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/loopnode.hpp"
30 #include "opto/movenode.hpp"
31 #include "opto/node.hpp"
32 #include "opto/opaquenode.hpp"
33 #include "opto/predicates.hpp"
34
35 //------------------------------split_thru_region------------------------------
36 // Split Node 'n' through merge point.
37 RegionNode* PhaseIdealLoop::split_thru_region(Node* n, RegionNode* region) {
38 assert(n->is_CFG(), "");
39 RegionNode* r = new RegionNode(region->req());
40 IdealLoopTree* loop = get_loop(n);
41 for (uint i = 1; i < region->req(); i++) {
42 Node* x = n->clone();
43 Node* in0 = n->in(0);
44 if (in0->in(0) == region) x->set_req(0, in0->in(i));
45 for (uint j = 1; j < n->req(); j++) {
46 Node* in = n->in(j);
47 if (get_ctrl(in) == region) {
48 x->set_req(j, in->in(i));
49 }
50 }
51 _igvn.register_new_node_with_optimizer(x);
52 set_loop(x, loop);
53 set_idom(x, x->in(0), dom_depth(x->in(0))+1);
54 r->init_req(i, x);
55 }
56
57 // Record region
58 r->set_req(0,region); // Not a TRUE RegionNode
59 _igvn.register_new_node_with_optimizer(r);
60 set_loop(r, loop);
61 if (!loop->_child) {
62 loop->_body.push(r);
63 }
64 return r;
65 }
66
67 //------------------------------split_up---------------------------------------
68 // Split block-local op up through the phis to empty the current block
69 bool PhaseIdealLoop::split_up( Node *n, Node *blk1, Node *blk2 ) {
70 if( n->is_CFG() ) {
71 assert( n->in(0) != blk1, "Lousy candidate for split-if" );
72 return false;
73 }
74 if (!at_relevant_ctrl(n, blk1, blk2))
75 return false; // Not block local
76 if( n->is_Phi() ) return false; // Local PHIs are expected
77
78 // Recursively split-up inputs
79 for (uint i = 1; i < n->req(); i++) {
80 if( split_up( n->in(i), blk1, blk2 ) ) {
81 // Got split recursively and self went dead?
82 if (n->outcnt() == 0)
83 _igvn.remove_dead_node(n);
84 return true;
85 }
86 }
87
88 if (clone_cmp_loadklass_down(n, blk1, blk2)) {
89 return true;
90 }
91
92 // Check for needing to clone-up a compare. Can't do that, it forces
93 // another (nested) split-if transform. Instead, clone it "down".
94 if (clone_cmp_down(n, blk1, blk2)) {
95 return true;
96 }
97
98 clone_template_assertion_expression_down(n);
99
100 if (n->Opcode() == Op_OpaqueZeroTripGuard) {
101 // If this Opaque1 is part of the zero trip guard for a loop:
102 // 1- it can't be shared
103 // 2- the zero trip guard can't be the if that's being split
104 // As a consequence, this node could be assigned control anywhere between its current control and the zero trip guard.
105 // Move it down to get it out of the way of split if and avoid breaking the zero trip guard shape.
106 Node* cmp = n->unique_out();
107 assert(cmp->Opcode() == Op_CmpI, "bad zero trip guard shape");
108 Node* bol = cmp->unique_out();
109 assert(bol->Opcode() == Op_Bool, "bad zero trip guard shape");
110 Node* iff = bol->unique_out();
111 assert(iff->Opcode() == Op_If, "bad zero trip guard shape");
112 set_ctrl(n, iff->in(0));
113 set_ctrl(cmp, iff->in(0));
114 set_ctrl(bol, iff->in(0));
115 return true;
116 }
117
118 // See if splitting-up a Store. Any anti-dep loads must go up as
119 // well. An anti-dep load might be in the wrong block, because in
120 // this particular layout/schedule we ignored anti-deps and allow
121 // memory to be alive twice. This only works if we do the same
122 // operations on anti-dep loads as we do their killing stores.
