Mercurial > hg > graal-compiler
annotate src/share/vm/opto/postaloc.cpp @ 404:78c058bc5cdc
6717150: improper constant folding of subnormal strictfp multiplications and divides
Summary: suppress constant folding of double divides and multiplications on ia32
Reviewed-by: never
author | rasbold |
---|---|
date | Tue, 14 Oct 2008 06:58:58 -0700 |
parents | cc80376deb0c |
children | 98cb887364d3 |
rev | line source |
---|---|
0 | 1 /* |
196 | 2 * Copyright 1998-2008 Sun Microsystems, Inc. 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 * | |
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 */ | |
24 | |
25 #include "incls/_precompiled.incl" | |
26 #include "incls/_postaloc.cpp.incl" | |
27 | |
28 // see if this register kind does not requires two registers | |
29 static bool is_single_register(uint x) { | |
30 #ifdef _LP64 | |
31 return (x != Op_RegD && x != Op_RegL && x != Op_RegP); | |
32 #else | |
33 return (x != Op_RegD && x != Op_RegL); | |
34 #endif | |
35 } | |
36 | |
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
37 //---------------------------may_be_copy_of_callee----------------------------- |
0 | 38 // Check to see if we can possibly be a copy of a callee-save value. |
39 bool PhaseChaitin::may_be_copy_of_callee( Node *def ) const { | |
40 // Short circuit if there are no callee save registers | |
41 if (_matcher.number_of_saved_registers() == 0) return false; | |
42 | |
43 // Expect only a spill-down and reload on exit for callee-save spills. | |
44 // Chains of copies cannot be deep. | |
45 // 5008997 - This is wishful thinking. Register allocator seems to | |
46 // be splitting live ranges for callee save registers to such | |
47 // an extent that in large methods the chains can be very long | |
48 // (50+). The conservative answer is to return true if we don't | |
49 // know as this prevents optimizations from occuring. | |
50 | |
51 const int limit = 60; | |
52 int i; | |
53 for( i=0; i < limit; i++ ) { | |
54 if( def->is_Proj() && def->in(0)->is_Start() && | |
55 _matcher.is_save_on_entry(lrgs(n2lidx(def)).reg()) ) | |
56 return true; // Direct use of callee-save proj | |
57 if( def->is_Copy() ) // Copies carry value through | |
58 def = def->in(def->is_Copy()); | |
59 else if( def->is_Phi() ) // Phis can merge it from any direction | |
60 def = def->in(1); | |
61 else | |
62 break; | |
63 guarantee(def != NULL, "must not resurrect dead copy"); | |
64 } | |
65 // If we reached the end and didn't find a callee save proj | |
66 // then this may be a callee save proj so we return true | |
67 // as the conservative answer. If we didn't reach then end | |
68 // we must have discovered that it was not a callee save | |
69 // else we would have returned. | |
70 return i == limit; | |
71 } | |
72 | |
73 | |
74 | |
75 //------------------------------yank_if_dead----------------------------------- | |
76 // Removed an edge from 'old'. Yank if dead. Return adjustment counts to | |
77 // iterators in the current block. | |
78 int PhaseChaitin::yank_if_dead( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) { | |
79 int blk_adjust=0; | |
80 while (old->outcnt() == 0 && old != C->top()) { | |
81 Block *oldb = _cfg._bbs[old->_idx]; | |
82 oldb->find_remove(old); | |
83 // Count 1 if deleting an instruction from the current block | |
84 if( oldb == current_block ) blk_adjust++; | |
85 _cfg._bbs.map(old->_idx,NULL); | |
86 OptoReg::Name old_reg = lrgs(n2lidx(old)).reg(); | |
87 if( regnd && (*regnd)[old_reg]==old ) { // Instruction is currently available? | |
88 value->map(old_reg,NULL); // Yank from value/regnd maps | |
89 regnd->map(old_reg,NULL); // This register's value is now unknown | |
90 } | |
91 Node *tmp = old->req() > 1 ? old->in(1) : NULL; | |
92 old->disconnect_inputs(NULL); | |
93 if( !tmp ) break; | |
94 old = tmp; | |
95 } | |
