1 /*
2 * Copyright (c) 1999, 2023, 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 "memory/allocation.inline.hpp"
27 #include "opto/addnode.hpp"
28 #include "opto/callnode.hpp"
29 #include "opto/inlinetypenode.hpp"
30 #include "opto/loopnode.hpp"
31 #include "opto/movenode.hpp"
32 #include "opto/node.hpp"
33 #include "opto/opaquenode.hpp"
34 #include "opto/predicates.hpp"
35
36 //------------------------------split_thru_region------------------------------
37 // Split Node 'n' through merge point.
38 RegionNode* PhaseIdealLoop::split_thru_region(Node* n, RegionNode* region) {
39 assert(n->is_CFG(), "");
40 RegionNode* r = new RegionNode(region->req());
41 IdealLoopTree* loop = get_loop(n);
42 for (uint i = 1; i < region->req(); i++) {
43 Node* x = n->clone();
44 Node* in0 = n->in(0);
45 if (in0->in(0) == region) x->set_req(0, in0->in(i));
46 for (uint j = 1; j < n->req(); j++) {
47 Node* in = n->in(j);
48 if (get_ctrl(in) == region) {
49 x->set_req(j, in->in(i));
50 }
51 }
52 _igvn.register_new_node_with_optimizer(x);
53 set_loop(x, loop);
54 set_idom(x, x->in(0), dom_depth(x->in(0))+1);
55 r->init_req(i, x);
56 }
57
58 // Record region
59 r->set_req(0,region); // Not a TRUE RegionNode
60 _igvn.register_new_node_with_optimizer(r);
61 set_loop(r, loop);
62 if (!loop->_child) {
63 loop->_body.push(r);
64 }
65 return r;
66 }
67
68 //------------------------------split_up---------------------------------------
69 // Split block-local op up through the phis to empty the current block
70 bool PhaseIdealLoop::split_up( Node *n, Node *blk1, Node *blk2 ) {
71 if( n->is_CFG() ) {
72 assert( n->in(0) != blk1, "Lousy candidate for split-if" );
73 return false;
74 }
75 if (!at_relevant_ctrl(n, blk1, blk2))
76 return false; // Not block local
77 if( n->is_Phi() ) return false; // Local PHIs are expected
78
79 // Recursively split-up inputs
80 for (uint i = 1; i < n->req(); i++) {
81 if( split_up( n->in(i), blk1, blk2 ) ) {
82 // Got split recursively and self went dead?
83 if (n->outcnt() == 0)
84 _igvn.remove_dead_node(n);
85 return true;
86 }
87 }
88
89 if (clone_cmp_loadklass_down(n, blk1, blk2)) {
90 return true;
91 }
92
93 // Check for needing to clone-up a compare. Can't do that, it forces
94 // another (nested) split-if transform. Instead, clone it "down".
95 if (clone_cmp_down(n, blk1, blk2)) {
96 return true;
97 }
98
99 clone_template_assertion_expression_down(n);
100
101 if (n->Opcode() == Op_OpaqueZeroTripGuard) {
102 // If this Opaque1 is part of the zero trip guard for a loop:
103 // 1- it can't be shared
104 // 2- the zero trip guard can't be the if that's being split
105 // As a consequence, this node could be assigned control anywhere between its current control and the zero trip guard.
106 // Move it down to get it out of the way of split if and avoid breaking the zero trip guard shape.
107 Node* cmp = n->unique_out();
108 assert(cmp->Opcode() == Op_CmpI, "bad zero trip guard shape");
109 Node* bol = cmp->unique_out();
110 assert(bol->Opcode() == Op_Bool, "bad zero trip guard shape");
111 Node* iff = bol->unique_out();
112 assert(iff->Opcode() == Op_If, "bad zero trip guard shape");
113 set_ctrl(n, iff->in(0));
114 set_ctrl(cmp, iff->in(0));
115 set_ctrl(bol, iff->in(0));
116 return true;
117 }
118
119 // See if splitting-up a Store. Any anti-dep loads must go up as
120 // well. An anti-dep load might be in the wrong block, because in
121 // this particular layout/schedule we ignored anti-deps and allow
122 // memory to be alive twice. This only works if we do the same
123 // operations on anti-dep loads as we do their killing stores.
