Mercurial > hg > graal-compiler
annotate src/share/vm/opto/reg_split.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 | 1c6e3bfb543a |
children | 9ee9cf798b59 |
rev | line source |
---|---|
0 | 1 /* |
2 * Copyright 2000-2006 Sun Microsystems, Inc. All Rights Reserved. | |
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
4 * | |
5 * This code is free software; you can redistribute it and/or modify it | |
6 * under the terms of the GNU General Public License version 2 only, as | |
7 * published by the Free Software Foundation. | |
8 * | |
9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
12 * version 2 for more details (a copy is included in the LICENSE file that | |
13 * accompanied this code). | |
14 * | |
15 * You should have received a copy of the GNU General Public License version | |
16 * 2 along with this work; if not, write to the Free Software Foundation, | |
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
18 * | |
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
21 * have any questions. | |
22 * | |
23 */ | |
24 | |
25 #include "incls/_precompiled.incl" | |
26 #include "incls/_reg_split.cpp.incl" | |
27 | |
28 //------------------------------Split-------------------------------------- | |
29 // Walk the graph in RPO and for each lrg which spills, propogate reaching | |
30 // definitions. During propogation, split the live range around regions of | |
31 // High Register Pressure (HRP). If a Def is in a region of Low Register | |
32 // Pressure (LRP), it will not get spilled until we encounter a region of | |
33 // HRP between it and one of its uses. We will spill at the transition | |
34 // point between LRP and HRP. Uses in the HRP region will use the spilled | |
35 // Def. The first Use outside the HRP region will generate a SpillCopy to | |
36 // hoist the live range back up into a register, and all subsequent uses | |
37 // will use that new Def until another HRP region is encountered. Defs in | |
38 // HRP regions will get trailing SpillCopies to push the LRG down into the | |
39 // stack immediately. | |
40 // | |
41 // As a side effect, unlink from (hence make dead) coalesced copies. | |
42 // | |
43 | |
44 static const char out_of_nodes[] = "out of nodes during split"; | |
45 | |
46 //------------------------------get_spillcopy_wide----------------------------- | |
47 // Get a SpillCopy node with wide-enough masks. Use the 'wide-mask', the | |
48 // wide ideal-register spill-mask if possible. If the 'wide-mask' does | |
49 // not cover the input (or output), use the input (or output) mask instead. | |
50 Node *PhaseChaitin::get_spillcopy_wide( Node *def, Node *use, uint uidx ) { | |
51 // If ideal reg doesn't exist we've got a bad schedule happening | |
52 // that is forcing us to spill something that isn't spillable. | |
53 // Bail rather than abort | |
54 int ireg = def->ideal_reg(); | |
55 if( ireg == 0 || ireg == Op_RegFlags ) { | |
56 C->record_method_not_compilable("attempted to spill a non-spillable item"); | |
57 return NULL; | |
58 } | |
59 if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { | |
60 return NULL; | |
61 } | |
62 const RegMask *i_mask = &def->out_RegMask(); | |
63 const RegMask *w_mask = C->matcher()->idealreg2spillmask[ireg]; | |
64 const RegMask *o_mask = use ? &use->in_RegMask(uidx) : w_mask; | |
65 const RegMask *w_i_mask = w_mask->overlap( *i_mask ) ? w_mask : i_mask; | |
66 const RegMask *w_o_mask; | |
67 | |
68 if( w_mask->overlap( *o_mask ) && // Overlap AND | |
69 ((ireg != Op_RegL && ireg != Op_RegD // Single use or aligned | |
70 #ifdef _LP64 | |
71 && ireg != Op_RegP | |
72 #endif | |
73 ) || o_mask->is_aligned_Pairs()) ) { | |
74 // Don't come here for mis-aligned doubles | |
75 w_o_mask = w_mask; | |
76 } else { // wide ideal mask does not overlap with o_mask | |
77 // Mis-aligned doubles come here and XMM->FPR moves on x86. | |
78 w_o_mask = o_mask; // Must target desired registers | |
79 // Does the ideal-reg-mask overlap with o_mask? I.e., can I use | |
80 // a reg-reg move or do I need a trip across register classes | |
81 // (and thus through memory)? | |
82 if( !C->matcher()->idealreg2regmask[ireg]->overlap( *o_mask) && o_mask->is_UP() ) | |
83 // Here we assume a trip through memory is required. | |
84 w_i_mask = &C->FIRST_STACK_mask(); | |
85 } | |
86 return new (C) MachSpillCopyNode( def, *w_i_mask, *w_o_mask ); | |
87 } | |
88 | |
89 //------------------------------insert_proj------------------------------------ | |
90 // Insert the spill at chosen location. Skip over any interveneing Proj's or | |
91 // Phis. Skip over a CatchNode and projs, inserting in the fall-through block | |
92 // instead. Update high-pressure indices. Create a new live range. | |
93 void PhaseChaitin::insert_proj( Block *b, uint i, Node *spill, uint maxlrg ) { | |
94 // Skip intervening ProjNodes. Do not insert between a ProjNode and | |
95 // its definer. | |
96 while( i < b->_nodes.size() && | |
97 (b->_nodes[i]->is_Proj() || | |
98 b->_nodes[i]->is_Phi() ) ) | |
99 i++; | |
100 | |
101 // Do not insert between a call and his Catch | |
102 if( b->_nodes[i]->is_Catch() ) { | |
103 // Put the instruction at the top of the fall-thru block. | |
104 // Find the fall-thru projection | |
105 while( 1 ) { | |
106 const CatchProjNode *cp = b->_nodes[++i]->as_CatchProj(); | |
107 if( cp->_con == CatchProjNode::fall_through_index ) | |
108 break; | |
109 } | |
110 int sidx = i - b->end_idx()-1; | |
111 b = b->_succs[sidx]; // Switch to successor block | |
112 i = 1; // Right at start of block | |
113 } | |
114 | |
115 b->_nodes.insert(i,spill); // Insert node in block | |
116 _cfg._bbs.map(spill->_idx,b); // Update node->block mapping to reflect | |
117 // Adjust the point where we go hi-pressure | |
118 if( i <= b->_ihrp_index ) b->_ihrp_index++; | |
119 if( i <= b->_fhrp_index ) b->_fhrp_index++; | |
120 | |
121 // Assign a new Live Range Number to the SpillCopy and grow | |
122 // the node->live range mapping. | |
123 new_lrg(spill,maxlrg); | |
124 } | |
125 | |
126 //------------------------------split_DEF-------------------------------------- | |
127 // There are four catagories of Split; UP/DOWN x DEF/USE | |
128 // Only three of these really occur as DOWN/USE will always color | |
129 // Any Split with a DEF cannot CISC-Spill now. Thus we need | |
130 // two helper routines, one for Split DEFS (insert after instruction), | |
131 // one for Split USES (insert before instruction). DEF insertion | |
132 // happens inside Split, where the Leaveblock array is updated. | |
133 uint PhaseChaitin::split_DEF( Node *def, Block *b, int loc, uint maxlrg, Node **Reachblock, Node **debug_defs, GrowableArray<uint> splits, int slidx ) { | |
134 #ifdef ASSERT | |
135 // Increment the counter for this lrg | |
136 splits.at_put(slidx, splits.at(slidx)+1); | |
137 #endif | |
138 // If we are spilling the memory op for an implicit null check, at the | |
139 // null check location (ie - null check is in HRP block) we need to do | |
140 // the null-check first, then spill-down in the following block. | |
141 // (The implicit_null_check function ensures the use is also dominated | |
