Mercurial > hg > truffle
annotate src/share/vm/opto/postaloc.cpp @ 2616:3558ca7088c0
FrameState and Graphviz changes:
* removed popx, pushx methods from GraphBuilder
* FrameState subclass of Value
* added String shortName() to Node
* added GraphvizPrinter option to use short names
* small hack in GraphvizPrinter: omit FrameState->Local connections
* added GraalGraphviz to implicit classpatch (read from GRAAL env var)
author | Lukas Stadler <lukas.stadler@jku.at> |
---|---|
date | Mon, 09 May 2011 17:00:25 +0200 |
parents | 2f644f85485d |
children | 8f47d8870d9a |
rev | line source |
---|---|
0 | 1 /* |
1972 | 2 * Copyright (c) 1998, 2010, Oracle and/or its affiliates. All rights reserved. |
0 | 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 * | |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
948
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
948
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
948
diff
changeset
|
21 * questions. |
0 | 22 * |
23 */ | |
24 | |
1972 | 25 #include "precompiled.hpp" |
26 #include "memory/allocation.inline.hpp" | |
27 #include "opto/chaitin.hpp" | |
28 #include "opto/machnode.hpp" | |
0 | 29 |
30 // see if this register kind does not requires two registers | |
31 static bool is_single_register(uint x) { | |
32 #ifdef _LP64 | |
33 return (x != Op_RegD && x != Op_RegL && x != Op_RegP); | |
34 #else | |
35 return (x != Op_RegD && x != Op_RegL); | |
36 #endif | |
37 } | |
38 | |
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
39 //---------------------------may_be_copy_of_callee----------------------------- |
0 | 40 // Check to see if we can possibly be a copy of a callee-save value. |
41 bool PhaseChaitin::may_be_copy_of_callee( Node *def ) const { | |
42 // Short circuit if there are no callee save registers | |
43 if (_matcher.number_of_saved_registers() == 0) return false; | |
44 | |
45 // Expect only a spill-down and reload on exit for callee-save spills. | |
46 // Chains of copies cannot be deep. | |
47 // 5008997 - This is wishful thinking. Register allocator seems to | |
48 // be splitting live ranges for callee save registers to such | |
49 // an extent that in large methods the chains can be very long | |
50 // (50+). The conservative answer is to return true if we don't | |
605 | 51 // know as this prevents optimizations from occurring. |
0 | 52 |
53 const int limit = 60; | |
54 int i; | |
55 for( i=0; i < limit; i++ ) { | |
56 if( def->is_Proj() && def->in(0)->is_Start() && | |
57 _matcher.is_save_on_entry(lrgs(n2lidx(def)).reg()) ) | |
58 return true; // Direct use of callee-save proj | |
59 if( def->is_Copy() ) // Copies carry value through | |
60 def = def->in(def->is_Copy()); | |
61 else if( def->is_Phi() ) // Phis can merge it from any direction | |
62 def = def->in(1); | |
63 else | |
64 break; | |
65 guarantee(def != NULL, "must not resurrect dead copy"); | |
66 } | |
67 // If we reached the end and didn't find a callee save proj | |
68 // then this may be a callee save proj so we return true | |
69 // as the conservative answer. If we didn't reach then end | |
70 // we must have discovered that it was not a callee save | |
71 // else we would have returned. | |
72 return i == limit; | |
73 } | |
74 | |
75 | |
76 | |
77 //------------------------------yank_if_dead----------------------------------- | |
78 // Removed an edge from 'old'. Yank if dead. Return adjustment counts to | |
79 // iterators in the current block. | |
80 int PhaseChaitin::yank_if_dead( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) { | |
81 int blk_adjust=0; | |
82 while (old->outcnt() == 0 && old != C->top()) { | |
83 Block *oldb = _cfg._bbs[old->_idx]; | |
84 oldb->find_remove(old); | |
85 // Count 1 if deleting an instruction from the current block | |
86 if( oldb == current_block ) blk_adjust++; | |
87 _cfg._bbs.map(old->_idx,NULL); | |
88 OptoReg::Name old_reg = lrgs(n2lidx(old)).reg(); | |
89 if( regnd && (*regnd)[old_reg]==old ) { // Instruction is currently available? | |
90 value->map(old_reg,NULL); // Yank from value/regnd maps | |
91 regnd->map(old_reg,NULL); // This register's value is now unknown | |
