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
|
|
287 if( lrgs(lidx)._def != NodeSentinel ) continue;
|
|
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
|
|
314 if( walkThru )
|
|
315 while ( in->is_SpillCopy() && lidx >= _maxlrg ) {
|
|
316 in = in->in(1);
|
|
317 lidx = Find_id(in);
|
|
318 }
|
|
319
|
|
320 if( lidx < _maxlrg && lrgs(lidx).reg() >= LRG::SPILL_REG ) {
|
|
321 Node *rdef = Reachblock[lrg2reach[lidx]];
|
|
322 if( rdef ) spill->set_req(i,rdef);
|
|
323 }
|
|
324 }
|
|
325 }
|
|
326
|
|
327
|
|
328 assert( spill->out_RegMask().is_UP(), "rematerialize to a reg" );
|
|
329 // Rematerialized op is def->spilled+1
|
|
330 set_was_spilled(spill);
|
|
331 if( _spilled_once.test(def->_idx) )
|
|
332 set_was_spilled(spill);
|
|
333
|
|
334 insert_proj( b, insidx, spill, maxlrg++ );
|
|
335 #ifdef ASSERT
|
|
336 // Increment the counter for this lrg
|
|
337 splits.at_put(slidx, splits.at(slidx)+1);
|
|
338 #endif
|
|
339 // See if the cloned def kills any flags, and copy those kills as well
|
|
340 uint i = insidx+1;
|
|
341 if( clone_projs( b, i, def, spill, maxlrg ) ) {
|
|
342 // Adjust the point where we go hi-pressure
|
|
343 if( i <= b->_ihrp_index ) b->_ihrp_index++;
|
|
344 if( i <= b->_fhrp_index ) b->_fhrp_index++;
|
|
345 }
|
|
346
|
|
347 return spill;
|
|
348 }
|
|
349
|
|
350 //------------------------------is_high_pressure-------------------------------
|
|
351 // Function to compute whether or not this live range is "high pressure"
|
|
352 // in this block - whether it spills eagerly or not.
|
|
353 bool PhaseChaitin::is_high_pressure( Block *b, LRG *lrg, uint insidx ) {
|
|
354 if( lrg->_was_spilled1 ) return true;
|
|
355 // Forced spilling due to conflict? Then split only at binding uses
|
|
356 // or defs, not for supposed capacity problems.
|
|
357 // CNC - Turned off 7/8/99, causes too much spilling
|
|
358 // if( lrg->_is_bound ) return false;
|
|
359
|
|
360 // Not yet reached the high-pressure cutoff point, so low pressure
|
|
361 uint hrp_idx = lrg->_is_float ? b->_fhrp_index : b->_ihrp_index;
|
|
362 if( insidx < hrp_idx ) return false;
|
|
363 // Register pressure for the block as a whole depends on reg class
|
|
364 int block_pres = lrg->_is_float ? b->_freg_pressure : b->_reg_pressure;
|
|
365 // Bound live ranges will split at the binding points first;
|
|
366 // Intermediate splits should assume the live range's register set
|
|
367 // got "freed up" and that num_regs will become INT_PRESSURE.
|
|
368 int bound_pres = lrg->_is_float ? FLOATPRESSURE : INTPRESSURE;
|
|
369 // Effective register pressure limit.
|
|
370 int lrg_pres = (lrg->get_invalid_mask_size() > lrg->num_regs())
|
|
371 ? (lrg->get_invalid_mask_size() >> (lrg->num_regs()-1)) : bound_pres;
|
|
372 // High pressure if block pressure requires more register freedom
|
|
373 // than live range has.
|
|
374 return block_pres >= lrg_pres;
|
|
375 }
|
|
376
|
|
377
|
|
378 //------------------------------prompt_use---------------------------------
|
|
379 // True if lidx is used before any real register is def'd in the block
|
|
380 bool PhaseChaitin::prompt_use( Block *b, uint lidx ) {
|
|
381 if( lrgs(lidx)._was_spilled2 ) return false;
|
|
382
|
|
383 // Scan block for 1st use.
|
|
384 for( uint i = 1; i <= b->end_idx(); i++ ) {
|
|
385 Node *n = b->_nodes[i];
|
|
386 // Ignore PHI use, these can be up or down
|
|
387 if( n->is_Phi() ) continue;
|
|
388 for( uint j = 1; j < n->req(); j++ )
|
|
389 if( Find_id(n->in(j)) == lidx )
|
|
390 return true; // Found 1st use!
|
|
391 if( n->out_RegMask().is_NotEmpty() ) return false;
|
|
392 }
|
|
393 return false;
|
|
394 }
|
|
395
|
|
396 //------------------------------Split--------------------------------------
|
|
397 //----------Split Routine----------
|
|
398 // ***** NEW SPLITTING HEURISTIC *****
|
|
399 // DEFS: If the DEF is in a High Register Pressure(HRP) Block, split there.
|
|
400 // Else, no split unless there is a HRP block between a DEF and
|
|
401 // one of its uses, and then split at the HRP block.
|
|
402 //
|
|
403 // USES: If USE is in HRP, split at use to leave main LRG on stack.
