Mercurial > hg > graal-compiler
annotate src/share/vm/opto/reg_split.cpp @ 3992:d1bdeef3e3e2
7098282: G1: assert(interval >= 0) failed: Sanity check, referencePolicy.cpp: 76
Summary: There is a race between one thread successfully forwarding and copying the klass mirror for the SoftReference class (including the static master clock) and another thread attempting to use the master clock while attempting to discover a soft reference object. Maintain a shadow copy of the soft reference master clock and use the shadow during reference discovery and reference processing.
Reviewed-by: tonyp, brutisso, ysr
author | johnc |
---|---|
date | Wed, 12 Oct 2011 10:25:51 -0700 |
parents | c7b60b601eb4 |
children | 8c92982cbbc4 |
rev | line source |
---|---|
0 | 1 /* |
1972 | 2 * Copyright (c) 2000, 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:
628
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
628
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:
628
diff
changeset
|
21 * questions. |
0 | 22 * |
23 */ | |
24 | |
1972 | 25 #include "precompiled.hpp" |
26 #include "libadt/vectset.hpp" | |
27 #include "memory/allocation.inline.hpp" | |
28 #include "opto/addnode.hpp" | |
29 #include "opto/c2compiler.hpp" | |
30 #include "opto/callnode.hpp" | |
31 #include "opto/cfgnode.hpp" | |
32 #include "opto/chaitin.hpp" | |
33 #include "opto/loopnode.hpp" | |
34 #include "opto/machnode.hpp" | |
0 | 35 |
36 //------------------------------Split-------------------------------------- | |
605 | 37 // Walk the graph in RPO and for each lrg which spills, propagate reaching |
38 // definitions. During propagation, split the live range around regions of | |
0 | 39 // High Register Pressure (HRP). If a Def is in a region of Low Register |
40 // Pressure (LRP), it will not get spilled until we encounter a region of | |
41 // HRP between it and one of its uses. We will spill at the transition | |
42 // point between LRP and HRP. Uses in the HRP region will use the spilled | |
43 // Def. The first Use outside the HRP region will generate a SpillCopy to | |
44 // hoist the live range back up into a register, and all subsequent uses | |
45 // will use that new Def until another HRP region is encountered. Defs in | |
46 // HRP regions will get trailing SpillCopies to push the LRG down into the | |
47 // stack immediately. | |
48 // | |
49 // As a side effect, unlink from (hence make dead) coalesced copies. | |
50 // | |
51 | |
52 static const char out_of_nodes[] = "out of nodes during split"; | |
53 | |
54 //------------------------------get_spillcopy_wide----------------------------- | |
55 // Get a SpillCopy node with wide-enough masks. Use the 'wide-mask', the | |
56 // wide ideal-register spill-mask if possible. If the 'wide-mask' does | |
57 // not cover the input (or output), use the input (or output) mask instead. | |
58 Node *PhaseChaitin::get_spillcopy_wide( Node *def, Node *use, uint uidx ) { | |
59 // If ideal reg doesn't exist we've got a bad schedule happening | |
60 // that is forcing us to spill something that isn't spillable. | |
61 // Bail rather than abort | |
62 int ireg = def->ideal_reg(); | |
63 if( ireg == 0 || ireg == Op_RegFlags ) { | |
415
4d9884b01ba6
6754519: don't emit flag fixup for NaN when condition being tested doesn't need it
never
parents:
337
diff
changeset
|
64 assert(false, "attempted to spill a non-spillable item"); |
0 | 65 C->record_method_not_compilable("attempted to spill a non-spillable item"); |
66 return NULL; | |
67 } | |
68 if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { | |
69 return NULL; | |
70 } | |
71 const RegMask *i_mask = &def->out_RegMask(); | |
72 const RegMask *w_mask = C->matcher()->idealreg2spillmask[ireg]; | |
73 const RegMask *o_mask = use ? &use->in_RegMask(uidx) : w_mask; | |
74 const RegMask *w_i_mask = w_mask->overlap( *i_mask ) ? w_mask : i_mask; | |
75 const RegMask *w_o_mask; | |
76 | |
77 if( w_mask->overlap( *o_mask ) && // Overlap AND | |
78 ((ireg != Op_RegL && ireg != Op_RegD // Single use or aligned | |
79 #ifdef _LP64 | |
80 && ireg != Op_RegP | |
81 #endif | |
82 ) || o_mask->is_aligned_Pairs()) ) { | |
83 // Don't come here for mis-aligned doubles | |
84 w_o_mask = w_mask; | |
85 } else { // wide ideal mask does not overlap with o_mask | |
86 // Mis-aligned doubles come here and XMM->FPR moves on x86. | |
87 w_o_mask = o_mask; // Must target desired registers | |
88 // Does the ideal-reg-mask overlap with o_mask? I.e., can I use | |
89 // a reg-reg move or do I need a trip across register classes | |
90 // (and thus through memory)? | |
91 if( !C->matcher()->idealreg2regmask[ireg]->overlap( *o_mask) && o_mask->is_UP() ) | |
92 // Here we assume a trip through memory is required. | |
93 w_i_mask = &C->FIRST_STACK_mask(); | |
94 } | |
95 return new (C) MachSpillCopyNode( def, *w_i_mask, *w_o_mask ); | |
96 } | |
97 | |
98 //------------------------------insert_proj------------------------------------ | |
605 | 99 // Insert the spill at chosen location. Skip over any intervening Proj's or |
0 | 100 // Phis. Skip over a CatchNode and projs, inserting in the fall-through block |
101 // instead. Update high-pressure indices. Create a new live range. | |
102 void PhaseChaitin::insert_proj( Block *b, uint i, Node *spill, uint maxlrg ) { | |
103 // Skip intervening ProjNodes. Do not insert between a ProjNode and | |
104 // its definer. | |
105 while( i < b->_nodes.size() && | |
106 (b->_nodes[i]->is_Proj() || | |
566
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
550
diff
changeset
|
107 b->_nodes[i]->is_Phi() ) ) |
0 | 108 i++; |
109 | |
110 // Do not insert between a call and his Catch | |
111 if( b->_nodes[i]->is_Catch() ) { | |
112 // Put the instruction at the top of the fall-thru block. | |
113 // Find the fall-thru projection | |
114 while( 1 ) { | |
115 const CatchProjNode *cp = b->_nodes[++i]->as_CatchProj(); | |
116 if( cp->_con == CatchProjNode::fall_through_index ) | |
117 break; | |
118 } | |
119 int sidx = i - b->end_idx()-1; | |
120 b = b->_succs[sidx]; // Switch to successor block | |
121 i = 1; // Right at start of block | |
122 } | |
123 | |
124 b->_nodes.insert(i,spill); // Insert node in block | |
125 _cfg._bbs.map(spill->_idx,b); // Update node->block mapping to reflect | |
126 // Adjust the point where we go hi-pressure | |
127 if( i <= b->_ihrp_index ) b->_ihrp_index++; | |
128 if( i <= b->_fhrp_index ) b->_fhrp_index++; | |
129 | |
130 // Assign a new Live Range Number to the SpillCopy and grow | |
131 // the node->live range mapping. | |
132 new_lrg(spill,maxlrg); | |
133 } | |
134 | |
135 //------------------------------split_DEF-------------------------------------- | |
605 | 136 // There are four categories of Split; UP/DOWN x DEF/USE |
0 | 137 // Only three of these really occur as DOWN/USE will always color |
138 // Any Split with a DEF cannot CISC-Spill now. Thus we need | |
139 // two helper routines, one for Split DEFS (insert after instruction), | |
140 // one for Split USES (insert before instruction). DEF insertion | |
141 // happens inside Split, where the Leaveblock array is updated. | |
142 uint PhaseChaitin::split_DEF( Node *def, Block *b, int loc, uint maxlrg, Node **Reachblock, Node **debug_defs, GrowableArray<uint> splits, int slidx ) { | |
143 #ifdef ASSERT | |
144 // Increment the counter for this lrg | |
145 splits.at_put(slidx, splits.at(slidx)+1); | |
146 #endif | |
147 // If we are spilling the memory op for an implicit null check, at the | |
148 // null check location (ie - null check is in HRP block) we need to do | |
149 // the null-check first, then spill-down in the following block. | |
150 // (The implicit_null_check function ensures the use is also dominated | |
151 // by the branch-not-taken block.) | |
152 Node *be = b->end(); | |
153 if( be->is_MachNullCheck() && be->in(1) == def && def == b->_nodes[loc] ) { | |
154 // Spill goes in the branch-not-taken block | |
