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 "opto/addnode.hpp"
26 #include "opto/node.hpp"
27 #include "precompiled.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/loopnode.hpp"
31 #include "opto/movenode.hpp"
32 #include "opto/opaquenode.hpp"
33
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 if (subgraph_has_opaque(n)) {
99 Unique_Node_List wq;
100 wq.push(n);
101 for (uint i = 0; i < wq.size(); i++) {
102 Node* m = wq.at(i);
103 if (m->is_If()) {
104 assert(skeleton_predicate_has_opaque(m->as_If()), "opaque node not reachable from if?");
105 Node* bol = clone_skeleton_predicate_bool(m, NULL, NULL, m->in(0));
106 _igvn.replace_input_of(m, 1, bol);
107 } else {
108 assert(!m->is_CFG(), "not CFG expected");
109 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
110 Node* u = m->fast_out(j);
111 wq.push(u);
112 }
113 }
114 }
115 }
116
117 if (n->Opcode() == Op_OpaqueZeroTripGuard) {
118 // If this Opaque1 is part of the zero trip guard for a loop:
119 // 1- it can't be shared
120 // 2- the zero trip guard can't be the if that's being split
121 // As a consequence, this node could be assigned control anywhere between its current control and the zero trip guard.
122 // Move it down to get it out of the way of split if and avoid breaking the zero trip guard shape.
123 Node* cmp = n->unique_out();
124 assert(cmp->Opcode() == Op_CmpI, "bad zero trip guard shape");
125 Node* bol = cmp->unique_out();
126 assert(bol->Opcode() == Op_Bool, "bad zero trip guard shape");
127 Node* iff = bol->unique_out();
128 assert(iff->Opcode() == Op_If, "bad zero trip guard shape");
129 set_ctrl(n, iff->in(0));
130 set_ctrl(cmp, iff->in(0));
131 set_ctrl(bol, iff->in(0));
132 return true;
133 }
134
135 // See if splitting-up a Store. Any anti-dep loads must go up as
136 // well. An anti-dep load might be in the wrong block, because in
137 // this particular layout/schedule we ignored anti-deps and allow
138 // memory to be alive twice. This only works if we do the same
139 // operations on anti-dep loads as we do their killing stores.
140 if( n->is_Store() && n->in(MemNode::Memory)->in(0) == n->in(0) ) {
141 // Get store's memory slice
142 int alias_idx = C->get_alias_index(_igvn.type(n->in(MemNode::Address))->is_ptr());
143
144 // Get memory-phi anti-dep loads will be using
145 Node *memphi = n->in(MemNode::Memory);
146 assert( memphi->is_Phi(), "" );
147 // Hoist any anti-dep load to the splitting block;
148 // it will then "split-up".
149 for (DUIterator_Fast imax,i = memphi->fast_outs(imax); i < imax; i++) {
150 Node *load = memphi->fast_out(i);
151 if( load->is_Load() && alias_idx == C->get_alias_index(_igvn.type(load->in(MemNode::Address))->is_ptr()) )
152 set_ctrl(load,blk1);
153 }
154 }
155
156 // Found some other Node; must clone it up
157 #ifndef PRODUCT
158 if( PrintOpto && VerifyLoopOptimizations ) {
159 tty->print("Cloning up: ");
160 n->dump();
161 }
162 #endif
163
164 // ConvI2L may have type information on it which becomes invalid if
165 // it moves up in the graph so change any clones so widen the type
166 // to TypeLong::INT when pushing it up.
167 const Type* rtype = NULL;
168 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::INT) {
169 rtype = TypeLong::INT;
170 }
171
172 // Now actually split-up this guy. One copy per control path merging.
173 Node *phi = PhiNode::make_blank(blk1, n);
174 for( uint j = 1; j < blk1->req(); j++ ) {
175 Node *x = n->clone();
176 // Widen the type of the ConvI2L when pushing up.