|
|
404 // Else, hoist LRG back up to register only (ie - split is also DEF)
|
|
405 // We will compute a new maxlrg as we go
|
|
406 uint PhaseChaitin::Split( uint maxlrg ) {
|
|
407 NOT_PRODUCT( Compile::TracePhase t3("regAllocSplit", &_t_regAllocSplit, TimeCompiler); )
|
|
408
|
|
409 uint bidx, pidx, slidx, insidx, inpidx, twoidx;
|
|
410 uint non_phi = 1, spill_cnt = 0;
|
|
411 Node **Reachblock;
|
|
412 Node *n1, *n2, *n3;
|
|
413 Node_List *defs,*phis;
|
|
414 bool *UPblock;
|
|
415 bool u1, u2, u3;
|
|
416 Block *b, *pred;
|
|
417 PhiNode *phi;
|
|
418 GrowableArray<uint> lidxs;
|
|
419
|
|
420 // Array of counters to count splits per live range
|
|
421 GrowableArray<uint> splits;
|
|
422
|
|
423 //----------Setup Code----------
|
|
424 // Create a convenient mapping from lrg numbers to reaches/leaves indices
|
|
425 uint *lrg2reach = NEW_RESOURCE_ARRAY( uint, _maxlrg );
|
|
426 // Keep track of DEFS & Phis for later passes
|
|
427 defs = new Node_List();
|
|
428 phis = new Node_List();
|
|
429 // Gather info on which LRG's are spilling, and build maps
|
|
430 for( bidx = 1; bidx < _maxlrg; bidx++ ) {
|
|
431 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) {
|
|
432 assert(!lrgs(bidx).mask().is_AllStack(),"AllStack should color");
|
|
433 lrg2reach[bidx] = spill_cnt;
|
|
434 spill_cnt++;
|
|
435 lidxs.append(bidx);
|
|
436 #ifdef ASSERT
|
|
437 // Initialize the split counts to zero
|
|
438 splits.append(0);
|
|
439 #endif
|
|
440 #ifndef PRODUCT
|
|
441 if( PrintOpto && WizardMode && lrgs(bidx)._was_spilled1 )
|
|
442 tty->print_cr("Warning, 2nd spill of L%d",bidx);
|
|
443 #endif
|
|
444 }
|
|
445 }
|
|
446
|
|
447 // Create side arrays for propagating reaching defs info.
|
|
448 // Each block needs a node pointer for each spilling live range for the
|
|
449 // Def which is live into the block. Phi nodes handle multiple input
|
|
450 // Defs by querying the output of their predecessor blocks and resolving
|
|
451 // them to a single Def at the phi. The pointer is updated for each
|
|
452 // Def in the block, and then becomes the output for the block when
|
|
453 // processing of the block is complete. We also need to track whether
|
|
454 // a Def is UP or DOWN. UP means that it should get a register (ie -
|
|
455 // it is always in LRP regions), and DOWN means that it is probably
|
|
456 // on the stack (ie - it crosses HRP regions).
|
|
457 Node ***Reaches = NEW_RESOURCE_ARRAY( Node**, _cfg._num_blocks+1 );
|
|
458 bool **UP = NEW_RESOURCE_ARRAY( bool*, _cfg._num_blocks+1 );
|
|
459 Node **debug_defs = NEW_RESOURCE_ARRAY( Node*, spill_cnt );
|
|
460 VectorSet **UP_entry= NEW_RESOURCE_ARRAY( VectorSet*, spill_cnt );
|
|
461
|
|
462 // Initialize Reaches & UP
|
|
463 for( bidx = 0; bidx < _cfg._num_blocks+1; bidx++ ) {
|
|
464 Reaches[bidx] = NEW_RESOURCE_ARRAY( Node*, spill_cnt );
|
|
465 UP[bidx] = NEW_RESOURCE_ARRAY( bool, spill_cnt );
|
|
466 Node **Reachblock = Reaches[bidx];
|
|
467 bool *UPblock = UP[bidx];
|
|
468 for( slidx = 0; slidx < spill_cnt; slidx++ ) {
|
|
469 UPblock[slidx] = true; // Assume they start in registers
|
|
470 Reachblock[slidx] = NULL; // Assume that no def is present
|
|
471 }
|
|
472 }
|
|
473
|
|
474 // Initialize to array of empty vectorsets
|
|
475 for( slidx = 0; slidx < spill_cnt; slidx++ )
|
|
476 UP_entry[slidx] = new VectorSet(Thread::current()->resource_area());
|
|
477
|
|
478 //----------PASS 1----------
|
|
479 //----------Propagation & Node Insertion Code----------
|
|
480 // Walk the Blocks in RPO for DEF & USE info
|
|
481 for( bidx = 0; bidx < _cfg._num_blocks; bidx++ ) {
|
|
482
|
|
483 if (C->check_node_count(spill_cnt, out_of_nodes)) {
|
|
484 return 0;
|
|
485 }
|
|
486
|
|
487 b = _cfg._blocks[bidx];
|
|
488 // Reaches & UP arrays for this block
|
|
489 Reachblock = Reaches[b->_pre_order];
|
|
490 UPblock = UP[b->_pre_order];
|
|
491 // Reset counter of start of non-Phi nodes in block
|
|
492 non_phi = 1;
|
|
493 //----------Block Entry Handling----------
|
|
494 // Check for need to insert a new phi
|
|
495 // Cycle through this block's predecessors, collecting Reaches
|
|
496 // info for each spilled LRG. If they are identical, no phi is
|
|
497 // needed. If they differ, check for a phi, and insert if missing,
|
|
498 // or update edges if present. Set current block's Reaches set to
|
|
499 // be either the phi's or the reaching def, as appropriate.
