Tue, 09 Mar 2010 20:16:19 +0100
6919934: JSR 292 needs to support x86 C1
Summary: This implements JSR 292 support for C1 x86.
Reviewed-by: never, jrose, kvn
1 /*
2 * Copyright 1999-2010 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_split_if.cpp.incl"
29 //------------------------------split_thru_region------------------------------
30 // Split Node 'n' through merge point.
31 Node *PhaseIdealLoop::split_thru_region( Node *n, Node *region ) {
32 uint wins = 0;
33 assert( n->is_CFG(), "" );
34 assert( region->is_Region(), "" );
35 Node *r = new (C, region->req()) RegionNode( region->req() );
36 IdealLoopTree *loop = get_loop( n );
37 for( uint i = 1; i < region->req(); i++ ) {
38 Node *x = n->clone();
39 Node *in0 = n->in(0);
40 if( in0->in(0) == region ) x->set_req( 0, in0->in(i) );
41 for( uint j = 1; j < n->req(); j++ ) {
42 Node *in = n->in(j);
43 if( get_ctrl(in) == region )
44 x->set_req( j, in->in(i) );
45 }
46 _igvn.register_new_node_with_optimizer(x);
47 set_loop(x, loop);
48 set_idom(x, x->in(0), dom_depth(x->in(0))+1);
49 r->init_req(i, x);
50 }
52 // Record region
53 r->set_req(0,region); // Not a TRUE RegionNode
54 _igvn.register_new_node_with_optimizer(r);
55 set_loop(r, loop);
56 if( !loop->_child )
57 loop->_body.push(r);
58 return r;
59 }
61 //------------------------------split_up---------------------------------------
62 // Split block-local op up through the phis to empty the current block
63 bool PhaseIdealLoop::split_up( Node *n, Node *blk1, Node *blk2 ) {
64 if( n->is_CFG() ) {
65 assert( n->in(0) != blk1, "Lousy candidate for split-if" );
66 return false;
67 }
68 if( get_ctrl(n) != blk1 && get_ctrl(n) != blk2 )
69 return false; // Not block local
70 if( n->is_Phi() ) return false; // Local PHIs are expected
72 // Recursively split-up inputs
73 for (uint i = 1; i < n->req(); i++) {
74 if( split_up( n->in(i), blk1, blk2 ) ) {
75 // Got split recursively and self went dead?
76 if (n->outcnt() == 0)
77 _igvn.remove_dead_node(n);
78 return true;
79 }
80 }
82 // Check for needing to clone-up a compare. Can't do that, it forces
83 // another (nested) split-if transform. Instead, clone it "down".
84 if( n->is_Cmp() ) {
85 assert(get_ctrl(n) == blk2 || get_ctrl(n) == blk1, "must be in block with IF");
86 // Check for simple Cmp/Bool/CMove which we can clone-up. Cmp/Bool/CMove
87 // sequence can have no other users and it must all reside in the split-if
88 // block. Non-simple Cmp/Bool/CMove sequences are 'cloned-down' below -
89 // private, per-use versions of the Cmp and Bool are made. These sink to
90 // the CMove block. If the CMove is in the split-if block, then in the
91 // next iteration this will become a simple Cmp/Bool/CMove set to clone-up.