177 if (rtype != NULL) x->as_Type()->set_type(rtype);
178 if( n->in(0) && n->in(0) == blk1 )
179 x->set_req( 0, blk1->in(j) );
180 for( uint i = 1; i < n->req(); i++ ) {
181 Node *m = n->in(i);
182 if( get_ctrl(m) == blk1 ) {
183 assert( m->in(0) == blk1, "" );
184 x->set_req( i, m->in(j) );
185 }
186 }
187 register_new_node( x, blk1->in(j) );
188 phi->init_req( j, x );
189 }
190 // Announce phi to optimizer
191 register_new_node(phi, blk1);
192
193 // Remove cloned-up value from optimizer; use phi instead
194 _igvn.replace_node( n, phi );
195
196 // (There used to be a self-recursive call to split_up() here,
197 // but it is not needed. All necessary forward walking is done
198 // by do_split_if() below.)
199
200 return true;
201 }
202
203 // Look for a (If .. (Bool(CmpP (LoadKlass .. (AddP obj ..)) ..))) and clone all of it down.
204 // There's likely a CheckCastPP on one of the branches of the If, with obj as input.
205 // If the (LoadKlass .. (AddP obj ..)) is not cloned down, then split if transforms this to: (If .. (Bool(CmpP phi1 ..)))
206 // and the CheckCastPP to (CheckCastPP phi2). It's possible then that phi2 is transformed to a CheckCastPP
207 // (through PhiNode::Ideal) and that that CheckCastPP is replaced by another narrower CheckCastPP at the same control
208 // (through ConstraintCastNode::Identity). That could cause the CheckCastPP at the If to become top while (CmpP phi1)
209 // wouldn't constant fold because it's using a different data path. Cloning the whole subgraph down guarantees both the
210 // AddP and CheckCastPP have the same obj input after split if.
211 bool PhaseIdealLoop::clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2) {
212 if (n->Opcode() == Op_AddP && at_relevant_ctrl(n, blk1, blk2)) {
213 Node_List cmp_nodes;
214 uint old = C->unique();
215 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
216 Node* u1 = n->fast_out(i);
217 if (u1->Opcode() == Op_LoadNKlass && at_relevant_ctrl(u1, blk1, blk2)) {
218 for (DUIterator_Fast jmax, j = u1->fast_outs(jmax); j < jmax; j++) {
219 Node* u2 = u1->fast_out(j);
220 if (u2->Opcode() == Op_DecodeNKlass && at_relevant_ctrl(u2, blk1, blk2)) {
221 for (DUIterator k = u2->outs(); u2->has_out(k); k++) {
222 Node* u3 = u2->out(k);
223 if (at_relevant_ctrl(u3, blk1, blk2) && clone_cmp_down(u3, blk1, blk2)) {
224 --k;
225 }
226 }
227 for (DUIterator_Fast kmax, k = u2->fast_outs(kmax); k < kmax; k++) {
228 Node* u3 = u2->fast_out(k);
229 if (u3->_idx >= old) {
230 cmp_nodes.push(u3);
231 }
232 }
233 }
234 }
235 } else if (u1->Opcode() == Op_LoadKlass && at_relevant_ctrl(u1, blk1, blk2)) {
236 for (DUIterator j = u1->outs(); u1->has_out(j); j++) {
237 Node* u2 = u1->out(j);
238 if (at_relevant_ctrl(u2, blk1, blk2) && clone_cmp_down(u2, blk1, blk2)) {
239 --j;
240 }
241 }
242 for (DUIterator_Fast kmax, k = u1->fast_outs(kmax); k < kmax; k++) {
243 Node* u2 = u1->fast_out(k);
244 if (u2->_idx >= old) {
245 cmp_nodes.push(u2);
246 }
247 }
248 }
249 }
250
251 for (uint i = 0; i < cmp_nodes.size(); ++i) {
252 Node* cmp = cmp_nodes.at(i);
253 clone_loadklass_nodes_at_cmp_index(n, cmp, 1);
254 clone_loadklass_nodes_at_cmp_index(n, cmp, 2);
255 }
256 if (n->outcnt() == 0) {
257 assert(n->is_dead(), "");
258 return true;
259 }
260 }
261 return false;
262 }
263
264 bool PhaseIdealLoop::at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2) {
265 return ctrl_or_self(n) == blk1 || ctrl_or_self(n) == blk2;