142 // by the branch-not-taken block.) | |
143 Node *be = b->end(); | |
144 if( be->is_MachNullCheck() && be->in(1) == def && def == b->_nodes[loc] ) { | |
145 // Spill goes in the branch-not-taken block | |
146 b = b->_succs[b->_nodes[b->end_idx()+1]->Opcode() == Op_IfTrue]; | |
147 loc = 0; // Just past the Region | |
148 } | |
149 assert( loc >= 0, "must insert past block head" ); | |
150 | |
151 // Get a def-side SpillCopy | |
152 Node *spill = get_spillcopy_wide(def,NULL,0); | |
153 // Did we fail to split?, then bail | |
154 if (!spill) { | |
155 return 0; | |
156 } | |
157 | |
158 // Insert the spill at chosen location | |
159 insert_proj( b, loc+1, spill, maxlrg++); | |
160 | |
161 // Insert new node into Reaches array | |
162 Reachblock[slidx] = spill; | |
163 // Update debug list of reaching down definitions by adding this one | |
164 debug_defs[slidx] = spill; | |
165 | |
166 // return updated count of live ranges | |
167 return maxlrg; | |
168 } | |
169 | |
170 //------------------------------split_USE-------------------------------------- | |
171 // Splits at uses can involve redeffing the LRG, so no CISC Spilling there. | |
172 // Debug uses want to know if def is already stack enabled. | |
173 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 ) { | |
174 #ifdef ASSERT | |
175 // Increment the counter for this lrg | |
176 splits.at_put(slidx, splits.at(slidx)+1); | |
177 #endif | |
178 | |
179 // Some setup stuff for handling debug node uses | |
180 JVMState* jvms = use->jvms(); | |
181 uint debug_start = jvms ? jvms->debug_start() : 999999; | |
182 uint debug_end = jvms ? jvms->debug_end() : 999999; | |
183 | |
184 //------------------------------------------- | |
185 // Check for use of debug info | |
186 if (useidx >= debug_start && useidx < debug_end) { | |
187 // Actually it's perfectly legal for constant debug info to appear | |
188 // just unlikely. In this case the optimizer left a ConI of a 4 | |
189 // as both inputs to a Phi with only a debug use. It's a single-def | |
190 // live range of a rematerializable value. The live range spills, | |
191 // rematerializes and now the ConI directly feeds into the debug info. | |
192 // assert(!def->is_Con(), "constant debug info already constructed directly"); | |
193 | |
194 // Special split handling for Debug Info | |
195 // If DEF is DOWN, just hook the edge and return | |
196 // If DEF is UP, Split it DOWN for this USE. | |
197 if( def->is_Mach() ) { | |
198 if( def_down ) { | |
199 // DEF is DOWN, so connect USE directly to the DEF | |
200 use->set_req(useidx, def); | |
201 } else { | |
202 // Block and index where the use occurs. | |
203 Block *b = _cfg._bbs[use->_idx]; | |
204 // Put the clone just prior to use | |
205 int bindex = b->find_node(use); | |
206 // DEF is UP, so must copy it DOWN and hook in USE | |
207 // Insert SpillCopy before the USE, which uses DEF as its input, | |
208 // and defs a new live range, which is used by this node. | |
209 Node *spill = get_spillcopy_wide(def,use,useidx); | |
210 // did we fail to split? | |
211 if (!spill) { | |
212 // Bail | |
213 return 0; | |
214 } | |
215 // insert into basic block | |
216 insert_proj( b, bindex, spill, maxlrg++ ); | |
217 // Use the new split | |
218 use->set_req(useidx,spill); | |
219 } | |
220 // No further split handling needed for this use | |
221 return maxlrg; | |
222 } // End special splitting for debug info live range | |
223 } // If debug info | |
224 | |
225 // CISC-SPILLING | |
226 // Finally, check to see if USE is CISC-Spillable, and if so, | |
227 // gather_lrg_masks will add the flags bit to its mask, and | |
228 // no use side copy is needed. This frees up the live range | |
229 // register choices without causing copy coalescing, etc. | |
230 if( UseCISCSpill && cisc_sp ) { | |
231 int inp = use->cisc_operand(); | |
232 if( inp != AdlcVMDeps::Not_cisc_spillable ) | |
233 // Convert operand number to edge index number | |
234 inp = use->as_Mach()->operand_index(inp); | |
235 if( inp == (int)useidx ) { | |
236 use->set_req(useidx, def); | |
237 #ifndef PRODUCT | |
238 if( TraceCISCSpill ) { | |
239 tty->print(" set_split: "); | |
240 use->dump(); | |
241 } | |
242 #endif | |
243 return maxlrg; | |
244 } | |
245 } | |
246 | |
247 //------------------------------------------- | |
248 // Insert a Copy before the use | |
249 | |
250 // Block and index where the use occurs. | |
251 int bindex; | |
252 // Phi input spill-copys belong at the end of the prior block | |
253 if( use->is_Phi() ) { | |
254 b = _cfg._bbs[b->pred(useidx)->_idx]; | |
255 bindex = b->end_idx(); | |
256 } else { | |
257 // Put the clone just prior to use | |
258 bindex = b->find_node(use); | |
259 } | |
260 | |
261 Node *spill = get_spillcopy_wide( def, use, useidx ); | |
262 if( !spill ) return 0; // Bailed out | |
263 // Insert SpillCopy before the USE, which uses the reaching DEF as | |
264 // its input, and defs a new live range, which is used by this node. | |
265 insert_proj( b, bindex, spill, maxlrg++ ); | |
266 // Use the spill/clone | |
267 use->set_req(useidx,spill); | |
268 | |
269 // return updated live range count | |
270 return maxlrg; | |
271 } | |
272 | |
273 //------------------------------split_Rematerialize---------------------------- | |
274 // Clone a local copy of the def. | |
275 Node *PhaseChaitin::split_Rematerialize( Node *def, Block *b, uint insidx, uint &maxlrg, GrowableArray<uint> splits, int slidx, uint *lrg2reach, Node **Reachblock, bool walkThru ) { | |
276 // The input live ranges will be stretched to the site of the new | |
277 // instruction. They might be stretched past a def and will thus | |
278 // have the old and new values of the same live range alive at the | |
279 // same time - a definite no-no. Split out private copies of | |
280 // the inputs. | |
281 if( def->req() > 1 ) { | |
282 for( uint i = 1; i < def->req(); i++ ) { | |
283 Node *in = def->in(i); | |
284 // Check for single-def (LRG cannot redefined) | |
285 uint lidx = n2lidx(in); | |
286 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
|
287 if (lrgs(lidx).is_singledef()) continue; |
0 | 288 |
289 Block *b_def = _cfg._bbs[def->_idx]; | |
290 int idx_def = b_def->find_node(def); | |
291 Node *in_spill = get_spillcopy_wide( in, def, i ); | |
292 if( !in_spill ) return 0; // Bailed out | |
293 insert_proj(b_def,idx_def,in_spill,maxlrg++); | |
294 if( b_def == b ) | |
295 insidx++; | |
296 def->set_req(i,in_spill); | |
297 } | |
298 } | |
299 | |
300 Node *spill = def->clone(); | |
301 if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { | |
302 // Check when generating nodes | |
303 return 0; | |
304 } | |
305 | |
306 // See if any inputs are currently being spilled, and take the | |
307 // latest copy of spilled inputs. | |
308 if( spill->req() > 1 ) { | |
309 for( uint i = 1; i < spill->req(); i++ ) { | |
310 Node *in = spill->in(i); | |
311 uint lidx = Find_id(in); | |
312 | |
313 // Walk backwards thru spill copy node intermediates | |
295
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
314 if (walkThru) { |
