List non_early_stores(area); // all relevant stores outside of early
723 bool must_raise_LCA = false;
724
725 // 'load' uses some memory state; look for users of the same state.
726 // Recurse through MergeMem nodes to the stores that use them.
727
728 // Each of these stores is a possible definition of memory
729 // that 'load' needs to use. We need to force 'load'
730 // to occur before each such store. When the store is in
731 // the same block as 'load', we insert an anti-dependence
732 // edge load->store.
733
734 // The relevant stores "nearby" the load consist of a tree rooted
735 // at initial_mem, with internal nodes of type MergeMem.
736 // Therefore, the branches visited by the worklist are of this form:
737 // initial_mem -> (MergeMem ->)* Memory state modifying node
738 // Memory state modifying nodes include Store and Phi nodes and any node for which needs_anti_dependence_check()
739 // returns false.
740 // The anti-dependence constraints apply only to the fringe of this tree.
741
742 Node* initial_mem = load->in(MemNode::Memory);
743
744 // We don't optimize the memory graph for pinned loads, so we may need to raise the
745 // root of our search tree through the corresponding slices of MergeMem nodes to
746 // get to the node that really creates the memory state for this slice.
747 if (load_alias_idx >= Compile::AliasIdxRaw) {
748 while (initial_mem->is_MergeMem()) {
749 MergeMemNode* mm = initial_mem->as_MergeMem();
750 Node* p = mm->memory_at(load_alias_idx);
751 if (p != mm->base_memory()) {
752 initial_mem = p;
753 } else {
754 break;
755 }
756 }
757 }
758 worklist_def_use_mem_states.push(nullptr, initial_mem);
759 while (worklist_def_use_mem_states.is_nonempty()) {
760 // Examine a nearby store to see if it might interfere with our load.
761 Node* def_mem_state = worklist_def_use_mem_states.top_def();
762 Node* use_mem_state = worklist_def_use_mem_states.top_use();
763 worklist_def_use_mem_states.pop();
764
765 uint op = use_mem_state->Opcode();
766
767 #ifdef ASSERT
768 // CacheWB nodes are peculiar in a sense that they both are anti-dependent and produce memory.
769 // Allow them to be treated as a store.
770 bool is_cache_wb = false;
771 if (use_mem_state->is_Mach()) {
772 int ideal_op = use_mem_state->as_Mach()->ideal_Opcode();
773 is_cache_wb = (ideal_op == Op_CacheWB);
774 }
775 assert(!use_mem_state->needs_anti_dependence_check() || is_cache_wb, "no loads");
776 #endif
777
778 // MergeMems do not directly have anti-deps.
779 // Treat them as internal nodes in a forward tree of memory states,
780 // the leaves of which are each a 'possible-def'.
781 if (use_mem_state == initial_mem // root (exclusive) of tree we are searching
782 || op == Op_MergeMem // internal node of tree we are searching
783 ) {
784 def_mem_state = use_mem_state; // It's not a possibly interfering store.
785 if (use_mem_state == initial_mem)
786 initial_mem = nullptr; // only process initial memory once
787
788 for (DUIterator_Fast imax, i = def_mem_state->fast_outs(imax); i < imax; i++) {
789 use_mem_state = def_mem_state->fast_out(i);
790 if (use_mem_state->needs_anti_dependence_check()) {
791 // use_mem_state is also a kind of load (i.e. needs_anti_dependence_check), and it is not a memory state
792 // modifying node (store, Phi or MergeMem). Hence, load can't be anti dependent on this node.
793 continue;
794 }
795 worklist_def_use_mem_states.push(def_mem_state, use_mem_state);
796 }
797 continue;
798 }
799
800 if (op == Op_MachProj || op == Op_Catch) continue;
801
802 // Compute the alias index. Loads and stores with different alias
803 // indices do not need anti-dependence edges. Wide MemBar's are
804 // anti-dependent on everything (except immutable memories).
805 const TypePtr* adr_type = use_mem_state->adr_type();
806 if (!C->can_alias(adr_type, load_alias_idx)) continue;
807
808 // Most slow-path runtime calls do NOT modify Java memory, but
809 // they can block and so write Raw memory.
810 if (use_mem_state->is_Mach()) {
811 MachNode* mstore = use_mem_state->as_Mach();
812 if (load_alias_idx != Compile::AliasIdxRaw) {
813 // Check for call into the runtime using the Java calling
814 // convention (and from there into a wrapper); it has no
815 // _method. Can't do this optimization for Native calls because
816 // they CAN write to Java memory.
817 if (mstore->ideal_Opcode() == Op_CallStaticJava) {
818 assert(mstore->is_MachSafePoint(), "");
819 MachSafePointNode* ms = (MachSafePointNode*) mstore;
820 assert(ms->is_MachCallJava(), "");
821 MachCallJavaNode* mcj = (MachCallJavaNode*) ms;
822 if (mcj->_method == nullptr) {
823 // These runtime calls do not write to Java visible memory
824 // (other than Raw) and so do not require anti-dependence edges.
825 continue;
826 }
827 }
828 // Same for SafePoints: they read/write Raw but only read otherwise.
829 // This is basically a workaround for SafePoints only defining control
830 // instead of control + memory.
831 if (mstore->ideal_Opcode() == Op_SafePoint)
832 continue;
833 } else {
834 // Some raw memory, such as the load of "top" at an allocation,
835 // can be control dependent on the previous safepoint. See
836 // comments in GraphKit::allocate_heap() about control input.
837 // Inserting an anti-dep between such a safepoint and a use
838 // creates a cycle, and will cause a subsequent failure in
839 // local scheduling. (BugId 4919904)
840 // (%%% How can a control input be a safepoint and not a projection??)
841 if (mstore->ideal_Opcode() == Op_SafePoint && load->in(0) == mstore)
842 continue;
843 }
844 }
845
846 // Identify a block that the current load must be above,
847 // or else observe that 'store' is all the way up in the
848 // earliest legal block for 'load'. In the latter case,
849 // immediately insert an anti-dependence edge.
850 Block* store_block = get_block_for_node(use_mem_state);
851 assert(store_block != nullptr, "unused killing projections skipped above");
852
853 if (use_mem_state->is_Phi()) {
854 // Loop-phis need to raise load before input. (Other phis are treated
855 // as store below.)
856 //
857 // 'load' uses memory which is one (or more) of the Phi's inputs.
858 // It must be scheduled not before the Phi, but rather before
859 // each of the relevant Phi inputs.
860 //
861 // Instead of finding the LCA of all inputs to a Phi that match 'mem',
862 // we mark each corresponding predecessor block and do a combined
863 // hoisting operation later (raise_LCA_above_marks).
864 //
865 // Do not assert(store_block != early, "Phi merging memory after access")
866 // PhiNode may be at start of block 'early' with backedge to 'early'
867 DEBUG_ONLY(bool found_match = false);
868 for (uint j = PhiNode::Input, jmax = use_mem_state->req(); j < jmax; j++) {
869 if (use_mem_state->in(j) == def_mem_state) { // Found matching input?
870 DEBUG_ONLY(found_match = true);
871 Block* pred_block = get_block_for_node(store_block->pred(j));
872 if (pred_block != early) {
873 // If any predecessor of the Phi matches the load's "early block",
874 // we do not need a precedence edge between the Phi and 'load'
875 // since the load will be forced into a block preceding the Phi.
876 pred_block->set_raise_LCA_mark(load_index);
877 assert(!LCA_orig->dominates(pred_block) ||
878 early->dominates(pred_block), "early is high enough");
879 must_raise_LCA = true;
880 } else {
881 // anti-dependent upon PHI pinned below 'early', no edge needed
882 LCA = early; // but can not schedule below 'early'
883 }
884 }
885 }
886 assert(found_match, "no worklist bug");
887 } else if (store_block != early) {
888 // 'store' is between the current LCA and earliest possible block.
889 // Label its block, and decide later on how to raise the LCA
890 // to include the effect on LCA of this store.
891 // If this store's block gets chosen as the raised LCA, we
892 // will find him on the non_early_stores list and stick him
893 // with a precedence edge.
894 // (But, don't bother if LCA is already raised all the way.)