155 b = b->_succs[b->_nodes[b->end_idx()+1]->Opcode() == Op_IfTrue]; | |
156 loc = 0; // Just past the Region | |
157 } | |
158 assert( loc >= 0, "must insert past block head" ); | |
159 | |
160 // Get a def-side SpillCopy | |
161 Node *spill = get_spillcopy_wide(def,NULL,0); | |
162 // Did we fail to split?, then bail | |
163 if (!spill) { | |
164 return 0; | |
165 } | |
166 | |
167 // Insert the spill at chosen location | |
168 insert_proj( b, loc+1, spill, maxlrg++); | |
169 | |
170 // Insert new node into Reaches array | |
171 Reachblock[slidx] = spill; | |
172 // Update debug list of reaching down definitions by adding this one | |
173 debug_defs[slidx] = spill; | |
174 | |
175 // return updated count of live ranges | |
176 return maxlrg; | |
177 } | |
178 | |
179 //------------------------------split_USE-------------------------------------- | |
180 // Splits at uses can involve redeffing the LRG, so no CISC Spilling there. | |
181 // Debug uses want to know if def is already stack enabled. | |
182 uint PhaseChaitin::split_USE( Node *def, Block *b, Node *use, uint useidx, uint maxlrg, bool def_down, bool cisc_sp, GrowableArray<uint> splits, int slidx ) { | |
183 #ifdef ASSERT | |
184 // Increment the counter for this lrg | |
185 splits.at_put(slidx, splits.at(slidx)+1); | |
186 #endif | |
187 | |
188 // Some setup stuff for handling debug node uses | |
189 JVMState* jvms = use->jvms(); | |
190 uint debug_start = jvms ? jvms->debug_start() : 999999; | |
191 uint debug_end = jvms ? jvms->debug_end() : 999999; | |
192 | |
193 //------------------------------------------- | |
194 // Check for use of debug info | |
195 if (useidx >= debug_start && useidx < debug_end) { | |
196 // Actually it's perfectly legal for constant debug info to appear | |
197 // just unlikely. In this case the optimizer left a ConI of a 4 | |
198 // as both inputs to a Phi with only a debug use. It's a single-def | |
199 // live range of a rematerializable value. The live range spills, | |
200 // rematerializes and now the ConI directly feeds into the debug info. | |
201 // assert(!def->is_Con(), "constant debug info already constructed directly"); | |
202 | |
203 // Special split handling for Debug Info | |
204 // If DEF is DOWN, just hook the edge and return | |
205 // If DEF is UP, Split it DOWN for this USE. | |
206 if( def->is_Mach() ) { | |
207 if( def_down ) { | |
208 // DEF is DOWN, so connect USE directly to the DEF | |
209 use->set_req(useidx, def); | |
210 } else { | |
211 // Block and index where the use occurs. | |
212 Block *b = _cfg._bbs[use->_idx]; | |
213 // Put the clone just prior to use | |
214 int bindex = b->find_node(use); | |
215 // DEF is UP, so must copy it DOWN and hook in USE | |
216 // Insert SpillCopy before the USE, which uses DEF as its input, | |
217 // and defs a new live range, which is used by this node. | |
218 Node *spill = get_spillcopy_wide(def,use,useidx); | |
219 // did we fail to split? | |
220 if (!spill) { | |
221 // Bail | |
222 return 0; | |
223 } | |
224 // insert into basic block | |
225 insert_proj( b, bindex, spill, maxlrg++ ); | |
226 // Use the new split | |
227 use->set_req(useidx,spill); | |
228 } | |
229 // No further split handling needed for this use | |
230 return maxlrg; | |
231 } // End special splitting for debug info live range | |
232 } // If debug info | |
233 | |
234 // CISC-SPILLING | |
235 // Finally, check to see if USE is CISC-Spillable, and if so, | |
236 // gather_lrg_masks will add the flags bit to its mask, and | |
237 // no use side copy is needed. This frees up the live range | |
238 // register choices without causing copy coalescing, etc. | |
239 if( UseCISCSpill && cisc_sp ) { | |
240 int inp = use->cisc_operand(); | |
241 if( inp != AdlcVMDeps::Not_cisc_spillable ) | |
242 // Convert operand number to edge index number | |
243 inp = use->as_Mach()->operand_index(inp); | |
244 if( inp == (int)useidx ) { | |
245 use->set_req(useidx, def); | |
246 #ifndef PRODUCT | |
247 if( TraceCISCSpill ) { | |
248 tty->print(" set_split: "); | |
249 use->dump(); | |
250 } | |
251 #endif | |
252 return maxlrg; | |
253 } | |
254 } | |
255 | |
256 //------------------------------------------- | |
257 // Insert a Copy before the use | |
258 | |
259 // Block and index where the use occurs. | |
260 int bindex; | |
261 // Phi input spill-copys belong at the end of the prior block | |
262 if( use->is_Phi() ) { | |
263 b = _cfg._bbs[b->pred(useidx)->_idx]; | |
264 bindex = b->end_idx(); | |
265 } else { | |
266 // Put the clone just prior to use | |
267 bindex = b->find_node(use); | |
268 } | |
269 | |
270 Node *spill = get_spillcopy_wide( def, use, useidx ); | |
271 if( !spill ) return 0; // Bailed out | |
272 // Insert SpillCopy before the USE, which uses the reaching DEF as | |
273 // its input, and defs a new live range, which is used by this node. | |
274 insert_proj( b, bindex, spill, maxlrg++ ); | |
275 // Use the spill/clone | |
276 use->set_req(useidx,spill); | |
277 | |
278 // return updated live range count | |
279 return maxlrg; | |
280 } | |
281 | |
1693
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
282 //------------------------------clone_node---------------------------- |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
283 // Clone node with anti dependence check. |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
284 Node* clone_node(Node* def, Block *b, Compile* C) { |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
285 if (def->needs_anti_dependence_check()) { |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
286 #ifdef ASSERT |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
287 if (Verbose) { |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
288 tty->print_cr("RA attempts to clone node with anti_dependence:"); |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
289 def->dump(-1); tty->cr(); |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
290 tty->print_cr("into block:"); |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
291 b->dump(); |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
292 } |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
293 #endif |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
294 if (C->subsume_loads() == true && !C->failing()) { |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
295 // Retry with subsume_loads == false |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
296 // If this is the first failure, the sentinel string will "stick" |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
297 // to the Compile object, and the C2Compiler will see it and retry. |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
298 C->record_failure(C2Compiler::retry_no_subsuming_loads()); |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
299 } else { |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
300 // Bailout without retry |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
301 C->record_method_not_compilable("RA Split failed: attempt to clone node with anti_dependence"); |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
302 } |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
303 return 0; |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
304 } |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
305 return def->clone(); |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
306 } |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
307 |
0 | 308 //------------------------------split_Rematerialize---------------------------- |
309 // Clone a local copy of the def. | |
310 Node *PhaseChaitin::split_Rematerialize( Node *def, Block *b, uint insidx, uint &maxlrg, GrowableArray<uint> splits, int slidx, uint *lrg2reach, Node **Reachblock, bool walkThru ) { | |
311 // The input live ranges will be stretched to the site of the new | |
312 // instruction. They might be stretched past a def and will thus | |