0 | 315 while ( in->is_SpillCopy() && lidx >= _maxlrg ) { |
316 in = in->in(1); | |
317 lidx = Find_id(in); | |
318 } | |
319 | |
295
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
320 if (lidx < _maxlrg && lrgs(lidx).is_multidef()) { |
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
321 // 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
|
322 // just keep the original def used in the clone. |
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
323 in = spill->in(i); |
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
324 lidx = Find_id(in); |
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
325 } |
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
326 } |
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
327 |
0 | 328 if( lidx < _maxlrg && lrgs(lidx).reg() >= LRG::SPILL_REG ) { |
329 Node *rdef = Reachblock[lrg2reach[lidx]]; | |
330 if( rdef ) spill->set_req(i,rdef); | |
331 } | |
332 } | |
333 } | |
334 | |
335 | |
336 assert( spill->out_RegMask().is_UP(), "rematerialize to a reg" ); | |
337 // Rematerialized op is def->spilled+1 | |
338 set_was_spilled(spill); | |
339 if( _spilled_once.test(def->_idx) ) | |
340 set_was_spilled(spill); | |
341 | |
342 insert_proj( b, insidx, spill, maxlrg++ ); | |
343 #ifdef ASSERT | |
344 // Increment the counter for this lrg | |
345 splits.at_put(slidx, splits.at(slidx)+1); | |
346 #endif | |
347 // See if the cloned def kills any flags, and copy those kills as well | |
348 uint i = insidx+1; | |
349 if( clone_projs( b, i, def, spill, maxlrg ) ) { | |
350 // Adjust the point where we go hi-pressure | |
351 if( i <= b->_ihrp_index ) b->_ihrp_index++; | |
352 if( i <= b->_fhrp_index ) b->_fhrp_index++; | |
353 } | |
354 | |
355 return spill; | |
356 } | |
357 | |
358 //------------------------------is_high_pressure------------------------------- | |
359 // Function to compute whether or not this live range is "high pressure" | |
360 // in this block - whether it spills eagerly or not. | |
361 bool PhaseChaitin::is_high_pressure( Block *b, LRG *lrg, uint insidx ) { | |
362 if( lrg->_was_spilled1 ) return true; | |
363 // Forced spilling due to conflict? Then split only at binding uses | |
364 // or defs, not for supposed capacity problems. | |
365 // CNC - Turned off 7/8/99, causes too much spilling | |
366 // if( lrg->_is_bound ) return false; | |
367 | |
368 // Not yet reached the high-pressure cutoff point, so low pressure | |
369 uint hrp_idx = lrg->_is_float ? b->_fhrp_index : b->_ihrp_index; | |
370 if( insidx < hrp_idx ) return false; | |
371 // Register pressure for the block as a whole depends on reg class | |
372 int block_pres = lrg->_is_float ? b->_freg_pressure : b->_reg_pressure; | |
373 // Bound live ranges will split at the binding points first; | |
374 // Intermediate splits should assume the live range's register set | |
375 // got "freed up" and that num_regs will become INT_PRESSURE. | |
376 int bound_pres = lrg->_is_float ? FLOATPRESSURE : INTPRESSURE; | |
377 // Effective register pressure limit. | |
378 int lrg_pres = (lrg->get_invalid_mask_size() > lrg->num_regs()) | |
379 ? (lrg->get_invalid_mask_size() >> (lrg->num_regs()-1)) : bound_pres; | |
380 // High pressure if block pressure requires more register freedom | |
381 // than live range has. | |
382 return block_pres >= lrg_pres; | |
383 } | |
384 | |
385 | |
386 //------------------------------prompt_use--------------------------------- | |
387 // True if lidx is used before any real register is def'd in the block | |
388 bool PhaseChaitin::prompt_use( Block *b, uint lidx ) { | |
389 if( lrgs(lidx)._was_spilled2 ) return false; | |
390 | |
391 // Scan block for 1st use. | |
392 for( uint i = 1; i <= b->end_idx(); i++ ) { | |
393 Node *n = b->_nodes[i]; | |
394 // Ignore PHI use, these can be up or down | |
395 if( n->is_Phi() ) continue; | |
396 for( uint j = 1; j < n->req(); j++ ) | |
397 if( Find_id(n->in(j)) == lidx ) | |
398 return true; // Found 1st use! | |
399 if( n->out_RegMask().is_NotEmpty() ) return false; | |
400 } | |
401 return false; | |
402 } | |
403 | |
404 //------------------------------Split-------------------------------------- | |
405 //----------Split Routine---------- | |
406 // ***** NEW SPLITTING HEURISTIC ***** | |
407 // DEFS: If the DEF is in a High Register Pressure(HRP) Block, split there. | |
408 // Else, no split unless there is a HRP block between a DEF and | |
409 // one of its uses, and then split at the HRP block. | |
410 // | |
411 // USES: If USE is in HRP, split at use to leave main LRG on stack. | |
412 // Else, hoist LRG back up to register only (ie - split is also DEF) | |
413 // We will compute a new maxlrg as we go | |
414 uint PhaseChaitin::Split( uint maxlrg ) { | |
415 NOT_PRODUCT( Compile::TracePhase t3("regAllocSplit", &_t_regAllocSplit, TimeCompiler); ) | |
416 | |
417 uint bidx, pidx, slidx, insidx, inpidx, twoidx; | |
418 uint non_phi = 1, spill_cnt = 0; | |
419 Node **Reachblock; | |
420 Node *n1, *n2, *n3; | |
421 Node_List *defs,*phis; | |
422 bool *UPblock; | |
423 bool u1, u2, u3; | |
424 Block *b, *pred; | |
425 PhiNode *phi; | |
426 GrowableArray<uint> lidxs; | |
427 | |
428 // Array of counters to count splits per live range | |
429 GrowableArray<uint> splits; | |
430 | |
431 //----------Setup Code---------- | |
432 // Create a convenient mapping from lrg numbers to reaches/leaves indices | |
433 uint *lrg2reach = NEW_RESOURCE_ARRAY( uint, _maxlrg ); | |
434 // Keep track of DEFS & Phis for later passes | |
435 defs = new Node_List(); | |
436 phis = new Node_List(); | |
437 // Gather info on which LRG's are spilling, and build maps | |
438 for( bidx = 1; bidx < _maxlrg; bidx++ ) { | |
439 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) { | |
440 assert(!lrgs(bidx).mask().is_AllStack(),"AllStack should color"); | |
441 lrg2reach[bidx] = spill_cnt; | |
442 spill_cnt++; | |
443 lidxs.append(bidx); | |
444 #ifdef ASSERT | |
445 // Initialize the split counts to zero | |
446 splits.append(0); | |
447 #endif | |
448 #ifndef PRODUCT | |
449 if( PrintOpto && WizardMode && lrgs(bidx)._was_spilled1 ) | |
450 tty->print_cr("Warning, 2nd spill of L%d",bidx); | |
451 #endif | |
452 } | |
453 } | |
454 | |
455 // Create side arrays for propagating reaching defs info. | |
456 // Each block needs a node pointer for each spilling live range for the | |
457 // Def which is live into the block. Phi nodes handle multiple input | |
458 // Defs by querying the output of their predecessor blocks and resolving | |
459 // them to a single Def at the phi. The pointer is updated for each | |
460 // Def in the block, and then becomes the output for the block when | |
461 // processing of the block is complete. We also need to track whether | |
462 // a Def is UP or DOWN. UP means that it should get a register (ie - | |
463 // it is always in LRP regions), and DOWN means that it is probably | |
464 // on the stack (ie - it crosses HRP regions). | |
465 Node ***Reaches = NEW_RESOURCE_ARRAY( Node**, _cfg._num_blocks+1 ); | |
466 bool **UP = NEW_RESOURCE_ARRAY( bool*, _cfg._num_blocks+1 ); | |