266 }
267
268 void PhaseIdealLoop::clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i) {
269 Node* decode = cmp->in(i);
270 if (decode->Opcode() == Op_DecodeNKlass) {
271 Node* loadklass = decode->in(1);
272 if (loadklass->Opcode() == Op_LoadNKlass) {
273 Node* addp = loadklass->in(MemNode::Address);
274 if (addp == n) {
275 Node* ctrl = get_ctrl(cmp);
276 Node* decode_clone = decode->clone();
277 Node* loadklass_clone = loadklass->clone();
278 Node* addp_clone = addp->clone();
279 register_new_node(decode_clone, ctrl);
280 register_new_node(loadklass_clone, ctrl);
281 register_new_node(addp_clone, ctrl);
282 _igvn.replace_input_of(cmp, i, decode_clone);
283 _igvn.replace_input_of(decode_clone, 1, loadklass_clone);
284 _igvn.replace_input_of(loadklass_clone, MemNode::Address, addp_clone);
285 if (decode->outcnt() == 0) {
286 _igvn.remove_dead_node(decode);
287 }
288 }
289 }
290 } else {
291 Node* loadklass = cmp->in(i);
292 if (loadklass->Opcode() == Op_LoadKlass) {
293 Node* addp = loadklass->in(MemNode::Address);
294 if (addp == n) {
295 Node* ctrl = get_ctrl(cmp);
296 Node* loadklass_clone = loadklass->clone();
297 Node* addp_clone = addp->clone();
298 register_new_node(loadklass_clone, ctrl);
299 register_new_node(addp_clone, ctrl);
300 _igvn.replace_input_of(cmp, i, loadklass_clone);
301 _igvn.replace_input_of(loadklass_clone, MemNode::Address, addp_clone);
302 if (loadklass->outcnt() == 0) {
303 _igvn.remove_dead_node(loadklass);
304 }
305 }
306 }
307 }
308 }
309
310 bool PhaseIdealLoop::clone_cmp_down(Node* n, const Node* blk1, const Node* blk2) {
311 if( n->is_Cmp() ) {
312 assert(get_ctrl(n) == blk2 || get_ctrl(n) == blk1, "must be in block with IF");
313 // Check for simple Cmp/Bool/CMove which we can clone-up. Cmp/Bool/CMove
314 // sequence can have no other users and it must all reside in the split-if
315 // block. Non-simple Cmp/Bool/CMove sequences are 'cloned-down' below -
316 // private, per-use versions of the Cmp and Bool are made. These sink to
317 // the CMove block. If the CMove is in the split-if block, then in the
318 // next iteration this will become a simple Cmp/Bool/CMove set to clone-up.
319 Node *bol, *cmov;
320 if (!(n->outcnt() == 1 && n->unique_out()->is_Bool() &&
321 (bol = n->unique_out()->as_Bool()) &&
322 (at_relevant_ctrl(bol, blk1, blk2) &&
323 bol->outcnt() == 1 &&
324 bol->unique_out()->is_CMove() &&
325 (cmov = bol->unique_out()->as_CMove()) &&
326 at_relevant_ctrl(cmov, blk1, blk2)))) {
327
328 // Must clone down
329 #ifndef PRODUCT
330 if( PrintOpto && VerifyLoopOptimizations ) {
331 tty->print("Cloning down: ");
332 n->dump();
333 }
334 #endif
335 if (!n->is_FastLock()) {
336 // Clone down any block-local BoolNode uses of this CmpNode
337 for (DUIterator i = n->outs(); n->has_out(i); i++) {
338 Node* bol = n->out(i);
339 assert( bol->is_Bool(), "" );
340 if (bol->outcnt() == 1) {
341 Node* use = bol->unique_out();
342 if (use->Opcode() == Op_Opaque4) {
343 if (use->outcnt() == 1) {
344 Node* iff = use->unique_out();
345 assert(iff->is_If(), "unexpected node type");
346 Node *use_c = iff->in(0);
347 if (use_c == blk1 || use_c == blk2) {
348 continue;
349 }
350 }
351 } else {
352 // We might see an Opaque1 from a loop limit check here
353 assert(use->is_If() || use->is_CMove() || use->Opcode() == Op_Opaque1 || use->is_AllocateArray(), "unexpected node type");
354 Node *use_c = (use->is_If() || use->is_AllocateArray()) ? use->in(0) : get_ctrl(use);