|
|
500 // If no Phi is needed, check if the LRG needs to spill on entry
|
|
501 // to the block due to HRP.
|
|
502 for( slidx = 0; slidx < spill_cnt; slidx++ ) {
|
|
503 // Grab the live range number
|
|
504 uint lidx = lidxs.at(slidx);
|
|
505 // Do not bother splitting or putting in Phis for single-def
|
|
506 // rematerialized live ranges. This happens alot to constants
|
|
507 // with long live ranges.
|
|
508 if( lrgs(lidx)._def != NodeSentinel &&
|
|
509 lrgs(lidx)._def->rematerialize() ) {
|
|
510 // reset the Reaches & UP entries
|
|
511 Reachblock[slidx] = lrgs(lidx)._def;
|
|
512 UPblock[slidx] = true;
|
|
513 // Record following instruction in case 'n' rematerializes and
|
|
514 // kills flags
|
|
515 Block *pred1 = _cfg._bbs[b->pred(1)->_idx];
|
|
516 continue;
|
|
517 }
|
|
518
|
|
519 // Initialize needs_phi and needs_split
|
|
520 bool needs_phi = false;
|
|
521 bool needs_split = false;
|
|
522 // Walk the predecessor blocks to check inputs for that live range
|
|
523 // Grab predecessor block header
|
|
524 n1 = b->pred(1);
|
|
525 // Grab the appropriate reaching def info for inpidx
|
|
526 pred = _cfg._bbs[n1->_idx];
|
|
527 pidx = pred->_pre_order;
|
|
528 Node **Ltmp = Reaches[pidx];
|
|
529 bool *Utmp = UP[pidx];
|
|
530 n1 = Ltmp[slidx];
|
|
531 u1 = Utmp[slidx];
|
|
532 // Initialize node for saving type info
|
|
533 n3 = n1;
|
|
534 u3 = u1;
|
|
535
|
|
536 // Compare inputs to see if a Phi is needed
|
|
537 for( inpidx = 2; inpidx < b->num_preds(); inpidx++ ) {
|
|
538 // Grab predecessor block headers
|
|
539 n2 = b->pred(inpidx);
|
|
540 // Grab the appropriate reaching def info for inpidx
|
|
541 pred = _cfg._bbs[n2->_idx];
|
|
542 pidx = pred->_pre_order;
|
|
543 Ltmp = Reaches[pidx];
|
|
544 Utmp = UP[pidx];
|
|
545 n2 = Ltmp[slidx];
|
|
546 u2 = Utmp[slidx];
|
|
547 // For each LRG, decide if a phi is necessary
|
|
548 if( n1 != n2 ) {
|
|
549 needs_phi = true;
|
|
550 }
|
|
551 // See if the phi has mismatched inputs, UP vs. DOWN
|
|
552 if( n1 && n2 && (u1 != u2) ) {
|
|
553 needs_split = true;
|
|
554 }
|
|
555 // Move n2/u2 to n1/u1 for next iteration
|
|
556 n1 = n2;
|
|
557 u1 = u2;
|
|
558 // Preserve a non-NULL predecessor for later type referencing
|
|
559 if( (n3 == NULL) && (n2 != NULL) ){
|
|
560 n3 = n2;
|
|
561 u3 = u2;
|
|
562 }
|
|
563 } // End for all potential Phi inputs
|
|
564
|
|
565 // If a phi is needed, check for it
|
|
566 if( needs_phi ) {
|
|
567 // check block for appropriate phinode & update edges
|
|
568 for( insidx = 1; insidx <= b->end_idx(); insidx++ ) {
|
|
569 n1 = b->_nodes[insidx];
|
|
570 // bail if this is not a phi
|
|
571 phi = n1->is_Phi() ? n1->as_Phi() : NULL;
|
|
572 if( phi == NULL ) {
|
|
573 // Keep track of index of first non-PhiNode instruction in block
|
|
574 non_phi = insidx;
|
|
575 // break out of the for loop as we have handled all phi nodes
|
|
576 break;
|
|
577 }
|
|
578 // must be looking at a phi
|
|
579 if( Find_id(n1) == lidxs.at(slidx) ) {
|
|
580 // found the necessary phi
|
|
581 needs_phi = false;
|
|
582 // initialize the Reaches entry for this LRG
|
|
583 Reachblock[slidx] = phi;
|
|
584 break;
|
|
585 } // end if found correct phi
|
|
586 } // end for all phi's
|
|
587 // add new phinode if one not already found
|
|
588 if( needs_phi ) {
|
|
589 // create a new phi node and insert it into the block
|
|
590 // type is taken from left over pointer to a predecessor
|
|
591 assert(n3,"No non-NULL reaching DEF for a Phi");
|
|
592 phi = new (C, b->num_preds()) PhiNode(b->head(), n3->bottom_type());
|
|
593 // initialize the Reaches entry for this LRG
|
|
594 Reachblock[slidx] = phi;
|
|
595
|
|
596 // add node to block & node_to_block mapping
|
|
597 insert_proj( b, insidx++, phi, maxlrg++ );
|
|
598 non_phi++;
|
|
599 // Reset new phi's mapping to be the spilling live range
|
|
600 _names.map(phi->_idx, lidx);
|
|
601 assert(Find_id(phi) == lidx,"Bad update on Union-Find mapping");
|
|
602 } // end if not found correct phi
|
|
603 // Here you have either found or created the Phi, so record it
|
|
604 assert(phi != NULL,"Must have a Phi Node here");
|
|
605 phis->push(phi);
|
|
606 // PhiNodes should either force the LRG UP or DOWN depending
|
|
607 // on its inputs and the register pressure in the Phi's block.