92 } | |
923 | 93 assert(old->req() <= 2, "can't handle more inputs"); |
0 | 94 Node *tmp = old->req() > 1 ? old->in(1) : NULL; |
95 old->disconnect_inputs(NULL); | |
96 if( !tmp ) break; | |
97 old = tmp; | |
98 } | |
99 return blk_adjust; | |
100 } | |
101 | |
102 //------------------------------use_prior_register----------------------------- | |
103 // Use the prior value instead of the current value, in an effort to make | |
104 // the current value go dead. Return block iterator adjustment, in case | |
105 // we yank some instructions from this block. | |
106 int PhaseChaitin::use_prior_register( Node *n, uint idx, Node *def, Block *current_block, Node_List &value, Node_List ®nd ) { | |
107 // No effect? | |
108 if( def == n->in(idx) ) return 0; | |
109 // Def is currently dead and can be removed? Do not resurrect | |
110 if( def->outcnt() == 0 ) return 0; | |
111 | |
112 // Not every pair of physical registers are assignment compatible, | |
113 // e.g. on sparc floating point registers are not assignable to integer | |
114 // registers. | |
115 const LRG &def_lrg = lrgs(n2lidx(def)); | |
116 OptoReg::Name def_reg = def_lrg.reg(); | |
117 const RegMask &use_mask = n->in_RegMask(idx); | |
118 bool can_use = ( RegMask::can_represent(def_reg) ? (use_mask.Member(def_reg) != 0) | |
119 : (use_mask.is_AllStack() != 0)); | |
120 // Check for a copy to or from a misaligned pair. | |
121 can_use = can_use && !use_mask.is_misaligned_Pair() && !def_lrg.mask().is_misaligned_Pair(); | |
122 | |
123 if (!can_use) | |
124 return 0; | |
125 | |
126 // Capture the old def in case it goes dead... | |
127 Node *old = n->in(idx); | |
128 | |
129 // Save-on-call copies can only be elided if the entire copy chain can go | |
130 // away, lest we get the same callee-save value alive in 2 locations at | |
131 // once. We check for the obvious trivial case here. Although it can | |
132 // sometimes be elided with cooperation outside our scope, here we will just | |
133 // miss the opportunity. :-( | |
134 if( may_be_copy_of_callee(def) ) { | |
135 if( old->outcnt() > 1 ) return 0; // We're the not last user | |
136 int idx = old->is_Copy(); | |
137 assert( idx, "chain of copies being removed" ); | |
138 Node *old2 = old->in(idx); // Chain of copies | |
139 if( old2->outcnt() > 1 ) return 0; // old is not the last user | |
140 int idx2 = old2->is_Copy(); | |
141 if( !idx2 ) return 0; // Not a chain of 2 copies | |
142 if( def != old2->in(idx2) ) return 0; // Chain of exactly 2 copies | |
143 } | |
144 | |
145 // Use the new def | |
146 n->set_req(idx,def); | |
147 _post_alloc++; | |
148 | |
149 // Is old def now dead? We successfully yanked a copy? | |
150 return yank_if_dead(old,current_block,&value,®nd); | |
151 } | |
152 | |
153 | |
154 //------------------------------skip_copies------------------------------------ | |
155 // Skip through any number of copies (that don't mod oop-i-ness) | |
156 Node *PhaseChaitin::skip_copies( Node *c ) { | |
157 int idx = c->is_Copy(); | |
158 uint is_oop = lrgs(n2lidx(c))._is_oop; | |
159 while (idx != 0) { | |
160 guarantee(c->in(idx) != NULL, "must not resurrect dead copy"); | |
161 if (lrgs(n2lidx(c->in(idx)))._is_oop != is_oop) | |
162 break; // casting copy, not the same value | |
163 c = c->in(idx); | |
164 idx = c->is_Copy(); | |
165 } | |
166 return c; | |
167 } | |
168 | |
169 //------------------------------elide_copy------------------------------------- | |
170 // Remove (bypass) copies along Node n, edge k. | |
171 int PhaseChaitin::elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List ®nd, bool can_change_regs ) { | |
172 int blk_adjust = 0; | |
173 | |
174 uint nk_idx = n2lidx(n->in(k)); | |
175 OptoReg::Name nk_reg = lrgs(nk_idx ).reg(); | |
176 | |
177 // Remove obvious same-register copies | |
178 Node *x = n->in(k); | |
179 int idx; | |
180 while( (idx=x->is_Copy()) != 0 ) { | |
181 Node *copy = x->in(idx); | |
182 guarantee(copy != NULL, "must not resurrect dead copy"); | |
183 if( lrgs(n2lidx(copy)).reg() != nk_reg ) break; | |
184 blk_adjust += use_prior_register(n,k,copy,current_block,value,regnd); | |