124 if( n->is_Store() && n->in(MemNode::Memory)->in(0) == n->in(0) ) {
125 // Get store's memory slice
126 int alias_idx = C->get_alias_index(_igvn.type(n->in(MemNode::Address))->is_ptr());
127
128 // Get memory-phi anti-dep loads will be using
129 Node *memphi = n->in(MemNode::Memory);
130 assert( memphi->is_Phi(), "" );
131 // Hoist any anti-dep load to the splitting block;
132 // it will then "split-up".
133 for (DUIterator_Fast imax,i = memphi->fast_outs(imax); i < imax; i++) {
134 Node *load = memphi->fast_out(i);
135 if( load->is_Load() && alias_idx == C->get_alias_index(_igvn.type(load->in(MemNode::Address))->is_ptr()) )
136 set_ctrl(load,blk1);
137 }
138 }
139
140 // Found some other Node; must clone it up
141 #ifndef PRODUCT
142 if( PrintOpto && VerifyLoopOptimizations ) {
143 tty->print("Cloning up: ");
144 n->dump();
145 }
146 #endif
147
148 // ConvI2L may have type information on it which becomes invalid if
149 // it moves up in the graph so change any clones so widen the type
150 // to TypeLong::INT when pushing it up.
151 const Type* rtype = nullptr;
152 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::INT) {
153 rtype = TypeLong::INT;
154 }
155
156 // Now actually split-up this guy. One copy per control path merging.
157 Node *phi = PhiNode::make_blank(blk1, n);
158 for( uint j = 1; j < blk1->req(); j++ ) {
159 Node *x = n->clone();
160 // Widen the type of the ConvI2L when pushing up.
161 if (rtype != nullptr) x->as_Type()->set_type(rtype);
162 if( n->in(0) && n->in(0) == blk1 )
163 x->set_req( 0, blk1->in(j) );
164 for( uint i = 1; i < n->req(); i++ ) {
165 Node *m = n->in(i);
166 if( get_ctrl(m) == blk1 ) {
167 assert( m->in(0) == blk1, "" );
168 x->set_req( i, m->in(j) );
169 }
170 }
171 register_new_node( x, blk1->in(j) );
172 phi->init_req( j, x );
173 }
174 // Announce phi to optimizer
175 register_new_node(phi, blk1);
176
177 // Remove cloned-up value from optimizer; use phi instead
178 _igvn.replace_node( n, phi );
179
180 // (There used to be a self-recursive call to split_up() here,
181 // but it is not needed. All necessary forward walking is done
182 // by do_split_if() below.)
183
184 return true;
185 }
186
187 // Look for a (If .. (Bool(CmpP (LoadKlass .. (AddP obj ..)) ..))) and clone all of it down.
188 // There's likely a CheckCastPP on one of the branches of the If, with obj as input.
189 // If the (LoadKlass .. (AddP obj ..)) is not cloned down, then split if transforms this to: (If .. (Bool(CmpP phi1 ..)))
190 // and the CheckCastPP to (CheckCastPP phi2). It's possible then that phi2 is transformed to a CheckCastPP
191 // (through PhiNode::Ideal) and that that CheckCastPP is replaced by another narrower CheckCastPP at the same control
192 // (through ConstraintCastNode::Identity). That could cause the CheckCastPP at the If to become top while (CmpP phi1)
193 // wouldn't constant fold because it's using a different data path. Cloning the whole subgraph down guarantees both the
194 // AddP and CheckCastPP have the same obj input after split if.
