Mercurial > hg > truffle
annotate src/share/vm/opto/postaloc.cpp @ 4597:8bc6f680a88d
fixed visual studio project files
author | Christian Haeubl <christian.haeubl@oracle.com> |
---|---|
date | Tue, 14 Feb 2012 15:01:36 -0800 |
parents | 2209834ccb59 |
children | 5da7201222d5 |
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 | |
3934
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
75 //------------------------------yank----------------------------------- |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
76 // Helper function for yank_if_dead |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
77 int PhaseChaitin::yank( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) { |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
78 int blk_adjust=0; |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
79 Block *oldb = _cfg._bbs[old->_idx]; |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
80 oldb->find_remove(old); |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
81 // Count 1 if deleting an instruction from the current block |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
82 if( oldb == current_block ) blk_adjust++; |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
83 _cfg._bbs.map(old->_idx,NULL); |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
84 OptoReg::Name old_reg = lrgs(n2lidx(old)).reg(); |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
85 if( regnd && (*regnd)[old_reg]==old ) { // Instruction is currently available? |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
86 value->map(old_reg,NULL); // Yank from value/regnd maps |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
87 regnd->map(old_reg,NULL); // This register's value is now unknown |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
88 } |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
89 return blk_adjust; |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
90 } |
0 | 91 |
92 //------------------------------yank_if_dead----------------------------------- | |
93 // Removed an edge from 'old'. Yank if dead. Return adjustment counts to | |
94 // iterators in the current block. | |
95 int PhaseChaitin::yank_if_dead( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) { | |
96 int blk_adjust=0; | |
97 while (old->outcnt() == 0 && old != C->top()) { | |
3934
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
98 blk_adjust += yank(old, current_block, value, regnd); |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
99 |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
100 Node *tmp = NULL; |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
101 for (uint i = 1; i < old->req(); i++) { |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
102 if (old->in(i)->is_MachTemp()) { |
3941
2209834ccb59
7089632: assert(machtmp->outcnt() == 1) failed: expected for a MachTemp
kvn
parents:
3934
diff
changeset
|
103 // handle TEMP inputs |
3934
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
104 Node* machtmp = old->in(i); |
3941
2209834ccb59
7089632: assert(machtmp->outcnt() == 1) failed: expected for a MachTemp
kvn
parents:
3934
diff
changeset
|
105 if (machtmp->outcnt() == 1) { |
2209834ccb59
7089632: assert(machtmp->outcnt() == 1) failed: expected for a MachTemp
kvn
parents:
3934
diff
changeset
|
106 assert(machtmp->unique_out() == old, "sanity"); |
2209834ccb59
7089632: assert(machtmp->outcnt() == 1) failed: expected for a MachTemp
kvn
parents:
3934
diff
changeset
|
107 blk_adjust += yank(machtmp, current_block, value, regnd); |
2209834ccb59
7089632: assert(machtmp->outcnt() == 1) failed: expected for a MachTemp
kvn
parents:
3934
diff
changeset
|
108 machtmp->disconnect_inputs(NULL); |
2209834ccb59
7089632: assert(machtmp->outcnt() == 1) failed: expected for a MachTemp
kvn
parents:
3934
diff
changeset
|
109 } |
3934
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
110 } else { |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
111 assert(tmp == NULL, "can't handle more non MachTemp inputs"); |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