92 Node *bol, *cmov;
93 if( !(n->outcnt() == 1 && n->unique_out()->is_Bool() &&
94 (bol = n->unique_out()->as_Bool()) &&
95 (get_ctrl(bol) == blk1 ||
96 get_ctrl(bol) == blk2) &&
97 bol->outcnt() == 1 &&
98 bol->unique_out()->is_CMove() &&
99 (cmov = bol->unique_out()->as_CMove()) &&
100 (get_ctrl(cmov) == blk1 ||
101 get_ctrl(cmov) == blk2) ) ) {
103 // Must clone down
104 #ifndef PRODUCT
105 if( PrintOpto && VerifyLoopOptimizations ) {
106 tty->print("Cloning down: ");
107 n->dump();
108 }
109 #endif
110 // Clone down any block-local BoolNode uses of this CmpNode
111 for (DUIterator i = n->outs(); n->has_out(i); i++) {
112 Node* bol = n->out(i);
113 assert( bol->is_Bool(), "" );
114 if (bol->outcnt() == 1) {
115 Node* use = bol->unique_out();
116 Node *use_c = use->is_If() ? use->in(0) : get_ctrl(use);
117 if (use_c == blk1 || use_c == blk2) {
118 continue;
119 }
120 }
121 if (get_ctrl(bol) == blk1 || get_ctrl(bol) == blk2) {
122 // Recursively sink any BoolNode
123 #ifndef PRODUCT
124 if( PrintOpto && VerifyLoopOptimizations ) {
125 tty->print("Cloning down: ");
126 bol->dump();
127 }
128 #endif
129 for (DUIterator_Last jmin, j = bol->last_outs(jmin); j >= jmin; --j) {
130 // Uses are either IfNodes or CMoves
131 Node* iff = bol->last_out(j);
132 assert( iff->in(1) == bol, "" );
133 // Get control block of either the CMove or the If input
134 Node *iff_ctrl = iff->is_If() ? iff->in(0) : get_ctrl(iff);
135 Node *x = bol->clone();
136 register_new_node(x, iff_ctrl);
137 _igvn.hash_delete(iff);
138 iff->set_req(1, x);
139 _igvn._worklist.push(iff);
140 }
141 _igvn.remove_dead_node( bol );
142 --i;
143 }
144 }
145 // Clone down this CmpNode
146 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; --j) {
147 Node* bol = n->last_out(j);
148 assert( bol->in(1) == n, "" );
149 Node *x = n->clone();
150 register_new_node(x, get_ctrl(bol));
151 _igvn.hash_delete(bol);
152 bol->set_req(1, x);
153 _igvn._worklist.push(bol);
154 }
155 _igvn.remove_dead_node( n );
157 return true;
158 }
159 }
161 // See if splitting-up a Store. Any anti-dep loads must go up as
162 // well. An anti-dep load might be in the wrong block, because in
163 // this particular layout/schedule we ignored anti-deps and allow
164 // memory to be alive twice. This only works if we do the same
165 // operations on anti-dep loads as we do their killing stores.
166 if( n->is_Store() && n->in(MemNode::Memory)->in(0) == n->in(0) ) {
167 // Get store's memory slice
168 int alias_idx = C->get_alias_index(_igvn.type(n->in(MemNode::Address))->is_ptr());
170 // Get memory-phi anti-dep loads will be using
171 Node *memphi = n->in(MemNode::Memory);
172 assert( memphi->is_Phi(), "" );
173 // Hoist any anti-dep load to the splitting block;
174 // it will then "split-up".
175 for (DUIterator_Fast imax,i = memphi->fast_outs(imax); i < imax; i++) {
176 Node *load = memphi->fast_out(i);
177 if( load->is_Load() && alias_idx == C->get_alias_index(_igvn.type(load->in(MemNode::Address))->is_ptr()) )
178 set_ctrl(load,blk1);
179 }
180 }
182 // Found some other Node; must clone it up
183 #ifndef PRODUCT
184 if( PrintOpto && VerifyLoopOptimizations ) {
185 tty->print("Cloning up: ");
186 n->dump();
187 }
188 #endif
190 // ConvI2L may have type information on it which becomes invalid if
191 // it moves up in the graph so change any clones so widen the type
192 // to TypeLong::INT when pushing it up.
193 const Type* rtype = NULL;
194 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::INT) {
195 rtype = TypeLong::INT;
196 }
198 // Now actually split-up this guy. One copy per control path merging.
199 Node *phi = PhiNode::make_blank(blk1, n);
200 for( uint j = 1; j < blk1->req(); j++ ) {
201 Node *x = n->clone();
202 // Widen the type of the ConvI2L when pushing up.
203 if (rtype != NULL) x->as_Type()->set_type(rtype);
204 if( n->in(0) && n->in(0) == blk1 )
205 x->set_req( 0, blk1->in(j) );
206 for( uint i = 1; i < n->req(); i++ ) {
207 Node *m = n->in(i);
208 if( get_ctrl(m) == blk1 ) {
209 assert( m->in(0) == blk1, "" );
210 x->set_req( i, m->in(j) );
211 }
212 }
213 register_new_node( x, blk1->in(j) );
214 phi->init_req( j, x );
215 }
216 // Announce phi to optimizer
217 register_new_node(phi, blk1);
219 // Remove cloned-up value from optimizer; use phi instead
220 _igvn.hash_delete(n);
221 _igvn.subsume_node( n, phi );
223 // (There used to be a self-recursive call to split_up() here,
224 // but it is not needed. All necessary forward walking is done
225 // by do_split_if() below.)