195 bool PhaseIdealLoop::clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2) {
196 if (n->Opcode() == Op_AddP && at_relevant_ctrl(n, blk1, blk2)) {
197 Node_List cmp_nodes;
198 uint old = C->unique();
199 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
200 Node* u1 = n->fast_out(i);
201 if (u1->Opcode() == Op_LoadNKlass && at_relevant_ctrl(u1, blk1, blk2)) {
202 for (DUIterator_Fast jmax, j = u1->fast_outs(jmax); j < jmax; j++) {
203 Node* u2 = u1->fast_out(j);
204 if (u2->Opcode() == Op_DecodeNKlass && at_relevant_ctrl(u2, blk1, blk2)) {
205 for (DUIterator k = u2->outs(); u2->has_out(k); k++) {
206 Node* u3 = u2->out(k);
207 if (at_relevant_ctrl(u3, blk1, blk2) && clone_cmp_down(u3, blk1, blk2)) {
208 --k;
209 }
210 }
211 for (DUIterator_Fast kmax, k = u2->fast_outs(kmax); k < kmax; k++) {
212 Node* u3 = u2->fast_out(k);
213 if (u3->_idx >= old) {
214 cmp_nodes.push(u3);
215 }
216 }
217 }
218 }
219 } else if (u1->Opcode() == Op_LoadKlass && at_relevant_ctrl(u1, blk1, blk2)) {
220 for (DUIterator j = u1->outs(); u1->has_out(j); j++) {
221 Node* u2 = u1->out(j);
222 if (at_relevant_ctrl(u2, blk1, blk2) && clone_cmp_down(u2, blk1, blk2)) {
223 --j;
224 }
225 }
226 for (DUIterator_Fast kmax, k = u1->fast_outs(kmax); k < kmax; k++) {
227 Node* u2 = u1->fast_out(k);
228 if (u2->_idx >= old) {
229 cmp_nodes.push(u2);
230 }
231 }
232 }
233 }
234
235 for (uint i = 0; i < cmp_nodes.size(); ++i) {
236 Node* cmp = cmp_nodes.at(i);
237 clone_loadklass_nodes_at_cmp_index(n, cmp, 1);
238 clone_loadklass_nodes_at_cmp_index(n, cmp, 2);
239 }
240 if (n->outcnt() == 0) {
241 assert(n->is_dead(), "");
242 return true;
243 }
244 }
245 return false;
246 }
247
248 bool PhaseIdealLoop::at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2) {
249 return ctrl_or_self(n) == blk1 || ctrl_or_self(n) == blk2;
250 }
251
252 void PhaseIdealLoop::clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i) {
253 Node* decode = cmp->in(i);
254 if (decode->Opcode() == Op_DecodeNKlass) {
255 Node* loadklass = decode->in(1);
256 if (loadklass->Opcode() == Op_LoadNKlass) {
257 Node* addp = loadklass->in(MemNode::Address);
258 if (addp == n) {
259 Node* ctrl = get_ctrl(cmp);
260 Node* decode_clone = decode->clone();
261 Node* loadklass_clone = loadklass->clone();
262 Node* addp_clone = addp->clone();
263 register_new_node(decode_clone, ctrl);
264 register_new_node(loadklass_clone, ctrl);
265 register_new_node(addp_clone, ctrl);
266 _igvn.replace_input_of(cmp, i, decode_clone);
267 _igvn.replace_input_of(decode_clone, 1, loadklass_clone);
268 _igvn.replace_input_of(loadklass_clone, MemNode::Address, addp_clone);
269 if (decode->outcnt() == 0) {
270 _igvn.remove_dead_node(decode);
271 }
272 }
273 }
274 } else {
275 Node* loadklass = cmp->in(i);
276 if (loadklass->Opcode() == Op_LoadKlass) {
277 Node* addp = loadklass->in(MemNode::Address);
278 if (addp == n) {
279 Node* ctrl = get_ctrl(cmp);
280 Node* loadklass_clone = loadklass->clone();
281 Node* addp_clone = addp->clone();
282 register_new_node(loadklass_clone, ctrl);
283 register_new_node(addp_clone, ctrl);
284 _igvn.replace_input_of(cmp, i, loadklass_clone);
285 _igvn.replace_input_of(loadklass_clone, MemNode::Address, addp_clone);
286 if (loadklass->outcnt() == 0) {
287 _igvn.remove_dead_node(loadklass);
288 }
289 }
290 }
291 }
292 }
293
294 bool PhaseIdealLoop::clone_cmp_down(Node* n, const Node* blk1, const Node* blk2) {
295 if( n->is_Cmp() ) {
296 assert(get_ctrl(n) == blk2 || get_ctrl(n) == blk1, "must be in block with IF");
297 // Check for simple Cmp/Bool/CMove which we can clone-up. Cmp/Bool/CMove
298 // sequence can have no other users and it must all reside in the split-if
299 // block. Non-simple Cmp/Bool/CMove sequences are 'cloned-down' below -
300 // private, per-use versions of the Cmp and Bool are made. These sink to
301 // the CMove block. If the CMove is in the split-if block, then in the
302 // next iteration this will become a simple Cmp/Bool/CMove set to clone-up.