313 // have the old and new values of the same live range alive at the | |
314 // same time - a definite no-no. Split out private copies of | |
315 // the inputs. | |
316 if( def->req() > 1 ) { | |
317 for( uint i = 1; i < def->req(); i++ ) { | |
318 Node *in = def->in(i); | |
319 // Check for single-def (LRG cannot redefined) | |
320 uint lidx = n2lidx(in); | |
321 if( lidx >= _maxlrg ) continue; // Value is a recent spill-copy | |
295
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
322 if (lrgs(lidx).is_singledef()) continue; |
0 | 323 |
324 Block *b_def = _cfg._bbs[def->_idx]; | |
325 int idx_def = b_def->find_node(def); | |
326 Node *in_spill = get_spillcopy_wide( in, def, i ); | |
327 if( !in_spill ) return 0; // Bailed out | |
328 insert_proj(b_def,idx_def,in_spill,maxlrg++); | |
329 if( b_def == b ) | |
330 insidx++; | |
331 def->set_req(i,in_spill); | |
332 } | |
333 } | |
334 | |
1693
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
335 Node *spill = clone_node(def, b, C); |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
336 if (spill == NULL || C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { |
0 | 337 // Check when generating nodes |
338 return 0; | |
339 } | |
340 | |
341 // See if any inputs are currently being spilled, and take the | |
342 // latest copy of spilled inputs. | |
343 if( spill->req() > 1 ) { | |
344 for( uint i = 1; i < spill->req(); i++ ) { | |
345 Node *in = spill->in(i); | |
346 uint lidx = Find_id(in); | |
347 | |
348 // Walk backwards thru spill copy node intermediates | |
295
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
349 if (walkThru) { |
0 | 350 while ( in->is_SpillCopy() && lidx >= _maxlrg ) { |
351 in = in->in(1); | |
352 lidx = Find_id(in); | |
353 } | |
354 | |
295
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
355 if (lidx < _maxlrg && lrgs(lidx).is_multidef()) { |
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
356 // walkThru found a multidef LRG, which is unsafe to use, so |
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
357 // just keep the original def used in the clone. |
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
358 in = spill->in(i); |
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
359 lidx = Find_id(in); |
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
360 } |
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
361 } |
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
362 |
0 | 363 if( lidx < _maxlrg && lrgs(lidx).reg() >= LRG::SPILL_REG ) { |
364 Node *rdef = Reachblock[lrg2reach[lidx]]; | |
365 if( rdef ) spill->set_req(i,rdef); | |
366 } | |
367 } | |
368 } | |
369 | |
370 | |
371 assert( spill->out_RegMask().is_UP(), "rematerialize to a reg" ); | |
372 // Rematerialized op is def->spilled+1 | |
373 set_was_spilled(spill); | |
374 if( _spilled_once.test(def->_idx) ) | |
375 set_was_spilled(spill); | |
376 | |
377 insert_proj( b, insidx, spill, maxlrg++ ); | |
378 #ifdef ASSERT | |
379 // Increment the counter for this lrg | |
380 splits.at_put(slidx, splits.at(slidx)+1); | |
381 #endif | |
382 // See if the cloned def kills any flags, and copy those kills as well | |
383 uint i = insidx+1; | |
384 if( clone_projs( b, i, def, spill, maxlrg ) ) { | |
385 // Adjust the point where we go hi-pressure | |
386 if( i <= b->_ihrp_index ) b->_ihrp_index++; | |
387 if( i <= b->_fhrp_index ) b->_fhrp_index++; | |
388 } | |
389 | |
390 return spill; | |
391 } | |
392 | |
393 //------------------------------is_high_pressure------------------------------- | |
394 // Function to compute whether or not this live range is "high pressure" | |
395 // in this block - whether it spills eagerly or not. | |
396 bool PhaseChaitin::is_high_pressure( Block *b, LRG *lrg, uint insidx ) { | |
397 if( lrg->_was_spilled1 ) return true; | |
398 // Forced spilling due to conflict? Then split only at binding uses | |
399 // or defs, not for supposed capacity problems. | |
400 // CNC - Turned off 7/8/99, causes too much spilling | |
401 // if( lrg->_is_bound ) return false; | |
402 | |
403 // Not yet reached the high-pressure cutoff point, so low pressure | |
404 uint hrp_idx = lrg->_is_float ? b->_fhrp_index : b->_ihrp_index; | |
405 if( insidx < hrp_idx ) return false; | |
406 // Register pressure for the block as a whole depends on reg class | |
407 int block_pres = lrg->_is_float ? b->_freg_pressure : b->_reg_pressure; | |
408 // Bound live ranges will split at the binding points first; | |
409 // Intermediate splits should assume the live range's register set | |
410 // got "freed up" and that num_regs will become INT_PRESSURE. | |
411 int bound_pres = lrg->_is_float ? FLOATPRESSURE : INTPRESSURE; | |
412 // Effective register pressure limit. | |
413 int lrg_pres = (lrg->get_invalid_mask_size() > lrg->num_regs()) | |
414 ? (lrg->get_invalid_mask_size() >> (lrg->num_regs()-1)) : bound_pres; | |
415 // High pressure if block pressure requires more register freedom | |
416 // than live range has. | |
417 return block_pres >= lrg_pres; | |
418 } | |
419 | |
420 | |
421 //------------------------------prompt_use--------------------------------- | |
422 // True if lidx is used before any real register is def'd in the block | |
423 bool PhaseChaitin::prompt_use( Block *b, uint lidx ) { | |
424 if( lrgs(lidx)._was_spilled2 ) return false; | |
425 | |
426 // Scan block for 1st use. | |
427 for( uint i = 1; i <= b->end_idx(); i++ ) { | |
428 Node *n = b->_nodes[i]; | |
429 // Ignore PHI use, these can be up or down | |
430 if( n->is_Phi() ) continue; | |
431 for( uint j = 1; j < n->req(); j++ ) | |
432 if( Find_id(n->in(j)) == lidx ) | |
433 return true; // Found 1st use! | |
434 if( n->out_RegMask().is_NotEmpty() ) return false; | |
435 } | |
436 return false; | |
437 } | |
438 | |
439 //------------------------------Split-------------------------------------- | |
440 //----------Split Routine---------- | |
441 // ***** NEW SPLITTING HEURISTIC ***** | |
442 // DEFS: If the DEF is in a High Register Pressure(HRP) Block, split there. | |
443 // Else, no split unless there is a HRP block between a DEF and | |
444 // one of its uses, and then split at the HRP block. | |
445 // | |
446 // USES: If USE is in HRP, split at use to leave main LRG on stack. | |
447 // Else, hoist LRG back up to register only (ie - split is also DEF) | |
448 // We will compute a new maxlrg as we go | |
449 uint PhaseChaitin::Split( uint maxlrg ) { | |
450 NOT_PRODUCT( Compile::TracePhase t3("regAllocSplit", &_t_regAllocSplit, TimeCompiler); ) | |
451 | |
452 uint bidx, pidx, slidx, insidx, inpidx, twoidx; | |
453 uint non_phi = 1, spill_cnt = 0; | |
454 Node **Reachblock; | |
455 Node *n1, *n2, *n3; | |
456 Node_List *defs,*phis; | |
457 bool *UPblock; | |
458 bool u1, u2, u3; | |
459 Block *b, *pred; | |
460 PhiNode *phi; | |
461 GrowableArray<uint> lidxs; | |
462 | |
463 // Array of counters to count splits per live range | |
464 GrowableArray<uint> splits; | |
465 | |
466 //----------Setup Code---------- | |
467 // Create a convenient mapping from lrg numbers to reaches/leaves indices | |
468 uint *lrg2reach = NEW_RESOURCE_ARRAY( uint, _maxlrg ); | |
469 // Keep track of DEFS & Phis for later passes | |
470 defs = new Node_List(); | |
471 phis = new Node_List(); | |
472 // Gather info on which LRG's are spilling, and build maps | |
473 for( bidx = 1; bidx < _maxlrg; bidx++ ) { | |
474 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) { | |
475 assert(!lrgs(bidx).mask().is_AllStack(),"AllStack should color"); | |
476 lrg2reach[bidx] = spill_cnt; | |
477 spill_cnt++; | |
478 lidxs.append(bidx); | |
479 #ifdef ASSERT | |
480 // Initialize the split counts to zero | |
481 splits.append(0); | |
482 #endif | |
483 #ifndef PRODUCT | |
484 if( PrintOpto && WizardMode && lrgs(bidx)._was_spilled1 ) | |
485 tty->print_cr("Warning, 2nd spill of L%d",bidx); | |
486 #endif | |
487 } | |
488 } | |
489 | |
490 // Create side arrays for propagating reaching defs info. | |