|
|
608 UPblock[slidx] = true; // Assume new DEF is UP
|
|
609 // If entering a high-pressure area with no immediate use,
|
|
610 // assume Phi is DOWN
|
|
611 if( is_high_pressure( b, &lrgs(lidx), b->end_idx()) && !prompt_use(b,lidx) )
|
|
612 UPblock[slidx] = false;
|
|
613 // If we are not split up/down and all inputs are down, then we
|
|
614 // are down
|
|
615 if( !needs_split && !u3 )
|
|
616 UPblock[slidx] = false;
|
|
617 } // end if phi is needed
|
|
618
|
|
619 // Do not need a phi, so grab the reaching DEF
|
|
620 else {
|
|
621 // Grab predecessor block header
|
|
622 n1 = b->pred(1);
|
|
623 // Grab the appropriate reaching def info for k
|
|
624 pred = _cfg._bbs[n1->_idx];
|
|
625 pidx = pred->_pre_order;
|
|
626 Node **Ltmp = Reaches[pidx];
|
|
627 bool *Utmp = UP[pidx];
|
|
628 // reset the Reaches & UP entries
|
|
629 Reachblock[slidx] = Ltmp[slidx];
|
|
630 UPblock[slidx] = Utmp[slidx];
|
|
631 } // end else no Phi is needed
|
|
632 } // end for all spilling live ranges
|
|
633 // DEBUG
|
|
634 #ifndef PRODUCT
|
|
635 if(trace_spilling()) {
|
|
636 tty->print("/`\nBlock %d: ", b->_pre_order);
|
|
637 tty->print("Reaching Definitions after Phi handling\n");
|
|
638 for( uint x = 0; x < spill_cnt; x++ ) {
|
|
639 tty->print("Spill Idx %d: UP %d: Node\n",x,UPblock[x]);
|
|
640 if( Reachblock[x] )
|
|
641 Reachblock[x]->dump();
|
|
642 else
|
|
643 tty->print("Undefined\n");
|
|
644 }
|
|
645 }
|
|
646 #endif
|
|
647
|
|
648 //----------Non-Phi Node Splitting----------
|
|
649 // Since phi-nodes have now been handled, the Reachblock array for this
|
|
650 // block is initialized with the correct starting value for the defs which
|
|
651 // reach non-phi instructions in this block. Thus, process non-phi
|
|
652 // instructions normally, inserting SpillCopy nodes for all spill
|
|
653 // locations.
|
|
654
|
|
655 // Memoize any DOWN reaching definitions for use as DEBUG info
|
|
656 for( insidx = 0; insidx < spill_cnt; insidx++ ) {
|
|
657 debug_defs[insidx] = (UPblock[insidx]) ? NULL : Reachblock[insidx];
|
|
658 if( UPblock[insidx] ) // Memoize UP decision at block start
|
|
659 UP_entry[insidx]->set( b->_pre_order );
|
|
660 }
|
|
661
|
|
662 //----------Walk Instructions in the Block and Split----------
|
|
663 // For all non-phi instructions in the block
|
|
664 for( insidx = 1; insidx <= b->end_idx(); insidx++ ) {
|
|
665 Node *n = b->_nodes[insidx];
|
|
666 // Find the defining Node's live range index
|
|
667 uint defidx = Find_id(n);
|
|
668 uint cnt = n->req();
|
|
669
|
|
670 if( n->is_Phi() ) {
|
|
671 // Skip phi nodes after removing dead copies.
|
|
672 if( defidx < _maxlrg ) {
|
|
673 // Check for useless Phis. These appear if we spill, then
|
|
674 // coalesce away copies. Dont touch Phis in spilling live
|
|
675 // ranges; they are busy getting modifed in this pass.