895 if (LCA != early && !unrelated_load_in_store_null_block(use_mem_state, load)) {
896 store_block->set_raise_LCA_mark(load_index);
897 must_raise_LCA = true;
898 non_early_stores.push(use_mem_state);
899 }
900 } else {
901 // Found a possibly-interfering store in the load's 'early' block.
902 // This means 'load' cannot sink at all in the dominator tree.
903 // Add an anti-dep edge, and squeeze 'load' into the highest block.
904 assert(use_mem_state != load->find_exact_control(load->in(0)), "dependence cycle found");
905 if (verify) {
906 assert(use_mem_state->find_edge(load) != -1 || unrelated_load_in_store_null_block(use_mem_state, load),
907 "missing precedence edge");
908 } else {
909 use_mem_state->add_prec(load);
910 }
911 LCA = early;
912 // This turns off the process of gathering non_early_stores.
913 }
914 }
915 // (Worklist is now empty; all nearby stores have been visited.)
916
917 // Finished if 'load' must be scheduled in its 'early' block.
918 // If we found any stores there, they have already been given
919 // precedence edges.
920 if (LCA == early) return LCA;
921
922 // We get here only if there are no possibly-interfering stores
923 // in the load's 'early' block. Move LCA up above all predecessors
924 // which contain stores we have noted.
925 //
926 // The raised LCA block can be a home to such interfering stores,
927 // but its predecessors must not contain any such stores.
928 //
929 // The raised LCA will be a lower bound for placing the load,
930 // preventing the load from sinking past any block containing
931 // a store that may invalidate the memory state required by 'load'.
932 if (must_raise_LCA)
933 LCA = raise_LCA_above_marks(LCA, load->_idx, early, this);
934 if (LCA == early) return LCA;
935
936 // Insert anti-dependence edges from 'load' to each store
937 // in the non-early LCA block.
938 // Mine the non_early_stores list for such stores.
939 if (LCA->raise_LCA_mark() == load_index) {
940 while (non_early_stores.size() > 0) {
941 Node* store = non_early_stores.pop();
942 Block* store_block = get_block_for_node(store);
943 if (store_block == LCA) {
944 // add anti_dependence from store to load in its own block
945 assert(store != load->find_exact_control(load->in(0)), "dependence cycle found");
946 if (verify) {
947 assert(store->find_edge(load) != -1, "missing precedence edge");
948 } else {
949 store->add_prec(load);
950 }
951 } else {
952 assert(store_block->raise_LCA_mark() == load_index, "block was marked");
953 // Any other stores we found must be either inside the new LCA
954 // or else outside the original LCA. In the latter case, they
955 // did not interfere with any use of 'load'.
956 assert(LCA->dominates(store_block)
957 || !LCA_orig->dominates(store_block), "no stray stores");
958 }
959 }
960 }
961
962 // Return the highest block containing stores; any stores
963 // within that block have been given anti-dependence edges.
964 return LCA;
965 }
966
967 // This class is used to iterate backwards over the nodes in the graph.
968
969 class Node_Backward_Iterator {
970
971 private:
972 Node_Backward_Iterator();
973
974 public:
975 // Constructor for the iterator
976 Node_Backward_Iterator(Node *root, VectorSet &visited, Node_Stack &stack, PhaseCFG &cfg);
977
978 // Postincrement operator to iterate over the nodes
979 Node *next();
980
981 private:
982 VectorSet &_visited;
983 Node_Stack &_stack;
984 PhaseCFG &_cfg;
985 };
986
987 // Constructor for the Node_Backward_Iterator
988 Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_Stack &stack, PhaseCFG &cfg)
989 : _visited(visited), _stack(stack), _cfg(cfg) {
990 // The stack should contain exactly the root
991 stack.clear();
992 stack.push(root, root->outcnt());
993
994 // Clear the visited bits
995 visited.clear();
996 }
997
998 // Iterator for the Node_Backward_Iterator
999 Node *Node_Backward_Iterator::next() {
1000
1001 // If the _stack is empty, then just return null: finished.
1002 if ( !_stack.size() )
1003 return nullptr;
1004
1005 // I visit unvisited not-anti-dependence users first, then anti-dependent
1006 // children next. I iterate backwards to support removal of nodes.
1007 // The stack holds states consisting of 3 values:
1008 // current Def node, flag which indicates 1st/2nd pass, index of current out edge
1009 Node *self = (Node*)(((uintptr_t)_stack.node()) & ~1);
1010 bool iterate_anti_dep = (((uintptr_t)_stack.node()) & 1);
1011 uint idx = MIN2(_stack.index(), self->outcnt()); // Support removal of nodes.
1012 _stack.pop();
1013
1014 // I cycle here when I am entering a deeper level of recursion.
1015 // The key variable 'self' was set prior to jumping here.
1016 while( 1 ) {
1017
1018 _visited.set(self->_idx);
1019
1020 // Now schedule all uses as late as possible.
1021 const Node* src = self->is_Proj() ? self->in(0) : self;
1022 uint src_rpo = _cfg.get_block_for_node(src)->_rpo;
1023
1024 // Schedule all nodes in a post-order visit
1025 Node *unvisited = nullptr; // Unvisited anti-dependent Node, if any
1026
1027 // Scan for unvisited nodes
1028 while (idx > 0) {
1029 // For all uses, schedule late
1030 Node* n = self->raw_out(--idx); // Use
1031
1032 // Skip already visited children
1033 if ( _visited.test(n->_idx) )
1034 continue;
1035
1036 // do not traverse backward control edges
1037 Node *use = n->is_Proj() ? n->in(0) : n;
1038 uint use_rpo = _cfg.get_block_for_node(use)->_rpo;
1039
1040 if ( use_rpo < src_rpo )
1041 continue;
1042
1043 // Phi nodes always precede uses in a basic block
1044 if ( use_rpo == src_rpo && use->is_Phi() )
1045 continue;
1046
1047 unvisited = n; // Found unvisited
1048
1049 // Check for possible-anti-dependent
1050 // 1st pass: No such nodes, 2nd pass: Only such nodes.
1051 if (n->needs_anti_dependence_check() == iterate_anti_dep) {
1052 unvisited = n; // Found unvisited
1053 break;
1054 }
1055 }
1056
1057 // Did I find an unvisited not-anti-dependent Node?
1058 if (!unvisited) {
1059 if (!iterate_anti_dep) {
1060 // 2nd pass: Iterate over nodes which needs_anti_dependence_check.
1061 iterate_anti_dep = true;
1062 idx = self->outcnt();
1063 continue;
1064 }
1065 break; // All done with children; post-visit 'self'
1066 }
1067
1068 // Visit the unvisited Node. Contains the obvious push to
1069 // indicate I'm entering a deeper level of recursion. I push the
1070 // old state onto the _stack and set a new state and loop (recurse).
1071 _stack.push((Node*)((uintptr_t)self | (uintptr_t)iterate_anti_dep), idx);
1072 self = unvisited;
1073 iterate_anti_dep = false;
1074 idx = self->outcnt();
1075 } // End recursion loop
1076
1077 return self;
1078 }
1079
1080 //------------------------------ComputeLatenciesBackwards----------------------
1081 // Compute the latency of all the instructions.
1082 void PhaseCFG::compute_latencies_backwards(VectorSet &visited, Node_Stack &stack) {
1083 #ifndef PRODUCT
1084 if (trace_opto_pipelining())
1085 tty->print("\n#---- ComputeLatenciesBackwards ----\n");
1086 #endif
1087
1088 Node_Backward_Iterator iter((Node *)_root, visited, stack, *this);
1089 Node *n;
1090
1091 // Walk over all the nodes from last to first
1092 while ((n = iter.next())) {
1093 // Set the latency for the definitions of this instruction
1094 partial_latency_of_defs(n);
1095 }
1096 } // end ComputeLatenciesBackwards
1097
1098 //------------------------------partial_latency_of_defs------------------------
1099 // Compute the latency impact of this node on all defs. This computes
1100 // a number that increases as we approach the beginning of the routine.