491 // Each block needs a node pointer for each spilling live range for the | |
492 // Def which is live into the block. Phi nodes handle multiple input | |
493 // Defs by querying the output of their predecessor blocks and resolving | |
494 // them to a single Def at the phi. The pointer is updated for each | |
495 // Def in the block, and then becomes the output for the block when | |
496 // processing of the block is complete. We also need to track whether | |
497 // a Def is UP or DOWN. UP means that it should get a register (ie - | |
498 // it is always in LRP regions), and DOWN means that it is probably | |
499 // on the stack (ie - it crosses HRP regions). | |
500 Node ***Reaches = NEW_RESOURCE_ARRAY( Node**, _cfg._num_blocks+1 ); | |
501 bool **UP = NEW_RESOURCE_ARRAY( bool*, _cfg._num_blocks+1 ); | |
502 Node **debug_defs = NEW_RESOURCE_ARRAY( Node*, spill_cnt ); | |
503 VectorSet **UP_entry= NEW_RESOURCE_ARRAY( VectorSet*, spill_cnt ); | |
504 | |
505 // Initialize Reaches & UP | |
506 for( bidx = 0; bidx < _cfg._num_blocks+1; bidx++ ) { | |
507 Reaches[bidx] = NEW_RESOURCE_ARRAY( Node*, spill_cnt ); | |
508 UP[bidx] = NEW_RESOURCE_ARRAY( bool, spill_cnt ); | |
509 Node **Reachblock = Reaches[bidx]; | |
510 bool *UPblock = UP[bidx]; | |
511 for( slidx = 0; slidx < spill_cnt; slidx++ ) { | |
512 UPblock[slidx] = true; // Assume they start in registers | |
513 Reachblock[slidx] = NULL; // Assume that no def is present | |
514 } | |
515 } | |
516 | |
517 // Initialize to array of empty vectorsets | |
518 for( slidx = 0; slidx < spill_cnt; slidx++ ) | |
519 UP_entry[slidx] = new VectorSet(Thread::current()->resource_area()); | |
520 | |
521 //----------PASS 1---------- | |
522 //----------Propagation & Node Insertion Code---------- | |
523 // Walk the Blocks in RPO for DEF & USE info | |
524 for( bidx = 0; bidx < _cfg._num_blocks; bidx++ ) { | |
525 | |
526 if (C->check_node_count(spill_cnt, out_of_nodes)) { | |
527 return 0; | |
528 } | |
529 | |
530 b = _cfg._blocks[bidx]; | |
531 // Reaches & UP arrays for this block | |
532 Reachblock = Reaches[b->_pre_order]; | |
533 UPblock = UP[b->_pre_order]; | |
534 // Reset counter of start of non-Phi nodes in block | |
535 non_phi = 1; | |
536 //----------Block Entry Handling---------- | |
537 // Check for need to insert a new phi | |
538 // Cycle through this block's predecessors, collecting Reaches | |
539 // info for each spilled LRG. If they are identical, no phi is | |
540 // needed. If they differ, check for a phi, and insert if missing, | |
541 // or update edges if present. Set current block's Reaches set to | |
542 // be either the phi's or the reaching def, as appropriate. | |
543 // If no Phi is needed, check if the LRG needs to spill on entry | |
544 // to the block due to HRP. | |
545 for( slidx = 0; slidx < spill_cnt; slidx++ ) { | |
546 // Grab the live range number | |
547 uint lidx = lidxs.at(slidx); | |
548 // Do not bother splitting or putting in Phis for single-def | |
549 // rematerialized live ranges. This happens alot to constants | |
550 // with long live ranges. | |
295
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
551 if( lrgs(lidx).is_singledef() && |
0 | 552 lrgs(lidx)._def->rematerialize() ) { |
553 // reset the Reaches & UP entries | |
554 Reachblock[slidx] = lrgs(lidx)._def; | |
555 UPblock[slidx] = true; | |
556 // Record following instruction in case 'n' rematerializes and | |
557 // kills flags | |
558 Block *pred1 = _cfg._bbs[b->pred(1)->_idx]; | |
559 continue; | |
560 } | |
561 | |
562 // Initialize needs_phi and needs_split | |
563 bool needs_phi = false; | |
564 bool needs_split = false; | |
330
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
565 bool has_phi = false; |
0 | 566 // Walk the predecessor blocks to check inputs for that live range |
567 // Grab predecessor block header | |
568 n1 = b->pred(1); | |
569 // Grab the appropriate reaching def info for inpidx | |
570 pred = _cfg._bbs[n1->_idx]; | |
571 pidx = pred->_pre_order; | |
572 Node **Ltmp = Reaches[pidx]; | |
573 bool *Utmp = UP[pidx]; | |
574 n1 = Ltmp[slidx]; | |
575 u1 = Utmp[slidx]; | |
576 // Initialize node for saving type info | |
577 n3 = n1; | |
578 u3 = u1; | |
579 | |
580 // Compare inputs to see if a Phi is needed | |
581 for( inpidx = 2; inpidx < b->num_preds(); inpidx++ ) { | |
582 // Grab predecessor block headers | |
583 n2 = b->pred(inpidx); | |
584 // Grab the appropriate reaching def info for inpidx | |
585 pred = _cfg._bbs[n2->_idx]; | |
586 pidx = pred->_pre_order; | |
587 Ltmp = Reaches[pidx]; | |
588 Utmp = UP[pidx]; | |
589 n2 = Ltmp[slidx]; | |
590 u2 = Utmp[slidx]; | |
591 // For each LRG, decide if a phi is necessary | |
592 if( n1 != n2 ) { | |
593 needs_phi = true; | |
594 } | |
595 // See if the phi has mismatched inputs, UP vs. DOWN | |
596 if( n1 && n2 && (u1 != u2) ) { | |
597 needs_split = true; | |
598 } | |
599 // Move n2/u2 to n1/u1 for next iteration | |
600 n1 = n2; | |
601 u1 = u2; | |
602 // Preserve a non-NULL predecessor for later type referencing | |
603 if( (n3 == NULL) && (n2 != NULL) ){ | |
604 n3 = n2; | |
605 u3 = u2; | |
606 } | |
607 } // End for all potential Phi inputs | |
608 | |
330
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
609 // check block for appropriate phinode & update edges |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
610 for( insidx = 1; insidx <= b->end_idx(); insidx++ ) { |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
611 n1 = b->_nodes[insidx]; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
612 // bail if this is not a phi |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
613 phi = n1->is_Phi() ? n1->as_Phi() : NULL; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
614 if( phi == NULL ) { |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
615 // Keep track of index of first non-PhiNode instruction in block |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
616 non_phi = insidx; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
617 // break out of the for loop as we have handled all phi nodes |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
618 break; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
619 } |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
620 // must be looking at a phi |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
621 if( Find_id(n1) == lidxs.at(slidx) ) { |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
622 // found the necessary phi |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
623 needs_phi = false; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
624 has_phi = true; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
625 // initialize the Reaches entry for this LRG |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
626 Reachblock[slidx] = phi; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
627 break; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
628 } // end if found correct phi |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
629 } // end for all phi's |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
630 |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
631 // If a phi is needed or exist, check for it |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
632 if( needs_phi || has_phi ) { |
0 | 633 // add new phinode if one not already found |
634 if( needs_phi ) { | |
635 // create a new phi node and insert it into the block | |
636 // type is taken from left over pointer to a predecessor | |
637 assert(n3,"No non-NULL reaching DEF for a Phi"); | |
638 phi = new (C, b->num_preds()) PhiNode(b->head(), n3->bottom_type()); | |
639 // initialize the Reaches entry for this LRG | |
640 Reachblock[slidx] = phi; | |
641 | |
642 // add node to block & node_to_block mapping | |
643 insert_proj( b, insidx++, phi, maxlrg++ ); | |
644 non_phi++; | |
645 // Reset new phi's mapping to be the spilling live range | |