467 Node **debug_defs = NEW_RESOURCE_ARRAY( Node*, spill_cnt ); | |
468 VectorSet **UP_entry= NEW_RESOURCE_ARRAY( VectorSet*, spill_cnt ); | |
469 | |
470 // Initialize Reaches & UP | |
471 for( bidx = 0; bidx < _cfg._num_blocks+1; bidx++ ) { | |
472 Reaches[bidx] = NEW_RESOURCE_ARRAY( Node*, spill_cnt ); | |
473 UP[bidx] = NEW_RESOURCE_ARRAY( bool, spill_cnt ); | |
474 Node **Reachblock = Reaches[bidx]; | |
475 bool *UPblock = UP[bidx]; | |
476 for( slidx = 0; slidx < spill_cnt; slidx++ ) { | |
477 UPblock[slidx] = true; // Assume they start in registers | |
478 Reachblock[slidx] = NULL; // Assume that no def is present | |
479 } | |
480 } | |
481 | |
482 // Initialize to array of empty vectorsets | |
483 for( slidx = 0; slidx < spill_cnt; slidx++ ) | |
484 UP_entry[slidx] = new VectorSet(Thread::current()->resource_area()); | |
485 | |
486 //----------PASS 1---------- | |
487 //----------Propagation & Node Insertion Code---------- | |
488 // Walk the Blocks in RPO for DEF & USE info | |
489 for( bidx = 0; bidx < _cfg._num_blocks; bidx++ ) { | |
490 | |
491 if (C->check_node_count(spill_cnt, out_of_nodes)) { | |
492 return 0; | |
493 } | |
494 | |
495 b = _cfg._blocks[bidx]; | |
496 // Reaches & UP arrays for this block | |
497 Reachblock = Reaches[b->_pre_order]; | |
498 UPblock = UP[b->_pre_order]; | |
499 // Reset counter of start of non-Phi nodes in block | |
500 non_phi = 1; | |
501 //----------Block Entry Handling---------- | |
502 // Check for need to insert a new phi | |
503 // Cycle through this block's predecessors, collecting Reaches | |
504 // info for each spilled LRG. If they are identical, no phi is | |
505 // needed. If they differ, check for a phi, and insert if missing, | |
506 // or update edges if present. Set current block's Reaches set to | |
507 // be either the phi's or the reaching def, as appropriate. | |
508 // If no Phi is needed, check if the LRG needs to spill on entry | |
509 // to the block due to HRP. | |
510 for( slidx = 0; slidx < spill_cnt; slidx++ ) { | |
511 // Grab the live range number | |
512 uint lidx = lidxs.at(slidx); | |
513 // Do not bother splitting or putting in Phis for single-def | |
514 // rematerialized live ranges. This happens alot to constants | |
515 // with long live ranges. | |
295
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
516 if( lrgs(lidx).is_singledef() && |
0 | 517 lrgs(lidx)._def->rematerialize() ) { |
518 // reset the Reaches & UP entries | |
519 Reachblock[slidx] = lrgs(lidx)._def; | |
520 UPblock[slidx] = true; | |
521 // Record following instruction in case 'n' rematerializes and | |
522 // kills flags | |
523 Block *pred1 = _cfg._bbs[b->pred(1)->_idx]; | |
524 continue; | |
525 } | |
526 | |
527 // Initialize needs_phi and needs_split | |
528 bool needs_phi = false; | |
529 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
|
530 bool has_phi = false; |
0 | 531 // Walk the predecessor blocks to check inputs for that live range |
532 // Grab predecessor block header | |
533 n1 = b->pred(1); | |
534 // Grab the appropriate reaching def info for inpidx | |
535 pred = _cfg._bbs[n1->_idx]; | |
536 pidx = pred->_pre_order; | |
537 Node **Ltmp = Reaches[pidx]; | |
538 bool *Utmp = UP[pidx]; | |
539 n1 = Ltmp[slidx]; | |
540 u1 = Utmp[slidx]; | |
541 // Initialize node for saving type info | |
542 n3 = n1; | |
543 u3 = u1; | |
544 | |
545 // Compare inputs to see if a Phi is needed | |
546 for( inpidx = 2; inpidx < b->num_preds(); inpidx++ ) { | |
547 // Grab predecessor block headers | |
548 n2 = b->pred(inpidx); | |
549 // Grab the appropriate reaching def info for inpidx | |
550 pred = _cfg._bbs[n2->_idx]; | |
551 pidx = pred->_pre_order; | |
552 Ltmp = Reaches[pidx]; | |
553 Utmp = UP[pidx]; | |
554 n2 = Ltmp[slidx]; | |
555 u2 = Utmp[slidx]; | |
556 // For each LRG, decide if a phi is necessary | |
557 if( n1 != n2 ) { | |
558 needs_phi = true; | |
559 } | |
560 // See if the phi has mismatched inputs, UP vs. DOWN | |
561 if( n1 && n2 && (u1 != u2) ) { | |
562 needs_split = true; | |
563 } | |
564 // Move n2/u2 to n1/u1 for next iteration | |
565 n1 = n2; | |
566 u1 = u2; | |
567 // Preserve a non-NULL predecessor for later type referencing | |
568 if( (n3 == NULL) && (n2 != NULL) ){ | |
569 n3 = n2; | |
570 u3 = u2; | |
571 } | |
572 } // End for all potential Phi inputs | |
573 | |
330
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
574 // 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
|
575 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
|
576 n1 = b->_nodes[insidx]; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
577 // 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
|
578 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
|
579 if( phi == NULL ) { |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
580 // 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
|
581 non_phi = insidx; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
582 // 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
|
583 break; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
584 } |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
585 // 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
|
586 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
|
587 // found the necessary phi |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
588 needs_phi = false; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
589 has_phi = true; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
590 // 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
|
591 Reachblock[slidx] = phi; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
592 break; |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
593 } // 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
|
594 } // 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
|
595 |
1c6e3bfb543a
6746892: Register Allocator does not process a data phi with one unique input correctly
kvn
parents:
295
diff
changeset
|
596 // 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
|
597 if( needs_phi || has_phi ) { |
0 | 598 // add new phinode if one not already found |
599 if( needs_phi ) { | |
600 // create a new phi node and insert it into the block | |
601 // type is taken from left over pointer to a predecessor | |
602 assert(n3,"No non-NULL reaching DEF for a Phi"); | |
603 phi = new (C, b->num_preds()) PhiNode(b->head(), n3->bottom_type()); | |
604 // initialize the Reaches entry for this LRG | |
605 Reachblock[slidx] = phi; | |
606 | |
607 // add node to block & node_to_block mapping | |
608 insert_proj( b, insidx++, phi, maxlrg++ ); | |
609 non_phi++; | |
610 // Reset new phi's mapping to be the spilling live range | |
611 _names.map(phi->_idx, lidx); | |
612 assert(Find_id(phi) == lidx,"Bad update on Union-Find mapping"); | |
613 } // end if not found correct phi | |
614 // Here you have either found or created the Phi, so record it | |
615 assert(phi != NULL,"Must have a Phi Node here"); | |
616 phis->push(phi); | |
617 // PhiNodes should either force the LRG UP or DOWN depending | |
618 // on its inputs and the register pressure in the Phi's block. | |