355 if (use_c == blk1 || use_c == blk2) {
356 assert(use->is_CMove(), "unexpected node type");
357 continue;
358 }
359 }
360 }
361 if (at_relevant_ctrl(bol, blk1, blk2)) {
362 // Recursively sink any BoolNode
363 #ifndef PRODUCT
364 if( PrintOpto && VerifyLoopOptimizations ) {
365 tty->print("Cloning down: ");
366 bol->dump();
367 }
368 #endif
369 for (DUIterator j = bol->outs(); bol->has_out(j); j++) {
370 Node* u = bol->out(j);
371 // Uses are either IfNodes, CMoves or Opaque4
372 if (u->Opcode() == Op_Opaque4) {
373 assert(u->in(1) == bol, "bad input");
374 for (DUIterator_Last kmin, k = u->last_outs(kmin); k >= kmin; --k) {
375 Node* iff = u->last_out(k);
376 assert(iff->is_If() || iff->is_CMove(), "unexpected node type");
377 assert( iff->in(1) == u, "" );
378 // Get control block of either the CMove or the If input
379 Node *iff_ctrl = iff->is_If() ? iff->in(0) : get_ctrl(iff);
380 Node *x1 = bol->clone();
381 Node *x2 = u->clone();
382 register_new_node(x1, iff_ctrl);
383 register_new_node(x2, iff_ctrl);
384 _igvn.replace_input_of(x2, 1, x1);
385 _igvn.replace_input_of(iff, 1, x2);
386 }
387 _igvn.remove_dead_node(u);
388 --j;
389 } else {
390 // We might see an Opaque1 from a loop limit check here
391 assert(u->is_If() || u->is_CMove() || u->Opcode() == Op_Opaque1 || u->is_AllocateArray(), "unexpected node type");
392 assert(u->is_AllocateArray() || u->in(1) == bol, "");
393 assert(!u->is_AllocateArray() || u->in(AllocateNode::ValidLengthTest) == bol, "wrong input to AllocateArray");
394 // Get control block of either the CMove or the If input
395 Node *u_ctrl = (u->is_If() || u->is_AllocateArray()) ? u->in(0) : get_ctrl(u);
396 assert((u_ctrl != blk1 && u_ctrl != blk2) || u->is_CMove(), "won't converge");
397 Node *x = bol->clone();
398 register_new_node(x, u_ctrl);
399 _igvn.replace_input_of(u, u->is_AllocateArray() ? AllocateNode::ValidLengthTest : 1, x);
400 --j;
401 }
402 }
403 _igvn.remove_dead_node(bol);
404 --i;
405 }
406 }
407 }
408 // Clone down this CmpNode
409 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; --j) {
410 Node* use = n->last_out(j);
411 uint pos = 1;
412 if (n->is_FastLock()) {
413 pos = TypeFunc::Parms + 2;
414 assert(use->is_Lock(), "FastLock only used by LockNode");
415 }
416 assert(use->in(pos) == n, "" );
417 Node *x = n->clone();
418 register_new_node(x, ctrl_or_self(use));
419 _igvn.replace_input_of(use, pos, x);
420 }
421 _igvn.remove_dead_node(n);
422
423 return true;
424 }
425 }
426 return false;
427 }
428
429 //------------------------------register_new_node------------------------------
430 void PhaseIdealLoop::register_new_node( Node *n, Node *blk ) {
431 assert(!n->is_CFG(), "must be data node");
432 _igvn.register_new_node_with_optimizer(n);
433 set_ctrl(n, blk);
434 IdealLoopTree *loop = get_loop(blk);
435 if( !loop->_child )
436 loop->_body.push(n);
437 }
438
439 //------------------------------small_cache------------------------------------
440 struct small_cache : public Dict {
441
442 small_cache() : Dict( cmpkey, hashptr ) {}
443 Node *probe( Node *use_blk ) { return (Node*)((*this)[use_blk]); }
444 void lru_insert( Node *use_blk, Node *new_def ) { Insert(use_blk,new_def); }
445 };
446
447 //------------------------------spinup-----------------------------------------
448 // "Spin up" the dominator tree, starting at the use site and stopping when we
449 // find the post-dominating point.