303 Node *bol, *cmov;
304 if (!(n->outcnt() == 1 && n->unique_out()->is_Bool() &&
305 (bol = n->unique_out()->as_Bool()) &&
306 (at_relevant_ctrl(bol, blk1, blk2) &&
307 bol->outcnt() == 1 &&
308 bol->unique_out()->is_CMove() &&
309 (cmov = bol->unique_out()->as_CMove()) &&
310 at_relevant_ctrl(cmov, blk1, blk2)))) {
311
312 // Must clone down
313 #ifndef PRODUCT
314 if( PrintOpto && VerifyLoopOptimizations ) {
315 tty->print("Cloning down: ");
316 n->dump();
317 }
318 #endif
319 if (!n->is_FastLock()) {
320 // Clone down any block-local BoolNode uses of this CmpNode
321 for (DUIterator i = n->outs(); n->has_out(i); i++) {
322 Node* bol = n->out(i);
323 assert( bol->is_Bool(), "" );
324 if (bol->outcnt() == 1) {
325 Node* use = bol->unique_out();
326 if (use->is_OpaqueNotNull() || use->is_OpaqueTemplateAssertionPredicate() ||
327 use->is_OpaqueInitializedAssertionPredicate()) {
328 if (use->outcnt() == 1) {
329 Node* iff = use->unique_out();
330 assert(iff->is_If(), "unexpected node type");
331 Node *use_c = iff->in(0);
332 if (use_c == blk1 || use_c == blk2) {
333 continue;
334 }
335 }
336 } else {
337 // We might see an Opaque1 from a loop limit check here
338 assert(use->is_If() || use->is_CMove() || use->Opcode() == Op_Opaque1 || use->is_AllocateArray(), "unexpected node type");
339 Node *use_c = (use->is_If() || use->is_AllocateArray()) ? use->in(0) : get_ctrl(use);
340 if (use_c == blk1 || use_c == blk2) {
341 assert(use->is_CMove(), "unexpected node type");
342 continue;
343 }
344 }
345 }
346 if (at_relevant_ctrl(bol, blk1, blk2)) {
347 // Recursively sink any BoolNode
348 #ifndef PRODUCT
349 if( PrintOpto && VerifyLoopOptimizations ) {
350 tty->print("Cloning down: ");
351 bol->dump();
352 }
353 #endif
354 for (DUIterator j = bol->outs(); bol->has_out(j); j++) {
355 Node* u = bol->out(j);
356 // Uses are either IfNodes, CMoves, OpaqueNotNull, or Opaque*AssertionPredicate
357 if (u->is_OpaqueNotNull() || u->is_OpaqueTemplateAssertionPredicate() ||
358 u->is_OpaqueInitializedAssertionPredicate()) {
359 assert(u->in(1) == bol, "bad input");
360 for (DUIterator_Last kmin, k = u->last_outs(kmin); k >= kmin; --k) {
361 Node* iff = u->last_out(k);
362 assert(iff->is_If() || iff->is_CMove(), "unexpected node type");
363 assert( iff->in(1) == u, "" );
364 // Get control block of either the CMove or the If input
365 Node *iff_ctrl = iff->is_If() ? iff->in(0) : get_ctrl(iff);
366 Node *x1 = bol->clone();
367 Node *x2 = u->clone();
368 register_new_node(x1, iff_ctrl);
369 register_new_node(x2, iff_ctrl);
370 _igvn.replace_input_of(x2, 1, x1);
371 _igvn.replace_input_of(iff, 1, x2);
372 }
373 _igvn.remove_dead_node(u);
374 --j;
375 } else {
376 // We might see an Opaque1 from a loop limit check here
377 assert(u->is_If() || u->is_CMove() || u->Opcode() == Op_Opaque1 || u->is_AllocateArray(), "unexpected node type");
378 assert(u->is_AllocateArray() || u->in(1) == bol, "");
379 assert(!u->is_AllocateArray() || u->in(AllocateNode::ValidLengthTest) == bol, "wrong input to AllocateArray");
380 // Get control block of either the CMove or the If input
381 Node *u_ctrl = (u->is_If() || u->is_AllocateArray()) ? u->in(0) : get_ctrl(u);
382 assert((u_ctrl != blk1 && u_ctrl != blk2) || u->is_CMove(), "won't converge");
383 Node *x = bol->clone();
384 register_new_node(x, u_ctrl);
385 _igvn.replace_input_of(u, u->is_AllocateArray() ? AllocateNode::ValidLengthTest : 1, x);
386 --j;
387 }
388 }
389 _igvn.remove_dead_node(bol);
390 --i;
391 }
392 }
393 }
394 // Clone down this CmpNode
395 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; --j) {