112 tmp = old->in(i); |
8f47d8870d9a
7087453: PhaseChaitin::yank_if_dead() should handle MachTemp inputs
roland
parents:
2008
diff
changeset
|
113 } |
0 | 114 } |
115 old->disconnect_inputs(NULL); | |
116 if( !tmp ) break; | |
117 old = tmp; | |
118 } | |
119 return blk_adjust; | |
120 } | |
121 | |
122 //------------------------------use_prior_register----------------------------- | |
123 // Use the prior value instead of the current value, in an effort to make | |
124 // the current value go dead. Return block iterator adjustment, in case | |
125 // we yank some instructions from this block. | |
126 int PhaseChaitin::use_prior_register( Node *n, uint idx, Node *def, Block *current_block, Node_List &value, Node_List ®nd ) { | |
127 // No effect? | |
128 if( def == n->in(idx) ) return 0; | |
129 // Def is currently dead and can be removed? Do not resurrect | |
130 if( def->outcnt() == 0 ) return 0; | |
131 | |
132 // Not every pair of physical registers are assignment compatible, | |
133 // e.g. on sparc floating point registers are not assignable to integer | |
134 // registers. | |
135 const LRG &def_lrg = lrgs(n2lidx(def)); | |
136 OptoReg::Name def_reg = def_lrg.reg(); | |
137 const RegMask &use_mask = n->in_RegMask(idx); | |
138 bool can_use = ( RegMask::can_represent(def_reg) ? (use_mask.Member(def_reg) != 0) | |
139 : (use_mask.is_AllStack() != 0)); | |
140 // Check for a copy to or from a misaligned pair. | |
141 can_use = can_use && !use_mask.is_misaligned_Pair() && !def_lrg.mask().is_misaligned_Pair(); | |
142 | |
143 if (!can_use) | |
144 return 0; | |
145 | |
146 // Capture the old def in case it goes dead... | |
147 Node *old = n->in(idx); | |
148 | |
149 // Save-on-call copies can only be elided if the entire copy chain can go | |
150 // away, lest we get the same callee-save value alive in 2 locations at | |
151 // once. We check for the obvious trivial case here. Although it can | |
152 // sometimes be elided with cooperation outside our scope, here we will just | |
153 // miss the opportunity. :-( | |
154 if( may_be_copy_of_callee(def) ) { | |
155 if( old->outcnt() > 1 ) return 0; // We're the not last user | |
156 int idx = old->is_Copy(); | |
157 assert( idx, "chain of copies being removed" ); | |
158 Node *old2 = old->in(idx); // Chain of copies | |
159 if( old2->outcnt() > 1 ) return 0; // old is not the last user | |
160 int idx2 = old2->is_Copy(); | |
161 if( !idx2 ) return 0; // Not a chain of 2 copies | |
162 if( def != old2->in(idx2) ) return 0; // Chain of exactly 2 copies | |
163 } | |
164 | |
165 // Use the new def | |
166 n->set_req(idx,def); | |
167 _post_alloc++; | |
168 | |
169 // Is old def now dead? We successfully yanked a copy? | |
170 return yank_if_dead(old,current_block,&value,®nd); | |
171 } | |
172 | |
173 | |
174 //------------------------------skip_copies------------------------------------ | |
175 // Skip through any number of copies (that don't mod oop-i-ness) | |
176 Node *PhaseChaitin::skip_copies( Node *c ) { | |
177 int idx = c->is_Copy(); | |
178 uint is_oop = lrgs(n2lidx(c))._is_oop; | |
179 while (idx != 0) { | |
180 guarantee(c->in(idx) != NULL, "must not resurrect dead copy"); | |
181 if (lrgs(n2lidx(c->in(idx)))._is_oop != is_oop) | |
182 break; // casting copy, not the same value | |
183 c = c->in(idx); | |
184 idx = c->is_Copy(); | |
185 } | |
186 return c; | |
187 } | |
188 | |
189 //------------------------------elide_copy------------------------------------- | |
190 // Remove (bypass) copies along Node n, edge k. | |
191 int PhaseChaitin::elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List ®nd, bool can_change_regs ) { | |
192 int blk_adjust = 0; | |
193 | |
194 uint nk_idx = n2lidx(n->in(k)); | |
195 OptoReg::Name nk_reg = lrgs(nk_idx ).reg(); | |
196 | |
197 // Remove obvious same-register copies | |
198 Node *x = n->in(k); | |
199 int idx; | |
200 while( (idx=x->is_Copy()) != 0 ) { | |
201 Node *copy = x->in(idx); | |
202 guarantee(copy != NULL, "must not resurrect dead copy"); | |