450
451 // We must be at the merge point which post-dominates 'new_false' and
452 // 'new_true'. Figure out which edges into the RegionNode eventually lead up
453 // to false and which to true. Put in a PhiNode to merge values; plug in
454 // the appropriate false-arm or true-arm values. If some path leads to the
455 // original IF, then insert a Phi recursively.
456 Node *PhaseIdealLoop::spinup( Node *iff_dom, Node *new_false, Node *new_true, Node *use_blk, Node *def, small_cache *cache ) {
457 if (use_blk->is_top()) // Handle dead uses
458 return use_blk;
459 Node *prior_n = (Node*)((intptr_t)0xdeadbeef);
460 Node *n = use_blk; // Get path input
461 assert( use_blk != iff_dom, "" );
462 // Here's the "spinup" the dominator tree loop. Do a cache-check
463 // along the way, in case we've come this way before.
464 while( n != iff_dom ) { // Found post-dominating point?
465 prior_n = n;
466 n = idom(n); // Search higher
467 Node *s = cache->probe( prior_n ); // Check cache
468 if( s ) return s; // Cache hit!
469 }
470
471 Node *phi_post;
472 if( prior_n == new_false || prior_n == new_true ) {
473 phi_post = def->clone();
474 phi_post->set_req(0, prior_n );
475 register_new_node(phi_post, prior_n);
476 } else {
477 // This method handles both control uses (looking for Regions) or data
478 // uses (looking for Phis). If looking for a control use, then we need
479 // to insert a Region instead of a Phi; however Regions always exist
480 // previously (the hash_find_insert below would always hit) so we can
481 // return the existing Region.
482 if( def->is_CFG() ) {
483 phi_post = prior_n; // If looking for CFG, return prior
484 } else {
485 assert( def->is_Phi(), "" );
486 assert( prior_n->is_Region(), "must be a post-dominating merge point" );
487
488 // Need a Phi here
489 phi_post = PhiNode::make_blank(prior_n, def);
490 // Search for both true and false on all paths till find one.
491 for( uint i = 1; i < phi_post->req(); i++ ) // For all paths
492 phi_post->init_req( i, spinup( iff_dom, new_false, new_true, prior_n->in(i), def, cache ) );
493 Node *t = _igvn.hash_find_insert(phi_post);
494 if( t ) { // See if we already have this one
495 // phi_post will not be used, so kill it
496 _igvn.remove_dead_node(phi_post);
497 phi_post->destruct(&_igvn);
498 phi_post = t;
499 } else {
500 register_new_node( phi_post, prior_n );
501 }
502 }
503 }
504
505 // Update cache everywhere
506 prior_n = (Node*)((intptr_t)0xdeadbeef); // Reset IDOM walk
507 n = use_blk; // Get path input
508 // Spin-up the idom tree again, basically doing path-compression.
509 // Insert cache entries along the way, so that if we ever hit this
510 // point in the IDOM tree again we'll stop immediately on a cache hit.
511 while( n != iff_dom ) { // Found post-dominating point?
512 prior_n = n;
513 n = idom(n); // Search higher
514 cache->lru_insert( prior_n, phi_post ); // Fill cache
515 } // End of while not gone high enough
516
517 return phi_post;
518 }
519
520 //------------------------------find_use_block---------------------------------
521 // Find the block a USE is in. Normally USE's are in the same block as the
522 // using instruction. For Phi-USE's, the USE is in the predecessor block
523 // along the corresponding path.
524 Node *PhaseIdealLoop::find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ) {
525 // CFG uses are their own block
526 if( use->is_CFG() )
527 return use;
528
529 if( use->is_Phi() ) { // Phi uses in prior block
530 // Grab the first Phi use; there may be many.
531 // Each will be handled as a separate iteration of
532 // the "while( phi->outcnt() )" loop.
533 uint j;
534 for( j = 1; j < use->req(); j++ )
535 if( use->in(j) == def )
536 break;
537 assert( j < use->req(), "def should be among use's inputs" );
538 return use->in(0)->in(j);
539 }
540 // Normal (non-phi) use
541 Node *use_blk = get_ctrl(use);
542 // Some uses are directly attached to the old (and going away)
543 // false and true branches.