646 _names.map(phi->_idx, lidx); | |
647 assert(Find_id(phi) == lidx,"Bad update on Union-Find mapping"); | |
648 } // end if not found correct phi | |
649 // Here you have either found or created the Phi, so record it | |
650 assert(phi != NULL,"Must have a Phi Node here"); | |
651 phis->push(phi); | |
652 // PhiNodes should either force the LRG UP or DOWN depending | |
653 // on its inputs and the register pressure in the Phi's block. | |
654 UPblock[slidx] = true; // Assume new DEF is UP | |
655 // If entering a high-pressure area with no immediate use, | |
656 // assume Phi is DOWN | |
657 if( is_high_pressure( b, &lrgs(lidx), b->end_idx()) && !prompt_use(b,lidx) ) | |
658 UPblock[slidx] = false; | |
659 // If we are not split up/down and all inputs are down, then we | |
660 // are down | |
661 if( !needs_split && !u3 ) | |
662 UPblock[slidx] = false; | |
663 } // end if phi is needed | |
664 | |
665 // Do not need a phi, so grab the reaching DEF | |
666 else { | |
667 // Grab predecessor block header | |
668 n1 = b->pred(1); | |
669 // Grab the appropriate reaching def info for k | |
670 pred = _cfg._bbs[n1->_idx]; | |
671 pidx = pred->_pre_order; | |
672 Node **Ltmp = Reaches[pidx]; | |
673 bool *Utmp = UP[pidx]; | |
674 // reset the Reaches & UP entries | |
675 Reachblock[slidx] = Ltmp[slidx]; | |
676 UPblock[slidx] = Utmp[slidx]; | |
677 } // end else no Phi is needed | |
678 } // end for all spilling live ranges | |
679 // DEBUG | |
680 #ifndef PRODUCT | |
681 if(trace_spilling()) { | |
682 tty->print("/`\nBlock %d: ", b->_pre_order); | |
683 tty->print("Reaching Definitions after Phi handling\n"); | |
684 for( uint x = 0; x < spill_cnt; x++ ) { | |
685 tty->print("Spill Idx %d: UP %d: Node\n",x,UPblock[x]); | |
686 if( Reachblock[x] ) | |
687 Reachblock[x]->dump(); | |
688 else | |
689 tty->print("Undefined\n"); | |
690 } | |
691 } | |
692 #endif | |
693 | |
694 //----------Non-Phi Node Splitting---------- | |
695 // Since phi-nodes have now been handled, the Reachblock array for this | |
696 // block is initialized with the correct starting value for the defs which | |
697 // reach non-phi instructions in this block. Thus, process non-phi | |
698 // instructions normally, inserting SpillCopy nodes for all spill | |
699 // locations. | |
700 | |
701 // Memoize any DOWN reaching definitions for use as DEBUG info | |
702 for( insidx = 0; insidx < spill_cnt; insidx++ ) { | |
703 debug_defs[insidx] = (UPblock[insidx]) ? NULL : Reachblock[insidx]; | |
704 if( UPblock[insidx] ) // Memoize UP decision at block start | |
705 UP_entry[insidx]->set( b->_pre_order ); | |
706 } | |
707 | |
708 //----------Walk Instructions in the Block and Split---------- | |
709 // For all non-phi instructions in the block | |
710 for( insidx = 1; insidx <= b->end_idx(); insidx++ ) { | |
711 Node *n = b->_nodes[insidx]; | |
712 // Find the defining Node's live range index | |
713 uint defidx = Find_id(n); | |
714 uint cnt = n->req(); | |
715 | |
716 if( n->is_Phi() ) { | |
717 // Skip phi nodes after removing dead copies. | |
718 if( defidx < _maxlrg ) { | |
719 // Check for useless Phis. These appear if we spill, then | |
720 // coalesce away copies. Dont touch Phis in spilling live | |
721 // ranges; they are busy getting modifed in this pass. | |
722 if( lrgs(defidx).reg() < LRG::SPILL_REG ) { | |
723 uint i; | |
724 Node *u = NULL; | |
725 // Look for the Phi merging 2 unique inputs | |
726 for( i = 1; i < cnt; i++ ) { | |
727 // Ignore repeats and self | |
728 if( n->in(i) != u && n->in(i) != n ) { | |
729 // Found a unique input | |
730 if( u != NULL ) // If it's the 2nd, bail out | |
731 break; | |
732 u = n->in(i); // Else record it | |
733 } | |
734 } | |
735 assert( u, "at least 1 valid input expected" ); | |
330
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
736 if( i >= cnt ) { // Found one unique input |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
737 assert(Find_id(n) == Find_id(u), "should be the same lrg"); |
0 | 738 n->replace_by(u); // Then replace with unique input |
739 n->disconnect_inputs(NULL); | |
740 b->_nodes.remove(insidx); | |
741 insidx--; | |
742 b->_ihrp_index--; | |
743 b->_fhrp_index--; | |
744 } | |
745 } | |
746 } | |
747 continue; | |
748 } | |
749 assert( insidx > b->_ihrp_index || | |
750 (b->_reg_pressure < (uint)INTPRESSURE) || | |
751 b->_ihrp_index > 4000000 || | |
752 b->_ihrp_index >= b->end_idx() || | |
753 !b->_nodes[b->_ihrp_index]->is_Proj(), "" ); | |
754 assert( insidx > b->_fhrp_index || | |
755 (b->_freg_pressure < (uint)FLOATPRESSURE) || | |
756 b->_fhrp_index > 4000000 || | |
757 b->_fhrp_index >= b->end_idx() || | |
758 !b->_nodes[b->_fhrp_index]->is_Proj(), "" ); | |
759 | |
760 // ********** Handle Crossing HRP Boundry ********** | |
761 if( (insidx == b->_ihrp_index) || (insidx == b->_fhrp_index) ) { | |
762 for( slidx = 0; slidx < spill_cnt; slidx++ ) { | |
605 | 763 // Check for need to split at HRP boundary - split if UP |
0 | 764 n1 = Reachblock[slidx]; |
765 // bail out if no reaching DEF | |
766 if( n1 == NULL ) continue; | |
767 // bail out if live range is 'isolated' around inner loop | |
768 uint lidx = lidxs.at(slidx); | |
769 // If live range is currently UP | |
770 if( UPblock[slidx] ) { | |
771 // set location to insert spills at | |
772 // SPLIT DOWN HERE - NO CISC SPILL | |
773 if( is_high_pressure( b, &lrgs(lidx), insidx ) && | |
774 !n1->rematerialize() ) { | |
775 // If there is already a valid stack definition available, use it | |
776 if( debug_defs[slidx] != NULL ) { | |
777 Reachblock[slidx] = debug_defs[slidx]; | |
778 } | |
779 else { | |
780 // Insert point is just past last use or def in the block | |
781 int insert_point = insidx-1; | |
782 while( insert_point > 0 ) { | |
783 Node *n = b->_nodes[insert_point]; | |
784 // Hit top of block? Quit going backwards | |
785 if( n->is_Phi() ) break; | |
786 // Found a def? Better split after it. | |
787 if( n2lidx(n) == lidx ) break; | |
788 // Look for a use | |
789 uint i; | |
790 for( i = 1; i < n->req(); i++ ) | |
791 if( n2lidx(n->in(i)) == lidx ) | |
792 break; | |
793 // Found a use? Better split after it. | |
794 if( i < n->req() ) break; | |
795 insert_point--; | |
796 } | |
797 maxlrg = split_DEF( n1, b, insert_point, maxlrg, Reachblock, debug_defs, splits, slidx); | |
798 // If it wasn't split bail | |
799 if (!maxlrg) { | |
800 return 0; | |
801 } | |
802 insidx++; | |
803 } | |
804 // This is a new DEF, so update UP | |
805 UPblock[slidx] = false; | |
806 #ifndef PRODUCT | |
807 // DEBUG | |
808 if( trace_spilling() ) { | |
809 tty->print("\nNew Split DOWN DEF of Spill Idx "); | |
810 tty->print("%d, UP %d:\n",slidx,false); | |
811 n1->dump(); | |
812 } | |
813 #endif | |
814 } | |
815 } // end if LRG is UP | |
816 } // end for all spilling live ranges | |
817 assert( b->_nodes[insidx] == n, "got insidx set incorrectly" ); | |
818 } // end if crossing HRP Boundry | |
819 | |
820 // If the LRG index is oob, then this is a new spillcopy, skip it. | |
821 if( defidx >= _maxlrg ) { | |
822 continue; | |
823 } | |
824 LRG &deflrg = lrgs(defidx); | |
825 uint copyidx = n->is_Copy(); | |
826 // Remove coalesced copy from CFG | |
827 if( copyidx && defidx == n2lidx(n->in(copyidx)) ) { | |
828 n->replace_by( n->in(copyidx) ); | |
829 n->set_req( copyidx, NULL ); | |
830 b->_nodes.remove(insidx--); | |
831 b->_ihrp_index--; // Adjust the point where we go hi-pressure | |
832 b->_fhrp_index--; | |
833 continue; | |
834 } | |
835 | |
836 #define DERIVED 0 | |
837 | |
838 // ********** Handle USES ********** | |
839 bool nullcheck = false; | |
840 // Implicit null checks never use the spilled value | |
841 if( n->is_MachNullCheck() ) | |
842 nullcheck = true; | |
843 if( !nullcheck ) { | |
844 // Search all inputs for a Spill-USE | |
845 JVMState* jvms = n->jvms(); | |
846 uint oopoff = jvms ? jvms->oopoff() : cnt; | |
847 uint old_last = cnt - 1; | |