203 if( lrgs(n2lidx(copy)).reg() != nk_reg ) break; | |
204 blk_adjust += use_prior_register(n,k,copy,current_block,value,regnd); | |
205 if( n->in(k) != copy ) break; // Failed for some cutout? | |
206 x = copy; // Progress, try again | |
207 } | |
208 | |
209 // Phis and 2-address instructions cannot change registers so easily - their | |
210 // outputs must match their input. | |
211 if( !can_change_regs ) | |
212 return blk_adjust; // Only check stupid copies! | |
213 | |
214 // Loop backedges won't have a value-mapping yet | |
215 if( &value == NULL ) return blk_adjust; | |
216 | |
217 // Skip through all copies to the _value_ being used. Do not change from | |
218 // int to pointer. This attempts to jump through a chain of copies, where | |
219 // intermediate copies might be illegal, i.e., value is stored down to stack | |
220 // then reloaded BUT survives in a register the whole way. | |
221 Node *val = skip_copies(n->in(k)); | |
222 | |
2008 | 223 if (val == x && nk_idx != 0 && |
224 regnd[nk_reg] != NULL && regnd[nk_reg] != x && | |
225 n2lidx(x) == n2lidx(regnd[nk_reg])) { | |
226 // When rematerialzing nodes and stretching lifetimes, the | |
227 // allocator will reuse the original def for multidef LRG instead | |
228 // of the current reaching def because it can't know it's safe to | |
229 // do so. After allocation completes if they are in the same LRG | |
230 // then it should use the current reaching def instead. | |
231 n->set_req(k, regnd[nk_reg]); | |
232 blk_adjust += yank_if_dead(val, current_block, &value, ®nd); | |
233 val = skip_copies(n->in(k)); | |
234 } | |
235 | |
0 | 236 if( val == x ) return blk_adjust; // No progress? |
237 | |
238 bool single = is_single_register(val->ideal_reg()); | |
239 uint val_idx = n2lidx(val); | |
240 OptoReg::Name val_reg = lrgs(val_idx).reg(); | |
241 | |
242 // See if it happens to already be in the correct register! | |
243 // (either Phi's direct register, or the common case of the name | |
244 // never-clobbered original-def register) | |
245 if( value[val_reg] == val && | |
246 // Doubles check both halves | |
247 ( single || value[val_reg-1] == val ) ) { | |
248 blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd); | |
249 if( n->in(k) == regnd[val_reg] ) // Success! Quit trying | |
250 return blk_adjust; | |
251 } | |
252 | |
253 // See if we can skip the copy by changing registers. Don't change from | |
254 // using a register to using the stack unless we know we can remove a | |
255 // copy-load. Otherwise we might end up making a pile of Intel cisc-spill | |
256 // ops reading from memory instead of just loading once and using the | |
257 // register. | |
258 | |
259 // Also handle duplicate copies here. | |
260 const Type *t = val->is_Con() ? val->bottom_type() : NULL; | |
261 | |
262 // Scan all registers to see if this value is around already | |
263 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
|
264 if (reg == (uint)nk_reg) { |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
265 // 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
|
266 // 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
|
267 bool ignore_self = true; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
268 x = n->in(k); |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
269 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
|
270 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
|
271 while (i < imax && ignore_self) { |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
272 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
|
273 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
|
274 } |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
275 if (ignore_self) continue; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
276 } |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
277 |
0 | 278 Node *vv = value[reg]; |
279 if( !single ) { // Doubles check for aligned-adjacent pair | |
280 if( (reg&1)==0 ) continue; // Wrong half of a pair | |