619 UPblock[slidx] = true; // Assume new DEF is UP | |
620 // If entering a high-pressure area with no immediate use, | |
621 // assume Phi is DOWN | |
622 if( is_high_pressure( b, &lrgs(lidx), b->end_idx()) && !prompt_use(b,lidx) ) | |
623 UPblock[slidx] = false; | |
624 // If we are not split up/down and all inputs are down, then we | |
625 // are down | |
626 if( !needs_split && !u3 ) | |
627 UPblock[slidx] = false; | |
628 } // end if phi is needed | |
629 | |
630 // Do not need a phi, so grab the reaching DEF | |
631 else { | |
632 // Grab predecessor block header | |
633 n1 = b->pred(1); | |
634 // Grab the appropriate reaching def info for k | |
635 pred = _cfg._bbs[n1->_idx]; | |
636 pidx = pred->_pre_order; | |
637 Node **Ltmp = Reaches[pidx]; | |
638 bool *Utmp = UP[pidx]; | |
639 // reset the Reaches & UP entries | |
640 Reachblock[slidx] = Ltmp[slidx]; | |
641 UPblock[slidx] = Utmp[slidx]; | |
642 } // end else no Phi is needed | |
643 } // end for all spilling live ranges | |
644 // DEBUG | |
645 #ifndef PRODUCT | |
646 if(trace_spilling()) { | |
647 tty->print("/`\nBlock %d: ", b->_pre_order); | |
648 tty->print("Reaching Definitions after Phi handling\n"); | |
649 for( uint x = 0; x < spill_cnt; x++ ) { | |
650 tty->print("Spill Idx %d: UP %d: Node\n",x,UPblock[x]); | |
651 if( Reachblock[x] ) | |
652 Reachblock[x]->dump(); | |
653 else | |
654 tty->print("Undefined\n"); | |
655 } | |
656 } | |
657 #endif | |
658 | |
659 //----------Non-Phi Node Splitting---------- | |
660 // Since phi-nodes have now been handled, the Reachblock array for this | |
661 // block is initialized with the correct starting value for the defs which | |
662 // reach non-phi instructions in this block. Thus, process non-phi | |
663 // instructions normally, inserting SpillCopy nodes for all spill | |
664 // locations. | |
665 | |
666 // Memoize any DOWN reaching definitions for use as DEBUG info | |
667 for( insidx = 0; insidx < spill_cnt; insidx++ ) { | |
668 debug_defs[insidx] = (UPblock[insidx]) ? NULL : Reachblock[insidx]; | |
669 if( UPblock[insidx] ) // Memoize UP decision at block start | |
670 UP_entry[insidx]->set( b->_pre_order ); | |
671 } | |
672 | |
673 //----------Walk Instructions in the Block and Split---------- | |
674 // For all non-phi instructions in the block | |
675 for( insidx = 1; insidx <= b->end_idx(); insidx++ ) { | |
676 Node *n = b->_nodes[insidx]; | |
677 // Find the defining Node's live range index | |
678 uint defidx = Find_id(n); | |
679 uint cnt = n->req(); | |
680 | |
681 if( n->is_Phi() ) { | |
682 // Skip phi nodes after removing dead copies. | |
683 if( defidx < _maxlrg ) { | |
684 // Check for useless Phis. These appear if we spill, then | |
685 // coalesce away copies. Dont touch Phis in spilling live | |
686 // ranges; they are busy getting modifed in this pass. | |
687 if( lrgs(defidx).reg() < LRG::SPILL_REG ) { | |
688 uint i; | |
689 Node *u = NULL; | |
690 // Look for the Phi merging 2 unique inputs | |
691 for( i = 1; i < cnt; i++ ) { | |
692 // Ignore repeats and self | |
693 if( n->in(i) != u && n->in(i) != n ) { | |
694 // Found a unique input | |
695 if( u != NULL ) // If it's the 2nd, bail out | |
696 break; | |
697 u = n->in(i); // Else record it | |
698 } | |
699 } | |
700 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
|
701 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
|
702 assert(Find_id(n) == Find_id(u), "should be the same lrg"); |
0 | 703 n->replace_by(u); // Then replace with unique input |
704 n->disconnect_inputs(NULL); | |
705 b->_nodes.remove(insidx); | |
706 insidx--; | |
707 b->_ihrp_index--; | |
708 b->_fhrp_index--; | |
709 } | |
710 } | |
711 } | |
712 continue; | |
713 } | |
714 assert( insidx > b->_ihrp_index || | |
715 (b->_reg_pressure < (uint)INTPRESSURE) || | |
716 b->_ihrp_index > 4000000 || | |
717 b->_ihrp_index >= b->end_idx() || | |
718 !b->_nodes[b->_ihrp_index]->is_Proj(), "" ); | |
719 assert( insidx > b->_fhrp_index || | |
720 (b->_freg_pressure < (uint)FLOATPRESSURE) || | |
721 b->_fhrp_index > 4000000 || | |
722 b->_fhrp_index >= b->end_idx() || | |
723 !b->_nodes[b->_fhrp_index]->is_Proj(), "" ); | |
724 | |
725 // ********** Handle Crossing HRP Boundry ********** | |
726 if( (insidx == b->_ihrp_index) || (insidx == b->_fhrp_index) ) { | |
727 for( slidx = 0; slidx < spill_cnt; slidx++ ) { | |
728 // Check for need to split at HRP boundry - split if UP | |
729 n1 = Reachblock[slidx]; | |
730 // bail out if no reaching DEF | |
731 if( n1 == NULL ) continue; | |
732 // bail out if live range is 'isolated' around inner loop | |
733 uint lidx = lidxs.at(slidx); | |
734 // If live range is currently UP | |
735 if( UPblock[slidx] ) { | |
736 // set location to insert spills at | |
737 // SPLIT DOWN HERE - NO CISC SPILL | |
738 if( is_high_pressure( b, &lrgs(lidx), insidx ) && | |
739 !n1->rematerialize() ) { | |
740 // If there is already a valid stack definition available, use it | |
741 if( debug_defs[slidx] != NULL ) { | |
742 Reachblock[slidx] = debug_defs[slidx]; | |
743 } | |
744 else { | |
745 // Insert point is just past last use or def in the block | |
746 int insert_point = insidx-1; | |
747 while( insert_point > 0 ) { | |
748 Node *n = b->_nodes[insert_point]; | |
749 // Hit top of block? Quit going backwards | |
750 if( n->is_Phi() ) break; | |
751 // Found a def? Better split after it. | |
752 if( n2lidx(n) == lidx ) break; | |
753 // Look for a use | |
754 uint i; | |
755 for( i = 1; i < n->req(); i++ ) | |
756 if( n2lidx(n->in(i)) == lidx ) | |
757 break; | |
758 // Found a use? Better split after it. | |
759 if( i < n->req() ) break; | |
760 insert_point--; | |
761 } | |
762 maxlrg = split_DEF( n1, b, insert_point, maxlrg, Reachblock, debug_defs, splits, slidx); | |
763 // If it wasn't split bail | |
764 if (!maxlrg) { | |
765 return 0; | |
766 } | |
767 insidx++; | |
768 } | |
769 // This is a new DEF, so update UP | |
770 UPblock[slidx] = false; | |
771 #ifndef PRODUCT | |
772 // DEBUG | |
773 if( trace_spilling() ) { | |
774 tty->print("\nNew Split DOWN DEF of Spill Idx "); | |
775 tty->print("%d, UP %d:\n",slidx,false); | |
776 n1->dump(); | |
777 } | |
778 #endif | |
779 } | |
780 } // end if LRG is UP | |
781 } // end for all spilling live ranges | |
782 assert( b->_nodes[insidx] == n, "got insidx set incorrectly" ); | |
783 } // end if crossing HRP Boundry | |
784 | |
785 // If the LRG index is oob, then this is a new spillcopy, skip it. | |
786 if( defidx >= _maxlrg ) { | |
787 continue; | |
788 } | |
789 LRG &deflrg = lrgs(defidx); | |
790 uint copyidx = n->is_Copy(); | |
791 // Remove coalesced copy from CFG | |
792 if( copyidx && defidx == n2lidx(n->in(copyidx)) ) { | |
793 n->replace_by( n->in(copyidx) ); | |
794 n->set_req( copyidx, NULL ); | |
795 b->_nodes.remove(insidx--); | |
796 b->_ihrp_index--; // Adjust the point where we go hi-pressure | |
797 b->_fhrp_index--; | |
798 continue; | |
799 } | |
800 | |
801 #define DERIVED 0 | |
802 | |
803 // ********** Handle USES ********** | |
804 bool nullcheck = false; | |
805 // Implicit null checks never use the spilled value | |
806 if( n->is_MachNullCheck() ) | |