123 if( n->is_Store() && n->in(MemNode::Memory)->in(0) == n->in(0) ) {
124 // Get store's memory slice
125 int alias_idx = C->get_alias_index(_igvn.type(n->in(MemNode::Address))->is_ptr());
126
127 // Get memory-phi anti-dep loads will be using
128 Node *memphi = n->in(MemNode::Memory);
129 assert( memphi->is_Phi(), "" );
130 // Hoist any anti-dep load to the splitting block;
131 // it will then "split-up".
132 for (DUIterator_Fast imax,i = memphi->fast_outs(imax); i < imax; i++) {
133 Node *load = memphi->fast_out(i);
134 if( load->is_Load() && alias_idx == C->get_alias_index(_igvn.type(load->in(MemNode::Address))->is_ptr()) )
135 set_ctrl(load,blk1);
136 }
137 }
138
139 // Found some other Node; must clone it up
140 #ifndef PRODUCT
141 if( PrintOpto && VerifyLoopOptimizations ) {
142 tty->print("Cloning up: ");
143 n->dump();
144 }
145 #endif
146
147 // ConvI2L may have type information on it which becomes invalid if
148 // it moves up in the graph so change any clones so widen the type
149 // to TypeLong::INT when pushing it up.
150 const Type* rtype = nullptr;
151 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::INT) {
152 rtype = TypeLong::INT;
153 }
154
155 // Now actually split-up this guy. One copy per control path merging.
156 Node *phi = PhiNode::make_blank(blk1, n);
157 for( uint j = 1; j < blk1->req(); j++ ) {
158 Node *x = n->clone();
159 // Widen the type of the ConvI2L when pushing up.
160 if (rtype != nullptr) x->as_Type()->set_type(rtype);
161 if( n->in(0) && n->in(0) == blk1 )
162 x->set_req( 0, blk1->in(j) );
163 for( uint i = 1; i < n->req(); i++ ) {
164 Node *m = n->in(i);
165 if( get_ctrl(m) == blk1 ) {
166 assert( m->in(0) == blk1, "" );
167 x->set_req( i, m->in(j) );
168 }
169 }
170 register_new_node( x, blk1->in(j) );
171 phi->init_req( j, x );
172 }
173 // Announce phi to optimizer
174 register_new_node(phi, blk1);
175
176 // Remove cloned-up value from optimizer; use phi instead
177 _igvn.replace_node( n, phi );
178
179 // (There used to be a self-recursive call to split_up() here,
180 // but it is not needed. All necessary forward walking is done
181 // by do_split_if() below.)
182
183 return true;
184 }
185
186 // Look for a (If .. (Bool(CmpP (LoadKlass .. (AddP obj ..)) ..))) and clone all of it down.
187 // There's likely a CheckCastPP on one of the branches of the If, with obj as input.
188 // If the (LoadKlass .. (AddP obj ..)) is not cloned down, then split if transforms this to: (If .. (Bool(CmpP phi1 ..)))
189 // and the CheckCastPP to (CheckCastPP phi2). It's possible then that phi2 is transformed to a CheckCastPP
190 // (through PhiNode::Ideal) and that that CheckCastPP is replaced by another narrower CheckCastPP at the same control
191 // (through ConstraintCastNode::Identity). That could cause the CheckCastPP at the If to become top while (CmpP phi1)
192 // wouldn't constant fold because it's using a different data path. Cloning the whole subgraph down guarantees both the
193 // AddP and CheckCastPP have the same obj input after split if.