96 return blk_adjust; | |
97 } | |
98 | |
99 //------------------------------use_prior_register----------------------------- | |
100 // Use the prior value instead of the current value, in an effort to make | |
101 // the current value go dead. Return block iterator adjustment, in case | |
102 // we yank some instructions from this block. | |
103 int PhaseChaitin::use_prior_register( Node *n, uint idx, Node *def, Block *current_block, Node_List &value, Node_List ®nd ) { | |
104 // No effect? | |
105 if( def == n->in(idx) ) return 0; | |
106 // Def is currently dead and can be removed? Do not resurrect | |
107 if( def->outcnt() == 0 ) return 0; | |
108 | |
109 // Not every pair of physical registers are assignment compatible, | |
110 // e.g. on sparc floating point registers are not assignable to integer | |
111 // registers. | |
112 const LRG &def_lrg = lrgs(n2lidx(def)); | |
113 OptoReg::Name def_reg = def_lrg.reg(); | |
114 const RegMask &use_mask = n->in_RegMask(idx); | |
115 bool can_use = ( RegMask::can_represent(def_reg) ? (use_mask.Member(def_reg) != 0) | |
116 : (use_mask.is_AllStack() != 0)); | |
117 // Check for a copy to or from a misaligned pair. | |
118 can_use = can_use && !use_mask.is_misaligned_Pair() && !def_lrg.mask().is_misaligned_Pair(); | |
119 | |
120 if (!can_use) | |
121 return 0; | |
122 | |
123 // Capture the old def in case it goes dead... | |
124 Node *old = n->in(idx); | |
125 | |
126 // Save-on-call copies can only be elided if the entire copy chain can go | |
127 // away, lest we get the same callee-save value alive in 2 locations at | |
128 // once. We check for the obvious trivial case here. Although it can | |
129 // sometimes be elided with cooperation outside our scope, here we will just | |
130 // miss the opportunity. :-( | |
131 if( may_be_copy_of_callee(def) ) { | |
132 if( old->outcnt() > 1 ) return 0; // We're the not last user | |
133 int idx = old->is_Copy(); | |
134 assert( idx, "chain of copies being removed" ); | |
135 Node *old2 = old->in(idx); // Chain of copies | |
136 if( old2->outcnt() > 1 ) return 0; // old is not the last user | |
137 int idx2 = old2->is_Copy(); | |
138 if( !idx2 ) return 0; // Not a chain of 2 copies | |
139 if( def != old2->in(idx2) ) return 0; // Chain of exactly 2 copies | |
140 } | |
141 | |
142 // Use the new def | |
143 n->set_req(idx,def); | |
144 _post_alloc++; | |
145 | |
146 // Is old def now dead? We successfully yanked a copy? | |
147 return yank_if_dead(old,current_block,&value,®nd); | |
148 } | |
149 | |
150 | |
151 //------------------------------skip_copies------------------------------------ | |
152 // Skip through any number of copies (that don't mod oop-i-ness) | |
153 Node *PhaseChaitin::skip_copies( Node *c ) { | |
154 int idx = c->is_Copy(); | |
155 uint is_oop = lrgs(n2lidx(c))._is_oop; | |
156 while (idx != 0) { | |
157 guarantee(c->in(idx) != NULL, "must not resurrect dead copy"); | |
158 if (lrgs(n2lidx(c->in(idx)))._is_oop != is_oop) | |
159 break; // casting copy, not the same value | |
160 c = c->in(idx); | |
161 idx = c->is_Copy(); | |
162 } | |
163 return c; | |
164 } | |
165 | |
166 //------------------------------elide_copy------------------------------------- | |
167 // Remove (bypass) copies along Node n, edge k. | |
168 int PhaseChaitin::elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List ®nd, bool can_change_regs ) { | |
169 int blk_adjust = 0; | |
170 | |
171 uint nk_idx = n2lidx(n->in(k)); | |
172 OptoReg::Name nk_reg = lrgs(nk_idx ).reg(); | |
173 | |
174 // Remove obvious same-register copies | |
175 Node *x = n->in(k); | |
176 int idx; | |
177 while( (idx=x->is_Copy()) != 0 ) { | |
178 Node *copy = x->in(idx); | |
179 guarantee(copy != NULL, "must not resurrect dead copy"); | |
180 if( lrgs(n2lidx(copy)).reg() != nk_reg ) break; | |