807 nullcheck = true; | |
808 if( !nullcheck ) { | |
809 // Search all inputs for a Spill-USE | |
810 JVMState* jvms = n->jvms(); | |
811 uint oopoff = jvms ? jvms->oopoff() : cnt; | |
812 uint old_last = cnt - 1; | |
813 for( inpidx = 1; inpidx < cnt; inpidx++ ) { | |
814 // Derived/base pairs may be added to our inputs during this loop. | |
815 // If inpidx > old_last, then one of these new inputs is being | |
816 // handled. Skip the derived part of the pair, but process | |
817 // the base like any other input. | |
818 if( inpidx > old_last && ((inpidx - oopoff) & 1) == DERIVED ) { | |
819 continue; // skip derived_debug added below | |
820 } | |
821 // Get lidx of input | |
822 uint useidx = Find_id(n->in(inpidx)); | |
823 // Not a brand-new split, and it is a spill use | |
824 if( useidx < _maxlrg && lrgs(useidx).reg() >= LRG::SPILL_REG ) { | |
825 // Check for valid reaching DEF | |
826 slidx = lrg2reach[useidx]; | |
827 Node *def = Reachblock[slidx]; | |
828 assert( def != NULL, "Using Undefined Value in Split()\n"); | |
829 | |
830 // (+++) %%%% remove this in favor of pre-pass in matcher.cpp | |
831 // monitor references do not care where they live, so just hook | |
832 if ( jvms && jvms->is_monitor_use(inpidx) ) { | |
833 // The effect of this clone is to drop the node out of the block, | |
834 // so that the allocator does not see it anymore, and therefore | |
835 // does not attempt to assign it a register. | |
836 def = def->clone(); | |
837 _names.extend(def->_idx,0); | |
838 _cfg._bbs.map(def->_idx,b); | |
839 n->set_req(inpidx, def); | |
840 if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { | |
841 return 0; | |
842 } | |
843 continue; | |
844 } | |
845 | |
846 // Rematerializable? Then clone def at use site instead | |
847 // of store/load | |
848 if( def->rematerialize() ) { | |
849 int old_size = b->_nodes.size(); | |
850 def = split_Rematerialize( def, b, insidx, maxlrg, splits, slidx, lrg2reach, Reachblock, true ); | |
851 if( !def ) return 0; // Bail out | |
852 insidx += b->_nodes.size()-old_size; | |
853 } | |
854 | |
855 MachNode *mach = n->is_Mach() ? n->as_Mach() : NULL; | |
856 // Base pointers and oopmap references do not care where they live. | |
857 if ((inpidx >= oopoff) || | |
858 (mach && mach->ideal_Opcode() == Op_AddP && inpidx == AddPNode::Base)) { | |
859 if (def->rematerialize() && lrgs(useidx)._was_spilled2) { | |
860 // This def has been rematerialized a couple of times without | |
861 // progress. It doesn't care if it lives UP or DOWN, so | |
862 // spill it down now. | |
863 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false,splits,slidx); | |
864 // If it wasn't split bail | |
865 if (!maxlrg) { | |
866 return 0; | |
867 } | |
868 insidx++; // Reset iterator to skip USE side split | |
869 } else { | |
870 // Just hook the def edge | |
871 n->set_req(inpidx, def); | |
872 } | |
873 | |
874 if (inpidx >= oopoff) { | |
875 // After oopoff, we have derived/base pairs. We must mention all | |
876 // derived pointers here as derived/base pairs for GC. If the | |
877 // derived value is spilling and we have a copy both in Reachblock | |
878 // (called here 'def') and debug_defs[slidx] we need to mention | |
879 // both in derived/base pairs or kill one. | |
880 Node *derived_debug = debug_defs[slidx]; | |
881 if( ((inpidx - oopoff) & 1) == DERIVED && // derived vs base? | |
882 mach && mach->ideal_Opcode() != Op_Halt && | |
883 derived_debug != NULL && | |
884 derived_debug != def ) { // Actual 2nd value appears | |
885 // We have already set 'def' as a derived value. | |
886 // Also set debug_defs[slidx] as a derived value. | |
887 uint k; | |
888 for( k = oopoff; k < cnt; k += 2 ) | |
889 if( n->in(k) == derived_debug ) | |
890 break; // Found an instance of debug derived | |
891 if( k == cnt ) {// No instance of debug_defs[slidx] | |
892 // Add a derived/base pair to cover the debug info. | |
893 // We have to process the added base later since it is not | |
894 // handled yet at this point but skip derived part. | |
895 assert(((n->req() - oopoff) & 1) == DERIVED, | |
896 "must match skip condition above"); | |
897 n->add_req( derived_debug ); // this will be skipped above | |
898 n->add_req( n->in(inpidx+1) ); // this will be processed | |
899 // Increment cnt to handle added input edges on | |
900 // subsequent iterations. | |
901 cnt += 2; | |
902 } | |
903 } | |
904 } | |
905 continue; | |
906 } | |
907 // Special logic for DEBUG info | |
908 if( jvms && b->_freq > BLOCK_FREQUENCY(0.5) ) { | |
909 uint debug_start = jvms->debug_start(); | |
910 // If this is debug info use & there is a reaching DOWN def | |
911 if ((debug_start <= inpidx) && (debug_defs[slidx] != NULL)) { | |
912 assert(inpidx < oopoff, "handle only debug info here"); | |
913 // Just hook it in & move on | |
914 n->set_req(inpidx, debug_defs[slidx]); | |
915 // (Note that this can make two sides of a split live at the | |
916 // same time: The debug def on stack, and another def in a | |
917 // register. The GC needs to know about both of them, but any | |
918 // derived pointers after oopoff will refer to only one of the | |
919 // two defs and the GC would therefore miss the other. Thus | |
920 // this hack is only allowed for debug info which is Java state | |
921 // and therefore never a derived pointer.) | |
922 continue; | |
923 } | |
924 } | |
925 // Grab register mask info | |
926 const RegMask &dmask = def->out_RegMask(); | |
927 const RegMask &umask = n->in_RegMask(inpidx); | |
928 | |
929 assert(inpidx < oopoff, "cannot use-split oop map info"); | |
930 | |
931 bool dup = UPblock[slidx]; | |
932 bool uup = umask.is_UP(); | |
933 | |
934 // Need special logic to handle bound USES. Insert a split at this | |
935 // bound use if we can't rematerialize the def, or if we need the | |
936 // split to form a misaligned pair. | |
937 if( !umask.is_AllStack() && | |
938 (int)umask.Size() <= lrgs(useidx).num_regs() && | |
939 (!def->rematerialize() || | |
940 umask.is_misaligned_Pair())) { | |
941 // These need a Split regardless of overlap or pressure | |
942 // SPLIT - NO DEF - NO CISC SPILL | |
943 maxlrg = split_USE(def,b,n,inpidx,maxlrg,dup,false, splits,slidx); | |
944 // If it wasn't split bail | |
945 if (!maxlrg) { | |
946 return 0; | |
947 } | |
948 insidx++; // Reset iterator to skip USE side split | |
949 continue; | |
950 } | |
951 // Here is the logic chart which describes USE Splitting: | |
952 // 0 = false or DOWN, 1 = true or UP | |
953 // | |
954 // Overlap | DEF | USE | Action | |
955 //------------------------------------------------------- | |
956 // 0 | 0 | 0 | Copy - mem -> mem | |
957 // 0 | 0 | 1 | Split-UP - Check HRP | |
958 // 0 | 1 | 0 | Split-DOWN - Debug Info? | |
959 // 0 | 1 | 1 | Copy - reg -> reg | |
960 // 1 | 0 | 0 | Reset Input Edge (no Split) | |
961 // 1 | 0 | 1 | Split-UP - Check HRP | |
962 // 1 | 1 | 0 | Split-DOWN - Debug Info? | |
963 // 1 | 1 | 1 | Reset Input Edge (no Split) | |
964 // | |
965 // So, if (dup == uup), then overlap test determines action, | |
966 // with true being no split, and false being copy. Else, | |