227 return true;
228 }
230 //------------------------------register_new_node------------------------------
231 void PhaseIdealLoop::register_new_node( Node *n, Node *blk ) {
232 assert(!n->is_CFG(), "must be data node");
233 _igvn.register_new_node_with_optimizer(n);
234 set_ctrl(n, blk);
235 IdealLoopTree *loop = get_loop(blk);
236 if( !loop->_child )
237 loop->_body.push(n);
238 }
240 //------------------------------small_cache------------------------------------
241 struct small_cache : public Dict {
243 small_cache() : Dict( cmpkey, hashptr ) {}
244 Node *probe( Node *use_blk ) { return (Node*)((*this)[use_blk]); }
245 void lru_insert( Node *use_blk, Node *new_def ) { Insert(use_blk,new_def); }
246 };
248 //------------------------------spinup-----------------------------------------
249 // "Spin up" the dominator tree, starting at the use site and stopping when we
250 // find the post-dominating point.
252 // We must be at the merge point which post-dominates 'new_false' and
253 // 'new_true'. Figure out which edges into the RegionNode eventually lead up
254 // to false and which to true. Put in a PhiNode to merge values; plug in
255 // the appropriate false-arm or true-arm values. If some path leads to the
256 // original IF, then insert a Phi recursively.
257 Node *PhaseIdealLoop::spinup( Node *iff_dom, Node *new_false, Node *new_true, Node *use_blk, Node *def, small_cache *cache ) {
258 if (use_blk->is_top()) // Handle dead uses
259 return use_blk;
260 Node *prior_n = (Node*)0xdeadbeef;
261 Node *n = use_blk; // Get path input
262 assert( use_blk != iff_dom, "" );
263 // Here's the "spinup" the dominator tree loop. Do a cache-check
264 // along the way, in case we've come this way before.
265 while( n != iff_dom ) { // Found post-dominating point?
266 prior_n = n;
267 n = idom(n); // Search higher
268 Node *s = cache->probe( prior_n ); // Check cache
269 if( s ) return s; // Cache hit!
270 }
272 Node *phi_post;
273 if( prior_n == new_false || prior_n == new_true ) {
274 phi_post = def->clone();
275 phi_post->set_req(0, prior_n );
276 register_new_node(phi_post, prior_n);
277 } else {
278 // This method handles both control uses (looking for Regions) or data
279 // uses (looking for Phis). If looking for a control use, then we need
280 // to insert a Region instead of a Phi; however Regions always exist
281 // previously (the hash_find_insert below would always hit) so we can
282 // return the existing Region.
283 if( def->is_CFG() ) {
284 phi_post = prior_n; // If looking for CFG, return prior
285 } else {
286 assert( def->is_Phi(), "" );
287 assert( prior_n->is_Region(), "must be a post-dominating merge point" );
289 // Need a Phi here
290 phi_post = PhiNode::make_blank(prior_n, def);
291 // Search for both true and false on all paths till find one.
292 for( uint i = 1; i < phi_post->req(); i++ ) // For all paths
293 phi_post->init_req( i, spinup( iff_dom, new_false, new_true, prior_n->in(i), def, cache ) );
294 Node *t = _igvn.hash_find_insert(phi_post);
295 if( t ) { // See if we already have this one
296 // phi_post will not be used, so kill it
297 _igvn.remove_dead_node(phi_post);
298 phi_post->destruct();
299 phi_post = t;
300 } else {
301 register_new_node( phi_post, prior_n );
302 }
303 }
304 }
306 // Update cache everywhere
307 prior_n = (Node*)0xdeadbeef; // Reset IDOM walk
308 n = use_blk; // Get path input
309 // Spin-up the idom tree again, basically doing path-compression.
310 // Insert cache entries along the way, so that if we ever hit this
311 // point in the IDOM tree again we'll stop immediately on a cache hit.
312 while( n != iff_dom ) { // Found post-dominating point?
313 prior_n = n;
314 n = idom(n); // Search higher
315 cache->lru_insert( prior_n, phi_post ); // Fill cache
316 } // End of while not gone high enough
318 return phi_post;
319 }
321 //------------------------------find_use_block---------------------------------
322 // Find the block a USE is in. Normally USE's are in the same block as the
323 // using instruction. For Phi-USE's, the USE is in the predecessor block
324 // along the corresponding path.