281 if( vv != value[reg-1] ) continue; // Not a complete pair | |
282 } | |
283 if( vv == val || // Got a direct hit? | |
284 (t && vv && vv->bottom_type() == t && vv->is_Mach() && | |
285 vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant? | |
286 assert( !n->is_Phi(), "cannot change registers at a Phi so easily" ); | |
287 if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR | |
288 OptoReg::is_reg(reg) || // turning into a register use OR | |
289 regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use | |
290 blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd); | |
291 if( n->in(k) == regnd[reg] ) // Success! Quit trying | |
292 return blk_adjust; | |
293 } // End of if not degrading to a stack | |
294 } // End of if found value in another register | |
295 } // End of scan all machine registers | |
296 return blk_adjust; | |
297 } | |
298 | |
299 | |
300 // | |
301 // Check if nreg already contains the constant value val. Normal copy | |
302 // elimination doesn't doesn't work on constants because multiple | |
303 // nodes can represent the same constant so the type and rule of the | |
304 // MachNode must be checked to ensure equivalence. | |
305 // | |
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
306 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
|
307 Block *current_block, |
0 | 308 Node_List& value, Node_List& regnd, |
309 OptoReg::Name nreg, OptoReg::Name nreg2) { | |
310 if (value[nreg] != val && val->is_Con() && | |
311 value[nreg] != NULL && value[nreg]->is_Con() && | |
312 (nreg2 == OptoReg::Bad || value[nreg] == value[nreg2]) && | |
313 value[nreg]->bottom_type() == val->bottom_type() && | |
314 value[nreg]->as_Mach()->rule() == val->as_Mach()->rule()) { | |
315 // This code assumes that two MachNodes representing constants | |
316 // which have the same rule and the same bottom type will produce | |
317 // identical effects into a register. This seems like it must be | |
318 // objectively true unless there are hidden inputs to the nodes | |
319 // but if that were to change this code would need to updated. | |
320 // Since they are equivalent the second one if redundant and can | |
321 // be removed. | |
322 // | |
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
323 // n will be replaced with the old value but n might have |
0 | 324 // kills projections associated with it so remove them now so that |
605 | 325 // yank_if_dead will be able to eliminate the copy once the uses |
0 | 326 // 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
|
327 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
|
328 Node* use = n->fast_out(i); |
0 | 329 if (use->is_Proj() && use->outcnt() == 0) { |
330 // Kill projections have no users and one input | |
331 use->set_req(0, C->top()); | |
332 yank_if_dead(use, current_block, &value, ®nd); | |
333 --i; --imax; | |
334 } | |
335 } | |
336 _post_alloc++; | |
337 return true; | |
338 } | |
339 return false; | |
340 } | |
341 | |
342 | |
343 //------------------------------post_allocate_copy_removal--------------------- | |
344 // Post-Allocation peephole copy removal. We do this in 1 pass over the | |
345 // basic blocks. We maintain a mapping of registers to Nodes (an array of | |
346 // Nodes indexed by machine register or stack slot number). NULL means that a | |
347 // register is not mapped to any Node. We can (want to have!) have several | |
348 // registers map to the same Node. We walk forward over the instructions | |
349 // updating the mapping as we go. At merge points we force a NULL if we have | |
350 // to merge 2 different Nodes into the same register. Phi functions will give | |
351 // us a new Node if there is a proper value merging. Since the blocks are | |
352 // arranged in some RPO, we will visit all parent blocks before visiting any | |
353 // successor blocks (except at loops). | |
354 // | |
355 // If we find a Copy we look to see if the Copy's source register is a stack | |
356 // slot and that value has already been loaded into some machine register; if | |