|
|
676 if( lrgs(defidx).reg() < LRG::SPILL_REG ) {
|
|
677 uint i;
|
|
678 Node *u = NULL;
|
|
679 // Look for the Phi merging 2 unique inputs
|
|
680 for( i = 1; i < cnt; i++ ) {
|
|
681 // Ignore repeats and self
|
|
682 if( n->in(i) != u && n->in(i) != n ) {
|
|
683 // Found a unique input
|
|
684 if( u != NULL ) // If it's the 2nd, bail out
|
|
685 break;
|
|
686 u = n->in(i); // Else record it
|
|
687 }
|
|
688 }
|
|
689 assert( u, "at least 1 valid input expected" );
|
|
690 if( i >= cnt ) { // Didn't find 2+ unique inputs?
|
|
691 n->replace_by(u); // Then replace with unique input
|
|
692 n->disconnect_inputs(NULL);
|
|
693 b->_nodes.remove(insidx);
|
|
694 insidx--;
|
|
695 b->_ihrp_index--;
|
|
696 b->_fhrp_index--;
|
|
697 }
|
|
698 }
|
|
699 }
|
|
700 continue;
|
|
701 }
|
|
702 assert( insidx > b->_ihrp_index ||
|
|
703 (b->_reg_pressure < (uint)INTPRESSURE) ||
|
|
704 b->_ihrp_index > 4000000 ||
|
|
705 b->_ihrp_index >= b->end_idx() ||
|
|
706 !b->_nodes[b->_ihrp_index]->is_Proj(), "" );
|
|
707 assert( insidx > b->_fhrp_index ||
|
|
708 (b->_freg_pressure < (uint)FLOATPRESSURE) ||
|
|
709 b->_fhrp_index > 4000000 ||
|
|
710 b->_fhrp_index >= b->end_idx() ||
|
|
711 !b->_nodes[b->_fhrp_index]->is_Proj(), "" );
|
|
712
|
|
713 // ********** Handle Crossing HRP Boundry **********
|
|
714 if( (insidx == b->_ihrp_index) || (insidx == b->_fhrp_index) ) {
|
|
715 for( slidx = 0; slidx < spill_cnt; slidx++ ) {
|
|
716 // Check for need to split at HRP boundry - split if UP
|
|
717 n1 = Reachblock[slidx];
|
|
718 // bail out if no reaching DEF
|
|
719 if( n1 == NULL ) continue;
|
|
720 // bail out if live range is 'isolated' around inner loop
|
|
721 uint lidx = lidxs.at(slidx);
|
|
722 // If live range is currently UP
|
|
723 if( UPblock[slidx] ) {
|
|
724 // set location to insert spills at
|
|
725 // SPLIT DOWN HERE - NO CISC SPILL
|
|
726 if( is_high_pressure( b, &lrgs(lidx), insidx ) &&
|
|
727 !n1->rematerialize() ) {
|
|
728 // If there is already a valid stack definition available, use it
|
|
729 if( debug_defs[slidx] != NULL ) {
|
|
730 Reachblock[slidx] = debug_defs[slidx];
|
|
731 }
|
|
732 else {
|
|
733 // Insert point is just past last use or def in the block
|
|
734 int insert_point = insidx-1;
|
|
735 while( insert_point > 0 ) {
|
|
736 Node *n = b->_nodes[insert_point];
|
|
737 // Hit top of block? Quit going backwards
|
|
738 if( n->is_Phi() ) break;
|
|
739 // Found a def? Better split after it.
|
|
740 if( n2lidx(n) == lidx ) break;
|
|
741 // Look for a use
|
|
742 uint i;
|
|
743 for( i = 1; i < n->req(); i++ )
|
|
744 if( n2lidx(n->in(i)) == lidx )
|
|
745 break;
|
|
746 // Found a use? Better split after it.
|
|
747 if( i < n->req() ) break;
|
|
748 insert_point--;
|
|
749 }
|
|
750 maxlrg = split_DEF( n1, b, insert_point, maxlrg, Reachblock, debug_defs, splits, slidx);
|
|
751 // If it wasn't split bail
|
|
752 if (!maxlrg) {
|
|
753 return 0;
|
|
754 }
|
|
755 insidx++;
|
|
756 }
|
|
757 // This is a new DEF, so update UP
|
|
758 UPblock[slidx] = false;
|
|
759 #ifndef PRODUCT
|
|
760 // DEBUG
|
|
761 if( trace_spilling() ) {
|
|
762 tty->print("\nNew Split DOWN DEF of Spill Idx ");
|
|
763 tty->print("%d, UP %d:\n",slidx,false);
|
|
764 n1->dump();
|
|
765 }
|
|
766 #endif
|
|
767 }
|
|
768 } // end if LRG is UP
|
|
769 } // end for all spilling live ranges
|
|
770 assert( b->_nodes[insidx] == n, "got insidx set incorrectly" );
|
|
771 } // end if crossing HRP Boundry
|
|
772
|
|
773 // If the LRG index is oob, then this is a new spillcopy, skip it.