1101 void PhaseCFG::partial_latency_of_defs(Node *n) {
1102 // Set the latency for this instruction
1103 #ifndef PRODUCT
1104 if (trace_opto_pipelining()) {
1105 tty->print("# latency_to_inputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n));
1106 dump();
1107 }
1108 #endif
1109
1110 if (n->is_Proj()) {
1111 n = n->in(0);
1112 }
1113
1114 if (n->is_Root()) {
1115 return;
1116 }
1117
1118 uint nlen = n->len();
1119 uint use_latency = get_latency_for_node(n);
1120 uint use_pre_order = get_block_for_node(n)->_pre_order;
1121
1122 for (uint j = 0; j < nlen; j++) {
1123 Node *def = n->in(j);
1124
1125 if (!def || def == n) {
1126 continue;
1127 }
1128
1129 // Walk backwards thru projections
1130 if (def->is_Proj()) {
1131 def = def->in(0);
1132 }
1133
1134 #ifndef PRODUCT
1135 if (trace_opto_pipelining()) {
1136 tty->print("# in(%2d): ", j);
1137 def->dump();
1138 }
1139 #endif
1140
1141 // If the defining block is not known, assume it is ok
1142 Block *def_block = get_block_for_node(def);
1143 uint def_pre_order = def_block ? def_block->_pre_order : 0;
1144
1145 if ((use_pre_order < def_pre_order) || (use_pre_order == def_pre_order && n->is_Phi())) {
1146 continue;
1147 }
1148
1149 uint delta_latency = n->latency(j);
1150 uint current_latency = delta_latency + use_latency;
1151
1152 if (get_latency_for_node(def) < current_latency) {
1153 set_latency_for_node(def, current_latency);
1154 }
1155
1156 #ifndef PRODUCT
1157 if (trace_opto_pipelining()) {
1158 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));
1159 }
1160 #endif
1161 }
1162 }
1163
1164 //------------------------------latency_from_use-------------------------------
1165 // Compute the latency of a specific use
1166 int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) {
1167 // If self-reference, return no latency
1168 if (use == n || use->is_Root()) {
1169 return 0;
1170 }
1171
1172 uint def_pre_order = get_block_for_node(def)->_pre_order;
1173 uint latency = 0;
1174
1175 // If the use is not a projection, then it is simple...
1176 if (!use->is_Proj()) {
1177 #ifndef PRODUCT
1178 if (trace_opto_pipelining()) {
1179 tty->print("# out(): ");
1180 use->dump();
1181 }
1182 #endif
1183
1184 uint use_pre_order = get_block_for_node(use)->_pre_order;
1185
1186 if (use_pre_order < def_pre_order)
1187 return 0;
1188
1189 if (use_pre_order == def_pre_order && use->is_Phi())
1190 return 0;
1191
1192 uint nlen = use->len();
1193 uint nl = get_latency_for_node(use);
1194
1195 for ( uint j=0; j<nlen; j++ ) {
1196 if (use->in(j) == n) {
1197 // Change this if we want local latencies
1198 uint ul = use->latency(j);
1199 uint l = ul + nl;
1200 if (latency < l) latency = l;
1201 #ifndef PRODUCT
1202 if (trace_opto_pipelining()) {
1203 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, latency = %d",
1204 nl, j, ul, l, latency);
1205 }
1206 #endif
1207 }
1208 }
1209 } else {
1210 // This is a projection, just grab the latency of the use(s)
1211 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
1212 uint l = latency_from_use(use, def, use->fast_out(j));
1213 if (latency < l) latency = l;
1214 }
1215 }
1216
1217 return latency;
1218 }
1219
1220 //------------------------------latency_from_uses------------------------------
1221 // Compute the latency of this instruction relative to all of it's uses.
1222 // This computes a number that increases as we approach the beginning of the
1223 // routine.
1224 void PhaseCFG::latency_from_uses(Node *n) {
1225 // Set the latency for this instruction
1226 #ifndef PRODUCT
1227 if (trace_opto_pipelining()) {
1228 tty->print("# latency_from_outputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n));
1229 dump();
1230 }
1231 #endif
1232 uint latency=0;
1233 const Node *def = n->is_Proj() ? n->in(0): n;
1234
1235 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1236 uint l = latency_from_use(n, def, n->fast_out(i));
1237
1238 if (latency < l) latency = l;
1239 }
1240
1241 set_latency_for_node(n, latency);
1242 }
1243
1244 //------------------------------is_cheaper_block-------------------------
1245 // Check if a block between early and LCA block of uses is cheaper by
1246 // frequency-based policy, latency-based policy and random-based policy
1247 bool PhaseCFG::is_cheaper_block(Block* LCA, Node* self, uint target_latency,
1248 uint end_latency, double least_freq,
1249 int cand_cnt, bool in_latency) {
1250 if (StressGCM) {
1251 // Should be randomly accepted in stress mode
1252 return C->randomized_select(cand_cnt);
1253 }
1254
1255 // Better Frequency
1256 if (LCA->_freq < least_freq) {
1257 return true;
1258 }
1259
1260 // Otherwise, choose with latency
1261 const double delta = 1 + PROB_UNLIKELY_MAG(4);
1262 if (!in_latency && // No block containing latency
1263 LCA->_freq < least_freq * delta && // No worse frequency
1264 target_latency >= end_latency && // within latency range
1265 !self->is_iteratively_computed() // But don't hoist IV increments
1266 // because they may end up above other uses of their phi forcing
1267 // their result register to be different from their input.
1268 ) {
1269 return true;
1270 }
1271
1272 return false;
1273 }
1274
1275 //------------------------------hoist_to_cheaper_block-------------------------
1276 // Pick a block for node self, between early and LCA block of uses, that is a
1277 // cheaper alternative to LCA.
1278 Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) {
1279 Block* least = LCA;
1280 double least_freq = least->_freq;
1281 uint target = get_latency_for_node(self);
1282 uint start_latency = get_latency_for_node(LCA->head());
1283 uint end_latency = get_latency_for_node(LCA->get_node(LCA->end_idx()));
1284 bool in_latency = (target <= start_latency);
1285 const Block* root_block = get_block_for_node(_root);
1286
1287 // Turn off latency scheduling if scheduling is just plain off
1288 if (!C->do_scheduling())
1289 in_latency = true;
1290
1291 // Do not hoist (to cover latency) instructions which target a
1292 // single register. Hoisting stretches the live range of the
1293 // single register and may force spilling.
1294 MachNode* mach = self->is_Mach() ? self->as_Mach() : nullptr;
1295 if (mach && mach->out_RegMask().is_bound1() && mach->out_RegMask().is_NotEmpty())
1296 in_latency = true;
1297
1298 #ifndef PRODUCT
1299 if (trace_opto_pipelining()) {
1300 tty->print("# Find cheaper block for latency %d: ", get_latency_for_node(self));
1301 self->dump();
1302 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1303 LCA->_pre_order,
1304 LCA->head()->_idx,
1305 start_latency,
1306 LCA->get_node(LCA->end_idx())->_idx,
1307 end_latency,
1308 least_freq);
1309 }
1310 #endif
1311
1312 int cand_cnt = 0; // number of candidates tried
1313
1314 // Walk up the dominator tree from LCA (Lowest common ancestor) to
1315 // the earliest legal location. Capture the least execution frequency,
1316 // or choose a random block if -XX:+StressGCM, or using latency-based policy
1317 while (LCA != early) {
1318 LCA = LCA->_idom; // Follow up the dominator tree
1319
1320 if (LCA == nullptr) {
1321 // Bailout without retry
1322 assert(false, "graph should be schedulable");
1323 C->record_method_not_compilable("late schedule failed: LCA is null");
1324 return least;
1325 }
1326
1327 // Don't hoist machine instructions to the root basic block
1328 if (mach && LCA == root_block)
1329 break;
1330
1331 if (self->is_memory_writer() &&
1332 (LCA->_loop->depth() > early->_loop->depth())) {
1333 // LCA is an invalid placement for a memory writer: choosing it would
1334 // cause memory interference, as illustrated in schedule_late().