396 Node* use = n->last_out(j);
397 uint pos = 1;
398 if (n->is_FastLock()) {
399 pos = TypeFunc::Parms + 2;
400 assert(use->is_Lock(), "FastLock only used by LockNode");
401 }
402 assert(use->in(pos) == n, "" );
403 Node *x = n->clone();
404 register_new_node(x, ctrl_or_self(use));
405 _igvn.replace_input_of(use, pos, x);
406 }
407 _igvn.remove_dead_node(n);
408
409 return true;
410 }
411 }
412 return false;
413 }
414
415 // 'n' could be a node belonging to a Template Assertion Expression (i.e. any node between a Template Assertion Predicate
416 // and its OpaqueLoop* nodes (included)). We cannot simply split this node up since this would create a phi node inside
417 // the Template Assertion Expression - making it unrecognizable as such. Therefore, we completely clone the entire
418 // Template Assertion Expression "down". This ensures that we have an untouched copy that is still recognized by the
419 // Template Assertion Predicate matching code.
420 void PhaseIdealLoop::clone_template_assertion_expression_down(Node* node) {
421 if (!TemplateAssertionExpressionNode::is_in_expression(node)) {
422 return;
423 }
424
425 TemplateAssertionExpressionNode template_assertion_expression_node(node);
426 auto clone_expression = [&](IfNode* template_assertion_predicate) {
427 OpaqueTemplateAssertionPredicateNode* opaque_node =
428 template_assertion_predicate->in(1)->as_OpaqueTemplateAssertionPredicate();
429 TemplateAssertionExpression template_assertion_expression(opaque_node);
430 Node* new_ctrl = template_assertion_predicate->in(0);
431 OpaqueTemplateAssertionPredicateNode* cloned_opaque_node = template_assertion_expression.clone(new_ctrl, this);
432 igvn().replace_input_of(template_assertion_predicate, 1, cloned_opaque_node);
433 };
434 template_assertion_expression_node.for_each_template_assertion_predicate(clone_expression);
435 }
436
437 //------------------------------register_new_node------------------------------
438 void PhaseIdealLoop::register_new_node( Node *n, Node *blk ) {
439 assert(!n->is_CFG(), "must be data node");
440 _igvn.register_new_node_with_optimizer(n);
441 set_ctrl(n, blk);
442 IdealLoopTree *loop = get_loop(blk);
443 if( !loop->_child )
444 loop->_body.push(n);
445 }
446
447 //------------------------------small_cache------------------------------------
448 struct small_cache : public Dict {
449
450 small_cache() : Dict( cmpkey, hashptr ) {}
451 Node *probe( Node *use_blk ) { return (Node*)((*this)[use_blk]); }
452 void lru_insert( Node *use_blk, Node *new_def ) { Insert(use_blk,new_def); }
453 };
454
455 //------------------------------spinup-----------------------------------------
456 // "Spin up" the dominator tree, starting at the use site and stopping when we
457 // find the post-dominating point.
458
459 // We must be at the merge point which post-dominates 'new_false' and
460 // 'new_true'. Figure out which edges into the RegionNode eventually lead up
461 // to false and which to true. Put in a PhiNode to merge values; plug in
462 // the appropriate false-arm or true-arm values. If some path leads to the
463 // original IF, then insert a Phi recursively.
464 Node *PhaseIdealLoop::spinup( Node *iff_dom, Node *new_false, Node *new_true, Node *use_blk, Node *def, small_cache *cache ) {
465 if (use_blk->is_top()) // Handle dead uses
466 return use_blk;
467 Node *prior_n = (Node*)((intptr_t)0xdeadbeef);
468 Node *n = use_blk; // Get path input
469 assert( use_blk != iff_dom, "" );
470 // Here's the "spinup" the dominator tree loop. Do a cache-check
471 // along the way, in case we've come this way before.