194 bool PhaseIdealLoop::clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2) {
195 if (n->Opcode() == Op_AddP && at_relevant_ctrl(n, blk1, blk2)) {
196 Node_List cmp_nodes;
197 uint old = C->unique();
198 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
199 Node* u1 = n->fast_out(i);
200 if (u1->Opcode() == Op_LoadNKlass && at_relevant_ctrl(u1, blk1, blk2)) {
201 for (DUIterator_Fast jmax, j = u1->fast_outs(jmax); j < jmax; j++) {
202 Node* u2 = u1->fast_out(j);
203 if (u2->Opcode() == Op_DecodeNKlass && at_relevant_ctrl(u2, blk1, blk2)) {
204 for (DUIterator k = u2->outs(); u2->has_out(k); k++) {
205 Node* u3 = u2->out(k);
206 if (at_relevant_ctrl(u3, blk1, blk2) && clone_cmp_down(u3, blk1, blk2)) {
207 --k;
208 }
209 }
210 for (DUIterator_Fast kmax, k = u2->fast_outs(kmax); k < kmax; k++) {
211 Node* u3 = u2->fast_out(k);
212 if (u3->_idx >= old) {
213 cmp_nodes.push(u3);
214 }
215 }
216 }
217 }
218 } else if (u1->Opcode() == Op_LoadKlass && at_relevant_ctrl(u1, blk1, blk2)) {
219 for (DUIterator j = u1->outs(); u1->has_out(j); j++) {
220 Node* u2 = u1->out(j);
221 if (at_relevant_ctrl(u2, blk1, blk2) && clone_cmp_down(u2, blk1, blk2)) {
222 --j;
223 }
224 }
225 for (DUIterator_Fast kmax, k = u1->fast_outs(kmax); k < kmax; k++) {
226 Node* u2 = u1->fast_out(k);
227 if (u2->_idx >= old) {
228 cmp_nodes.push(u2);
229 }
230 }
231 }
232 }
233
234 for (uint i = 0; i < cmp_nodes.size(); ++i) {
235 Node* cmp = cmp_nodes.at(i);
236 clone_loadklass_nodes_at_cmp_index(n, cmp, 1);
237 clone_loadklass_nodes_at_cmp_index(n, cmp, 2);
238 }
239 if (n->outcnt() == 0) {
240 assert(n->is_dead(), "");
241 return true;
242 }
243 }
244 return false;
245 }
246
247 bool PhaseIdealLoop::at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2) {
248 return ctrl_or_self(n) == blk1 || ctrl_or_self(n) == blk2;
249 }
250
251 void PhaseIdealLoop::clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i) {
252 Node* decode = cmp->in(i);
253 if (decode->Opcode() == Op_DecodeNKlass) {
254 Node* loadklass = decode->in(1);
255 if (loadklass->Opcode() == Op_LoadNKlass) {
256 Node* addp = loadklass->in(MemNode::Address);
257 if (addp == n) {
258 Node* ctrl = get_ctrl(cmp);
259 Node* decode_clone = decode->clone();
260 Node* loadklass_clone = loadklass->clone();
261 Node* addp_clone = addp->clone();
262 register_new_node(decode_clone, ctrl);
263 register_new_node(loadklass_clone, ctrl);
264 register_new_node(addp_clone, ctrl);
265 _igvn.replace_input_of(cmp, i, decode_clone);
266 _igvn.replace_input_of(decode_clone, 1, loadklass_clone);
267 _igvn.replace_input_of(loadklass_clone, MemNode::Address, addp_clone);
268 if (decode->outcnt() == 0) {
269 _igvn.remove_dead_node(decode);
270 }
271 }
272 }
273 } else {
274 Node* loadklass = cmp->in(i);
275 if (loadklass->Opcode() == Op_LoadKlass) {
276 Node* addp = loadklass->in(MemNode::Address);
277 if (addp == n) {
278 Node* ctrl = get_ctrl(cmp);
279 Node* loadklass_clone = loadklass->clone();
280 Node* addp_clone = addp->clone();
281 register_new_node(loadklass_clone, ctrl);
282 register_new_node(addp_clone, ctrl);
283 _igvn.replace_input_of(cmp, i, loadklass_clone);
284 _igvn.replace_input_of(loadklass_clone, MemNode::Address, addp_clone);
285 if (loadklass->outcnt() == 0) {
286 _igvn.remove_dead_node(loadklass);
287 }
288 }
289 }
290 }
291 }
292
293 bool PhaseIdealLoop::clone_cmp_down(Node* n, const Node* blk1, const Node* blk2) {
294 if( n->is_Cmp() ) {
295 assert(get_ctrl(n) == blk2 || get_ctrl(n) == blk1, "must be in block with IF");
296 // Check for simple Cmp/Bool/CMove which we can clone-up. Cmp/Bool/CMove
297 // sequence can have no other users and it must all reside in the split-if
298 // block. Non-simple Cmp/Bool/CMove sequences are 'cloned-down' below -
299 // private, per-use versions of the Cmp and Bool are made. These sink to
300 // the CMove block. If the CMove is in the split-if block, then in the
301 // next iteration this will become a simple Cmp/Bool/CMove set to clone-up.