1335 continue;
1336 }
1337 verify_memory_writer_placement(LCA, self);
1338
1339 uint start_lat = get_latency_for_node(LCA->head());
1340 uint end_idx = LCA->end_idx();
1341 uint end_lat = get_latency_for_node(LCA->get_node(end_idx));
1342 double LCA_freq = LCA->_freq;
1343 #ifndef PRODUCT
1344 if (trace_opto_pipelining()) {
1345 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1346 LCA->_pre_order, LCA->head()->_idx, start_lat, end_idx, end_lat, LCA_freq);
1347 }
1348 #endif
1349 cand_cnt++;
1350 if (is_cheaper_block(LCA, self, target, end_lat, least_freq, cand_cnt, in_latency)) {
1351 least = LCA; // Found cheaper block
1352 least_freq = LCA_freq;
1353 start_latency = start_lat;
1354 end_latency = end_lat;
1355 if (target <= start_lat)
1356 in_latency = true;
1357 }
1358 }
1359
1360 #ifndef PRODUCT
1361 if (trace_opto_pipelining()) {
1362 tty->print_cr("# Choose block B%d with start latency=%d and freq=%g",
1363 least->_pre_order, start_latency, least_freq);
1364 }
1365 #endif
1366
1367 // See if the latency needs to be updated
1368 if (target < end_latency) {
1369 #ifndef PRODUCT
1370 if (trace_opto_pipelining()) {
1371 tty->print_cr("# Change latency for [%4d] from %d to %d", self->_idx, target, end_latency);
1372 }
1373 #endif
1374 set_latency_for_node(self, end_latency);
1375 partial_latency_of_defs(self);
1376 }
1377
1378 return least;
1379 }
1380
1381
1382 //------------------------------schedule_late-----------------------------------
1383 // Now schedule all codes as LATE as possible. This is the LCA in the
1384 // dominator tree of all USES of a value. Pick the block with the least
1385 // loop nesting depth that is lowest in the dominator tree.
1386 extern const char must_clone[];
1387 void PhaseCFG::schedule_late(VectorSet &visited, Node_Stack &stack) {
1388 #ifndef PRODUCT
1389 if (trace_opto_pipelining())
1390 tty->print("\n#---- schedule_late ----\n");
1391 #endif
1392
1393 Node_Backward_Iterator iter((Node *)_root, visited, stack, *this);
1394 Node *self;
1395
1396 // Walk over all the nodes from last to first
1397 while ((self = iter.next())) {
1398 Block* early = get_block_for_node(self); // Earliest legal placement
1399
1400 if (self->is_top()) {
1401 // Top node goes in bb #2 with other constants.
1402 // It must be special-cased, because it has no out edges.
1403 early->add_inst(self);
1404 continue;
1405 }
1406
1407 // No uses, just terminate
1408 if (self->outcnt() == 0) {
1409 assert(self->is_MachProj(), "sanity");
1410 continue; // Must be a dead machine projection
1411 }
1412
1413 // If node is pinned in the block, then no scheduling can be done.
1414 if( self->pinned() ) // Pinned in block?
1415 continue;
1416
1417 #ifdef ASSERT
1418 // Assert that memory writers (e.g. stores) have a "home" block (the block
1419 // given by their control input), and that this block corresponds to their
1420 // earliest possible placement. This guarantees that
1421 // hoist_to_cheaper_block() will always have at least one valid choice.
1422 if (self->is_memory_writer()) {
1423 assert(find_block_for_node(self->in(0)) == early,
1424 "The home of a memory writer must also be its earliest placement");
1425 }
1426 #endif
1427
1428 MachNode* mach = self->is_Mach() ? self->as_Mach() : nullptr;
1429 if (mach) {
1430 switch (mach->ideal_Opcode()) {
1431 case Op_CreateEx:
1432 // Don't move exception creation
1433 early->add_inst(self);
1434 continue;
1435 break;
1436 case Op_CheckCastPP: {
1437 // Don't move CheckCastPP nodes away from their input, if the input
1438 // is a rawptr (5071820).
1439 Node *def = self->in(1);
1440 if (def != nullptr && def->bottom_type()->base() == Type::RawPtr) {
1441 early->add_inst(self);
1442 #ifdef ASSERT
1443 _raw_oops.push(def);
1444 #endif
1445 continue;
1446 }
1447 break;
1448 }
1449 default:
1450 break;
1451 }
1452 if (C->has_irreducible_loop() && self->is_memory_writer()) {
1453 // If the CFG is irreducible, place memory writers in their home block.
1454 // This prevents hoist_to_cheaper_block() from accidentally placing such
1455 // nodes into deeper loops, as in the following example:
1456 //
1457 // Home placement of store in B1 (loop L1):
1458 //
1459 // B1 (L1):
1460 // m1 <- ..
1461 // m2 <- store m1, ..
1462 // B2 (L2):
1463 // jump B2
1464 // B3 (L1):
1465 // .. <- .. m2, ..
1466 //
1467 // Wrong "hoisting" of store to B2 (in loop L2, child of L1):
1468 //
1469 // B1 (L1):
1470 // m1 <- ..
1471 // B2 (L2):
1472 // m2 <- store m1, ..
1473 // # Wrong: m1 and m2 interfere at this point.
1474 // jump B2
1475 // B3 (L1):
1476 // .. <- .. m2, ..
1477 //
1478 // This "hoist inversion" can happen due to different factors such as
1479 // inaccurate estimation of frequencies for irreducible CFGs, and loops
1480 // with always-taken exits in reducible CFGs. In the reducible case,
1481 // hoist inversion is prevented by discarding invalid blocks (those in
1482 // deeper loops than the home block). In the irreducible case, the
1483 // invalid blocks cannot be identified due to incomplete loop nesting
1484 // information, hence a conservative solution is taken.
1485 #ifndef PRODUCT
1486 if (trace_opto_pipelining()) {
1487 tty->print_cr("# Irreducible loops: schedule in home block B%d:",
1488 early->_pre_order);
1489 self->dump();
1490 }
1491 #endif
1492 schedule_node_into_block(self, early);
1493 continue;
1494 }
1495 }
1496
1497 // Gather LCA of all uses
1498 Block *LCA = nullptr;
1499 {
1500 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
1501 // For all uses, find LCA
1502 Node* use = self->fast_out(i);
1503 LCA = raise_LCA_above_use(LCA, use, self, this);
1504 }
1505 guarantee(LCA != nullptr, "There must be a LCA");
1506 } // (Hide defs of imax, i from rest of block.)
1507
1508 // Place temps in the block of their use. This isn't a
1509 // requirement for correctness but it reduces useless
1510 // interference between temps and other nodes.
1511 if (mach != nullptr && mach->is_MachTemp()) {
1512 map_node_to_block(self, LCA);
1513 LCA->add_inst(self);
1514 continue;
1515 }
1516
1517 // Check if 'self' could be anti-dependent on memory
1518 if (self->needs_anti_dependence_check()) {
1519 // Hoist LCA above possible-defs and insert anti-dependences to
1520 // defs in new LCA block.
1521 LCA = insert_anti_dependences(LCA, self);
1522 }
1523
1524 if (early->_dom_depth > LCA->_dom_depth) {
1525 // Somehow the LCA has moved above the earliest legal point.
1526 // (One way this can happen is via memory_early_block.)
1527 if (C->subsume_loads() == true && !C->failing()) {
1528 // Retry with subsume_loads == false
1529 // If this is the first failure, the sentinel string will "stick"
1530 // to the Compile object, and the C2Compiler will see it and retry.
1531 C->record_failure(C2Compiler::retry_no_subsuming_loads());
1532 } else {
1533 // Bailout without retry when (early->_dom_depth > LCA->_dom_depth)
1534 assert(C->failure_is_artificial(), "graph should be schedulable");
1535 C->record_method_not_compilable("late schedule failed: incorrect graph" DEBUG_ONLY(COMMA true));
1536 }
1537 return;
1538 }
1539
1540 if (self->is_memory_writer()) {
1541 // If the LCA of a memory writer is a descendant of its home loop, hoist
1542 // it into a valid placement.
1543 while (LCA->_loop->depth() > early->_loop->depth()) {
1544 LCA = LCA->_idom;
1545 }
1546 assert(LCA != nullptr, "a valid LCA must exist");
1547 verify_memory_writer_placement(LCA, self);
1548 }
1549
1550 // If there is no opportunity to hoist, then we're done.