|
|
774 if( defidx >= _maxlrg ) {
|
|
775 continue;
|
|
776 }
|
|
777 LRG &deflrg = lrgs(defidx);
|
|
778 uint copyidx = n->is_Copy();
|
|
779 // Remove coalesced copy from CFG
|
|
780 if( copyidx && defidx == n2lidx(n->in(copyidx)) ) {
|
|
781 n->replace_by( n->in(copyidx) );
|
|
782 n->set_req( copyidx, NULL );
|
|
783 b->_nodes.remove(insidx--);
|
|
784 b->_ihrp_index--; // Adjust the point where we go hi-pressure
|
|
785 b->_fhrp_index--;
|
|
786 continue;
|
|
787 }
|
|
788
|
|
789 #define DERIVED 0
|
|
790
|
|
791 // ********** Handle USES **********
|
|
792 bool nullcheck = false;
|
|
793 // Implicit null checks never use the spilled value
|
|
794 if( n->is_MachNullCheck() )
|
|
795 nullcheck = true;
|
|
796 if( !nullcheck ) {
|
|
797 // Search all inputs for a Spill-USE
|
|
798 JVMState* jvms = n->jvms();
|
|
799 uint oopoff = jvms ? jvms->oopoff() : cnt;
|
|
800 uint old_last = cnt - 1;
|
|
801 for( inpidx = 1; inpidx < cnt; inpidx++ ) {
|
|
802 // Derived/base pairs may be added to our inputs during this loop.
|
|
803 // If inpidx > old_last, then one of these new inputs is being
|
|
804 // handled. Skip the derived part of the pair, but process
|
|
805 // the base like any other input.
|
|
806 if( inpidx > old_last && ((inpidx - oopoff) & 1) == DERIVED ) {
|
|
807 continue; // skip derived_debug added below
|
|
808 }
|
|
809 // Get lidx of input
|
|
810 uint useidx = Find_id(n->in(inpidx));
|
|
811 // Not a brand-new split, and it is a spill use
|
|
812 if( useidx < _maxlrg && lrgs(useidx).reg() >= LRG::SPILL_REG ) {
|
|
813 // Check for valid reaching DEF
|
|
814 slidx = lrg2reach[useidx];
|
|
815 Node *def = Reachblock[slidx];
|
|
816 assert( def != NULL, "Using Undefined Value in Split()\n");
|
|
817
|
|
818 // (+++) %%%% remove this in favor of pre-pass in matcher.cpp
|
|
819 // monitor references do not care where they live, so just hook
|
|
820 if ( jvms && jvms->is_monitor_use(inpidx) ) {
|
|
821 // The effect of this clone is to drop the node out of the block,
|
|
822 // so that the allocator does not see it anymore, and therefore
|
|
823 // does not attempt to assign it a register.
|
|
824 def = def->clone();
|
|
825 _names.extend(def->_idx,0);
|
|
826 _cfg._bbs.map(def->_idx,b);
|
|
827 n->set_req(inpidx, def);
|
|
828 if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) {
|
|
829 return 0;
|
|
830 }
|
|
831 continue;
|
|
832 }
|
|
833
|
|
834 // Rematerializable? Then clone def at use site instead
|
|
835 // of store/load
|
|
836 if( def->rematerialize() ) {
|
|
837 int old_size = b->_nodes.size();
|
|
838 def = split_Rematerialize( def, b, insidx, maxlrg, splits, slidx, lrg2reach, Reachblock, true );
|
|
839 if( !def ) return 0; // Bail out
|
|
840 insidx += b->_nodes.size()-old_size;
|
|
841 }
|
|
842
|
|
843 MachNode *mach = n->is_Mach() ? n->as_Mach() : NULL;
|
|
844 // Base pointers and oopmap references do not care where they live.
|
|
845 if ((inpidx >= oopoff) ||
|
|
846 (mach && mach->ideal_Opcode() == Op_AddP && inpidx == AddPNode::Base)) {
|
|
847 if (def->rematerialize() && lrgs(useidx)._was_spilled2) {
|
|
848 // This def has been rematerialized a couple of times without
|
|
849 // progress. It doesn't care if it lives UP or DOWN, so
|
|
850 // spill it down now.
|
|
851 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false,splits,slidx);
|
|
852 // If it wasn't split bail
|
|
853 if (!maxlrg) {
|
|
854 return 0;
|
|
855 }
|
|
856 insidx++; // Reset iterator to skip USE side split
|
|
857 } else {
|
|
858 // Just hook the def edge
|
|
859 n->set_req(inpidx, def);
|
|
860 }
|
|
861
|
|
862 if (inpidx >= oopoff) {
|
|
863 // After oopoff, we have derived/base pairs. We must mention all
|
|
864 // derived pointers here as derived/base pairs for GC. If the
|
|
865 // derived value is spilling and we have a copy both in Reachblock
|
|
866 // (called here 'def') and debug_defs[slidx] we need to mention
|
|
867 // both in derived/base pairs or kill one.
|
|
868 Node *derived_debug = debug_defs[slidx];
|
|
869 if( ((inpidx - oopoff) & 1) == DERIVED && // derived vs base?
|
|
870 mach && mach->ideal_Opcode() != Op_Halt &&
|
|
871 derived_debug != NULL &&
|
|
872 derived_debug != def ) { // Actual 2nd value appears
|
|
873 // We have already set 'def' as a derived value.