302 Node *bol, *cmov;
303 if (!(n->outcnt() == 1 && n->unique_out()->is_Bool() &&
304 (bol = n->unique_out()->as_Bool()) &&
305 (at_relevant_ctrl(bol, blk1, blk2) &&
306 bol->outcnt() == 1 &&
307 bol->unique_out()->is_CMove() &&
308 (cmov = bol->unique_out()->as_CMove()) &&
309 at_relevant_ctrl(cmov, blk1, blk2)))) {
310
311 // Must clone down
312 #ifndef PRODUCT
313 if( PrintOpto && VerifyLoopOptimizations ) {
314 tty->print("Cloning down: ");
315 n->dump();
316 }
317 #endif
318 if (!n->is_FastLock()) {
319 // Clone down any block-local BoolNode uses of this CmpNode
320 for (DUIterator i = n->outs(); n->has_out(i); i++) {
321 Node* bol = n->out(i);
322 assert( bol->is_Bool(), "" );
323 if (bol->outcnt() == 1) {
324 Node* use = bol->unique_out();
325 if (use->is_Opaque4() || use->is_OpaqueInitializedAssertionPredicate()) {
326 if (use->outcnt() == 1) {
327 Node* iff = use->unique_out();
328 assert(iff->is_If(), "unexpected node type");
329 Node *use_c = iff->in(0);
330 if (use_c == blk1 || use_c == blk2) {
331 continue;
332 }
333 }
334 } else {
335 // We might see an Opaque1 from a loop limit check here
336 assert(use->is_If() || use->is_CMove() || use->Opcode() == Op_Opaque1 || use->is_AllocateArray(), "unexpected node type");
337 Node *use_c = (use->is_If() || use->is_AllocateArray()) ? use->in(0) : get_ctrl(use);
338 if (use_c == blk1 || use_c == blk2) {
339 assert(use->is_CMove(), "unexpected node type");
340 continue;
341 }
342 }
343 }
344 if (at_relevant_ctrl(bol, blk1, blk2)) {
345 // Recursively sink any BoolNode
346 #ifndef PRODUCT
347 if( PrintOpto && VerifyLoopOptimizations ) {
348 tty->print("Cloning down: ");
349 bol->dump();
350 }
351 #endif
352 for (DUIterator j = bol->outs(); bol->has_out(j); j++) {
353 Node* u = bol->out(j);
354 // Uses are either IfNodes, CMoves, Opaque4, or OpaqueInitializedAssertionPredicates
355 if (u->is_Opaque4() || u->is_OpaqueInitializedAssertionPredicate()) {
356 assert(u->in(1) == bol, "bad input");
357 for (DUIterator_Last kmin, k = u->last_outs(kmin); k >= kmin; --k) {
358 Node* iff = u->last_out(k);
359 assert(iff->is_If() || iff->is_CMove(), "unexpected node type");
360 assert( iff->in(1) == u, "" );
361 // Get control block of either the CMove or the If input
362 Node *iff_ctrl = iff->is_If() ? iff->in(0) : get_ctrl(iff);
363 Node *x1 = bol->clone();
364 Node *x2 = u->clone();
365 register_new_node(x1, iff_ctrl);
366 register_new_node(x2, iff_ctrl);
367 _igvn.replace_input_of(x2, 1, x1);
368 _igvn.replace_input_of(iff, 1, x2);
369 }
370 _igvn.remove_dead_node(u);
371 --j;
372 } else {
373 // We might see an Opaque1 from a loop limit check here
374 assert(u->is_If() || u->is_CMove() || u->Opcode() == Op_Opaque1 || u->is_AllocateArray(), "unexpected node type");
375 assert(u->is_AllocateArray() || u->in(1) == bol, "");
376 assert(!u->is_AllocateArray() || u->in(AllocateNode::ValidLengthTest) == bol, "wrong input to AllocateArray");
377 // Get control block of either the CMove or the If input
378 Node *u_ctrl = (u->is_If() || u->is_AllocateArray()) ? u->in(0) : get_ctrl(u);
379 assert((u_ctrl != blk1 && u_ctrl != blk2) || u->is_CMove(), "won't converge");
380 Node *x = bol->clone();
381 register_new_node(x, u_ctrl);
382 _igvn.replace_input_of(u, u->is_AllocateArray() ? AllocateNode::ValidLengthTest : 1, x);
383 --j;
384 }
385 }
386 _igvn.remove_dead_node(bol);
387 --i;
388 }
389 }
390 }
391 // Clone down this CmpNode
392 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; --j) {