1551 // In stress mode, try to hoist even the single operations.
1552 bool try_to_hoist = StressGCM || (LCA != early);
1553
1554 // Must clone guys stay next to use; no hoisting allowed.
1555 // Also cannot hoist guys that alter memory or are otherwise not
1556 // allocatable (hoisting can make a value live longer, leading to
1557 // anti and output dependency problems which are normally resolved
1558 // by the register allocator giving everyone a different register).
1559 if (mach != nullptr && must_clone[mach->ideal_Opcode()])
1560 try_to_hoist = false;
1561
1562 Block* late = nullptr;
1563 if (try_to_hoist) {
1564 // Now find the block with the least execution frequency.
1565 // Start at the latest schedule and work up to the earliest schedule
1566 // in the dominator tree. Thus the Node will dominate all its uses.
1567 late = hoist_to_cheaper_block(LCA, early, self);
1568 } else {
1569 // Just use the LCA of the uses.
1570 late = LCA;
1571 }
1572
1573 // Put the node into target block
1574 schedule_node_into_block(self, late);
1575
1576 #ifdef ASSERT
1577 if (self->needs_anti_dependence_check()) {
1578 // since precedence edges are only inserted when we're sure they
1579 // are needed make sure that after placement in a block we don't
1580 // need any new precedence edges.
1581 verify_anti_dependences(late, self);
1582 }
1583 #endif
1584 } // Loop until all nodes have been visited
1585
1586 } // end ScheduleLate
1587
1588 //------------------------------GlobalCodeMotion-------------------------------
1589 void PhaseCFG::global_code_motion() {
1590 ResourceMark rm;
1591
1592 #ifndef PRODUCT
1593 if (trace_opto_pipelining()) {
1594 tty->print("\n---- Start GlobalCodeMotion ----\n");
1595 }
1596 #endif
1597
1598 // Initialize the node to block mapping for things on the proj_list
1599 for (uint i = 0; i < _matcher.number_of_projections(); i++) {
1600 unmap_node_from_block(_matcher.get_projection(i));
1601 }
1602
1603 // Set the basic block for Nodes pinned into blocks
1604 VectorSet visited;
1605 schedule_pinned_nodes(visited);
1606
1607 // Find the earliest Block any instruction can be placed in. Some
1608 // instructions are pinned into Blocks. Unpinned instructions can
1609 // appear in last block in which all their inputs occur.
1610 visited.clear();
1611 Node_Stack stack((C->live_nodes() >> 2) + 16); // pre-grow
1612 if (!schedule_early(visited, stack)) {
1613 // Bailout without retry
1614 assert(false, "early schedule failed");
1615 C->record_method_not_compilable("early schedule failed");
1616 return;
1617 }
1618
1619 // Build Def-Use edges.
1620 // Compute the latency information (via backwards walk) for all the
1621 // instructions in the graph
1622 _node_latency = new GrowableArray<uint>(); // resource_area allocation
1623
1624 if (C->do_scheduling()) {
1625 compute_latencies_backwards(visited, stack);
1626 }
1627
1628 // Now schedule all codes as LATE as possible. This is the LCA in the
1629 // dominator tree of all USES of a value. Pick the block with the least
1630 // loop nesting depth that is lowest in the dominator tree.
1631 // ( visited.clear() called in schedule_late()->Node_Backward_Iterator() )
1632 schedule_late(visited, stack);
1633 if (C->failing()) {
1634 return;
1635 }
1636
1637 #ifndef PRODUCT
1638 if (trace_opto_pipelining()) {
1639 tty->print("\n---- Detect implicit null checks ----\n");
1640 }
1641 #endif
1642
1643 // Detect implicit-null-check opportunities. Basically, find null checks
1644 // with suitable memory ops nearby. Use the memory op to do the null check.
1645 // I can generate a memory op if there is not one nearby.
1646 if (C->is_method_compilation()) {
1647 // By reversing the loop direction we get a very minor gain on mpegaudio.
1648 // Feel free to revert to a forward loop for clarity.
1649 // for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) {
1650 for (int i = _matcher._null_check_tests.size() - 2; i >= 0; i -= 2) {
1651 Node* proj = _matcher._null_check_tests[i];
1652 Node* val = _matcher._null_check_tests[i + 1];
1653 Block* block = get_block_for_node(proj);
1654 implicit_null_check(block, proj, val, C->allowed_deopt_reasons());
1655 // The implicit_null_check will only perform the transformation
1656 // if the null branch is truly uncommon, *and* it leads to an
1657 // uncommon trap. Combined with the too_many_traps guards
1658 // above, this prevents SEGV storms reported in 6366351,
1659 // by recompiling offending methods without this optimization.
1660 }
1661 }
1662
1663 bool block_size_threshold_ok = false;
1664 intptr_t *recalc_pressure_nodes = nullptr;
1665 if (OptoRegScheduling) {
1666 for (uint i = 0; i < number_of_blocks(); i++) {
1667 Block* block = get_block(i);
1668 if (block->number_of_nodes() > 10) {
1669 block_size_threshold_ok = true;
1670 break;
1671 }
1672 }
1673 }
1674
1675 // Enabling the scheduler for register pressure plus finding blocks of size to schedule for it
1676 // is key to enabling this feature.
1677 PhaseChaitin regalloc(C->unique(), *this, _matcher, true);
1678 ResourceArea live_arena(mtCompiler); // Arena for liveness
1679 ResourceMark rm_live(&live_arena);
1680 PhaseLive live(*this, regalloc._lrg_map.names(), &live_arena, true);
1681 PhaseIFG ifg(&live_arena);
1682 if (OptoRegScheduling && block_size_threshold_ok) {
1683 regalloc.mark_ssa();
1684 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]);
1685 rm_live.reset_to_mark(); // Reclaim working storage
1686 IndexSet::reset_memory(C, &live_arena);
1687 uint node_size = regalloc._lrg_map.max_lrg_id();
1688 ifg.init(node_size); // Empty IFG
1689 regalloc.set_ifg(ifg);
1690 regalloc.set_live(live);
1691 regalloc.gather_lrg_masks(false); // Collect LRG masks
1692 live.compute(node_size); // Compute liveness
1693
1694 recalc_pressure_nodes = NEW_RESOURCE_ARRAY(intptr_t, node_size);
1695 for (uint i = 0; i < node_size; i++) {
1696 recalc_pressure_nodes[i] = 0;
1697 }
1698 }
1699 _regalloc = ®alloc;
1700
1701 #ifndef PRODUCT
1702 if (trace_opto_pipelining()) {
1703 tty->print("\n---- Start Local Scheduling ----\n");
1704 }
1705 #endif
1706
1707 // Schedule locally. Right now a simple topological sort.
1708 // Later, do a real latency aware scheduler.
1709 GrowableArray<int> ready_cnt(C->unique(), C->unique(), -1);
1710 visited.reset();
1711 for (uint i = 0; i < number_of_blocks(); i++) {
1712 Block* block = get_block(i);
1713 if (!schedule_local(block, ready_cnt, visited, recalc_pressure_nodes)) {
1714 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
1715 assert(C->failure_is_artificial(), "local schedule failed");
1716 C->record_method_not_compilable("local schedule failed" DEBUG_ONLY(COMMA true));
1717 }
1718 _regalloc = nullptr;
1719 return;
1720 }
1721 }
1722 _regalloc = nullptr;
1723
1724 // If we inserted any instructions between a Call and his CatchNode,
1725 // clone the instructions on all paths below the Catch.
1726 for (uint i = 0; i < number_of_blocks(); i++) {
1727 Block* block = get_block(i);
1728 call_catch_cleanup(block);
1729 }
1730
1731 #ifndef PRODUCT
1732 if (trace_opto_pipelining()) {
1733 tty->print("\n---- After GlobalCodeMotion ----\n");
1734 for (uint i = 0; i < number_of_blocks(); i++) {
1735 Block* block = get_block(i);
1736 block->dump();
1737 }
1738 }
1739 #endif
1740 // Dead.