848 for( inpidx = 1; inpidx < cnt; inpidx++ ) { | |
849 // Derived/base pairs may be added to our inputs during this loop. | |
850 // If inpidx > old_last, then one of these new inputs is being | |
851 // handled. Skip the derived part of the pair, but process | |
852 // the base like any other input. | |
853 if( inpidx > old_last && ((inpidx - oopoff) & 1) == DERIVED ) { | |
854 continue; // skip derived_debug added below | |
855 } | |
856 // Get lidx of input | |
857 uint useidx = Find_id(n->in(inpidx)); | |
858 // Not a brand-new split, and it is a spill use | |
859 if( useidx < _maxlrg && lrgs(useidx).reg() >= LRG::SPILL_REG ) { | |
860 // Check for valid reaching DEF | |
861 slidx = lrg2reach[useidx]; | |
862 Node *def = Reachblock[slidx]; | |
863 assert( def != NULL, "Using Undefined Value in Split()\n"); | |
864 | |
865 // (+++) %%%% remove this in favor of pre-pass in matcher.cpp | |
866 // monitor references do not care where they live, so just hook | |
867 if ( jvms && jvms->is_monitor_use(inpidx) ) { | |
868 // The effect of this clone is to drop the node out of the block, | |
869 // so that the allocator does not see it anymore, and therefore | |
870 // does not attempt to assign it a register. | |
1693
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
871 def = clone_node(def, b, C); |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
872 if (def == NULL || C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
873 return 0; |
6c9cc03d8726
6973329: C2 with Zero based COOP produces code with broken anti-dependency on x86
kvn
parents:
1552
diff
changeset
|
874 } |
0 | 875 _names.extend(def->_idx,0); |
876 _cfg._bbs.map(def->_idx,b); | |
877 n->set_req(inpidx, def); | |
878 continue; | |
879 } | |
880 | |
881 // Rematerializable? Then clone def at use site instead | |
882 // of store/load | |
883 if( def->rematerialize() ) { | |
884 int old_size = b->_nodes.size(); | |
885 def = split_Rematerialize( def, b, insidx, maxlrg, splits, slidx, lrg2reach, Reachblock, true ); | |
886 if( !def ) return 0; // Bail out | |
887 insidx += b->_nodes.size()-old_size; | |
888 } | |
889 | |
890 MachNode *mach = n->is_Mach() ? n->as_Mach() : NULL; | |
891 // Base pointers and oopmap references do not care where they live. | |
892 if ((inpidx >= oopoff) || | |
893 (mach && mach->ideal_Opcode() == Op_AddP && inpidx == AddPNode::Base)) { | |
894 if (def->rematerialize() && lrgs(useidx)._was_spilled2) { | |
895 // This def has been rematerialized a couple of times without | |
896 // progress. It doesn't care if it lives UP or DOWN, so | |
897 // spill it down now. | |
898 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false,splits,slidx); | |
899 // If it wasn't split bail | |
900 if (!maxlrg) { | |
901 return 0; | |
902 } | |
903 insidx++; // Reset iterator to skip USE side split | |
904 } else { | |
905 // Just hook the def edge | |
906 n->set_req(inpidx, def); | |
907 } | |
908 | |
909 if (inpidx >= oopoff) { | |
910 // After oopoff, we have derived/base pairs. We must mention all | |
911 // derived pointers here as derived/base pairs for GC. If the | |
912 // derived value is spilling and we have a copy both in Reachblock | |
913 // (called here 'def') and debug_defs[slidx] we need to mention | |
914 // both in derived/base pairs or kill one. | |
915 Node *derived_debug = debug_defs[slidx]; | |
916 if( ((inpidx - oopoff) & 1) == DERIVED && // derived vs base? | |
917 mach && mach->ideal_Opcode() != Op_Halt && | |
918 derived_debug != NULL && | |
919 derived_debug != def ) { // Actual 2nd value appears | |
920 // We have already set 'def' as a derived value. | |
921 // Also set debug_defs[slidx] as a derived value. | |
922 uint k; | |
923 for( k = oopoff; k < cnt; k += 2 ) | |
924 if( n->in(k) == derived_debug ) | |
925 break; // Found an instance of debug derived | |
926 if( k == cnt ) {// No instance of debug_defs[slidx] | |
927 // Add a derived/base pair to cover the debug info. | |
928 // We have to process the added base later since it is not | |
929 // handled yet at this point but skip derived part. | |
930 assert(((n->req() - oopoff) & 1) == DERIVED, | |
931 "must match skip condition above"); | |
932 n->add_req( derived_debug ); // this will be skipped above | |
933 n->add_req( n->in(inpidx+1) ); // this will be processed | |
934 // Increment cnt to handle added input edges on | |
935 // subsequent iterations. | |
936 cnt += 2; | |
937 } | |
938 } | |
939 } | |
940 continue; | |
941 } | |
942 // Special logic for DEBUG info | |
943 if( jvms && b->_freq > BLOCK_FREQUENCY(0.5) ) { | |
944 uint debug_start = jvms->debug_start(); | |
945 // If this is debug info use & there is a reaching DOWN def | |
946 if ((debug_start <= inpidx) && (debug_defs[slidx] != NULL)) { | |
947 assert(inpidx < oopoff, "handle only debug info here"); | |
948 // Just hook it in & move on | |
949 n->set_req(inpidx, debug_defs[slidx]); | |
950 // (Note that this can make two sides of a split live at the | |
951 // same time: The debug def on stack, and another def in a | |
952 // register. The GC needs to know about both of them, but any | |
953 // derived pointers after oopoff will refer to only one of the | |
954 // two defs and the GC would therefore miss the other. Thus | |
955 // this hack is only allowed for debug info which is Java state | |
956 // and therefore never a derived pointer.) | |
957 continue; | |
958 } | |
959 } | |
960 // Grab register mask info | |
961 const RegMask &dmask = def->out_RegMask(); | |
962 const RegMask &umask = n->in_RegMask(inpidx); | |
963 | |
964 assert(inpidx < oopoff, "cannot use-split oop map info"); | |
965 | |
966 bool dup = UPblock[slidx]; | |
967 bool uup = umask.is_UP(); | |
968 | |
969 // Need special logic to handle bound USES. Insert a split at this | |
970 // bound use if we can't rematerialize the def, or if we need the | |
971 // split to form a misaligned pair. | |
972 if( !umask.is_AllStack() && | |
973 (int)umask.Size() <= lrgs(useidx).num_regs() && | |
974 (!def->rematerialize() || | |
975 umask.is_misaligned_Pair())) { | |
976 // These need a Split regardless of overlap or pressure | |
977 // SPLIT - NO DEF - NO CISC SPILL | |
978 maxlrg = split_USE(def,b,n,inpidx,maxlrg,dup,false, splits,slidx); | |
979 // If it wasn't split bail | |
980 if (!maxlrg) { | |
981 return 0; | |
982 } | |
983 insidx++; // Reset iterator to skip USE side split | |
984 continue; | |
985 } | |
1730
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
986 |
3842 | 987 if (UseFPUForSpilling && n->is_MachCall() && !uup && !dup ) { |
1730
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
988 // The use at the call can force the def down so insert |
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
989 // a split before the use to allow the def more freedom. |
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
990 maxlrg = split_USE(def,b,n,inpidx,maxlrg,dup,false, splits,slidx); |
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
991 // If it wasn't split bail |
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
992 if (!maxlrg) { |
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
993 return 0; |
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
994 } |
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
995 insidx++; // Reset iterator to skip USE side split |
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
996 continue; |
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
997 } |
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1693
diff
changeset
|
998 |
0 | 999 // Here is the logic chart which describes USE Splitting: |
1000 // 0 = false or DOWN, 1 = true or UP | |
1001 // | |
1002 // Overlap | DEF | USE | Action | |
1003 //------------------------------------------------------- | |
1004 // 0 | 0 | 0 | Copy - mem -> mem | |
1005 // 0 | 0 | 1 | Split-UP - Check HRP | |
1006 // 0 | 1 | 0 | Split-DOWN - Debug Info? | |
1007 // 0 | 1 | 1 | Copy - reg -> reg | |