185 if( n->in(k) != copy ) break; // Failed for some cutout? | |
186 x = copy; // Progress, try again | |
187 } | |
188 | |
189 // Phis and 2-address instructions cannot change registers so easily - their | |
190 // outputs must match their input. | |
191 if( !can_change_regs ) | |
192 return blk_adjust; // Only check stupid copies! | |
193 | |
194 // Loop backedges won't have a value-mapping yet | |
195 if( &value == NULL ) return blk_adjust; | |
196 | |
197 // Skip through all copies to the _value_ being used. Do not change from | |
198 // int to pointer. This attempts to jump through a chain of copies, where | |
199 // intermediate copies might be illegal, i.e., value is stored down to stack | |
200 // then reloaded BUT survives in a register the whole way. | |
201 Node *val = skip_copies(n->in(k)); | |
202 | |
2008 | 203 if (val == x && nk_idx != 0 && |
204 regnd[nk_reg] != NULL && regnd[nk_reg] != x && | |
205 n2lidx(x) == n2lidx(regnd[nk_reg])) { | |
206 // When rematerialzing nodes and stretching lifetimes, the | |
207 // allocator will reuse the original def for multidef LRG instead | |
208 // of the current reaching def because it can't know it's safe to | |
209 // do so. After allocation completes if they are in the same LRG | |
210 // then it should use the current reaching def instead. | |
211 n->set_req(k, regnd[nk_reg]); | |
212 blk_adjust += yank_if_dead(val, current_block, &value, ®nd); | |
213 val = skip_copies(n->in(k)); | |
214 } | |
215 | |
0 | 216 if( val == x ) return blk_adjust; // No progress? |
217 | |
218 bool single = is_single_register(val->ideal_reg()); | |
219 uint val_idx = n2lidx(val); | |
220 OptoReg::Name val_reg = lrgs(val_idx).reg(); | |
221 | |
222 // See if it happens to already be in the correct register! | |
223 // (either Phi's direct register, or the common case of the name | |
224 // never-clobbered original-def register) | |
225 if( value[val_reg] == val && | |
226 // Doubles check both halves | |
227 ( single || value[val_reg-1] == val ) ) { | |
228 blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd); | |
229 if( n->in(k) == regnd[val_reg] ) // Success! Quit trying | |
230 return blk_adjust; | |
231 } | |
232 | |
233 // See if we can skip the copy by changing registers. Don't change from | |
234 // using a register to using the stack unless we know we can remove a | |
235 // copy-load. Otherwise we might end up making a pile of Intel cisc-spill | |
236 // ops reading from memory instead of just loading once and using the | |
237 // register. | |
238 | |
239 // Also handle duplicate copies here. | |
240 const Type *t = val->is_Con() ? val->bottom_type() : NULL; | |
241 | |
242 // Scan all registers to see if this value is around already | |
243 for( uint reg = 0; reg < (uint)_max_reg; reg++ ) { | |
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
244 if (reg == (uint)nk_reg) { |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
245 // Found ourselves so check if there is only one user of this |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
246 // copy and keep on searching for a better copy if so. |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
247 bool ignore_self = true; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
248 x = n->in(k); |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
249 DUIterator_Fast imax, i = x->fast_outs(imax); |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
250 Node* first = x->fast_out(i); i++; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
251 while (i < imax && ignore_self) { |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
252 Node* use = x->fast_out(i); i++; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
253 if (use != first) ignore_self = false; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
254 } |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
255 if (ignore_self) continue; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
256 } |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
257 |
0 | 258 Node *vv = value[reg]; |
259 if( !single ) { // Doubles check for aligned-adjacent pair | |