1741 _node_latency = (GrowableArray<uint> *)((intptr_t)0xdeadbeef);
1742 }
1743
1744 bool PhaseCFG::do_global_code_motion() {
1745
1746 build_dominator_tree();
1747 if (C->failing()) {
1748 return false;
1749 }
1750
1751 NOT_PRODUCT( C->verify_graph_edges(); )
1752
1753 estimate_block_frequency();
1754
1755 global_code_motion();
1756
1757 if (C->failing()) {
1758 return false;
1759 }
1760
1761 return true;
1762 }
1763
1764 //------------------------------Estimate_Block_Frequency-----------------------
1765 // Estimate block frequencies based on IfNode probabilities.
1766 void PhaseCFG::estimate_block_frequency() {
1767
1768 // Force conditional branches leading to uncommon traps to be unlikely,
1769 // not because we get to the uncommon_trap with less relative frequency,
1770 // but because an uncommon_trap typically causes a deopt, so we only get
1771 // there once.
1772 if (C->do_freq_based_layout()) {
1773 Block_List worklist;
1774 Block* root_blk = get_block(0);
1775 for (uint i = 1; i < root_blk->num_preds(); i++) {
1776 Block *pb = get_block_for_node(root_blk->pred(i));
1777 if (pb->has_uncommon_code()) {
1778 worklist.push(pb);
1779 }
1780 }
1781 while (worklist.size() > 0) {
1782 Block* uct = worklist.pop();
1783 if (uct == get_root_block()) {
1784 continue;
1785 }
1786 for (uint i = 1; i < uct->num_preds(); i++) {
1787 Block *pb = get_block_for_node(uct->pred(i));
1788 if (pb->_num_succs == 1) {
1789 worklist.push(pb);
1790 } else if (pb->num_fall_throughs() == 2) {
1791 pb->update_uncommon_branch(uct);
1792 }
1793 }
1794 }
1795 }
1796
1797 // Create the loop tree and calculate loop depth.
1798 _root_loop = create_loop_tree();
1799 _root_loop->compute_loop_depth(0);
1800
1801 // Compute block frequency of each block, relative to a single loop entry.
1802 _root_loop->compute_freq();
1803
1804 // Adjust all frequencies to be relative to a single method entry
1805 _root_loop->_freq = 1.0;
1806 _root_loop->scale_freq();
1807
1808 // Save outmost loop frequency for LRG frequency threshold
1809 _outer_loop_frequency = _root_loop->outer_loop_freq();
1810
1811 // force paths ending at uncommon traps to be infrequent
1812 if (!C->do_freq_based_layout()) {
1813 Block_List worklist;
1814 Block* root_blk = get_block(0);
1815 for (uint i = 1; i < root_blk->num_preds(); i++) {
1816 Block *pb = get_block_for_node(root_blk->pred(i));
1817 if (pb->has_uncommon_code()) {
1818 worklist.push(pb);
1819 }
1820 }
1821 while (worklist.size() > 0) {
1822 Block* uct = worklist.pop();
1823 uct->_freq = PROB_MIN;
1824 for (uint i = 1; i < uct->num_preds(); i++) {
1825 Block *pb = get_block_for_node(uct->pred(i));
1826 if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) {
1827 worklist.push(pb);
1828 }
1829 }
1830 }
1831 }
1832
1833 #ifdef ASSERT
1834 for (uint i = 0; i < number_of_blocks(); i++) {
1835 Block* b = get_block(i);
1836 assert(b->_freq >= MIN_BLOCK_FREQUENCY, "Register Allocator requires meaningful block frequency");
1837 }
1838 #endif
1839
1840 #ifndef PRODUCT
1841 if (PrintCFGBlockFreq) {
1842 tty->print_cr("CFG Block Frequencies");
1843 _root_loop->dump_tree();
1844 if (Verbose) {
1845 tty->print_cr("PhaseCFG dump");
1846 dump();
1847 tty->print_cr("Node dump");
1848 _root->dump(99999);
1849 }
1850 }
1851 #endif
1852 }
1853
1854 //----------------------------create_loop_tree--------------------------------
1855 // Create a loop tree from the CFG
1856 CFGLoop* PhaseCFG::create_loop_tree() {
1857
1858 #ifdef ASSERT
1859 assert(get_block(0) == get_root_block(), "first block should be root block");
1860 for (uint i = 0; i < number_of_blocks(); i++) {
1861 Block* block = get_block(i);
1862 // Check that _loop field are clear...we could clear them if not.
1863 assert(block->_loop == nullptr, "clear _loop expected");
1864 // Sanity check that the RPO numbering is reflected in the _blocks array.
1865 // It doesn't have to be for the loop tree to be built, but if it is not,
1866 // then the blocks have been reordered since dom graph building...which
1867 // may question the RPO numbering
1868 assert(block->_rpo == i, "unexpected reverse post order number");
1869 }
1870 #endif
1871
1872 int idct = 0;
1873 CFGLoop* root_loop = new CFGLoop(idct++);
1874
1875 Block_List worklist;
1876
1877 // Assign blocks to loops
1878 for(uint i = number_of_blocks() - 1; i > 0; i-- ) { // skip Root block
1879 Block* block = get_block(i);
1880
1881 if (block->head()->is_Loop()) {
1882 Block* loop_head = block;
1883 assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1884 Node* tail_n = loop_head->pred(LoopNode::LoopBackControl);
1885 Block* tail = get_block_for_node(tail_n);
1886
1887 // Defensively filter out Loop nodes for non-single-entry loops.
1888 // For all reasonable loops, the head occurs before the tail in RPO.
1889 if (i <= tail->_rpo) {
1890
1891 // The tail and (recursive) predecessors of the tail
1892 // are made members of a new loop.
1893
1894 assert(worklist.size() == 0, "nonempty worklist");
1895 CFGLoop* nloop = new CFGLoop(idct++);
1896 assert(loop_head->_loop == nullptr, "just checking");
1897 loop_head->_loop = nloop;
1898 // Add to nloop so push_pred() will skip over inner loops
1899 nloop->add_member(loop_head);
1900 nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, this);
1901
1902 while (worklist.size() > 0) {
1903 Block* member = worklist.pop();
1904 if (member != loop_head) {
1905 for (uint j = 1; j < member->num_preds(); j++) {
1906 nloop->push_pred(member, j, worklist, this);
1907 }
1908 }
1909 }
1910 }
1911 }
1912 }
1913
1914 // Create a member list for each loop consisting
1915 // of both blocks and (immediate child) loops.
1916 for (uint i = 0; i < number_of_blocks(); i++) {
1917 Block* block = get_block(i);
1918 CFGLoop* lp = block->_loop;
1919 if (lp == nullptr) {
1920 // Not assigned to a loop. Add it to the method's pseudo loop.
1921 block->_loop = root_loop;
1922 lp = root_loop;
1923 }
1924 if (lp == root_loop || block != lp->head()) { // loop heads are already members
1925 lp->add_member(block);
1926 }
1927 if (lp != root_loop) {
1928 if (lp->parent() == nullptr) {
1929 // Not a nested loop. Make it a child of the method's pseudo loop.
1930 root_loop->add_nested_loop(lp);
1931 }
1932 if (block == lp->head()) {
1933 // Add nested loop to member list of parent loop.
1934 lp->parent()->add_member(lp);
1935 }
1936 }
1937 }
1938
1939 return root_loop;
1940 }
1941
1942 //------------------------------push_pred--------------------------------------
1943 void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg) {
1944 Node* pred_n = blk->pred(i);
1945 Block* pred = cfg->get_block_for_node(pred_n);
1946 CFGLoop *pred_loop = pred->_loop;
1947 if (pred_loop == nullptr) {
1948 // Filter out blocks for non-single-entry loops.
1949 // For all reasonable loops, the head occurs before the tail in RPO.
1950 if (pred->_rpo > head()->_rpo) {
1951 pred->_loop = this;
1952 worklist.push(pred);
1953 }
1954 } else if (pred_loop != this) {
1955 // Nested loop.
1956 while (pred_loop->_parent != nullptr && pred_loop->_parent != this) {
1957 pred_loop = pred_loop->_parent;
1958 }
1959 // Make pred's loop be a child
1960 if (pred_loop->_parent == nullptr) {
1961 add_nested_loop(pred_loop);
1962 // Continue with loop entry predecessor.