357 // so we use machine register directly. This turns a Load into a reg-reg | |
358 // Move. We also look for reloads of identical constants. | |
359 // | |
360 // When we see a use from a reg-reg Copy, we will attempt to use the copy's | |
361 // source directly and make the copy go dead. | |
362 void PhaseChaitin::post_allocate_copy_removal() { | |
363 NOT_PRODUCT( Compile::TracePhase t3("postAllocCopyRemoval", &_t_postAllocCopyRemoval, TimeCompiler); ) | |
364 ResourceMark rm; | |
365 | |
366 // Need a mapping from basic block Node_Lists. We need a Node_List to | |
367 // map from register number to value-producing Node. | |
368 Node_List **blk2value = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); | |
369 memset( blk2value, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); | |
370 // Need a mapping from basic block Node_Lists. We need a Node_List to | |
371 // map from register number to register-defining Node. | |
372 Node_List **blk2regnd = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); | |
373 memset( blk2regnd, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); | |
374 | |
375 // We keep unused Node_Lists on a free_list to avoid wasting | |
376 // memory. | |
377 GrowableArray<Node_List*> free_list = GrowableArray<Node_List*>(16); | |
378 | |
379 // For all blocks | |
380 for( uint i = 0; i < _cfg._num_blocks; i++ ) { | |
381 uint j; | |
382 Block *b = _cfg._blocks[i]; | |
383 | |
384 // Count of Phis in block | |
385 uint phi_dex; | |
386 for( phi_dex = 1; phi_dex < b->_nodes.size(); phi_dex++ ) { | |
387 Node *phi = b->_nodes[phi_dex]; | |
388 if( !phi->is_Phi() ) | |
389 break; | |
390 } | |
391 | |
392 // If any predecessor has not been visited, we do not know the state | |
393 // of registers at the start. Check for this, while updating copies | |
394 // along Phi input edges | |
395 bool missing_some_inputs = false; | |
396 Block *freed = NULL; | |
397 for( j = 1; j < b->num_preds(); j++ ) { | |
398 Block *pb = _cfg._bbs[b->pred(j)->_idx]; | |
399 // Remove copies along phi edges | |
400 for( uint k=1; k<phi_dex; k++ ) | |
401 elide_copy( b->_nodes[k], j, b, *blk2value[pb->_pre_order], *blk2regnd[pb->_pre_order], false ); | |
402 if( blk2value[pb->_pre_order] ) { // Have a mapping on this edge? | |
403 // See if this predecessor's mappings have been used by everybody | |
404 // who wants them. If so, free 'em. | |
405 uint k; | |
406 for( k=0; k<pb->_num_succs; k++ ) { | |
407 Block *pbsucc = pb->_succs[k]; | |
408 if( !blk2value[pbsucc->_pre_order] && pbsucc != b ) | |
409 break; // Found a future user | |
410 } | |
411 if( k >= pb->_num_succs ) { // No more uses, free! | |
412 freed = pb; // Record last block freed | |
413 free_list.push(blk2value[pb->_pre_order]); | |
414 free_list.push(blk2regnd[pb->_pre_order]); | |
415 } | |
416 } else { // This block has unvisited (loopback) inputs | |
417 missing_some_inputs = true; | |
418 } | |
419 } | |
420 | |
421 | |
422 // Extract Node_List mappings. If 'freed' is non-zero, we just popped | |
423 // 'freed's blocks off the list | |
424 Node_List ®nd = *(free_list.is_empty() ? new Node_List() : free_list.pop()); | |
425 Node_List &value = *(free_list.is_empty() ? new Node_List() : free_list.pop()); | |
426 assert( !freed || blk2value[freed->_pre_order] == &value, "" ); | |
427 value.map(_max_reg,NULL); | |
428 regnd.map(_max_reg,NULL); | |
429 // Set mappings as OUR mappings | |
430 blk2value[b->_pre_order] = &value; | |
431 blk2regnd[b->_pre_order] = ®nd; | |
432 | |
433 // Initialize value & regnd for this block | |
434 if( missing_some_inputs ) { | |
435 // Some predecessor has not yet been visited; zap map to empty | |
436 for( uint k = 0; k < (uint)_max_reg; k++ ) { | |
437 value.map(k,NULL); | |
438 regnd.map(k,NULL); | |
439 } | |
440 } else { | |
441 if( !freed ) { // Didn't get a freebie prior block | |
442 // Must clone some data | |
443 freed = _cfg._bbs[b->pred(1)->_idx]; | |
444 Node_List &f_value = *blk2value[freed->_pre_order]; | |
445 Node_List &f_regnd = *blk2regnd[freed->_pre_order]; | |