544 if( use_blk == old_false ) {
545 use_blk = new_false;
546 set_ctrl(use, new_false);
547 }
548 if( use_blk == old_true ) {
549 use_blk = new_true;
550 set_ctrl(use, new_true);
551 }
552
553 if (use_blk == NULL) { // He's dead, Jim
554 _igvn.replace_node(use, C->top());
555 }
556
557 return use_blk;
558 }
559
560 //------------------------------handle_use-------------------------------------
561 // Handle uses of the merge point. Basically, split-if makes the merge point
562 // go away so all uses of the merge point must go away as well. Most block
563 // local uses have already been split-up, through the merge point. Uses from
564 // far below the merge point can't always be split up (e.g., phi-uses are
565 // pinned) and it makes too much stuff live. Instead we use a path-based
566 // solution to move uses down.
567 //
568 // If the use is along the pre-split-CFG true branch, then the new use will
569 // be from the post-split-CFG true merge point. Vice-versa for the false
570 // path. Some uses will be along both paths; then we sink the use to the
571 // post-dominating location; we may need to insert a Phi there.
572 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 ) {
573
574 Node *use_blk = find_use_block(use,def,old_false,new_false,old_true,new_true);
575 if( !use_blk ) return; // He's dead, Jim
576
577 // Walk up the dominator tree until I hit either the old IfFalse, the old
578 // IfTrue or the old If. Insert Phis where needed.
579 Node *new_def = spinup( region_dom, new_false, new_true, use_blk, def, cache );
580
581 // Found where this USE goes. Re-point him.
582 uint i;
583 for( i = 0; i < use->req(); i++ )
584 if( use->in(i) == def )
585 break;
586 assert( i < use->req(), "def should be among use's inputs" );
587 _igvn.replace_input_of(use, i, new_def);
588 }
589
590 //------------------------------do_split_if------------------------------------
591 // Found an If getting its condition-code input from a Phi in the same block.
592 // Split thru the Region.
593 void PhaseIdealLoop::do_split_if(Node* iff, RegionNode** new_false_region, RegionNode** new_true_region) {
594 if (PrintOpto && VerifyLoopOptimizations) {
595 tty->print_cr("Split-if");
596 }
597 if (TraceLoopOpts) {
598 tty->print_cr("SplitIf");
599 }
600
601 C->set_major_progress();
602 RegionNode *region = iff->in(0)->as_Region();
603 Node *region_dom = idom(region);
604
605 // We are going to clone this test (and the control flow with it) up through
606 // the incoming merge point. We need to empty the current basic block.
607 // Clone any instructions which must be in this block up through the merge
608 // point.
609 DUIterator i, j;
610 bool progress = true;
611 while (progress) {
612 progress = false;
613 for (i = region->outs(); region->has_out(i); i++) {
614 Node* n = region->out(i);
615 if( n == region ) continue;
616 // The IF to be split is OK.
617 if( n == iff ) continue;
618 if( !n->is_Phi() ) { // Found pinned memory op or such
619 if (split_up(n, region, iff)) {
620 i = region->refresh_out_pos(i);
621 progress = true;
622 }
623 continue;
624 }
625 assert( n->in(0) == region, "" );
626
627 // Recursively split up all users of a Phi
628 for (j = n->outs(); n->has_out(j); j++) {
629 Node* m = n->out(j);
630 // If m is dead, throw it away, and declare progress
631 if (_nodes[m->_idx] == NULL) {
632 _igvn.remove_dead_node(m);
633 // fall through
634 }
635 else if (m != iff && split_up(m, region, iff)) {
636 // fall through
637 } else {
638 continue;
639 }
640 // Something unpredictable changed.
641 // Tell the iterators to refresh themselves, and rerun the loop.
642 i = region->refresh_out_pos(i);
643 j = region->refresh_out_pos(j);
644 progress = true;
645 }
646 }
647 }
648
649 // Now we have no instructions in the block containing the IF.
650 // Split the IF.
651 RegionNode *new_iff = split_thru_region(iff, region);
652
653 // Replace both uses of 'new_iff' with Regions merging True/False
654 // paths. This makes 'new_iff' go dead.