325 Node *PhaseIdealLoop::find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ) {
326 // CFG uses are their own block
327 if( use->is_CFG() )
328 return use;
330 if( use->is_Phi() ) { // Phi uses in prior block
331 // Grab the first Phi use; there may be many.
332 // Each will be handled as a separate iteration of
333 // the "while( phi->outcnt() )" loop.
334 uint j;
335 for( j = 1; j < use->req(); j++ )
336 if( use->in(j) == def )
337 break;
338 assert( j < use->req(), "def should be among use's inputs" );
339 return use->in(0)->in(j);
340 }
341 // Normal (non-phi) use
342 Node *use_blk = get_ctrl(use);
343 // Some uses are directly attached to the old (and going away)
344 // false and true branches.
345 if( use_blk == old_false ) {
346 use_blk = new_false;
347 set_ctrl(use, new_false);
348 }
349 if( use_blk == old_true ) {
350 use_blk = new_true;
351 set_ctrl(use, new_true);
352 }
354 if (use_blk == NULL) { // He's dead, Jim
355 _igvn.hash_delete(use);
356 _igvn.subsume_node(use, C->top());
357 }
359 return use_blk;
360 }
362 //------------------------------handle_use-------------------------------------
363 // Handle uses of the merge point. Basically, split-if makes the merge point
364 // go away so all uses of the merge point must go away as well. Most block
365 // local uses have already been split-up, through the merge point. Uses from
366 // far below the merge point can't always be split up (e.g., phi-uses are
367 // pinned) and it makes too much stuff live. Instead we use a path-based
368 // solution to move uses down.
369 //
370 // If the use is along the pre-split-CFG true branch, then the new use will
371 // be from the post-split-CFG true merge point. Vice-versa for the false
372 // path. Some uses will be along both paths; then we sink the use to the
373 // post-dominating location; we may need to insert a Phi there.
374 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 ) {
376 Node *use_blk = find_use_block(use,def,old_false,new_false,old_true,new_true);
377 if( !use_blk ) return; // He's dead, Jim
379 // Walk up the dominator tree until I hit either the old IfFalse, the old
380 // IfTrue or the old If. Insert Phis where needed.
381 Node *new_def = spinup( region_dom, new_false, new_true, use_blk, def, cache );
383 // Found where this USE goes. Re-point him.
384 uint i;
385 for( i = 0; i < use->req(); i++ )
386 if( use->in(i) == def )
387 break;
388 assert( i < use->req(), "def should be among use's inputs" );
389 _igvn.hash_delete(use);
390 use->set_req(i, new_def);
391 _igvn._worklist.push(use);
392 }
394 //------------------------------do_split_if------------------------------------
395 // Found an If getting its condition-code input from a Phi in the same block.
396 // Split thru the Region.
397 void PhaseIdealLoop::do_split_if( Node *iff ) {
398 #ifndef PRODUCT
399 if( PrintOpto && VerifyLoopOptimizations )
400 tty->print_cr("Split-if");
401 #endif
402 C->set_major_progress();
403 Node *region = iff->in(0);
404 Node *region_dom = idom(region);
406 // We are going to clone this test (and the control flow with it) up through
407 // the incoming merge point. We need to empty the current basic block.
408 // Clone any instructions which must be in this block up through the merge
409 // point.
410 DUIterator i, j;
411 bool progress = true;
412 while (progress) {
413 progress = false;
414 for (i = region->outs(); region->has_out(i); i++) {
415 Node* n = region->out(i);
416 if( n == region ) continue;
417 // The IF to be split is OK.
418 if( n == iff ) continue;
419 if( !n->is_Phi() ) { // Found pinned memory op or such
420 if (split_up(n, region, iff)) {
421 i = region->refresh_out_pos(i);
422 progress = true;
423 }
424 continue;
425 }
426 assert( n->in(0) == region, "" );
428 // Recursively split up all users of a Phi
429 for (j = n->outs(); n->has_out(j); j++) {
430 Node* m = n->out(j);
431 // If m is dead, throw it away, and declare progress
432 if (_nodes[m->_idx] == NULL) {
433 _igvn.remove_dead_node(m);
434 // fall through
435 }
436 else if (m != iff && split_up(m, region, iff)) {
437 // fall through
438 } else {
439 continue;
440 }
441 // Something unpredictable changed.