472 while( n != iff_dom ) { // Found post-dominating point?
473 prior_n = n;
474 n = idom(n); // Search higher
475 Node *s = cache->probe( prior_n ); // Check cache
476 if( s ) return s; // Cache hit!
477 }
478
479 Node *phi_post;
480 if( prior_n == new_false || prior_n == new_true ) {
481 phi_post = def->clone();
482 phi_post->set_req(0, prior_n );
483 register_new_node(phi_post, prior_n);
484 } else {
485 // This method handles both control uses (looking for Regions) or data
486 // uses (looking for Phis). If looking for a control use, then we need
487 // to insert a Region instead of a Phi; however Regions always exist
488 // previously (the hash_find_insert below would always hit) so we can
489 // return the existing Region.
490 if( def->is_CFG() ) {
491 phi_post = prior_n; // If looking for CFG, return prior
492 } else {
493 assert( def->is_Phi(), "" );
494 assert( prior_n->is_Region(), "must be a post-dominating merge point" );
495
496 // Need a Phi here
497 phi_post = PhiNode::make_blank(prior_n, def);
498 // Search for both true and false on all paths till find one.
499 for( uint i = 1; i < phi_post->req(); i++ ) // For all paths
500 phi_post->init_req( i, spinup( iff_dom, new_false, new_true, prior_n->in(i), def, cache ) );
501 Node *t = _igvn.hash_find_insert(phi_post);
502 if( t ) { // See if we already have this one
503 // phi_post will not be used, so kill it
504 _igvn.remove_dead_node(phi_post);
505 phi_post->destruct(&_igvn);
506 phi_post = t;
507 } else {
508 register_new_node( phi_post, prior_n );
509 }
510 }
511 }
512
513 // Update cache everywhere
514 prior_n = (Node*)((intptr_t)0xdeadbeef); // Reset IDOM walk
515 n = use_blk; // Get path input
516 // Spin-up the idom tree again, basically doing path-compression.
517 // Insert cache entries along the way, so that if we ever hit this
518 // point in the IDOM tree again we'll stop immediately on a cache hit.
519 while( n != iff_dom ) { // Found post-dominating point?
520 prior_n = n;
521 n = idom(n); // Search higher
522 cache->lru_insert( prior_n, phi_post ); // Fill cache
523 } // End of while not gone high enough
524
525 return phi_post;
526 }
527
528 //------------------------------find_use_block---------------------------------
529 // Find the block a USE is in. Normally USE's are in the same block as the
530 // using instruction. For Phi-USE's, the USE is in the predecessor block
531 // along the corresponding path.
532 Node *PhaseIdealLoop::find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ) {
533 // CFG uses are their own block
534 if( use->is_CFG() )
535 return use;
536
537 if( use->is_Phi() ) { // Phi uses in prior block
538 // Grab the first Phi use; there may be many.
539 // Each will be handled as a separate iteration of
540 // the "while( phi->outcnt() )" loop.
541 uint j;
542 for( j = 1; j < use->req(); j++ )
543 if( use->in(j) == def )
544 break;
545 assert( j < use->req(), "def should be among use's inputs" );
546 return use->in(0)->in(j);
547 }
548 // Normal (non-phi) use
549 Node *use_blk = get_ctrl(use);
550 // Some uses are directly attached to the old (and going away)
551 // false and true branches.
552 if( use_blk == old_false ) {
553 use_blk = new_false;
554 set_ctrl(use, new_false);
555 }
556 if( use_blk == old_true ) {
557 use_blk = new_true;
558 set_ctrl(use, new_true);
559 }
560
561 if (use_blk == nullptr) { // He's dead, Jim
562 _igvn.replace_node(use, C->top());
563 }
564
565 return use_blk;
566 }
567
568 //------------------------------handle_use-------------------------------------
569 // Handle uses of the merge point. Basically, split-if makes the merge point
570 // go away so all uses of the merge point must go away as well. Most block
571 // local uses have already been split-up, through the merge point. Uses from
572 // far below the merge point can't always be split up (e.g., phi-uses are
573 // pinned) and it makes too much stuff live. Instead we use a path-based
574 // solution to move uses down.