260 if( (reg&1)==0 ) continue; // Wrong half of a pair | |
261 if( vv != value[reg-1] ) continue; // Not a complete pair | |
262 } | |
263 if( vv == val || // Got a direct hit? | |
264 (t && vv && vv->bottom_type() == t && vv->is_Mach() && | |
265 vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant? | |
266 assert( !n->is_Phi(), "cannot change registers at a Phi so easily" ); | |
267 if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR | |
268 OptoReg::is_reg(reg) || // turning into a register use OR | |
269 regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use | |
270 blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd); | |
271 if( n->in(k) == regnd[reg] ) // Success! Quit trying | |
272 return blk_adjust; | |
273 } // End of if not degrading to a stack | |
274 } // End of if found value in another register | |
275 } // End of scan all machine registers | |
276 return blk_adjust; | |
277 } | |
278 | |
279 | |
280 // | |
281 // Check if nreg already contains the constant value val. Normal copy | |
282 // elimination doesn't doesn't work on constants because multiple | |
283 // nodes can represent the same constant so the type and rule of the | |
284 // MachNode must be checked to ensure equivalence. | |
285 // | |
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
286 bool PhaseChaitin::eliminate_copy_of_constant(Node* val, Node* n, |
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
287 Block *current_block, |
0 | 288 Node_List& value, Node_List& regnd, |
289 OptoReg::Name nreg, OptoReg::Name nreg2) { | |
290 if (value[nreg] != val && val->is_Con() && | |
291 value[nreg] != NULL && value[nreg]->is_Con() && | |
292 (nreg2 == OptoReg::Bad || value[nreg] == value[nreg2]) && | |
293 value[nreg]->bottom_type() == val->bottom_type() && | |
294 value[nreg]->as_Mach()->rule() == val->as_Mach()->rule()) { | |
295 // This code assumes that two MachNodes representing constants | |
296 // which have the same rule and the same bottom type will produce | |
297 // identical effects into a register. This seems like it must be | |
298 // objectively true unless there are hidden inputs to the nodes | |
299 // but if that were to change this code would need to updated. | |
300 // Since they are equivalent the second one if redundant and can | |
301 // be removed. | |
302 // | |
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
303 // n will be replaced with the old value but n might have |
0 | 304 // kills projections associated with it so remove them now so that |
605 | 305 // yank_if_dead will be able to eliminate the copy once the uses |
0 | 306 // have been transferred to the old[value]. |
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
307 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { |
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
308 Node* use = n->fast_out(i); |
0 | 309 if (use->is_Proj() && use->outcnt() == 0) { |
310 // Kill projections have no users and one input | |
311 use->set_req(0, C->top()); | |
312 yank_if_dead(use, current_block, &value, ®nd); | |
313 --i; --imax; | |
314 } | |
315 } | |
316 _post_alloc++; | |
317 return true; | |
318 } | |
319 return false; | |
320 } | |
321 | |
322 | |
323 //------------------------------post_allocate_copy_removal--------------------- | |
324 // Post-Allocation peephole copy removal. We do this in 1 pass over the | |
325 // basic blocks. We maintain a mapping of registers to Nodes (an array of | |
326 // Nodes indexed by machine register or stack slot number). NULL means that a | |
327 // register is not mapped to any Node. We can (want to have!) have several | |
328 // registers map to the same Node. We walk forward over the instructions | |
329 // updating the mapping as we go. At merge points we force a NULL if we have | |
330 // to merge 2 different Nodes into the same register. Phi functions will give | |
331 // us a new Node if there is a proper value merging. Since the blocks are | |
332 // arranged in some RPO, we will visit all parent blocks before visiting any | |
333 // successor blocks (except at loops). | |
334 // | |