442 // Tell the iterators to refresh themselves, and rerun the loop.
443 i = region->refresh_out_pos(i);
444 j = region->refresh_out_pos(j);
445 progress = true;
446 }
447 }
448 }
450 // Now we have no instructions in the block containing the IF.
451 // Split the IF.
452 Node *new_iff = split_thru_region( iff, region );
454 // Replace both uses of 'new_iff' with Regions merging True/False
455 // paths. This makes 'new_iff' go dead.
456 Node *old_false, *old_true;
457 Node *new_false, *new_true;
458 for (DUIterator_Last j2min, j2 = iff->last_outs(j2min); j2 >= j2min; --j2) {
459 Node *ifp = iff->last_out(j2);
460 assert( ifp->Opcode() == Op_IfFalse || ifp->Opcode() == Op_IfTrue, "" );
461 ifp->set_req(0, new_iff);
462 Node *ifpx = split_thru_region( ifp, region );
464 // Replace 'If' projection of a Region with a Region of
465 // 'If' projections.
466 ifpx->set_req(0, ifpx); // A TRUE RegionNode
468 // Setup dominator info
469 set_idom(ifpx, region_dom, dom_depth(region_dom) + 1);
471 // Check for splitting loop tails
472 if( get_loop(iff)->tail() == ifp )
473 get_loop(iff)->_tail = ifpx;
475 // Replace in the graph with lazy-update mechanism
476 new_iff->set_req(0, new_iff); // hook self so it does not go dead
477 lazy_replace_proj( ifp, ifpx );
478 new_iff->set_req(0, region);
480 // Record bits for later xforms
481 if( ifp->Opcode() == Op_IfFalse ) {
482 old_false = ifp;
483 new_false = ifpx;
484 } else {
485 old_true = ifp;
486 new_true = ifpx;
487 }
488 }
489 _igvn.remove_dead_node(new_iff);
490 // Lazy replace IDOM info with the region's dominator
491 lazy_replace( iff, region_dom );
493 // Now make the original merge point go dead, by handling all its uses.
494 small_cache region_cache;
495 // Preload some control flow in region-cache
496 region_cache.lru_insert( new_false, new_false );
497 region_cache.lru_insert( new_true , new_true );
498 // Now handle all uses of the splitting block
499 for (DUIterator_Last kmin, k = region->last_outs(kmin); k >= kmin; --k) {
500 Node* phi = region->last_out(k);
501 if( !phi->in(0) ) { // Dead phi? Remove it
502 _igvn.remove_dead_node(phi);
503 continue;
504 }
505 assert( phi->in(0) == region, "" );
506 if( phi == region ) { // Found the self-reference
507 phi->set_req(0, NULL);
508 continue; // Break the self-cycle
509 }
510 // Expected common case: Phi hanging off of Region
511 if( phi->is_Phi() ) {
512 // Need a per-def cache. Phi represents a def, so make a cache
513 small_cache phi_cache;
515 // Inspect all Phi uses to make the Phi go dead
516 for (DUIterator_Last lmin, l = phi->last_outs(lmin); l >= lmin; --l) {
517 Node* use = phi->last_out(l);
518 // Compute the new DEF for this USE. New DEF depends on the path
519 // taken from the original DEF to the USE. The new DEF may be some
520 // collection of PHI's merging values from different paths. The Phis
521 // inserted depend only on the location of the USE. We use a
522 // 2-element cache to handle multiple uses from the same block.
523 handle_use( use, phi, &phi_cache, region_dom, new_false, new_true, old_false, old_true );
524 } // End of while phi has uses
526 // Because handle_use might relocate region->_out,
527 // we must refresh the iterator.
528 k = region->last_outs(kmin);
530 // Remove the dead Phi
531 _igvn.remove_dead_node( phi );
533 } else {
534 // Random memory op guarded by Region. Compute new DEF for USE.
535 handle_use( phi, region, ®ion_cache, region_dom, new_false, new_true, old_false, old_true );
536 }
538 } // End of while merge point has phis
540 // Any leftover bits in the splitting block must not have depended on local
541 // Phi inputs (these have already been split-up). Hence it's safe to hoist
542 // these guys to the dominating point.
543 lazy_replace( region, region_dom );
544 #ifndef PRODUCT
545 if( VerifyLoopOptimizations ) verify();
546 #endif
547 }