575 //
576 // If the use is along the pre-split-CFG true branch, then the new use will
577 // be from the post-split-CFG true merge point. Vice-versa for the false
578 // path. Some uses will be along both paths; then we sink the use to the
579 // post-dominating location; we may need to insert a Phi there.
580 void PhaseIdealLoop::handle_use( Node *use, Node *def, small_cache *cache, Node *region_dom, Node *new_false, Node *new_true, Node *old_false, Node *old_true ) {
581
582 Node *use_blk = find_use_block(use,def,old_false,new_false,old_true,new_true);
583 if( !use_blk ) return; // He's dead, Jim
584
585 // Walk up the dominator tree until I hit either the old IfFalse, the old
586 // IfTrue or the old If. Insert Phis where needed.
587 Node *new_def = spinup( region_dom, new_false, new_true, use_blk, def, cache );
588
589 // Found where this USE goes. Re-point him.
590 uint i;
591 for( i = 0; i < use->req(); i++ )
592 if( use->in(i) == def )
593 break;
594 assert( i < use->req(), "def should be among use's inputs" );
595 _igvn.replace_input_of(use, i, new_def);
596 }
597
598 //------------------------------do_split_if------------------------------------
599 // Found an If getting its condition-code input from a Phi in the same block.
600 // Split thru the Region.
601 void PhaseIdealLoop::do_split_if(Node* iff, RegionNode** new_false_region, RegionNode** new_true_region) {
602
603 C->set_major_progress();
604 RegionNode *region = iff->in(0)->as_Region();
605 Node *region_dom = idom(region);
606
607 // We are going to clone this test (and the control flow with it) up through
608 // the incoming merge point. We need to empty the current basic block.
609 // Clone any instructions which must be in this block up through the merge
610 // point.
611 DUIterator i, j;
612 bool progress = true;
613 while (progress) {
614 progress = false;
615 for (i = region->outs(); region->has_out(i); i++) {
616 Node* n = region->out(i);
617 if( n == region ) continue;
618 // The IF to be split is OK.
619 if( n == iff ) continue;
620 if( !n->is_Phi() ) { // Found pinned memory op or such
621 if (split_up(n, region, iff)) {
622 i = region->refresh_out_pos(i);
623 progress = true;
624 }
625 continue;
626 }
627 assert( n->in(0) == region, "" );
628
629 // Recursively split up all users of a Phi
630 for (j = n->outs(); n->has_out(j); j++) {
631 Node* m = n->out(j);
632 // If m is dead, throw it away, and declare progress
633 if (_loop_or_ctrl[m->_idx] == nullptr) {
634 _igvn.remove_dead_node(m);
635 // fall through
636 } else if (m != iff && split_up(m, region, iff)) {
637 // fall through
638 } else {
639 continue;
640 }
641 // Something unpredictable changed.
642 // Tell the iterators to refresh themselves, and rerun the loop.
643 i = region->refresh_out_pos(i);
644 j = region->refresh_out_pos(j);
645 progress = true;
646 }
647 }
648 }
649
650 // Now we have no instructions in the block containing the IF.
651 // Split the IF.
652 RegionNode *new_iff = split_thru_region(iff, region);
653
654 // Replace both uses of 'new_iff' with Regions merging True/False
655 // paths. This makes 'new_iff' go dead.
656 Node *old_false = nullptr, *old_true = nullptr;
657 RegionNode* new_false = nullptr;
658 RegionNode* new_true = nullptr;
659 for (DUIterator_Last j2min, j2 = iff->last_outs(j2min); j2 >= j2min; --j2) {
660 Node *ifp = iff->last_out(j2);
661 assert( ifp->Opcode() == Op_IfFalse || ifp->Opcode() == Op_IfTrue, "" );
662 ifp->set_req(0, new_iff);
663 RegionNode* ifpx = split_thru_region(ifp, region);
664
665 // Replace 'If' projection of a Region with a Region of
666 // 'If' projections.