393 Node* use = n->last_out(j);
394 uint pos = 1;
395 if (n->is_FastLock()) {
396 pos = TypeFunc::Parms + 2;
397 assert(use->is_Lock(), "FastLock only used by LockNode");
398 }
399 assert(use->in(pos) == n, "" );
400 Node *x = n->clone();
401 register_new_node(x, ctrl_or_self(use));
402 _igvn.replace_input_of(use, pos, x);
403 }
404 _igvn.remove_dead_node(n);
405
406 return true;
407 }
408 }
409 return false;
410 }
411
412 // 'n' could be a node belonging to a Template Assertion Expression (i.e. any node between a Template Assertion Predicate
413 // and its OpaqueLoop* nodes (included)). We cannot simply split this node up since this would create a phi node inside
414 // the Template Assertion Expression - making it unrecognizable as such. Therefore, we completely clone the entire
415 // Template Assertion Expression "down". This ensures that we have an untouched copy that is still recognized by the
416 // Template Assertion Predicate matching code.
417 void PhaseIdealLoop::clone_template_assertion_expression_down(Node* node) {
418 if (!TemplateAssertionExpressionNode::is_in_expression(node)) {
419 return;
420 }
421
422 TemplateAssertionExpressionNode template_assertion_expression_node(node);
423 auto clone_expression = [&](IfNode* template_assertion_predicate) {
424 Opaque4Node* opaque4_node = template_assertion_predicate->in(1)->as_Opaque4();
425 TemplateAssertionExpression template_assertion_expression(opaque4_node);
426 Node* new_ctrl = template_assertion_predicate->in(0);
427 Opaque4Node* cloned_opaque4_node = template_assertion_expression.clone(new_ctrl, this);
428 igvn().replace_input_of(template_assertion_predicate, 1, cloned_opaque4_node);
429 };
430 template_assertion_expression_node.for_each_template_assertion_predicate(clone_expression);
431 }
432
433 //------------------------------register_new_node------------------------------
434 void PhaseIdealLoop::register_new_node( Node *n, Node *blk ) {
435 assert(!n->is_CFG(), "must be data node");
436 _igvn.register_new_node_with_optimizer(n);
437 set_ctrl(n, blk);
438 IdealLoopTree *loop = get_loop(blk);
439 if( !loop->_child )
440 loop->_body.push(n);
441 }
442
443 //------------------------------small_cache------------------------------------
444 struct small_cache : public Dict {
445
446 small_cache() : Dict( cmpkey, hashptr ) {}
447 Node *probe( Node *use_blk ) { return (Node*)((*this)[use_blk]); }
448 void lru_insert( Node *use_blk, Node *new_def ) { Insert(use_blk,new_def); }
449 };
450
451 //------------------------------spinup-----------------------------------------
452 // "Spin up" the dominator tree, starting at the use site and stopping when we
453 // find the post-dominating point.
454
455 // We must be at the merge point which post-dominates 'new_false' and
456 // 'new_true'. Figure out which edges into the RegionNode eventually lead up
457 // to false and which to true. Put in a PhiNode to merge values; plug in
458 // the appropriate false-arm or true-arm values. If some path leads to the
459 // original IF, then insert a Phi recursively.
460 Node *PhaseIdealLoop::spinup( Node *iff_dom, Node *new_false, Node *new_true, Node *use_blk, Node *def, small_cache *cache ) {
461 if (use_blk->is_top()) // Handle dead uses
462 return use_blk;
463 Node *prior_n = (Node*)((intptr_t)0xdeadbeef);
464 Node *n = use_blk; // Get path input
465 assert( use_blk != iff_dom, "" );
466 // Here's the "spinup" the dominator tree loop. Do a cache-check
467 // along the way, in case we've come this way before.