181 blk_adjust += use_prior_register(n,k,copy,current_block,value,regnd); | |
182 if( n->in(k) != copy ) break; // Failed for some cutout? | |
183 x = copy; // Progress, try again | |
184 } | |
185 | |
186 // Phis and 2-address instructions cannot change registers so easily - their | |
187 // outputs must match their input. | |
188 if( !can_change_regs ) | |
189 return blk_adjust; // Only check stupid copies! | |
190 | |
191 // Loop backedges won't have a value-mapping yet | |
192 if( &value == NULL ) return blk_adjust; | |
193 | |
194 // Skip through all copies to the _value_ being used. Do not change from | |
195 // int to pointer. This attempts to jump through a chain of copies, where | |
196 // intermediate copies might be illegal, i.e., value is stored down to stack | |
197 // then reloaded BUT survives in a register the whole way. | |
198 Node *val = skip_copies(n->in(k)); | |
199 | |
200 if( val == x ) return blk_adjust; // No progress? | |
201 | |
202 bool single = is_single_register(val->ideal_reg()); | |
203 uint val_idx = n2lidx(val); | |
204 OptoReg::Name val_reg = lrgs(val_idx).reg(); | |
205 | |
206 // See if it happens to already be in the correct register! | |
207 // (either Phi's direct register, or the common case of the name | |
208 // never-clobbered original-def register) | |
209 if( value[val_reg] == val && | |
210 // Doubles check both halves | |
211 ( single || value[val_reg-1] == val ) ) { | |
212 blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd); | |
213 if( n->in(k) == regnd[val_reg] ) // Success! Quit trying | |
214 return blk_adjust; | |
215 } | |
216 | |
217 // See if we can skip the copy by changing registers. Don't change from | |
218 // using a register to using the stack unless we know we can remove a | |
219 // copy-load. Otherwise we might end up making a pile of Intel cisc-spill | |
220 // ops reading from memory instead of just loading once and using the | |
221 // register. | |
222 | |
223 // Also handle duplicate copies here. | |
224 const Type *t = val->is_Con() ? val->bottom_type() : NULL; | |
225 | |
226 // Scan all registers to see if this value is around already | |
227 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
|
228 if (reg == (uint)nk_reg) { |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
229 // 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
|
230 // 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
|
231 bool ignore_self = true; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
232 x = n->in(k); |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
233 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
|
234 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
|
235 while (i < imax && ignore_self) { |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
236 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
|
237 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
|
238 } |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
239 if (ignore_self) continue; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
240 } |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
196
diff
changeset
|
241 |
0 | 242 Node *vv = value[reg]; |
243 if( !single ) { // Doubles check for aligned-adjacent pair | |
244 if( (reg&1)==0 ) continue; // Wrong half of a pair | |
245 if( vv != value[reg-1] ) continue; // Not a complete pair | |
246 } | |
247 if( vv == val || // Got a direct hit? | |
248 (t && vv && vv->bottom_type() == t && vv->is_Mach() && | |
249 vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant? | |
250 assert( !n->is_Phi(), "cannot change registers at a Phi so easily" ); | |
251 if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR | |
252 OptoReg::is_reg(reg) || // turning into a register use OR | |
253 regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use | |