667 ifpx->set_req(0, ifpx); // A TRUE RegionNode
668
669 // Setup dominator info
670 set_idom(ifpx, region_dom, dom_depth(region_dom) + 1);
671
672 // Check for splitting loop tails
673 if( get_loop(iff)->tail() == ifp )
674 get_loop(iff)->_tail = ifpx;
675
676 // Replace in the graph with lazy-update mechanism
677 new_iff->set_req(0, new_iff); // hook self so it does not go dead
678 lazy_replace(ifp, ifpx);
679 new_iff->set_req(0, region);
680
681 // Record bits for later xforms
682 if( ifp->Opcode() == Op_IfFalse ) {
683 old_false = ifp;
684 new_false = ifpx;
685 } else {
686 old_true = ifp;
687 new_true = ifpx;
688 }
689 }
690 _igvn.remove_dead_node(new_iff);
691 // Lazy replace IDOM info with the region's dominator
692 lazy_replace(iff, region_dom);
693 lazy_update(region, region_dom); // idom must be update before handle_uses
694 region->set_req(0, nullptr); // Break the self-cycle. Required for lazy_update to work on region
695
696 // Now make the original merge point go dead, by handling all its uses.
697 small_cache region_cache;
698 // Preload some control flow in region-cache
699 region_cache.lru_insert( new_false, new_false );
700 region_cache.lru_insert( new_true , new_true );
701 // Now handle all uses of the splitting block
702 for (DUIterator k = region->outs(); region->has_out(k); k++) {
703 Node* phi = region->out(k);
704 if (!phi->in(0)) { // Dead phi? Remove it
705 _igvn.remove_dead_node(phi);
706 } else if (phi == region) { // Found the self-reference
707 continue; // No roll-back of DUIterator
708 } else if (phi->is_Phi()) { // Expected common case: Phi hanging off of Region
709 assert(phi->in(0) == region, "Inconsistent graph");
710 // Need a per-def cache. Phi represents a def, so make a cache
711 small_cache phi_cache;
712
713 // Inspect all Phi uses to make the Phi go dead
714 for (DUIterator_Last lmin, l = phi->last_outs(lmin); l >= lmin; --l) {
715 Node* use = phi->last_out(l);
716 // Compute the new DEF for this USE. New DEF depends on the path
717 // taken from the original DEF to the USE. The new DEF may be some
718 // collection of PHI's merging values from different paths. The Phis
719 // inserted depend only on the location of the USE. We use a
720 // 2-element cache to handle multiple uses from the same block.
721 handle_use(use, phi, &phi_cache, region_dom, new_false, new_true, old_false, old_true);
722 } // End of while phi has uses
723 // Remove the dead Phi
724 _igvn.remove_dead_node( phi );
725 } else {
726 assert(phi->in(0) == region, "Inconsistent graph");
727 // Random memory op guarded by Region. Compute new DEF for USE.
728 handle_use(phi, region, ®ion_cache, region_dom, new_false, new_true, old_false, old_true);
729 }
730 // Every path above deletes a use of the region, except for the region
731 // self-cycle (which is needed by handle_use calling find_use_block
732 // calling get_ctrl calling get_ctrl_no_update looking for dead
733 // regions). So roll back the DUIterator innards.
734 --k;
735 } // End of while merge point has phis
736
737 _igvn.remove_dead_node(region);
738 if (iff->Opcode() == Op_RangeCheck) {
739 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path
740 // into the region is left, an array load could become dependent on a condition that's not a range check for
741 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
742 // floating above its range check.
743 pin_array_access_nodes_dependent_on(new_true);
744 pin_array_access_nodes_dependent_on(new_false);
745 }
746
747 if (new_false_region != nullptr) {
748 *new_false_region = new_false;
749 }
750 if (new_true_region != nullptr) {
751 *new_true_region = new_true;
752 }
753
754 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
755 }
756
757 void PhaseIdealLoop::pin_array_access_nodes_dependent_on(Node* ctrl) {
758 for (DUIterator i = ctrl->outs(); ctrl->has_out(i); i++) {
759 Node* use = ctrl->out(i);
760 if (!use->depends_only_on_test()) {
761 continue;
762 }
763 Node* pinned_clone = use->pin_array_access_node();
764 if (pinned_clone != nullptr) {
765 register_new_node_with_ctrl_of(pinned_clone, use);
766 _igvn.replace_node(use, pinned_clone);
767 --i;
768 }
769 }
770 }