468 while( n != iff_dom ) { // Found post-dominating point?
469 prior_n = n;
470 n = idom(n); // Search higher
471 Node *s = cache->probe( prior_n ); // Check cache
472 if( s ) return s; // Cache hit!
473 }
474
475 Node *phi_post;
476 if( prior_n == new_false || prior_n == new_true ) {
477 phi_post = def->clone();
478 phi_post->set_req(0, prior_n );
479 register_new_node(phi_post, prior_n);
480 } else {
481 // This method handles both control uses (looking for Regions) or data
482 // uses (looking for Phis). If looking for a control use, then we need
483 // to insert a Region instead of a Phi; however Regions always exist
484 // previously (the hash_find_insert below would always hit) so we can
485 // return the existing Region.
486 if( def->is_CFG() ) {
487 phi_post = prior_n; // If looking for CFG, return prior
488 } else {
489 assert( def->is_Phi(), "" );
490 assert( prior_n->is_Region(), "must be a post-dominating merge point" );
491
492 // Need a Phi here
493 phi_post = PhiNode::make_blank(prior_n, def);
494 // Search for both true and false on all paths till find one.
495 for( uint i = 1; i < phi_post->req(); i++ ) // For all paths
496 phi_post->init_req( i, spinup( iff_dom, new_false, new_true, prior_n->in(i), def, cache ) );
497 Node *t = _igvn.hash_find_insert(phi_post);
498 if( t ) { // See if we already have this one
499 // phi_post will not be used, so kill it
500 _igvn.remove_dead_node(phi_post);
501 phi_post->destruct(&_igvn);
502 phi_post = t;
503 } else {
504 register_new_node( phi_post, prior_n );
505 }
506 }
507 }
508
509 // Update cache everywhere
510 prior_n = (Node*)((intptr_t)0xdeadbeef); // Reset IDOM walk
511 n = use_blk; // Get path input
512 // Spin-up the idom tree again, basically doing path-compression.
513 // Insert cache entries along the way, so that if we ever hit this
514 // point in the IDOM tree again we'll stop immediately on a cache hit.
515 while( n != iff_dom ) { // Found post-dominating point?
516 prior_n = n;
517 n = idom(n); // Search higher
518 cache->lru_insert( prior_n, phi_post ); // Fill cache
519 } // End of while not gone high enough
520
521 return phi_post;
522 }
523
524 //------------------------------find_use_block---------------------------------
525 // Find the block a USE is in. Normally USE's are in the same block as the
526 // using instruction. For Phi-USE's, the USE is in the predecessor block
527 // along the corresponding path.
528 Node *PhaseIdealLoop::find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ) {
529 // CFG uses are their own block
530 if( use->is_CFG() )
531 return use;
532
533 if( use->is_Phi() ) { // Phi uses in prior block
534 // Grab the first Phi use; there may be many.
535 // Each will be handled as a separate iteration of
536 // the "while( phi->outcnt() )" loop.
537 uint j;
538 for( j = 1; j < use->req(); j++ )
539 if( use->in(j) == def )
540 break;
541 assert( j < use->req(), "def should be among use's inputs" );
542 return use->in(0)->in(j);
543 }
544 // Normal (non-phi) use
545 Node *use_blk = get_ctrl(use);
546 // Some uses are directly attached to the old (and going away)
547 // false and true branches.
548 if( use_blk == old_false ) {
549 use_blk = new_false;
550 set_ctrl(use, new_false);
551 }
552 if( use_blk == old_true ) {
553 use_blk = new_true;
554 set_ctrl(use, new_true);
555 }
556
557 if (use_blk == nullptr) { // He's dead, Jim
558 _igvn.replace_node(use, C->top());
559 }
560
561 return use_blk;
562 }
563
564 //------------------------------handle_use-------------------------------------
565 // Handle uses of the merge point. Basically, split-if makes the merge point
566 // go away so all uses of the merge point must go away as well. Most block
567 // local uses have already been split-up, through the merge point. Uses from
568 // far below the merge point can't always be split up (e.g., phi-uses are
569 // pinned) and it makes too much stuff live. Instead we use a path-based
570 // solution to move uses down.