254 blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd); | |
255 if( n->in(k) == regnd[reg] ) // Success! Quit trying | |
256 return blk_adjust; | |
257 } // End of if not degrading to a stack | |
258 } // End of if found value in another register | |
259 } // End of scan all machine registers | |
260 return blk_adjust; | |
261 } | |
262 | |
263 | |
264 // | |
265 // Check if nreg already contains the constant value val. Normal copy | |
266 // elimination doesn't doesn't work on constants because multiple | |
267 // nodes can represent the same constant so the type and rule of the | |
268 // MachNode must be checked to ensure equivalence. | |
269 // | |
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
270 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
|
271 Block *current_block, |
0 | 272 Node_List& value, Node_List& regnd, |
273 OptoReg::Name nreg, OptoReg::Name nreg2) { | |
274 if (value[nreg] != val && val->is_Con() && | |
275 value[nreg] != NULL && value[nreg]->is_Con() && | |
276 (nreg2 == OptoReg::Bad || value[nreg] == value[nreg2]) && | |
277 value[nreg]->bottom_type() == val->bottom_type() && | |
278 value[nreg]->as_Mach()->rule() == val->as_Mach()->rule()) { | |
279 // This code assumes that two MachNodes representing constants | |
280 // which have the same rule and the same bottom type will produce | |
281 // identical effects into a register. This seems like it must be | |
282 // objectively true unless there are hidden inputs to the nodes | |
283 // but if that were to change this code would need to updated. | |
284 // Since they are equivalent the second one if redundant and can | |
285 // be removed. | |
286 // | |
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
287 // n will be replaced with the old value but n might have |
0 | 288 // kills projections associated with it so remove them now so that |
289 // yank_if_dead will be able to elminate the copy once the uses | |
290 // 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
|
291 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
|
292 Node* use = n->fast_out(i); |
0 | 293 if (use->is_Proj() && use->outcnt() == 0) { |
294 // Kill projections have no users and one input | |
295 use->set_req(0, C->top()); | |
296 yank_if_dead(use, current_block, &value, ®nd); | |
297 --i; --imax; | |
298 } | |
299 } | |
300 _post_alloc++; | |
301 return true; | |
302 } | |
303 return false; | |
304 } | |
305 | |
306 | |
307 //------------------------------post_allocate_copy_removal--------------------- | |
308 // Post-Allocation peephole copy removal. We do this in 1 pass over the | |
309 // basic blocks. We maintain a mapping of registers to Nodes (an array of | |
310 // Nodes indexed by machine register or stack slot number). NULL means that a | |
311 // register is not mapped to any Node. We can (want to have!) have several | |
312 // registers map to the same Node. We walk forward over the instructions | |
313 // updating the mapping as we go. At merge points we force a NULL if we have | |
314 // to merge 2 different Nodes into the same register. Phi functions will give | |
315 // us a new Node if there is a proper value merging. Since the blocks are | |
316 // arranged in some RPO, we will visit all parent blocks before visiting any | |
317 // successor blocks (except at loops). | |
318 // | |
319 // If we find a Copy we look to see if the Copy's source register is a stack | |
320 // slot and that value has already been loaded into some machine register; if | |
321 // so we use machine register directly. This turns a Load into a reg-reg | |
322 // Move. We also look for reloads of identical constants. | |
323 // | |
324 // When we see a use from a reg-reg Copy, we will attempt to use the copy's | |
325 // source directly and make the copy go dead. | |