335 // If we find a Copy we look to see if the Copy's source register is a stack | |
336 // slot and that value has already been loaded into some machine register; if | |
337 // so we use machine register directly. This turns a Load into a reg-reg | |
338 // Move. We also look for reloads of identical constants. | |
339 // | |
340 // When we see a use from a reg-reg Copy, we will attempt to use the copy's | |
341 // source directly and make the copy go dead. | |
342 void PhaseChaitin::post_allocate_copy_removal() { | |
343 NOT_PRODUCT( Compile::TracePhase t3("postAllocCopyRemoval", &_t_postAllocCopyRemoval, TimeCompiler); ) | |
344 ResourceMark rm; | |
345 | |
346 // Need a mapping from basic block Node_Lists. We need a Node_List to | |
347 // map from register number to value-producing Node. | |
348 Node_List **blk2value = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); | |
349 memset( blk2value, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); | |
350 // Need a mapping from basic block Node_Lists. We need a Node_List to | |
351 // map from register number to register-defining Node. | |
352 Node_List **blk2regnd = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); | |
353 memset( blk2regnd, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); | |
354 | |
355 // We keep unused Node_Lists on a free_list to avoid wasting | |
356 // memory. | |
357 GrowableArray<Node_List*> free_list = GrowableArray<Node_List*>(16); | |
358 | |
359 // For all blocks | |
360 for( uint i = 0; i < _cfg._num_blocks; i++ ) { | |
361 uint j; | |
362 Block *b = _cfg._blocks[i]; | |
363 | |
364 // Count of Phis in block | |
365 uint phi_dex; | |
366 for( phi_dex = 1; phi_dex < b->_nodes.size(); phi_dex++ ) { | |
367 Node *phi = b->_nodes[phi_dex]; | |
368 if( !phi->is_Phi() ) | |
369 break; | |
370 } | |
371 | |
372 // If any predecessor has not been visited, we do not know the state | |
373 // of registers at the start. Check for this, while updating copies | |
374 // along Phi input edges | |
375 bool missing_some_inputs = false; | |
376 Block *freed = NULL; | |
377 for( j = 1; j < b->num_preds(); j++ ) { | |
378 Block *pb = _cfg._bbs[b->pred(j)->_idx]; | |
379 // Remove copies along phi edges | |
380 for( uint k=1; k<phi_dex; k++ ) | |
381 elide_copy( b->_nodes[k], j, b, *blk2value[pb->_pre_order], *blk2regnd[pb->_pre_order], false ); | |
382 if( blk2value[pb->_pre_order] ) { // Have a mapping on this edge? | |
383 // See if this predecessor's mappings have been used by everybody | |
384 // who wants them. If so, free 'em. | |
385 uint k; | |
386 for( k=0; k<pb->_num_succs; k++ ) { | |
387 Block *pbsucc = pb->_succs[k]; | |
388 if( !blk2value[pbsucc->_pre_order] && pbsucc != b ) | |
389 break; // Found a future user | |
390 } | |
391 if( k >= pb->_num_succs ) { // No more uses, free! | |
392 freed = pb; // Record last block freed | |
393 free_list.push(blk2value[pb->_pre_order]); | |
394 free_list.push(blk2regnd[pb->_pre_order]); | |
395 } | |
396 } else { // This block has unvisited (loopback) inputs | |
397 missing_some_inputs = true; | |
398 } | |
399 } | |
400 | |
401 | |
402 // Extract Node_List mappings. If 'freed' is non-zero, we just popped | |
403 // 'freed's blocks off the list | |
404 Node_List ®nd = *(free_list.is_empty() ? new Node_List() : free_list.pop()); | |
405 Node_List &value = *(free_list.is_empty() ? new Node_List() : free_list.pop()); | |
406 assert( !freed || blk2value[freed->_pre_order] == &value, "" ); | |
407 value.map(_max_reg,NULL); | |
408 regnd.map(_max_reg,NULL); | |
409 // Set mappings as OUR mappings | |
410 blk2value[b->_pre_order] = &value; | |
411 blk2regnd[b->_pre_order] = ®nd; | |
412 | |
413 // Initialize value & regnd for this block | |
414 if( missing_some_inputs ) { | |
415 // Some predecessor has not yet been visited; zap map to empty | |
416 for( uint k = 0; k < (uint)_max_reg; k++ ) { | |
417 value.map(k,NULL); | |
418 regnd.map(k,NULL); | |
419 } | |
420 } else { | |
421 if( !freed ) { // Didn't get a freebie prior block | |
422 // Must clone some data | |
423 freed = _cfg._bbs[b->pred(1)->_idx]; | |