655 Node *old_false = NULL, *old_true = NULL;
656 RegionNode* new_false = NULL;
657 RegionNode* new_true = NULL;
658 for (DUIterator_Last j2min, j2 = iff->last_outs(j2min); j2 >= j2min; --j2) {
659 Node *ifp = iff->last_out(j2);
660 assert( ifp->Opcode() == Op_IfFalse || ifp->Opcode() == Op_IfTrue, "" );
661 ifp->set_req(0, new_iff);
662 RegionNode* ifpx = split_thru_region(ifp, region);
663
664 // Replace 'If' projection of a Region with a Region of
665 // 'If' projections.
666 ifpx->set_req(0, ifpx); // A TRUE RegionNode
667
668 // Setup dominator info
669 set_idom(ifpx, region_dom, dom_depth(region_dom) + 1);
670
671 // Check for splitting loop tails
672 if( get_loop(iff)->tail() == ifp )
673 get_loop(iff)->_tail = ifpx;
674
675 // Replace in the graph with lazy-update mechanism
676 new_iff->set_req(0, new_iff); // hook self so it does not go dead
677 lazy_replace(ifp, ifpx);
678 new_iff->set_req(0, region);
679
680 // Record bits for later xforms
681 if( ifp->Opcode() == Op_IfFalse ) {
682 old_false = ifp;
683 new_false = ifpx;
684 } else {
685 old_true = ifp;
686 new_true = ifpx;
687 }
688 }
689 _igvn.remove_dead_node(new_iff);
690 // Lazy replace IDOM info with the region's dominator
691 lazy_replace(iff, region_dom);
692 lazy_update(region, region_dom); // idom must be update before handle_uses
693 region->set_req(0, NULL); // Break the self-cycle. Required for lazy_update to work on region
694
695 // Now make the original merge point go dead, by handling all its uses.
696 small_cache region_cache;
697 // Preload some control flow in region-cache
698 region_cache.lru_insert( new_false, new_false );
699 region_cache.lru_insert( new_true , new_true );
700 // Now handle all uses of the splitting block
701 for (DUIterator k = region->outs(); region->has_out(k); k++) {
702 Node* phi = region->out(k);
703 if (!phi->in(0)) { // Dead phi? Remove it
704 _igvn.remove_dead_node(phi);
705 } else if (phi == region) { // Found the self-reference
706 continue; // No roll-back of DUIterator
707 } else if (phi->is_Phi()) { // Expected common case: Phi hanging off of Region
708 assert(phi->in(0) == region, "Inconsistent graph");
709 // Need a per-def cache. Phi represents a def, so make a cache
710 small_cache phi_cache;
711
712 // Inspect all Phi uses to make the Phi go dead
713 for (DUIterator_Last lmin, l = phi->last_outs(lmin); l >= lmin; --l) {
714 Node* use = phi->last_out(l);
715 // Compute the new DEF for this USE. New DEF depends on the path
716 // taken from the original DEF to the USE. The new DEF may be some
717 // collection of PHI's merging values from different paths. The Phis
718 // inserted depend only on the location of the USE. We use a
719 // 2-element cache to handle multiple uses from the same block.
720 handle_use(use, phi, &phi_cache, region_dom, new_false, new_true, old_false, old_true);
721 } // End of while phi has uses
722 // Remove the dead Phi
723 _igvn.remove_dead_node( phi );
724 } else {
725 assert(phi->in(0) == region, "Inconsistent graph");
726 // Random memory op guarded by Region. Compute new DEF for USE.
727 handle_use(phi, region, ®ion_cache, region_dom, new_false, new_true, old_false, old_true);
728 }
729 // Every path above deletes a use of the region, except for the region
730 // self-cycle (which is needed by handle_use calling find_use_block
731 // calling get_ctrl calling get_ctrl_no_update looking for dead
732 // regions). So roll back the DUIterator innards.
733 --k;
734 } // End of while merge point has phis
735
736 _igvn.remove_dead_node(region);
737
738 if (new_false_region != NULL) {
739 *new_false_region = new_false;
740 }
741 if (new_true_region != NULL) {
742 *new_true_region = new_true;
743 }
744
745 #ifndef PRODUCT
746 if( VerifyLoopOptimizations ) verify();
747 #endif
748 }