446 for( uint k = 0; k < (uint)_max_reg; k++ ) { | |
447 value.map(k,f_value[k]); | |
448 regnd.map(k,f_regnd[k]); | |
449 } | |
450 } | |
451 // Merge all inputs together, setting to NULL any conflicts. | |
452 for( j = 1; j < b->num_preds(); j++ ) { | |
453 Block *pb = _cfg._bbs[b->pred(j)->_idx]; | |
454 if( pb == freed ) continue; // Did self already via freelist | |
455 Node_List &p_regnd = *blk2regnd[pb->_pre_order]; | |
456 for( uint k = 0; k < (uint)_max_reg; k++ ) { | |
457 if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs? | |
458 value.map(k,NULL); // Then no value handy | |
459 regnd.map(k,NULL); | |
460 } | |
461 } | |
462 } | |
463 } | |
464 | |
465 // For all Phi's | |
466 for( j = 1; j < phi_dex; j++ ) { | |
467 uint k; | |
468 Node *phi = b->_nodes[j]; | |
469 uint pidx = n2lidx(phi); | |
470 OptoReg::Name preg = lrgs(n2lidx(phi)).reg(); | |
471 | |
472 // Remove copies remaining on edges. Check for junk phi. | |
473 Node *u = NULL; | |
474 for( k=1; k<phi->req(); k++ ) { | |
475 Node *x = phi->in(k); | |
476 if( phi != x && u != x ) // Found a different input | |
477 u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input | |
478 } | |
479 if( u != NodeSentinel ) { // Junk Phi. Remove | |
480 b->_nodes.remove(j--); phi_dex--; | |
481 _cfg._bbs.map(phi->_idx,NULL); | |
482 phi->replace_by(u); | |
483 phi->disconnect_inputs(NULL); | |
484 continue; | |
485 } | |
486 // Note that if value[pidx] exists, then we merged no new values here | |
487 // and the phi is useless. This can happen even with the above phi | |
488 // removal for complex flows. I cannot keep the better known value here | |
489 // because locally the phi appears to define a new merged value. If I | |
490 // keep the better value then a copy of the phi, being unable to use the | |
491 // global flow analysis, can't "peek through" the phi to the original | |
492 // reaching value and so will act like it's defining a new value. This | |
493 // can lead to situations where some uses are from the old and some from | |
494 // the new values. Not illegal by itself but throws the over-strong | |
495 // assert in scheduling. | |
496 if( pidx ) { | |
497 value.map(preg,phi); | |
498 regnd.map(preg,phi); | |
499 OptoReg::Name preg_lo = OptoReg::add(preg,-1); | |
500 if( !is_single_register(phi->ideal_reg()) ) { | |
501 value.map(preg_lo,phi); | |
502 regnd.map(preg_lo,phi); | |
503 } | |
504 } | |
505 } | |
506 | |
507 // For all remaining instructions | |
508 for( j = phi_dex; j < b->_nodes.size(); j++ ) { | |
509 Node *n = b->_nodes[j]; | |
510 | |
511 if( n->outcnt() == 0 && // Dead? | |
512 n != C->top() && // (ignore TOP, it has no du info) | |
513 !n->is_Proj() ) { // fat-proj kills | |
514 j -= yank_if_dead(n,b,&value,®nd); | |
515 continue; | |
516 } | |
517 | |
518 // Improve reaching-def info. Occasionally post-alloc's liveness gives | |
519 // up (at loop backedges, because we aren't doing a full flow pass). | |
520 // The presence of a live use essentially asserts that the use's def is | |
521 // alive and well at the use (or else the allocator fubar'd). Take | |
522 // advantage of this info to set a reaching def for the use-reg. | |
523 uint k; | |
524 for( k = 1; k < n->req(); k++ ) { | |
525 Node *def = n->in(k); // n->in(k) is a USE; def is the DEF for this USE | |
526 guarantee(def != NULL, "no disconnected nodes at this point"); | |
527 uint useidx = n2lidx(def); // useidx is the live range index for this USE | |
528 | |
529 if( useidx ) { | |
530 OptoReg::Name ureg = lrgs(useidx).reg(); | |
531 if( !value[ureg] ) { | |
532 int idx; // Skip occasional useless copy | |
533 while( (idx=def->is_Copy()) != 0 && | |
534 def->in(idx) != NULL && // NULL should not happen | |
535 ureg == lrgs(n2lidx(def->in(idx))).reg() ) | |
536 def = def->in(idx); | |
537 Node *valdef = skip_copies(def); // tighten up val through non-useless copies | |
538 value.map(ureg,valdef); // record improved reaching-def info | |
539 regnd.map(ureg, def); | |
540 // Record other half of doubles | |