424 Node_List &f_value = *blk2value[freed->_pre_order]; | |
425 Node_List &f_regnd = *blk2regnd[freed->_pre_order]; | |
426 for( uint k = 0; k < (uint)_max_reg; k++ ) { | |
427 value.map(k,f_value[k]); | |
428 regnd.map(k,f_regnd[k]); | |
429 } | |
430 } | |
431 // Merge all inputs together, setting to NULL any conflicts. | |
432 for( j = 1; j < b->num_preds(); j++ ) { | |
433 Block *pb = _cfg._bbs[b->pred(j)->_idx]; | |
434 if( pb == freed ) continue; // Did self already via freelist | |
435 Node_List &p_regnd = *blk2regnd[pb->_pre_order]; | |
436 for( uint k = 0; k < (uint)_max_reg; k++ ) { | |
437 if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs? | |
438 value.map(k,NULL); // Then no value handy | |
439 regnd.map(k,NULL); | |
440 } | |
441 } | |
442 } | |
443 } | |
444 | |
445 // For all Phi's | |
446 for( j = 1; j < phi_dex; j++ ) { | |
447 uint k; | |
448 Node *phi = b->_nodes[j]; | |
449 uint pidx = n2lidx(phi); | |
450 OptoReg::Name preg = lrgs(n2lidx(phi)).reg(); | |
451 | |
452 // Remove copies remaining on edges. Check for junk phi. | |
453 Node *u = NULL; | |
454 for( k=1; k<phi->req(); k++ ) { | |
455 Node *x = phi->in(k); | |
456 if( phi != x && u != x ) // Found a different input | |
457 u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input | |
458 } | |
459 if( u != NodeSentinel ) { // Junk Phi. Remove | |
460 b->_nodes.remove(j--); phi_dex--; | |
461 _cfg._bbs.map(phi->_idx,NULL); | |
462 phi->replace_by(u); | |
463 phi->disconnect_inputs(NULL); | |
464 continue; | |
465 } | |
466 // Note that if value[pidx] exists, then we merged no new values here | |
467 // and the phi is useless. This can happen even with the above phi | |
468 // removal for complex flows. I cannot keep the better known value here | |
469 // because locally the phi appears to define a new merged value. If I | |
470 // keep the better value then a copy of the phi, being unable to use the | |
471 // global flow analysis, can't "peek through" the phi to the original | |
472 // reaching value and so will act like it's defining a new value. This | |
473 // can lead to situations where some uses are from the old and some from | |
474 // the new values. Not illegal by itself but throws the over-strong | |
475 // assert in scheduling. | |
476 if( pidx ) { | |
477 value.map(preg,phi); | |
478 regnd.map(preg,phi); | |
479 OptoReg::Name preg_lo = OptoReg::add(preg,-1); | |
480 if( !is_single_register(phi->ideal_reg()) ) { | |
481 value.map(preg_lo,phi); | |
482 regnd.map(preg_lo,phi); | |
483 } | |
484 } | |
485 } | |
486 | |
487 // For all remaining instructions | |
488 for( j = phi_dex; j < b->_nodes.size(); j++ ) { | |
489 Node *n = b->_nodes[j]; | |
490 | |
491 if( n->outcnt() == 0 && // Dead? | |
492 n != C->top() && // (ignore TOP, it has no du info) | |
493 !n->is_Proj() ) { // fat-proj kills | |
494 j -= yank_if_dead(n,b,&value,®nd); | |
495 continue; | |
496 } | |
497 | |
498 // Improve reaching-def info. Occasionally post-alloc's liveness gives | |
499 // up (at loop backedges, because we aren't doing a full flow pass). | |
500 // The presence of a live use essentially asserts that the use's def is | |
501 // alive and well at the use (or else the allocator fubar'd). Take | |
502 // advantage of this info to set a reaching def for the use-reg. | |
503 uint k; | |
504 for( k = 1; k < n->req(); k++ ) { | |
505 Node *def = n->in(k); // n->in(k) is a USE; def is the DEF for this USE | |
506 guarantee(def != NULL, "no disconnected nodes at this point"); | |
507 uint useidx = n2lidx(def); // useidx is the live range index for this USE | |
508 | |
509 if( useidx ) { | |
510 OptoReg::Name ureg = lrgs(useidx).reg(); | |
511 if( !value[ureg] ) { | |
512 int idx; // Skip occasional useless copy | |
513 while( (idx=def->is_Copy()) != 0 && | |
514 def->in(idx) != NULL && // NULL should not happen | |
515 ureg == lrgs(n2lidx(def->in(idx))).reg() ) | |
516 def = def->in(idx); | |
517 Node *valdef = skip_copies(def); // tighten up val through non-useless copies | |
518 value.map(ureg,valdef); // record improved reaching-def info | |
519 regnd.map(ureg, def); | |