1008 // 1 | 0 | 0 | Reset Input Edge (no Split) | |
1009 // 1 | 0 | 1 | Split-UP - Check HRP | |
1010 // 1 | 1 | 0 | Split-DOWN - Debug Info? | |
1011 // 1 | 1 | 1 | Reset Input Edge (no Split) | |
1012 // | |
1013 // So, if (dup == uup), then overlap test determines action, | |
1014 // with true being no split, and false being copy. Else, | |
1015 // if DEF is DOWN, Split-UP, and check HRP to decide on | |
1016 // resetting DEF. Finally if DEF is UP, Split-DOWN, with | |
1017 // special handling for Debug Info. | |
1018 if( dup == uup ) { | |
1019 if( dmask.overlap(umask) ) { | |
1020 // Both are either up or down, and there is overlap, No Split | |
1021 n->set_req(inpidx, def); | |
1022 } | |
1023 else { // Both are either up or down, and there is no overlap | |
1024 if( dup ) { // If UP, reg->reg copy | |
1025 // COPY ACROSS HERE - NO DEF - NO CISC SPILL | |
1026 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false, splits,slidx); | |
1027 // If it wasn't split bail | |
1028 if (!maxlrg) { | |
1029 return 0; | |
1030 } | |
1031 insidx++; // Reset iterator to skip USE side split | |
1032 } | |
1033 else { // DOWN, mem->mem copy | |
1034 // COPY UP & DOWN HERE - NO DEF - NO CISC SPILL | |
1035 // First Split-UP to move value into Register | |
1036 uint def_ideal = def->ideal_reg(); | |
1037 const RegMask* tmp_rm = Matcher::idealreg2regmask[def_ideal]; | |
1038 Node *spill = new (C) MachSpillCopyNode(def, dmask, *tmp_rm); | |
1039 insert_proj( b, insidx, spill, maxlrg ); | |
1040 // Then Split-DOWN as if previous Split was DEF | |
1041 maxlrg = split_USE(spill,b,n,inpidx,maxlrg,false,false, splits,slidx); | |
1042 // If it wasn't split bail | |
1043 if (!maxlrg) { | |
1044 return 0; | |
1045 } | |
1046 insidx += 2; // Reset iterator to skip USE side splits | |
1047 } | |
1048 } // End else no overlap | |
1049 } // End if dup == uup | |
1050 // dup != uup, so check dup for direction of Split | |
1051 else { | |
1052 if( dup ) { // If UP, Split-DOWN and check Debug Info | |
1053 // If this node is already a SpillCopy, just patch the edge | |
1054 // except the case of spilling to stack. | |
1055 if( n->is_SpillCopy() ) { | |
1056 RegMask tmp_rm(umask); | |
1057 tmp_rm.SUBTRACT(Matcher::STACK_ONLY_mask); | |
1058 if( dmask.overlap(tmp_rm) ) { | |
1059 if( def != n->in(inpidx) ) { | |
1060 n->set_req(inpidx, def); | |
1061 } | |
1062 continue; | |
1063 } | |
1064 } | |
1065 // COPY DOWN HERE - NO DEF - NO CISC SPILL | |
1066 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false, splits,slidx); | |
1067 // If it wasn't split bail | |
1068 if (!maxlrg) { | |
1069 return 0; | |
1070 } | |
1071 insidx++; // Reset iterator to skip USE side split | |
1072 // Check for debug-info split. Capture it for later | |
1073 // debug splits of the same value | |
1074 if (jvms && jvms->debug_start() <= inpidx && inpidx < oopoff) | |
1075 debug_defs[slidx] = n->in(inpidx); | |
1076 | |
1077 } | |
1078 else { // DOWN, Split-UP and check register pressure | |
1079 if( is_high_pressure( b, &lrgs(useidx), insidx ) ) { | |
1080 // COPY UP HERE - NO DEF - CISC SPILL | |
1081 maxlrg = split_USE(def,b,n,inpidx,maxlrg,true,true, splits,slidx); | |
1082 // If it wasn't split bail | |
1083 if (!maxlrg) { | |
1084 return 0; | |
1085 } | |
1086 insidx++; // Reset iterator to skip USE side split | |
1087 } else { // LRP | |
1088 // COPY UP HERE - WITH DEF - NO CISC SPILL | |
1089 maxlrg = split_USE(def,b,n,inpidx,maxlrg,true,false, splits,slidx); | |
1090 // If it wasn't split bail | |
1091 if (!maxlrg) { | |
1092 return 0; | |
1093 } | |
1094 // Flag this lift-up in a low-pressure block as | |
1095 // already-spilled, so if it spills again it will | |
1096 // spill hard (instead of not spilling hard and | |
1097 // coalescing away). | |
1098 set_was_spilled(n->in(inpidx)); | |
1099 // Since this is a new DEF, update Reachblock & UP | |
1100 Reachblock[slidx] = n->in(inpidx); | |
1101 UPblock[slidx] = true; | |
1102 insidx++; // Reset iterator to skip USE side split | |
1103 } | |
1104 } // End else DOWN | |
1105 } // End dup != uup | |
1106 } // End if Spill USE | |
1107 } // End For All Inputs | |
1108 } // End If not nullcheck | |
1109 | |
1110 // ********** Handle DEFS ********** | |
1111 // DEFS either Split DOWN in HRP regions or when the LRG is bound, or | |
1112 // just reset the Reaches info in LRP regions. DEFS must always update | |
1113 // UP info. | |
1114 if( deflrg.reg() >= LRG::SPILL_REG ) { // Spilled? | |
1115 uint slidx = lrg2reach[defidx]; | |
1116 // Add to defs list for later assignment of new live range number | |
1117 defs->push(n); | |
1118 // Set a flag on the Node indicating it has already spilled. | |
1119 // Only do it for capacity spills not conflict spills. | |
1120 if( !deflrg._direct_conflict ) | |
1121 set_was_spilled(n); | |
1122 assert(!n->is_Phi(),"Cannot insert Phi into DEFS list"); | |
1123 // Grab UP info for DEF | |
1124 const RegMask &dmask = n->out_RegMask(); | |
1125 bool defup = dmask.is_UP(); | |
1126 // Only split at Def if this is a HRP block or bound (and spilled once) | |
1127 if( !n->rematerialize() && | |
1128 (((dmask.is_bound1() || dmask.is_bound2() || dmask.is_misaligned_Pair()) && | |
1129 (deflrg._direct_conflict || deflrg._must_spill)) || | |
1130 // Check for LRG being up in a register and we are inside a high | |
1131 // pressure area. Spill it down immediately. | |
1132 (defup && is_high_pressure(b,&deflrg,insidx))) ) { | |
1133 assert( !n->rematerialize(), "" ); | |
1134 assert( !n->is_SpillCopy(), "" ); | |
1135 // Do a split at the def site. | |
1136 maxlrg = split_DEF( n, b, insidx, maxlrg, Reachblock, debug_defs, splits, slidx ); | |
1137 // If it wasn't split bail | |
1138 if (!maxlrg) { | |
1139 return 0; | |
1140 } | |
1141 // Split DEF's Down | |
1142 UPblock[slidx] = 0; | |
1143 #ifndef PRODUCT | |
1144 // DEBUG | |
1145 if( trace_spilling() ) { | |
1146 tty->print("\nNew Split DOWN DEF of Spill Idx "); | |
1147 tty->print("%d, UP %d:\n",slidx,false); | |
1148 n->dump(); | |
1149 } | |
1150 #endif | |
1151 } | |
1152 else { // Neither bound nor HRP, must be LRP | |
1153 // otherwise, just record the def | |
1154 Reachblock[slidx] = n; | |
1155 // UP should come from the outRegmask() of the DEF | |
1156 UPblock[slidx] = defup; | |
1157 // Update debug list of reaching down definitions, kill if DEF is UP | |
1158 debug_defs[slidx] = defup ? NULL : n; | |
1159 #ifndef PRODUCT | |
1160 // DEBUG | |
1161 if( trace_spilling() ) { | |
1162 tty->print("\nNew DEF of Spill Idx "); | |
1163 tty->print("%d, UP %d:\n",slidx,defup); | |
1164 n->dump(); | |
1165 } | |
1166 #endif | |
1167 } // End else LRP | |
1168 } // End if spill def | |
1169 | |
1170 // ********** Split Left Over Mem-Mem Moves ********** | |
1171 // Check for mem-mem copies and split them now. Do not do this | |
1172 // to copies about to be spilled; they will be Split shortly. | |
1173 if( copyidx ) { | |
1174 Node *use = n->in(copyidx); | |
1175 uint useidx = Find_id(use); | |
1176 if( useidx < _maxlrg && // This is not a new split | |
1177 OptoReg::is_stack(deflrg.reg()) && | |
1178 deflrg.reg() < LRG::SPILL_REG ) { // And DEF is from stack | |
1179 LRG &uselrg = lrgs(useidx); | |
1180 if( OptoReg::is_stack(uselrg.reg()) && | |
1181 uselrg.reg() < LRG::SPILL_REG && // USE is from stack | |
1182 deflrg.reg() != uselrg.reg() ) { // Not trivially removed | |
1183 uint def_ideal_reg = Matcher::base2reg[n->bottom_type()->base()]; | |
1184 const RegMask &def_rm = *Matcher::idealreg2regmask[def_ideal_reg]; | |
1185 const RegMask &use_rm = n->in_RegMask(copyidx); | |
1186 if( def_rm.overlap(use_rm) && n->is_SpillCopy() ) { // Bug 4707800, 'n' may be a storeSSL | |
1187 if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { // Check when generating nodes | |
1188 return 0; | |
1189 } | |
1190 Node *spill = new (C) MachSpillCopyNode(use,use_rm,def_rm); | |