967 // if DEF is DOWN, Split-UP, and check HRP to decide on | |
968 // resetting DEF. Finally if DEF is UP, Split-DOWN, with | |
969 // special handling for Debug Info. | |
970 if( dup == uup ) { | |
971 if( dmask.overlap(umask) ) { | |
972 // Both are either up or down, and there is overlap, No Split | |
973 n->set_req(inpidx, def); | |
974 } | |
975 else { // Both are either up or down, and there is no overlap | |
976 if( dup ) { // If UP, reg->reg copy | |
977 // COPY ACROSS HERE - NO DEF - NO CISC SPILL | |
978 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,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 } | |
985 else { // DOWN, mem->mem copy | |
986 // COPY UP & DOWN HERE - NO DEF - NO CISC SPILL | |
987 // First Split-UP to move value into Register | |
988 uint def_ideal = def->ideal_reg(); | |
989 const RegMask* tmp_rm = Matcher::idealreg2regmask[def_ideal]; | |
990 Node *spill = new (C) MachSpillCopyNode(def, dmask, *tmp_rm); | |
991 insert_proj( b, insidx, spill, maxlrg ); | |
992 // Then Split-DOWN as if previous Split was DEF | |
993 maxlrg = split_USE(spill,b,n,inpidx,maxlrg,false,false, splits,slidx); | |
994 // If it wasn't split bail | |
995 if (!maxlrg) { | |
996 return 0; | |
997 } | |
998 insidx += 2; // Reset iterator to skip USE side splits | |
999 } | |
1000 } // End else no overlap | |
1001 } // End if dup == uup | |
1002 // dup != uup, so check dup for direction of Split | |
1003 else { | |
1004 if( dup ) { // If UP, Split-DOWN and check Debug Info | |
1005 // If this node is already a SpillCopy, just patch the edge | |
1006 // except the case of spilling to stack. | |
1007 if( n->is_SpillCopy() ) { | |
1008 RegMask tmp_rm(umask); | |
1009 tmp_rm.SUBTRACT(Matcher::STACK_ONLY_mask); | |
1010 if( dmask.overlap(tmp_rm) ) { | |
1011 if( def != n->in(inpidx) ) { | |
1012 n->set_req(inpidx, def); | |
1013 } | |
1014 continue; | |
1015 } | |
1016 } | |
1017 // COPY DOWN HERE - NO DEF - NO CISC SPILL | |
1018 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false, splits,slidx); | |
1019 // If it wasn't split bail | |
1020 if (!maxlrg) { | |
1021 return 0; | |
1022 } | |
1023 insidx++; // Reset iterator to skip USE side split | |
1024 // Check for debug-info split. Capture it for later | |
1025 // debug splits of the same value | |
1026 if (jvms && jvms->debug_start() <= inpidx && inpidx < oopoff) | |
1027 debug_defs[slidx] = n->in(inpidx); | |
1028 | |
1029 } | |
1030 else { // DOWN, Split-UP and check register pressure | |
1031 if( is_high_pressure( b, &lrgs(useidx), insidx ) ) { | |
1032 // COPY UP HERE - NO DEF - CISC SPILL | |
1033 maxlrg = split_USE(def,b,n,inpidx,maxlrg,true,true, splits,slidx); | |
1034 // If it wasn't split bail | |
1035 if (!maxlrg) { | |
1036 return 0; | |
1037 } | |
1038 insidx++; // Reset iterator to skip USE side split | |
1039 } else { // LRP | |
1040 // COPY UP HERE - WITH DEF - NO CISC SPILL | |
1041 maxlrg = split_USE(def,b,n,inpidx,maxlrg,true,false, splits,slidx); | |
1042 // If it wasn't split bail | |
1043 if (!maxlrg) { | |
1044 return 0; | |
1045 } | |
1046 // Flag this lift-up in a low-pressure block as | |
1047 // already-spilled, so if it spills again it will | |
1048 // spill hard (instead of not spilling hard and | |
1049 // coalescing away). | |
1050 set_was_spilled(n->in(inpidx)); | |
1051 // Since this is a new DEF, update Reachblock & UP | |
1052 Reachblock[slidx] = n->in(inpidx); | |
1053 UPblock[slidx] = true; | |
1054 insidx++; // Reset iterator to skip USE side split | |
1055 } | |
1056 } // End else DOWN | |
1057 } // End dup != uup | |
1058 } // End if Spill USE | |
1059 } // End For All Inputs | |
1060 } // End If not nullcheck | |
1061 | |
1062 // ********** Handle DEFS ********** | |
1063 // DEFS either Split DOWN in HRP regions or when the LRG is bound, or | |
1064 // just reset the Reaches info in LRP regions. DEFS must always update | |
1065 // UP info. | |
1066 if( deflrg.reg() >= LRG::SPILL_REG ) { // Spilled? | |
1067 uint slidx = lrg2reach[defidx]; | |
1068 // Add to defs list for later assignment of new live range number | |
1069 defs->push(n); | |
1070 // Set a flag on the Node indicating it has already spilled. | |
1071 // Only do it for capacity spills not conflict spills. | |
1072 if( !deflrg._direct_conflict ) | |
1073 set_was_spilled(n); | |
1074 assert(!n->is_Phi(),"Cannot insert Phi into DEFS list"); | |
1075 // Grab UP info for DEF | |
1076 const RegMask &dmask = n->out_RegMask(); | |
1077 bool defup = dmask.is_UP(); | |
1078 // Only split at Def if this is a HRP block or bound (and spilled once) | |
1079 if( !n->rematerialize() && | |
1080 (((dmask.is_bound1() || dmask.is_bound2() || dmask.is_misaligned_Pair()) && | |
1081 (deflrg._direct_conflict || deflrg._must_spill)) || | |
1082 // Check for LRG being up in a register and we are inside a high | |
1083 // pressure area. Spill it down immediately. | |
1084 (defup && is_high_pressure(b,&deflrg,insidx))) ) { | |
1085 assert( !n->rematerialize(), "" ); | |
1086 assert( !n->is_SpillCopy(), "" ); | |
1087 // Do a split at the def site. | |
1088 maxlrg = split_DEF( n, b, insidx, maxlrg, Reachblock, debug_defs, splits, slidx ); | |
1089 // If it wasn't split bail | |
1090 if (!maxlrg) { | |
1091 return 0; | |
1092 } | |
1093 // Split DEF's Down | |
1094 UPblock[slidx] = 0; | |
1095 #ifndef PRODUCT | |
1096 // DEBUG | |
1097 if( trace_spilling() ) { | |
1098 tty->print("\nNew Split DOWN DEF of Spill Idx "); | |
1099 tty->print("%d, UP %d:\n",slidx,false); | |
1100 n->dump(); | |
1101 } | |
1102 #endif | |
1103 } | |
1104 else { // Neither bound nor HRP, must be LRP | |
1105 // otherwise, just record the def | |
1106 Reachblock[slidx] = n; | |
1107 // UP should come from the outRegmask() of the DEF | |
1108 UPblock[slidx] = defup; | |
1109 // Update debug list of reaching down definitions, kill if DEF is UP | |
1110 debug_defs[slidx] = defup ? NULL : n; | |
1111 #ifndef PRODUCT | |
1112 // DEBUG | |
1113 if( trace_spilling() ) { | |
1114 tty->print("\nNew DEF of Spill Idx "); | |
1115 tty->print("%d, UP %d:\n",slidx,defup); | |
1116 n->dump(); | |
1117 } | |
1118 #endif | |
1119 } // End else LRP | |
1120 } // End if spill def | |
1121 | |
1122 // ********** Split Left Over Mem-Mem Moves ********** | |
1123 // Check for mem-mem copies and split them now. Do not do this | |
1124 // to copies about to be spilled; they will be Split shortly. | |
1125 if( copyidx ) { | |
1126 Node *use = n->in(copyidx); | |
1127 uint useidx = Find_id(use); | |
1128 if( useidx < _maxlrg && // This is not a new split | |
1129 OptoReg::is_stack(deflrg.reg()) && | |
1130 deflrg.reg() < LRG::SPILL_REG ) { // And DEF is from stack | |
1131 LRG &uselrg = lrgs(useidx); | |
1132 if( OptoReg::is_stack(uselrg.reg()) && | |
1133 uselrg.reg() < LRG::SPILL_REG && // USE is from stack | |
1134 deflrg.reg() != uselrg.reg() ) { // Not trivially removed | |
1135 uint def_ideal_reg = Matcher::base2reg[n->bottom_type()->base()]; | |
1136 const RegMask &def_rm = *Matcher::idealreg2regmask[def_ideal_reg]; | |
1137 const RegMask &use_rm = n->in_RegMask(copyidx); | |