571 //
572 // If the use is along the pre-split-CFG true branch, then the new use will
573 // be from the post-split-CFG true merge point. Vice-versa for the false
574 // path. Some uses will be along both paths; then we sink the use to the
575 // post-dominating location; we may need to insert a Phi there.
576 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 ) {
577
578 Node *use_blk = find_use_block(use,def,old_false,new_false,old_true,new_true);
579 if( !use_blk ) return; // He's dead, Jim
580
581 // Walk up the dominator tree until I hit either the old IfFalse, the old
582 // IfTrue or the old If. Insert Phis where needed.
583 Node *new_def = spinup( region_dom, new_false, new_true, use_blk, def, cache );
584
585 // Found where this USE goes. Re-point him.
586 uint i;
587 for( i = 0; i < use->req(); i++ )
588 if( use->in(i) == def )
589 break;
590 assert( i < use->req(), "def should be among use's inputs" );
591 _igvn.replace_input_of(use, i, new_def);
592 }
593
594 //------------------------------do_split_if------------------------------------
595 // Found an If getting its condition-code input from a Phi in the same block.
596 // Split thru the Region.
597 void PhaseIdealLoop::do_split_if(Node* iff, RegionNode** new_false_region, RegionNode** new_true_region) {
598
599 C->set_major_progress();
600 RegionNode *region = iff->in(0)->as_Region();
601 Node *region_dom = idom(region);
602
603 // We are going to clone this test (and the control flow with it) up through
604 // the incoming merge point. We need to empty the current basic block.
605 // Clone any instructions which must be in this block up through the merge
606 // point.
607 DUIterator i, j;
608 bool progress = true;
609 while (progress) {
610 progress = false;
611 for (i = region->outs(); region->has_out(i); i++) {
612 Node* n = region->out(i);
613 if( n == region ) continue;
614 // The IF to be split is OK.
615 if( n == iff ) continue;
616 if( !n->is_Phi() ) { // Found pinned memory op or such
617 if (split_up(n, region, iff)) {
618 i = region->refresh_out_pos(i);
619 progress = true;
620 }
621 continue;
622 }
623 assert( n->in(0) == region, "" );
624
625 // Recursively split up all users of a Phi
626 for (j = n->outs(); n->has_out(j); j++) {
627 Node* m = n->out(j);
628 // If m is dead, throw it away, and declare progress
629 if (_loop_or_ctrl[m->_idx] == nullptr) {
630 _igvn.remove_dead_node(m);
631 // fall through
632 }
633 else if (m != iff && split_up(m, region, iff)) {
634 // fall through
635 } else {
636 continue;
637 }
638 // Something unpredictable changed.
639 // Tell the iterators to refresh themselves, and rerun the loop.
640 i = region->refresh_out_pos(i);
641 j = region->refresh_out_pos(j);
642 progress = true;
643 }
644 }
645 }
646
647 // Now we have no instructions in the block containing the IF.
648 // Split the IF.
649 RegionNode *new_iff = split_thru_region(iff, region);
650
651 // Replace both uses of 'new_iff' with Regions merging True/False
652 // paths. This makes 'new_iff' go dead.
653 Node *old_false = nullptr, *old_true = nullptr;
654 RegionNode* new_false = nullptr;
655 RegionNode* new_true = nullptr;
656 for (DUIterator_Last j2min, j2 = iff->last_outs(j2min); j2 >= j2min; --j2) {
657 Node *ifp = iff->last_out(j2);
658 assert( ifp->Opcode() == Op_IfFalse || ifp->Opcode() == Op_IfTrue, "" );
659 ifp->set_req(0, new_iff);
660 RegionNode* ifpx = split_thru_region(ifp, region);
661
662 // Replace 'If' projection of a Region with a Region of
663 // 'If' projections.