326 void PhaseChaitin::post_allocate_copy_removal() { | |
327 NOT_PRODUCT( Compile::TracePhase t3("postAllocCopyRemoval", &_t_postAllocCopyRemoval, TimeCompiler); ) | |
328 ResourceMark rm; | |
329 | |
330 // Need a mapping from basic block Node_Lists. We need a Node_List to | |
331 // map from register number to value-producing Node. | |
332 Node_List **blk2value = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); | |
333 memset( blk2value, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); | |
334 // Need a mapping from basic block Node_Lists. We need a Node_List to | |
335 // map from register number to register-defining Node. | |
336 Node_List **blk2regnd = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); | |
337 memset( blk2regnd, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); | |
338 | |
339 // We keep unused Node_Lists on a free_list to avoid wasting | |
340 // memory. | |
341 GrowableArray<Node_List*> free_list = GrowableArray<Node_List*>(16); | |
342 | |
343 // For all blocks | |
344 for( uint i = 0; i < _cfg._num_blocks; i++ ) { | |
345 uint j; | |
346 Block *b = _cfg._blocks[i]; | |
347 | |
348 // Count of Phis in block | |
349 uint phi_dex; | |
350 for( phi_dex = 1; phi_dex < b->_nodes.size(); phi_dex++ ) { | |
351 Node *phi = b->_nodes[phi_dex]; | |
352 if( !phi->is_Phi() ) | |
353 break; | |
354 } | |
355 | |
356 // If any predecessor has not been visited, we do not know the state | |
357 // of registers at the start. Check for this, while updating copies | |
358 // along Phi input edges | |
359 bool missing_some_inputs = false; | |
360 Block *freed = NULL; | |
361 for( j = 1; j < b->num_preds(); j++ ) { | |
362 Block *pb = _cfg._bbs[b->pred(j)->_idx]; | |
363 // Remove copies along phi edges | |
364 for( uint k=1; k<phi_dex; k++ ) | |
365 elide_copy( b->_nodes[k], j, b, *blk2value[pb->_pre_order], *blk2regnd[pb->_pre_order], false ); | |
366 if( blk2value[pb->_pre_order] ) { // Have a mapping on this edge? | |
367 // See if this predecessor's mappings have been used by everybody | |
368 // who wants them. If so, free 'em. | |
369 uint k; | |
370 for( k=0; k<pb->_num_succs; k++ ) { | |
371 Block *pbsucc = pb->_succs[k]; | |
372 if( !blk2value[pbsucc->_pre_order] && pbsucc != b ) | |
373 break; // Found a future user | |
374 } | |
375 if( k >= pb->_num_succs ) { // No more uses, free! | |
376 freed = pb; // Record last block freed | |
377 free_list.push(blk2value[pb->_pre_order]); | |
378 free_list.push(blk2regnd[pb->_pre_order]); | |
379 } | |
380 } else { // This block has unvisited (loopback) inputs | |
381 missing_some_inputs = true; | |
382 } | |
383 } | |
384 | |
385 | |
386 // Extract Node_List mappings. If 'freed' is non-zero, we just popped | |
387 // 'freed's blocks off the list | |
388 Node_List ®nd = *(free_list.is_empty() ? new Node_List() : free_list.pop()); | |
389 Node_List &value = *(free_list.is_empty() ? new Node_List() : free_list.pop()); | |
390 assert( !freed || blk2value[freed->_pre_order] == &value, "" ); | |
391 value.map(_max_reg,NULL); | |
392 regnd.map(_max_reg,NULL); | |
393 // Set mappings as OUR mappings | |
394 blk2value[b->_pre_order] = &value; | |
395 blk2regnd[b->_pre_order] = ®nd; | |
396 | |
397 // Initialize value & regnd for this block | |
398 if( missing_some_inputs ) { | |
399 // Some predecessor has not yet been visited; zap map to empty | |
400 for( uint k = 0; k < (uint)_max_reg; k++ ) { | |
401 value.map(k,NULL); | |
402 regnd.map(k,NULL); | |
403 } | |
404 } else { | |
405 if( !freed ) { // Didn't get a freebie prior block | |
406 // Must clone some data | |
407 freed = _cfg._bbs[b->pred(1)->_idx]; | |
408 Node_List &f_value = *blk2value[freed->_pre_order]; | |
409 Node_List &f_regnd = *blk2regnd[freed->_pre_order]; | |
410 for( uint k = 0; k < (uint)_max_reg; k++ ) { | |
411 value.map(k,f_value[k]); | |