|
|
874 // Also set debug_defs[slidx] as a derived value.
|
|
875 uint k;
|
|
876 for( k = oopoff; k < cnt; k += 2 )
|
|
877 if( n->in(k) == derived_debug )
|
|
878 break; // Found an instance of debug derived
|
|
879 if( k == cnt ) {// No instance of debug_defs[slidx]
|
|
880 // Add a derived/base pair to cover the debug info.
|
|
881 // We have to process the added base later since it is not
|
|
882 // handled yet at this point but skip derived part.
|
|
883 assert(((n->req() - oopoff) & 1) == DERIVED,
|
|
884 "must match skip condition above");
|
|
885 n->add_req( derived_debug ); // this will be skipped above
|
|
886 n->add_req( n->in(inpidx+1) ); // this will be processed
|
|
887 // Increment cnt to handle added input edges on
|
|
888 // subsequent iterations.
|
|
889 cnt += 2;
|
|
890 }
|
|
891 }
|
|
892 }
|
|
893 continue;
|
|
894 }
|
|
895 // Special logic for DEBUG info
|
|
896 if( jvms && b->_freq > BLOCK_FREQUENCY(0.5) ) {
|
|
897 uint debug_start = jvms->debug_start();
|
|
898 // If this is debug info use & there is a reaching DOWN def
|
|
899 if ((debug_start <= inpidx) && (debug_defs[slidx] != NULL)) {
|
|
900 assert(inpidx < oopoff, "handle only debug info here");
|
|
901 // Just hook it in & move on
|
|
902 n->set_req(inpidx, debug_defs[slidx]);
|
|
903 // (Note that this can make two sides of a split live at the
|
|
904 // same time: The debug def on stack, and another def in a
|
|
905 // register. The GC needs to know about both of them, but any
|
|
906 // derived pointers after oopoff will refer to only one of the
|
|
907 // two defs and the GC would therefore miss the other. Thus
|
|
908 // this hack is only allowed for debug info which is Java state
|
|
909 // and therefore never a derived pointer.)
|
|
910 continue;
|
|
911 }
|
|
912 }
|
|
913 // Grab register mask info
|
|
914 const RegMask &dmask = def->out_RegMask();
|
|
915 const RegMask &umask = n->in_RegMask(inpidx);
|
|
916
|
|
917 assert(inpidx < oopoff, "cannot use-split oop map info");
|
|
918
|
|
919 bool dup = UPblock[slidx];
|
|
920 bool uup = umask.is_UP();
|
|
921
|
|
922 // Need special logic to handle bound USES. Insert a split at this
|
|
923 // bound use if we can't rematerialize the def, or if we need the
|
|
924 // split to form a misaligned pair.
|
|
925 if( !umask.is_AllStack() &&
|
|
926 (int)umask.Size() <= lrgs(useidx).num_regs() &&
|
|
927 (!def->rematerialize() ||
|
|
928 umask.is_misaligned_Pair())) {
|
|
929 // These need a Split regardless of overlap or pressure
|
|
930 // SPLIT - NO DEF - NO CISC SPILL
|
|
931 maxlrg = split_USE(def,b,n,inpidx,maxlrg,dup,false, splits,slidx);
|
|
932 // If it wasn't split bail
|
|
933 if (!maxlrg) {
|
|
934 return 0;
|
|
935 }
|
|
936 insidx++; // Reset iterator to skip USE side split
|
|
937 continue;
|
|
938 }
|
|
939 // Here is the logic chart which describes USE Splitting:
|
|
940 // 0 = false or DOWN, 1 = true or UP
|
|
941 //
|
|
942 // Overlap | DEF | USE | Action
|
|
943 //-------------------------------------------------------
|
|
944 // 0 | 0 | 0 | Copy - mem -> mem
|
|
945 // 0 | 0 | 1 | Split-UP - Check HRP
|
|
946 // 0 | 1 | 0 | Split-DOWN - Debug Info?
|
|
947 // 0 | 1 | 1 | Copy - reg -> reg
|
|
948 // 1 | 0 | 0 | Reset Input Edge (no Split)
|
|
949 // 1 | 0 | 1 | Split-UP - Check HRP
|
|
950 // 1 | 1 | 0 | Split-DOWN - Debug Info?
|
|
951 // 1 | 1 | 1 | Reset Input Edge (no Split)
|
|
952 //
|
|
953 // So, if (dup == uup), then overlap test determines action,
|
|
954 // with true being no split, and false being copy. Else,
|
|
955 // if DEF is DOWN, Split-UP, and check HRP to decide on
|
|
956 // resetting DEF. Finally if DEF is UP, Split-DOWN, with
|
|
957 // special handling for Debug Info.