520 // Record other half of doubles | |
521 OptoReg::Name ureg_lo = OptoReg::add(ureg,-1); | |
522 if( !is_single_register(def->ideal_reg()) && | |
523 ( !RegMask::can_represent(ureg_lo) || | |
524 lrgs(useidx).mask().Member(ureg_lo) ) && // Nearly always adjacent | |
525 !value[ureg_lo] ) { | |
526 value.map(ureg_lo,valdef); // record improved reaching-def info | |
527 regnd.map(ureg_lo, def); | |
528 } | |
529 } | |
530 } | |
531 } | |
532 | |
533 const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0; | |
534 | |
535 // Remove copies along input edges | |
536 for( k = 1; k < n->req(); k++ ) | |
537 j -= elide_copy( n, k, b, value, regnd, two_adr!=k ); | |
538 | |
539 // Unallocated Nodes define no registers | |
540 uint lidx = n2lidx(n); | |
541 if( !lidx ) continue; | |
542 | |
543 // Update the register defined by this instruction | |
544 OptoReg::Name nreg = lrgs(lidx).reg(); | |
545 // Skip through all copies to the _value_ being defined. | |
546 // Do not change from int to pointer | |
547 Node *val = skip_copies(n); | |
548 | |
923 | 549 // Clear out a dead definition before starting so that the |
550 // elimination code doesn't have to guard against it. The | |
551 // definition could in fact be a kill projection with a count of | |
552 // 0 which is safe but since those are uninteresting for copy | |
553 // elimination just delete them as well. | |
554 if (regnd[nreg] != NULL && regnd[nreg]->outcnt() == 0) { | |
555 regnd.map(nreg, NULL); | |
556 value.map(nreg, NULL); | |
557 } | |
558 | |
0 | 559 uint n_ideal_reg = n->ideal_reg(); |
560 if( is_single_register(n_ideal_reg) ) { | |
561 // If Node 'n' does not change the value mapped by the register, | |
562 // then 'n' is a useless copy. Do not update the register->node | |
563 // mapping so 'n' will go dead. | |
564 if( value[nreg] != val ) { | |
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
565 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, OptoReg::Bad)) { |
923 | 566 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); |
0 | 567 } else { |
568 // Update the mapping: record new Node defined by the register | |
569 regnd.map(nreg,n); | |
570 // Update mapping for defined *value*, which is the defined | |
571 // Node after skipping all copies. | |
572 value.map(nreg,val); | |
573 } | |
923 | 574 } else if( !may_be_copy_of_callee(n) ) { |
0 | 575 assert( n->is_Copy(), "" ); |
923 | 576 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); |
0 | 577 } |
578 } else { | |
579 // If the value occupies a register pair, record same info | |
580 // in both registers. | |
581 OptoReg::Name nreg_lo = OptoReg::add(nreg,-1); | |
582 if( RegMask::can_represent(nreg_lo) && // Either a spill slot, or | |
583 !lrgs(lidx).mask().Member(nreg_lo) ) { // Nearly always adjacent | |
584 // Sparc occasionally has non-adjacent pairs. | |
585 // Find the actual other value | |
586 RegMask tmp = lrgs(lidx).mask(); | |
587 tmp.Remove(nreg); | |
588 nreg_lo = tmp.find_first_elem(); | |
589 } | |
590 if( value[nreg] != val || value[nreg_lo] != val ) { | |
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
591 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, nreg_lo)) { |
923 | 592 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); |
0 | 593 } else { |
594 regnd.map(nreg , n ); | |
595 regnd.map(nreg_lo, n ); | |
596 value.map(nreg ,val); | |
597 value.map(nreg_lo,val); | |
598 } | |
923 | 599 } else if( !may_be_copy_of_callee(n) ) { |
0 | 600 assert( n->is_Copy(), "" ); |
923 | 601 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); |
0 | 602 } |
603 } | |
604 | |
605 // Fat projections kill many registers | |
606 if( n_ideal_reg == MachProjNode::fat_proj ) { | |
607 RegMask rm = n->out_RegMask(); | |
608 // wow, what an expensive iterator... | |
609 nreg = rm.find_first_elem(); | |
610 while( OptoReg::is_valid(nreg)) { | |
611 rm.Remove(nreg); | |
612 value.map(nreg,n); | |
613 regnd.map(nreg,n); | |
614 nreg = rm.find_first_elem(); | |
615 } | |
616 } | |
617 | |
618 } // End of for all instructions in the block | |
619 | |
620 } // End for all blocks | |
621 } |