412 regnd.map(k,f_regnd[k]); | |
413 } | |
414 } | |
415 // Merge all inputs together, setting to NULL any conflicts. | |
416 for( j = 1; j < b->num_preds(); j++ ) { | |
417 Block *pb = _cfg._bbs[b->pred(j)->_idx]; | |
418 if( pb == freed ) continue; // Did self already via freelist | |
419 Node_List &p_regnd = *blk2regnd[pb->_pre_order]; | |
420 for( uint k = 0; k < (uint)_max_reg; k++ ) { | |
421 if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs? | |
422 value.map(k,NULL); // Then no value handy | |
423 regnd.map(k,NULL); | |
424 } | |
425 } | |
426 } | |
427 } | |
428 | |
429 // For all Phi's | |
430 for( j = 1; j < phi_dex; j++ ) { | |
431 uint k; | |
432 Node *phi = b->_nodes[j]; | |
433 uint pidx = n2lidx(phi); | |
434 OptoReg::Name preg = lrgs(n2lidx(phi)).reg(); | |
435 | |
436 // Remove copies remaining on edges. Check for junk phi. | |
437 Node *u = NULL; | |
438 for( k=1; k<phi->req(); k++ ) { | |
439 Node *x = phi->in(k); | |
440 if( phi != x && u != x ) // Found a different input | |
441 u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input | |
442 } | |
443 if( u != NodeSentinel ) { // Junk Phi. Remove | |
444 b->_nodes.remove(j--); phi_dex--; | |
445 _cfg._bbs.map(phi->_idx,NULL); | |
446 phi->replace_by(u); | |
447 phi->disconnect_inputs(NULL); | |
448 continue; | |
449 } | |
450 // Note that if value[pidx] exists, then we merged no new values here | |
451 // and the phi is useless. This can happen even with the above phi | |
452 // removal for complex flows. I cannot keep the better known value here | |
453 // because locally the phi appears to define a new merged value. If I | |
454 // keep the better value then a copy of the phi, being unable to use the | |
455 // global flow analysis, can't "peek through" the phi to the original | |
456 // reaching value and so will act like it's defining a new value. This | |
457 // can lead to situations where some uses are from the old and some from | |
458 // the new values. Not illegal by itself but throws the over-strong | |
459 // assert in scheduling. | |
460 if( pidx ) { | |
461 value.map(preg,phi); | |
462 regnd.map(preg,phi); | |
463 OptoReg::Name preg_lo = OptoReg::add(preg,-1); | |
464 if( !is_single_register(phi->ideal_reg()) ) { | |
465 value.map(preg_lo,phi); | |
466 regnd.map(preg_lo,phi); | |
467 } | |
468 } | |
469 } | |
470 | |
471 // For all remaining instructions | |
472 for( j = phi_dex; j < b->_nodes.size(); j++ ) { | |
473 Node *n = b->_nodes[j]; | |
474 | |
475 if( n->outcnt() == 0 && // Dead? | |
476 n != C->top() && // (ignore TOP, it has no du info) | |
477 !n->is_Proj() ) { // fat-proj kills | |
478 j -= yank_if_dead(n,b,&value,®nd); | |
479 continue; | |
480 } | |
481 | |
482 // Improve reaching-def info. Occasionally post-alloc's liveness gives | |
483 // up (at loop backedges, because we aren't doing a full flow pass). | |
484 // The presence of a live use essentially asserts that the use's def is | |
485 // alive and well at the use (or else the allocator fubar'd). Take | |
486 // advantage of this info to set a reaching def for the use-reg. | |
487 uint k; | |
488 for( k = 1; k < n->req(); k++ ) { | |
489 Node *def = n->in(k); // n->in(k) is a USE; def is the DEF for this USE | |
490 guarantee(def != NULL, "no disconnected nodes at this point"); | |
491 uint useidx = n2lidx(def); // useidx is the live range index for this USE | |
492 | |
493 if( useidx ) { | |
494 OptoReg::Name ureg = lrgs(useidx).reg(); | |
495 if( !value[ureg] ) { | |
496 int idx; // Skip occasional useless copy | |
497 while( (idx=def->is_Copy()) != 0 && | |
498 def->in(idx) != NULL && // NULL should not happen | |
499 ureg == lrgs(n2lidx(def->in(idx))).reg() ) | |
500 def = def->in(idx); | |
501 Node *valdef = skip_copies(def); // tighten up val through non-useless copies | |