1138 if( def_rm.overlap(use_rm) && n->is_SpillCopy() ) { // Bug 4707800, 'n' may be a storeSSL | |
1139 if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { // Check when generating nodes | |
1140 return 0; | |
1141 } | |
1142 Node *spill = new (C) MachSpillCopyNode(use,use_rm,def_rm); | |
1143 n->set_req(copyidx,spill); | |
1144 n->as_MachSpillCopy()->set_in_RegMask(def_rm); | |
1145 // Put the spill just before the copy | |
1146 insert_proj( b, insidx++, spill, maxlrg++ ); | |
1147 } | |
1148 } | |
1149 } | |
1150 } | |
1151 } // End For All Instructions in Block - Non-PHI Pass | |
1152 | |
1153 // Check if each LRG is live out of this block so as not to propagate | |
1154 // beyond the last use of a LRG. | |
1155 for( slidx = 0; slidx < spill_cnt; slidx++ ) { | |
1156 uint defidx = lidxs.at(slidx); | |
1157 IndexSet *liveout = _live->live(b); | |
1158 if( !liveout->member(defidx) ) { | |
1159 #ifdef ASSERT | |
1160 // The index defidx is not live. Check the liveout array to ensure that | |
1161 // it contains no members which compress to defidx. Finding such an | |
1162 // instance may be a case to add liveout adjustment in compress_uf_map(). | |
1163 // See 5063219. | |
1164 uint member; | |
1165 IndexSetIterator isi(liveout); | |
1166 while ((member = isi.next()) != 0) { | |
1167 assert(defidx != Find_const(member), "Live out member has not been compressed"); | |
1168 } | |
1169 #endif | |
1170 Reachblock[slidx] = NULL; | |
1171 } else { | |
1172 assert(Reachblock[slidx] != NULL,"No reaching definition for liveout value"); | |
1173 } | |
1174 } | |
1175 #ifndef PRODUCT | |
1176 if( trace_spilling() ) | |
1177 b->dump(); | |
1178 #endif | |
1179 } // End For All Blocks | |
1180 | |
1181 //----------PASS 2---------- | |
1182 // Reset all DEF live range numbers here | |
1183 for( insidx = 0; insidx < defs->size(); insidx++ ) { | |
1184 // Grab the def | |
1185 n1 = defs->at(insidx); | |
1186 // Set new lidx for DEF | |
1187 new_lrg(n1, maxlrg++); | |
1188 } | |
1189 //----------Phi Node Splitting---------- | |
1190 // Clean up a phi here, and assign a new live range number | |
1191 // Cycle through this block's predecessors, collecting Reaches | |
1192 // info for each spilled LRG and update edges. | |
1193 // Walk the phis list to patch inputs, split phis, and name phis | |
1194 for( insidx = 0; insidx < phis->size(); insidx++ ) { | |
1195 Node *phi = phis->at(insidx); | |
1196 assert(phi->is_Phi(),"This list must only contain Phi Nodes"); | |
1197 Block *b = _cfg._bbs[phi->_idx]; | |
1198 // Grab the live range number | |
1199 uint lidx = Find_id(phi); | |
1200 uint slidx = lrg2reach[lidx]; | |
1201 // Update node to lidx map | |
1202 new_lrg(phi, maxlrg++); | |
1203 // Get PASS1's up/down decision for the block. | |
1204 int phi_up = !!UP_entry[slidx]->test(b->_pre_order); | |
1205 | |
1206 // Force down if double-spilling live range | |
1207 if( lrgs(lidx)._was_spilled1 ) | |
1208 phi_up = false; | |
1209 | |
1210 // When splitting a Phi we an split it normal or "inverted". | |
1211 // An inverted split makes the splits target the Phi's UP/DOWN | |
1212 // sense inverted; then the Phi is followed by a final def-side | |
1213 // split to invert back. It changes which blocks the spill code | |
1214 // goes in. | |
1215 | |
1216 // Walk the predecessor blocks and assign the reaching def to the Phi. | |
1217 // Split Phi nodes by placing USE side splits wherever the reaching | |
1218 // DEF has the wrong UP/DOWN value. | |
1219 for( uint i = 1; i < b->num_preds(); i++ ) { | |
1220 // Get predecessor block pre-order number | |
1221 Block *pred = _cfg._bbs[b->pred(i)->_idx]; | |
1222 pidx = pred->_pre_order; | |
1223 // Grab reaching def | |
1224 Node *def = Reaches[pidx][slidx]; | |
1225 assert( def, "must have reaching def" ); | |
1226 // If input up/down sense and reg-pressure DISagree | |
1227 if( def->rematerialize() ) { | |
1228 def = split_Rematerialize( def, pred, pred->end_idx(), maxlrg, splits, slidx, lrg2reach, Reachblock, false ); | |
1229 if( !def ) return 0; // Bail out | |
1230 } | |
1231 // Update the Phi's input edge array | |
1232 phi->set_req(i,def); | |
1233 // Grab the UP/DOWN sense for the input | |
1234 u1 = UP[pidx][slidx]; | |
1235 if( u1 != (phi_up != 0)) { | |
1236 maxlrg = split_USE(def, b, phi, i, maxlrg, !u1, false, splits,slidx); | |
1237 // If it wasn't split bail | |
1238 if (!maxlrg) { | |
1239 return 0; | |
1240 } | |
1241 } | |
1242 } // End for all inputs to the Phi | |
1243 } // End for all Phi Nodes | |
1244 // Update _maxlrg to save Union asserts | |
1245 _maxlrg = maxlrg; | |
1246 | |
1247 | |
1248 //----------PASS 3---------- | |
1249 // Pass over all Phi's to union the live ranges | |
1250 for( insidx = 0; insidx < phis->size(); insidx++ ) { | |
1251 Node *phi = phis->at(insidx); | |
1252 assert(phi->is_Phi(),"This list must only contain Phi Nodes"); | |
1253 // Walk all inputs to Phi and Union input live range with Phi live range | |
1254 for( uint i = 1; i < phi->req(); i++ ) { | |
1255 // Grab the input node | |
1256 Node *n = phi->in(i); | |
1257 assert( n, "" ); | |
1258 uint lidx = Find(n); | |
1259 uint pidx = Find(phi); | |
1260 if( lidx < pidx ) | |
1261 Union(n, phi); | |
1262 else if( lidx > pidx ) | |
1263 Union(phi, n); | |
1264 } // End for all inputs to the Phi Node | |
1265 } // End for all Phi Nodes | |
1266 // Now union all two address instructions | |
1267 for( insidx = 0; insidx < defs->size(); insidx++ ) { | |
1268 // Grab the def | |
1269 n1 = defs->at(insidx); | |
1270 // Set new lidx for DEF & handle 2-addr instructions | |
1271 if( n1->is_Mach() && ((twoidx = n1->as_Mach()->two_adr()) != 0) ) { | |
1272 assert( Find(n1->in(twoidx)) < maxlrg,"Assigning bad live range index"); | |
1273 // Union the input and output live ranges | |
1274 uint lr1 = Find(n1); | |
1275 uint lr2 = Find(n1->in(twoidx)); | |
1276 if( lr1 < lr2 ) | |
1277 Union(n1, n1->in(twoidx)); | |
1278 else if( lr1 > lr2 ) | |
1279 Union(n1->in(twoidx), n1); | |
1280 } // End if two address | |
1281 } // End for all defs | |
1282 // DEBUG | |
1283 #ifdef ASSERT | |
1284 // Validate all live range index assignments | |
1285 for( bidx = 0; bidx < _cfg._num_blocks; bidx++ ) { | |
1286 b = _cfg._blocks[bidx]; | |
1287 for( insidx = 0; insidx <= b->end_idx(); insidx++ ) { | |
1288 Node *n = b->_nodes[insidx]; | |
1289 uint defidx = Find(n); | |
1290 assert(defidx < _maxlrg,"Bad live range index in Split"); | |
1291 assert(defidx < maxlrg,"Bad live range index in Split"); | |
1292 } | |
1293 } | |
1294 // Issue a warning if splitting made no progress | |
1295 int noprogress = 0; | |
1296 for( slidx = 0; slidx < spill_cnt; slidx++ ) { | |
1297 if( PrintOpto && WizardMode && splits.at(slidx) == 0 ) { | |
1298 tty->print_cr("Failed to split live range %d", lidxs.at(slidx)); | |
1299 //BREAKPOINT; | |
1300 } | |
1301 else { | |
1302 noprogress++; | |
1303 } | |
1304 } | |
1305 if(!noprogress) { | |
1306 tty->print_cr("Failed to make progress in Split"); | |
1307 //BREAKPOINT; | |
1308 } | |
1309 #endif | |
1310 // Return updated count of live ranges | |
1311 return maxlrg; | |
1312 } |