1191 n->set_req(copyidx,spill); | |
1192 n->as_MachSpillCopy()->set_in_RegMask(def_rm); | |
1193 // Put the spill just before the copy | |
1194 insert_proj( b, insidx++, spill, maxlrg++ ); | |
1195 } | |
1196 } | |
1197 } | |
1198 } | |
1199 } // End For All Instructions in Block - Non-PHI Pass | |
1200 | |
1201 // Check if each LRG is live out of this block so as not to propagate | |
1202 // beyond the last use of a LRG. | |
1203 for( slidx = 0; slidx < spill_cnt; slidx++ ) { | |
1204 uint defidx = lidxs.at(slidx); | |
1205 IndexSet *liveout = _live->live(b); | |
1206 if( !liveout->member(defidx) ) { | |
1207 #ifdef ASSERT | |
1208 // The index defidx is not live. Check the liveout array to ensure that | |
1209 // it contains no members which compress to defidx. Finding such an | |
1210 // instance may be a case to add liveout adjustment in compress_uf_map(). | |
1211 // See 5063219. | |
1212 uint member; | |
1213 IndexSetIterator isi(liveout); | |
1214 while ((member = isi.next()) != 0) { | |
1215 assert(defidx != Find_const(member), "Live out member has not been compressed"); | |
1216 } | |
1217 #endif | |
1218 Reachblock[slidx] = NULL; | |
1219 } else { | |
1220 assert(Reachblock[slidx] != NULL,"No reaching definition for liveout value"); | |
1221 } | |
1222 } | |
1223 #ifndef PRODUCT | |
1224 if( trace_spilling() ) | |
1225 b->dump(); | |
1226 #endif | |
1227 } // End For All Blocks | |
1228 | |
1229 //----------PASS 2---------- | |
1230 // Reset all DEF live range numbers here | |
1231 for( insidx = 0; insidx < defs->size(); insidx++ ) { | |
1232 // Grab the def | |
1233 n1 = defs->at(insidx); | |
1234 // Set new lidx for DEF | |
1235 new_lrg(n1, maxlrg++); | |
1236 } | |
1237 //----------Phi Node Splitting---------- | |
1238 // Clean up a phi here, and assign a new live range number | |
1239 // Cycle through this block's predecessors, collecting Reaches | |
1240 // info for each spilled LRG and update edges. | |
1241 // Walk the phis list to patch inputs, split phis, and name phis | |
2016
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1242 uint lrgs_before_phi_split = maxlrg; |
0 | 1243 for( insidx = 0; insidx < phis->size(); insidx++ ) { |
1244 Node *phi = phis->at(insidx); | |
1245 assert(phi->is_Phi(),"This list must only contain Phi Nodes"); | |
1246 Block *b = _cfg._bbs[phi->_idx]; | |
1247 // Grab the live range number | |
1248 uint lidx = Find_id(phi); | |
1249 uint slidx = lrg2reach[lidx]; | |
1250 // Update node to lidx map | |
1251 new_lrg(phi, maxlrg++); | |
1252 // Get PASS1's up/down decision for the block. | |
1253 int phi_up = !!UP_entry[slidx]->test(b->_pre_order); | |
1254 | |
1255 // Force down if double-spilling live range | |
1256 if( lrgs(lidx)._was_spilled1 ) | |
1257 phi_up = false; | |
1258 | |
1259 // When splitting a Phi we an split it normal or "inverted". | |
1260 // An inverted split makes the splits target the Phi's UP/DOWN | |
1261 // sense inverted; then the Phi is followed by a final def-side | |
1262 // split to invert back. It changes which blocks the spill code | |
1263 // goes in. | |
1264 | |
1265 // Walk the predecessor blocks and assign the reaching def to the Phi. | |
1266 // Split Phi nodes by placing USE side splits wherever the reaching | |
1267 // DEF has the wrong UP/DOWN value. | |
1268 for( uint i = 1; i < b->num_preds(); i++ ) { | |
1269 // Get predecessor block pre-order number | |
1270 Block *pred = _cfg._bbs[b->pred(i)->_idx]; | |
1271 pidx = pred->_pre_order; | |
1272 // Grab reaching def | |
1273 Node *def = Reaches[pidx][slidx]; | |
1274 assert( def, "must have reaching def" ); | |
1275 // If input up/down sense and reg-pressure DISagree | |
1276 if( def->rematerialize() ) { | |
2016
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1277 // Place the rematerialized node above any MSCs created during |
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1278 // phi node splitting. end_idx points at the insertion point |
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1279 // so look at the node before it. |
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1280 int insert = pred->end_idx(); |
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1281 while (insert >= 1 && |
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1282 pred->_nodes[insert - 1]->is_SpillCopy() && |
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1283 Find(pred->_nodes[insert - 1]) >= lrgs_before_phi_split) { |
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1284 insert--; |
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1285 } |
361783318e7e
7004940: CTW: assert(!def_outside->member(r)) failed: Use of external LRG overlaps the same LRG
never
parents:
1972
diff
changeset
|
1286 def = split_Rematerialize( def, pred, insert, maxlrg, splits, slidx, lrg2reach, Reachblock, false ); |
0 | 1287 if( !def ) return 0; // Bail out |
1288 } | |
1289 // Update the Phi's input edge array | |
1290 phi->set_req(i,def); | |
1291 // Grab the UP/DOWN sense for the input | |
1292 u1 = UP[pidx][slidx]; | |
1293 if( u1 != (phi_up != 0)) { | |
1294 maxlrg = split_USE(def, b, phi, i, maxlrg, !u1, false, splits,slidx); | |
1295 // If it wasn't split bail | |
1296 if (!maxlrg) { | |
1297 return 0; | |
1298 } | |
1299 } | |
1300 } // End for all inputs to the Phi | |
1301 } // End for all Phi Nodes | |
1302 // Update _maxlrg to save Union asserts | |
1303 _maxlrg = maxlrg; | |
1304 | |
1305 | |
1306 //----------PASS 3---------- | |
1307 // Pass over all Phi's to union the live ranges | |
1308 for( insidx = 0; insidx < phis->size(); insidx++ ) { | |
1309 Node *phi = phis->at(insidx); | |
1310 assert(phi->is_Phi(),"This list must only contain Phi Nodes"); | |
1311 // Walk all inputs to Phi and Union input live range with Phi live range | |
1312 for( uint i = 1; i < phi->req(); i++ ) { | |
1313 // Grab the input node | |
1314 Node *n = phi->in(i); | |
1315 assert( n, "" ); | |
1316 uint lidx = Find(n); | |
1317 uint pidx = Find(phi); | |
1318 if( lidx < pidx ) | |
1319 Union(n, phi); | |
1320 else if( lidx > pidx ) | |
1321 Union(phi, n); | |
1322 } // End for all inputs to the Phi Node | |
1323 } // End for all Phi Nodes | |
1324 // Now union all two address instructions | |
1325 for( insidx = 0; insidx < defs->size(); insidx++ ) { | |
1326 // Grab the def | |
1327 n1 = defs->at(insidx); | |
1328 // Set new lidx for DEF & handle 2-addr instructions | |
1329 if( n1->is_Mach() && ((twoidx = n1->as_Mach()->two_adr()) != 0) ) { | |
1330 assert( Find(n1->in(twoidx)) < maxlrg,"Assigning bad live range index"); | |
1331 // Union the input and output live ranges | |
1332 uint lr1 = Find(n1); | |
1333 uint lr2 = Find(n1->in(twoidx)); | |
1334 if( lr1 < lr2 ) | |
1335 Union(n1, n1->in(twoidx)); | |
1336 else if( lr1 > lr2 ) | |
1337 Union(n1->in(twoidx), n1); | |
1338 } // End if two address | |
1339 } // End for all defs | |
1340 // DEBUG | |
1341 #ifdef ASSERT | |
1342 // Validate all live range index assignments | |
1343 for( bidx = 0; bidx < _cfg._num_blocks; bidx++ ) { | |
1344 b = _cfg._blocks[bidx]; | |
1345 for( insidx = 0; insidx <= b->end_idx(); insidx++ ) { | |
1346 Node *n = b->_nodes[insidx]; | |
1347 uint defidx = Find(n); | |
1348 assert(defidx < _maxlrg,"Bad live range index in Split"); | |
1349 assert(defidx < maxlrg,"Bad live range index in Split"); | |
1350 } | |
1351 } | |
1352 // Issue a warning if splitting made no progress | |
1353 int noprogress = 0; | |
1354 for( slidx = 0; slidx < spill_cnt; slidx++ ) { | |
1355 if( PrintOpto && WizardMode && splits.at(slidx) == 0 ) { | |
1356 tty->print_cr("Failed to split live range %d", lidxs.at(slidx)); | |
1357 //BREAKPOINT; | |
1358 } | |
1359 else { | |
1360 noprogress++; | |
1361 } | |
1362 } | |
1363 if(!noprogress) { | |
1364 tty->print_cr("Failed to make progress in Split"); | |
1365 //BREAKPOINT; | |
1366 } | |
1367 #endif | |
1368 // Return updated count of live ranges | |
1369 return maxlrg; | |
1370 } |