1963 Block* pred_head = pred_loop->head();
1964 assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1965 assert(pred_head != head(), "loop head in only one loop");
1966 push_pred(pred_head, LoopNode::EntryControl, worklist, cfg);
1967 } else {
1968 assert(pred_loop->_parent == this && _parent == nullptr, "just checking");
1969 }
1970 }
1971 }
1972
1973 //------------------------------add_nested_loop--------------------------------
1974 // Make cl a child of the current loop in the loop tree.
1975 void CFGLoop::add_nested_loop(CFGLoop* cl) {
1976 assert(_parent == nullptr, "no parent yet");
1977 assert(cl != this, "not my own parent");
1978 cl->_parent = this;
1979 CFGLoop* ch = _child;
1980 if (ch == nullptr) {
1981 _child = cl;
1982 } else {
1983 while (ch->_sibling != nullptr) { ch = ch->_sibling; }
1984 ch->_sibling = cl;
1985 }
1986 }
1987
1988 //------------------------------compute_loop_depth-----------------------------
1989 // Store the loop depth in each CFGLoop object.
1990 // Recursively walk the children to do the same for them.
1991 void CFGLoop::compute_loop_depth(int depth) {
1992 _depth = depth;
1993 CFGLoop* ch = _child;
1994 while (ch != nullptr) {
1995 ch->compute_loop_depth(depth + 1);
1996 ch = ch->_sibling;
1997 }
1998 }
1999
2000 //------------------------------compute_freq-----------------------------------
2001 // Compute the frequency of each block and loop, relative to a single entry
2002 // into the dominating loop head.
2003 void CFGLoop::compute_freq() {
2004 // Bottom up traversal of loop tree (visit inner loops first.)
2005 // Set loop head frequency to 1.0, then transitively
2006 // compute frequency for all successors in the loop,
2007 // as well as for each exit edge. Inner loops are
2008 // treated as single blocks with loop exit targets
2009 // as the successor blocks.
2010
2011 // Nested loops first
2012 CFGLoop* ch = _child;
2013 while (ch != nullptr) {
2014 ch->compute_freq();
2015 ch = ch->_sibling;
2016 }
2017 assert (_members.length() > 0, "no empty loops");
2018 Block* hd = head();
2019 hd->_freq = 1.0;
2020 for (int i = 0; i < _members.length(); i++) {
2021 CFGElement* s = _members.at(i);
2022 double freq = s->_freq;
2023 if (s->is_block()) {
2024 Block* b = s->as_Block();
2025 for (uint j = 0; j < b->_num_succs; j++) {
2026 Block* sb = b->_succs[j];
2027 update_succ_freq(sb, freq * b->succ_prob(j));
2028 }
2029 } else {
2030 CFGLoop* lp = s->as_CFGLoop();
2031 assert(lp->_parent == this, "immediate child");
2032 for (int k = 0; k < lp->_exits.length(); k++) {
2033 Block* eb = lp->_exits.at(k).get_target();
2034 double prob = lp->_exits.at(k).get_prob();
2035 update_succ_freq(eb, freq * prob);
2036 }
2037 }
2038 }
2039
2040 // For all loops other than the outer, "method" loop,
2041 // sum and normalize the exit probability. The "method" loop
2042 // should keep the initial exit probability of 1, so that
2043 // inner blocks do not get erroneously scaled.
2044 if (_depth != 0) {
2045 // Total the exit probabilities for this loop.
2046 double exits_sum = 0.0f;
2047 for (int i = 0; i < _exits.length(); i++) {
2048 exits_sum += _exits.at(i).get_prob();
2049 }
2050
2051 // Normalize the exit probabilities. Until now, the
2052 // probabilities estimate the possibility of exit per
2053 // a single loop iteration; afterward, they estimate
2054 // the probability of exit per loop entry.
2055 for (int i = 0; i < _exits.length(); i++) {
2056 Block* et = _exits.at(i).get_target();
2057 float new_prob = 0.0f;
2058 if (_exits.at(i).get_prob() > 0.0f) {
2059 new_prob = _exits.at(i).get_prob() / exits_sum;
2060 }
2061 BlockProbPair bpp(et, new_prob);
2062 _exits.at_put(i, bpp);
2063 }
2064
2065 // Save the total, but guard against unreasonable probability,
2066 // as the value is used to estimate the loop trip count.
2067 // An infinite trip count would blur relative block
2068 // frequencies.
2069 if (exits_sum > 1.0f) exits_sum = 1.0;
2070 if (exits_sum < PROB_MIN) exits_sum = PROB_MIN;
2071 _exit_prob = exits_sum;
2072 }
2073 }
2074
2075 //------------------------------succ_prob-------------------------------------
2076 // Determine the probability of reaching successor 'i' from the receiver block.
2077 float Block::succ_prob(uint i) {
2078 int eidx = end_idx();
2079 Node *n = get_node(eidx); // Get ending Node
2080
2081 int op = n->Opcode();
2082 if (n->is_Mach()) {
2083 if (n->is_MachNullCheck()) {
2084 // Can only reach here if called after lcm. The original Op_If is gone,
2085 // so we attempt to infer the probability from one or both of the
2086 // successor blocks.
2087 assert(_num_succs == 2, "expecting 2 successors of a null check");
2088 // If either successor has only one predecessor, then the
2089 // probability estimate can be derived using the
2090 // relative frequency of the successor and this block.
2091 if (_succs[i]->num_preds() == 2) {
2092 return _succs[i]->_freq / _freq;
2093 } else if (_succs[1-i]->num_preds() == 2) {
2094 return 1 - (_succs[1-i]->_freq / _freq);
2095 } else {
2096 // Estimate using both successor frequencies
2097 float freq = _succs[i]->_freq;
2098 return freq / (freq + _succs[1-i]->_freq);
2099 }
2100 }
2101 op = n->as_Mach()->ideal_Opcode();
2102 }
2103
2104
2105 // Switch on branch type
2106 switch( op ) {
2107 case Op_CountedLoopEnd:
2108 case Op_If: {
2109 assert (i < 2, "just checking");
2110 // Conditionals pass on only part of their frequency
2111 float prob = n->as_MachIf()->_prob;
2112 assert(prob >= 0.0 && prob <= 1.0, "out of range probability");
2113 // If succ[i] is the FALSE branch, invert path info
2114 if( get_node(i + eidx + 1)->Opcode() == Op_IfFalse ) {
2115 return 1.0f - prob; // not taken
2116 } else {
2117 return prob; // taken
2118 }
2119 }
2120
2121 case Op_Jump:
2122 return n->as_MachJump()->_probs[get_node(i + eidx + 1)->as_JumpProj()->_con];
2123
2124 case Op_Catch: {
2125 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
2126 if (ci->_con == CatchProjNode::fall_through_index) {
2127 // Fall-thru path gets the lion's share.
2128 return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs;
2129 } else {
2130 // Presume exceptional paths are equally unlikely
2131 return PROB_UNLIKELY_MAG(5);
2132 }
2133 }
2134
2135 case Op_Root:
2136 case Op_Goto:
2137 // Pass frequency straight thru to target
2138 return 1.0f;
2139
2140 case Op_NeverBranch:
2141 return 0.0f;
2142
2143 case Op_TailCall:
2144 case Op_TailJump:
2145 case Op_ForwardException:
2146 case Op_Return:
2147 case Op_Halt:
2148 case Op_Rethrow:
2149 // Do not push out freq to root block
2150 return 0.0f;
2151
2152 default:
2153 ShouldNotReachHere();
2154 }
2155
2156 return 0.0f;
2157 }
2158
2159 //------------------------------num_fall_throughs-----------------------------
2160 // Return the number of fall-through candidates for a block
2161 int Block::num_fall_throughs() {
2162 int eidx = end_idx();
2163 Node *n = get_node(eidx); // Get ending Node
2164
2165 int op = n->Opcode();
2166 if (n->is_Mach()) {
2167 if (n->is_MachNullCheck()) {
2168 // In theory, either side can fall-thru, for simplicity sake,
2169 // let's say only the false branch can now.