664 ifpx->set_req(0, ifpx); // A TRUE RegionNode
665
666 // Setup dominator info
667 set_idom(ifpx, region_dom, dom_depth(region_dom) + 1);
668
669 // Check for splitting loop tails
670 if( get_loop(iff)->tail() == ifp )
671 get_loop(iff)->_tail = ifpx;
672
673 // Replace in the graph with lazy-update mechanism
674 new_iff->set_req(0, new_iff); // hook self so it does not go dead
675 lazy_replace(ifp, ifpx);
676 new_iff->set_req(0, region);
677
678 // Record bits for later xforms
679 if( ifp->Opcode() == Op_IfFalse ) {
680 old_false = ifp;
681 new_false = ifpx;
682 } else {
683 old_true = ifp;
684 new_true = ifpx;
685 }
686 }
687 _igvn.remove_dead_node(new_iff);
688 // Lazy replace IDOM info with the region's dominator
689 lazy_replace(iff, region_dom);
690 lazy_update(region, region_dom); // idom must be update before handle_uses
691 region->set_req(0, nullptr); // Break the self-cycle. Required for lazy_update to work on region
692
693 // Now make the original merge point go dead, by handling all its uses.
694 small_cache region_cache;
695 // Preload some control flow in region-cache
696 region_cache.lru_insert( new_false, new_false );
697 region_cache.lru_insert( new_true , new_true );
698 // Now handle all uses of the splitting block
699 for (DUIterator k = region->outs(); region->has_out(k); k++) {
700 Node* phi = region->out(k);
701 if (!phi->in(0)) { // Dead phi? Remove it
702 _igvn.remove_dead_node(phi);
703 } else if (phi == region) { // Found the self-reference
704 continue; // No roll-back of DUIterator
705 } else if (phi->is_Phi()) { // Expected common case: Phi hanging off of Region
706 assert(phi->in(0) == region, "Inconsistent graph");
707 // Need a per-def cache. Phi represents a def, so make a cache
708 small_cache phi_cache;
709
710 // Inspect all Phi uses to make the Phi go dead
711 for (DUIterator_Last lmin, l = phi->last_outs(lmin); l >= lmin; --l) {
712 Node* use = phi->last_out(l);
713 // Compute the new DEF for this USE. New DEF depends on the path
714 // taken from the original DEF to the USE. The new DEF may be some
715 // collection of PHI's merging values from different paths. The Phis
716 // inserted depend only on the location of the USE. We use a
717 // 2-element cache to handle multiple uses from the same block.
718 handle_use(use, phi, &phi_cache, region_dom, new_false, new_true, old_false, old_true);
719 } // End of while phi has uses
720 // Remove the dead Phi
721 _igvn.remove_dead_node( phi );
722 } else {
723 assert(phi->in(0) == region, "Inconsistent graph");
724 // Random memory op guarded by Region. Compute new DEF for USE.
725 handle_use(phi, region, ®ion_cache, region_dom, new_false, new_true, old_false, old_true);
726 }
727 // Every path above deletes a use of the region, except for the region
728 // self-cycle (which is needed by handle_use calling find_use_block
729 // calling get_ctrl calling get_ctrl_no_update looking for dead
730 // regions). So roll back the DUIterator innards.
731 --k;
732 } // End of while merge point has phis
733
734 _igvn.remove_dead_node(region);
735 if (iff->Opcode() == Op_RangeCheck) {
736 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path
737 // into the region is left, an array load could become dependent on a condition that's not a range check for
738 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk
739 // floating above its range check.
740 pin_array_access_nodes_dependent_on(new_true);
741 pin_array_access_nodes_dependent_on(new_false);
742 }
743
744 if (new_false_region != nullptr) {
745 *new_false_region = new_false;
746 }
747 if (new_true_region != nullptr) {
748 *new_true_region = new_true;
749 }
750
751 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
752 }
753
754 void PhaseIdealLoop::pin_array_access_nodes_dependent_on(Node* ctrl) {
755 for (DUIterator i = ctrl->outs(); ctrl->has_out(i); i++) {
756 Node* use = ctrl->out(i);
757 if (!use->depends_only_on_test()) {
758 continue;
759 }
760 Node* pinned_clone = use->pin_array_access_node();
761 if (pinned_clone != nullptr) {
762 register_new_node_with_ctrl_of(pinned_clone, use);
763 _igvn.replace_node(use, pinned_clone);
764 --i;
765 }
766 }
767 }