541 OptoReg::Name ureg_lo = OptoReg::add(ureg,-1); | |
542 if( !is_single_register(def->ideal_reg()) && | |
543 ( !RegMask::can_represent(ureg_lo) || | |
544 lrgs(useidx).mask().Member(ureg_lo) ) && // Nearly always adjacent | |
545 !value[ureg_lo] ) { | |
546 value.map(ureg_lo,valdef); // record improved reaching-def info | |
547 regnd.map(ureg_lo, def); | |
548 } | |
549 } | |
550 } | |
551 } | |
552 | |
553 const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0; | |
554 | |
555 // Remove copies along input edges | |
556 for( k = 1; k < n->req(); k++ ) | |
557 j -= elide_copy( n, k, b, value, regnd, two_adr!=k ); | |
558 | |
559 // Unallocated Nodes define no registers | |
560 uint lidx = n2lidx(n); | |
561 if( !lidx ) continue; | |
562 | |
563 // Update the register defined by this instruction | |
564 OptoReg::Name nreg = lrgs(lidx).reg(); | |
565 // Skip through all copies to the _value_ being defined. | |
566 // Do not change from int to pointer | |
567 Node *val = skip_copies(n); | |
568 | |
923 | 569 // Clear out a dead definition before starting so that the |
570 // elimination code doesn't have to guard against it. The | |
571 // definition could in fact be a kill projection with a count of | |
572 // 0 which is safe but since those are uninteresting for copy | |
573 // elimination just delete them as well. | |
574 if (regnd[nreg] != NULL && regnd[nreg]->outcnt() == 0) { | |
575 regnd.map(nreg, NULL); | |
576 value.map(nreg, NULL); | |
577 } | |
578 | |
0 | 579 uint n_ideal_reg = n->ideal_reg(); |
580 if( is_single_register(n_ideal_reg) ) { | |
581 // If Node 'n' does not change the value mapped by the register, | |
582 // then 'n' is a useless copy. Do not update the register->node | |
583 // mapping so 'n' will go dead. | |
584 if( value[nreg] != val ) { | |
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
585 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, OptoReg::Bad)) { |
923 | 586 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); |
0 | 587 } else { |
588 // Update the mapping: record new Node defined by the register | |
589 regnd.map(nreg,n); | |
590 // Update mapping for defined *value*, which is the defined | |
591 // Node after skipping all copies. | |
592 value.map(nreg,val); | |
593 } | |
923 | 594 } else if( !may_be_copy_of_callee(n) ) { |
0 | 595 assert( n->is_Copy(), "" ); |
923 | 596 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); |
0 | 597 } |
598 } else { | |
599 // If the value occupies a register pair, record same info | |
600 // in both registers. | |
601 OptoReg::Name nreg_lo = OptoReg::add(nreg,-1); | |
602 if( RegMask::can_represent(nreg_lo) && // Either a spill slot, or | |
603 !lrgs(lidx).mask().Member(nreg_lo) ) { // Nearly always adjacent | |
604 // Sparc occasionally has non-adjacent pairs. | |
605 // Find the actual other value | |
606 RegMask tmp = lrgs(lidx).mask(); | |
607 tmp.Remove(nreg); | |
608 nreg_lo = tmp.find_first_elem(); | |
609 } | |
610 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
|
611 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, nreg_lo)) { |
923 | 612 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); |
0 | 613 } else { |
614 regnd.map(nreg , n ); | |
615 regnd.map(nreg_lo, n ); | |
616 value.map(nreg ,val); | |
617 value.map(nreg_lo,val); | |
618 } | |
923 | 619 } else if( !may_be_copy_of_callee(n) ) { |
0 | 620 assert( n->is_Copy(), "" ); |
923 | 621 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); |
0 | 622 } |
623 } | |
624 | |
625 // Fat projections kill many registers | |
626 if( n_ideal_reg == MachProjNode::fat_proj ) { | |
627 RegMask rm = n->out_RegMask(); | |
628 // wow, what an expensive iterator... | |
629 nreg = rm.find_first_elem(); | |
630 while( OptoReg::is_valid(nreg)) { | |
631 rm.Remove(nreg); | |
632 value.map(nreg,n); | |
633 regnd.map(nreg,n); | |
634 nreg = rm.find_first_elem(); | |
635 } | |
636 } | |
637 | |
638 } // End of for all instructions in the block | |
639 | |
640 } // End for all blocks | |
641 } |