2170 return 1;
2171 }
2172 op = n->as_Mach()->ideal_Opcode();
2173 }
2174
2175 // Switch on branch type
2176 switch( op ) {
2177 case Op_CountedLoopEnd:
2178 case Op_If:
2179 return 2;
2180
2181 case Op_Root:
2182 case Op_Goto:
2183 return 1;
2184
2185 case Op_Catch: {
2186 for (uint i = 0; i < _num_succs; i++) {
2187 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
2188 if (ci->_con == CatchProjNode::fall_through_index) {
2189 return 1;
2190 }
2191 }
2192 return 0;
2193 }
2194
2195 case Op_Jump:
2196 case Op_NeverBranch:
2197 case Op_TailCall:
2198 case Op_TailJump:
2199 case Op_ForwardException:
2200 case Op_Return:
2201 case Op_Halt:
2202 case Op_Rethrow:
2203 return 0;
2204
2205 default:
2206 ShouldNotReachHere();
2207 }
2208
2209 return 0;
2210 }
2211
2212 //------------------------------succ_fall_through-----------------------------
2213 // Return true if a specific successor could be fall-through target.
2214 bool Block::succ_fall_through(uint i) {
2215 int eidx = end_idx();
2216 Node *n = get_node(eidx); // Get ending Node
2217
2218 int op = n->Opcode();
2219 if (n->is_Mach()) {
2220 if (n->is_MachNullCheck()) {
2221 // In theory, either side can fall-thru, for simplicity sake,
2222 // let's say only the false branch can now.
2223 return get_node(i + eidx + 1)->Opcode() == Op_IfFalse;
2224 }
2225 op = n->as_Mach()->ideal_Opcode();
2226 }
2227
2228 // Switch on branch type
2229 switch( op ) {
2230 case Op_CountedLoopEnd:
2231 case Op_If:
2232 case Op_Root:
2233 case Op_Goto:
2234 return true;
2235
2236 case Op_Catch: {
2237 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
2238 return ci->_con == CatchProjNode::fall_through_index;
2239 }
2240
2241 case Op_Jump:
2242 case Op_NeverBranch:
2243 case Op_TailCall:
2244 case Op_TailJump:
2245 case Op_ForwardException:
2246 case Op_Return:
2247 case Op_Halt:
2248 case Op_Rethrow:
2249 return false;
2250
2251 default:
2252 ShouldNotReachHere();
2253 }
2254
2255 return false;
2256 }
2257
2258 //------------------------------update_uncommon_branch------------------------
2259 // Update the probability of a two-branch to be uncommon
2260 void Block::update_uncommon_branch(Block* ub) {
2261 int eidx = end_idx();
2262 Node *n = get_node(eidx); // Get ending Node
2263
2264 int op = n->as_Mach()->ideal_Opcode();
2265
2266 assert(op == Op_CountedLoopEnd || op == Op_If, "must be a If");
2267 assert(num_fall_throughs() == 2, "must be a two way branch block");
2268
2269 // Which successor is ub?
2270 uint s;
2271 for (s = 0; s <_num_succs; s++) {
2272 if (_succs[s] == ub) break;
2273 }
2274 assert(s < 2, "uncommon successor must be found");
2275
2276 // If ub is the true path, make the proability small, else
2277 // ub is the false path, and make the probability large
2278 bool invert = (get_node(s + eidx + 1)->Opcode() == Op_IfFalse);
2279
2280 // Get existing probability
2281 float p = n->as_MachIf()->_prob;
2282
2283 if (invert) p = 1.0 - p;
2284 if (p > PROB_MIN) {
2285 p = PROB_MIN;
2286 }
2287 if (invert) p = 1.0 - p;
2288
2289 n->as_MachIf()->_prob = p;
2290 }
2291
2292 //------------------------------update_succ_freq-------------------------------
2293 // Update the appropriate frequency associated with block 'b', a successor of
2294 // a block in this loop.
2295 void CFGLoop::update_succ_freq(Block* b, double freq) {
2296 if (b->_loop == this) {
2297 if (b == head()) {
2298 // back branch within the loop
2299 // Do nothing now, the loop carried frequency will be
2300 // adjust later in scale_freq().
2301 } else {
2302 // simple branch within the loop
2303 b->_freq += freq;
2304 }
2305 } else if (!in_loop_nest(b)) {
2306 // branch is exit from this loop
2307 BlockProbPair bpp(b, freq);
2308 _exits.append(bpp);
2309 } else {
2310 // branch into nested loop
2311 CFGLoop* ch = b->_loop;
2312 ch->_freq += freq;
2313 }
2314 }
2315
2316 //------------------------------in_loop_nest-----------------------------------
2317 // Determine if block b is in the receiver's loop nest.
2318 bool CFGLoop::in_loop_nest(Block* b) {
2319 int depth = _depth;
2320 CFGLoop* b_loop = b->_loop;
2321 int b_depth = b_loop->_depth;
2322 if (depth == b_depth) {
2323 return true;
2324 }
2325 while (b_depth > depth) {
2326 b_loop = b_loop->_parent;
2327 b_depth = b_loop->_depth;
2328 }
2329 return b_loop == this;
2330 }
2331
2332 //------------------------------scale_freq-------------------------------------
2333 // Scale frequency of loops and blocks by trip counts from outer loops
2334 // Do a top down traversal of loop tree (visit outer loops first.)
2335 void CFGLoop::scale_freq() {
2336 double loop_freq = _freq * trip_count();
2337 _freq = loop_freq;
2338 for (int i = 0; i < _members.length(); i++) {
2339 CFGElement* s = _members.at(i);
2340 double block_freq = s->_freq * loop_freq;
2341 if (g_isnan(block_freq) || block_freq < MIN_BLOCK_FREQUENCY)
2342 block_freq = MIN_BLOCK_FREQUENCY;
2343 s->_freq = block_freq;
2344 }
2345 CFGLoop* ch = _child;
2346 while (ch != nullptr) {
2347 ch->scale_freq();
2348 ch = ch->_sibling;
2349 }
2350 }
2351
2352 // Frequency of outer loop
2353 double CFGLoop::outer_loop_freq() const {
2354 if (_child != nullptr) {
2355 return _child->_freq;
2356 }
2357 return _freq;
2358 }
2359
2360 #ifndef PRODUCT
2361 //------------------------------dump_tree--------------------------------------
2362 void CFGLoop::dump_tree() const {
2363 dump();
2364 if (_child != nullptr) _child->dump_tree();
2365 if (_sibling != nullptr) _sibling->dump_tree();
2366 }
2367
2368 //------------------------------dump-------------------------------------------
2369 void CFGLoop::dump() const {
2370 for (int i = 0; i < _depth; i++) tty->print(" ");
2371 tty->print("%s: %d trip_count: %6.0f freq: %6.0f\n",
2372 _depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq);
2373 for (int i = 0; i < _depth; i++) tty->print(" ");
2374 tty->print(" members:");
2375 int k = 0;
2376 for (int i = 0; i < _members.length(); i++) {
2377 if (k++ >= 6) {
2378 tty->print("\n ");
2379 for (int j = 0; j < _depth+1; j++) tty->print(" ");
2380 k = 0;
2381 }
2382 CFGElement *s = _members.at(i);
2383 if (s->is_block()) {
2384 Block *b = s->as_Block();
2385 tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq);
2386 } else {
2387 CFGLoop* lp = s->as_CFGLoop();
2388 tty->print(" L%d(%6.3f)", lp->_id, lp->_freq);
2389 }
2390 }
2391 tty->print("\n");
2392 for (int i = 0; i < _depth; i++) tty->print(" ");
2393 tty->print(" exits: ");
2394 k = 0;
2395 for (int i = 0; i < _exits.length(); i++) {
2396 if (k++ >= 7) {
2397 tty->print("\n ");
2398 for (int j = 0; j < _depth+1; j++) tty->print(" ");
2399 k = 0;
2400 }
2401 Block *blk = _exits.at(i).get_target();
2402 double prob = _exits.at(i).get_prob();
2403 tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100));
2404 }
2405 tty->print("\n");
2406 }
2407 #endif