502 value.map(ureg,valdef); // record improved reaching-def info | |
503 regnd.map(ureg, def); | |
504 // Record other half of doubles | |
505 OptoReg::Name ureg_lo = OptoReg::add(ureg,-1); | |
506 if( !is_single_register(def->ideal_reg()) && | |
507 ( !RegMask::can_represent(ureg_lo) || | |
508 lrgs(useidx).mask().Member(ureg_lo) ) && // Nearly always adjacent | |
509 !value[ureg_lo] ) { | |
510 value.map(ureg_lo,valdef); // record improved reaching-def info | |
511 regnd.map(ureg_lo, def); | |
512 } | |
513 } | |
514 } | |
515 } | |
516 | |
517 const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0; | |
518 | |
519 // Remove copies along input edges | |
520 for( k = 1; k < n->req(); k++ ) | |
521 j -= elide_copy( n, k, b, value, regnd, two_adr!=k ); | |
522 | |
523 // Unallocated Nodes define no registers | |
524 uint lidx = n2lidx(n); | |
525 if( !lidx ) continue; | |
526 | |
527 // Update the register defined by this instruction | |
528 OptoReg::Name nreg = lrgs(lidx).reg(); | |
529 // Skip through all copies to the _value_ being defined. | |
530 // Do not change from int to pointer | |
531 Node *val = skip_copies(n); | |
532 | |
533 uint n_ideal_reg = n->ideal_reg(); | |
534 if( is_single_register(n_ideal_reg) ) { | |
535 // If Node 'n' does not change the value mapped by the register, | |
536 // then 'n' is a useless copy. Do not update the register->node | |
537 // mapping so 'n' will go dead. | |
538 if( value[nreg] != val ) { | |
70
b683f557224b
6661247: Internal bug in 32-bit HotSpot optimizer while bit manipulations
never
parents:
0
diff
changeset
|
539 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, OptoReg::Bad)) { |
0 | 540 n->replace_by(regnd[nreg]); |
541 j -= yank_if_dead(n,b,&value,®nd); | |
542 } else { | |
543 // Update the mapping: record new Node defined by the register | |
544 regnd.map(nreg,n); | |
545 // Update mapping for defined *value*, which is the defined | |
546 // Node after skipping all copies. | |
547 value.map(nreg,val); | |
548 } | |
549 } else if( !may_be_copy_of_callee(n) && regnd[nreg]->outcnt() != 0 ) { | |
550 assert( n->is_Copy(), "" ); | |
551 n->replace_by(regnd[nreg]); | |
552 j -= yank_if_dead(n,b,&value,®nd); | |
553 } | |
554 } else { | |
555 // If the value occupies a register pair, record same info | |
556 // in both registers. | |
557 OptoReg::Name nreg_lo = OptoReg::add(nreg,-1); | |
558 if( RegMask::can_represent(nreg_lo) && // Either a spill slot, or | |
559 !lrgs(lidx).mask().Member(nreg_lo) ) { // Nearly always adjacent | |
560 // Sparc occasionally has non-adjacent pairs. | |
561 // Find the actual other value | |
562 RegMask tmp = lrgs(lidx).mask(); | |
563 tmp.Remove(nreg); | |
564 nreg_lo = tmp.find_first_elem(); | |
565 } | |
566 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
|
567 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, nreg_lo)) { |
0 | 568 n->replace_by(regnd[nreg]); |
569 j -= yank_if_dead(n,b,&value,®nd); | |
570 } else { | |
571 regnd.map(nreg , n ); | |
572 regnd.map(nreg_lo, n ); | |
573 value.map(nreg ,val); | |
574 value.map(nreg_lo,val); | |
575 } | |
576 } else if( !may_be_copy_of_callee(n) && regnd[nreg]->outcnt() != 0 ) { | |
577 assert( n->is_Copy(), "" ); | |
578 n->replace_by(regnd[nreg]); | |
579 j -= yank_if_dead(n,b,&value,®nd); | |
580 } | |
581 } | |
582 | |
583 // Fat projections kill many registers | |
584 if( n_ideal_reg == MachProjNode::fat_proj ) { | |
585 RegMask rm = n->out_RegMask(); | |
586 // wow, what an expensive iterator... | |
587 nreg = rm.find_first_elem(); | |
588 while( OptoReg::is_valid(nreg)) { | |
589 rm.Remove(nreg); | |
590 value.map(nreg,n); | |
591 regnd.map(nreg,n); | |
592 nreg = rm.find_first_elem(); | |
593 } | |
594 } | |
595 | |
596 } // End of for all instructions in the block | |
597 | |
598 } // End for all blocks | |
599 } |