|
|
958 if( dup == uup ) {
|
|
959 if( dmask.overlap(umask) ) {
|
|
960 // Both are either up or down, and there is overlap, No Split
|
|
961 n->set_req(inpidx, def);
|
|
962 }
|
|
963 else { // Both are either up or down, and there is no overlap
|
|
964 if( dup ) { // If UP, reg->reg copy
|
|
965 // COPY ACROSS HERE - NO DEF - NO CISC SPILL
|
|
966 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false, splits,slidx);
|
|
967 // If it wasn't split bail
|
|
968 if (!maxlrg) {
|
|
969 return 0;
|
|
970 }
|
|
971 insidx++; // Reset iterator to skip USE side split
|
|
972 }
|
|
973 else { // DOWN, mem->mem copy
|
|
974 // COPY UP & DOWN HERE - NO DEF - NO CISC SPILL
|
|
975 // First Split-UP to move value into Register
|
|
976 uint def_ideal = def->ideal_reg();
|
|
977 const RegMask* tmp_rm = Matcher::idealreg2regmask[def_ideal];
|
|
978 Node *spill = new (C) MachSpillCopyNode(def, dmask, *tmp_rm);
|
|
979 insert_proj( b, insidx, spill, maxlrg );
|
|
980 // Then Split-DOWN as if previous Split was DEF
|
|
981 maxlrg = split_USE(spill,b,n,inpidx,maxlrg,false,false, splits,slidx);
|
|
982 // If it wasn't split bail
|
|
983 if (!maxlrg) {
|
|
984 return 0;
|
|
985 }
|
|
986 insidx += 2; // Reset iterator to skip USE side splits
|
|
987 }
|
|
988 } // End else no overlap
|
|
989 } // End if dup == uup
|
|
990 // dup != uup, so check dup for direction of Split
|
|
991 else {
|
|
992 if( dup ) { // If UP, Split-DOWN and check Debug Info
|
|
993 // If this node is already a SpillCopy, just patch the edge
|
|
994 // except the case of spilling to stack.
|
|
995 if( n->is_SpillCopy() ) {
|
|
996 RegMask tmp_rm(umask);
|
|
997 tmp_rm.SUBTRACT(Matcher::STACK_ONLY_mask);
|
|
998 if( dmask.overlap(tmp_rm) ) {
|
|
999 if( def != n->in(inpidx) ) {
|
|
1000 n->set_req(inpidx, def);
|
|
1001 }
|
|
1002 continue;
|
|
1003 }
|
|
1004 }
|
|
1005 // COPY DOWN HERE - NO DEF - NO CISC SPILL
|
|
1006 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false, splits,slidx);
|
|
1007 // If it wasn't split bail
|
|
1008 if (!maxlrg) {
|
|
1009 return 0;
|
|
1010 }
|
|
1011 insidx++; // Reset iterator to skip USE side split
|
|
1012 // Check for debug-info split. Capture it for later
|
|
1013 // debug splits of the same value
|
|
1014 if (jvms && jvms->debug_start() <= inpidx && inpidx < oopoff)
|
|
1015 debug_defs[slidx] = n->in(inpidx);
|
|
1016
|
|
1017 }
|
|
1018 else { // DOWN, Split-UP and check register pressure
|
|
1019 if( is_high_pressure( b, &lrgs(useidx), insidx ) ) {
|
|
1020 // COPY UP HERE - NO DEF - CISC SPILL
|
|
1021 maxlrg = split_USE(def,b,n,inpidx,maxlrg,true,true, splits,slidx);
|
|
1022 // If it wasn't split bail
|
|
1023 if (!maxlrg) {
|
|
1024 return 0;
|
|
1025 }
|
|
1026 insidx++; // Reset iterator to skip USE side split
|
|
1027 } else { // LRP
|
|
1028 // COPY UP HERE - WITH DEF - NO CISC SPILL
|
|
1029 maxlrg = split_USE(def,b,n,inpidx,maxlrg,true,false, splits,slidx);
|
|
1030 // If it wasn't split bail
|
|
1031 if (!maxlrg) {
|
|
1032 return 0;
|
|
1033 }
|
|
1034 // Flag this lift-up in a low-pressure block as
|
|
1035 // already-spilled, so if it spills again it will
|
|
1036 // spill hard (instead of not spilling hard and
|
|
1037 // coalescing away).
|
|
1038 set_was_spilled(n->in(inpidx));
|
|
1039 // Since this is a new DEF, update Reachblock & UP
|
|
1040 Reachblock[slidx] = n->in(inpidx);
|
|
1041 UPblock[slidx] = true;
|
|
1042 insidx++; // Reset iterator to skip USE side split
|
|
1043 }
|
|
1044 } // End else DOWN
|
|
1045 } // End dup != uup
|
|
1046 } // End if Spill USE
|
|
1047 } // End For All Inputs
|
|
1048 } // End If not nullcheck
|
|
1049
|
|
1050 // ********** Handle DEFS **********
|
|
1051 // DEFS either Split DOWN in HRP regions or when the LRG is bound, or
|
|
1052 // just reset the Reaches info in LRP regions. DEFS must always update
|
|
1053 // UP info.
|
|
1054 if( deflrg.reg() >= LRG::SPILL_REG ) { // Spilled?
|
|
1055 uint slidx = lrg2reach[defidx];
|
|
1056 // Add to defs list for later assignment of new live range number
|
|
1057 defs->push(n);
|
|
1058 // Set a flag on the Node indicating it has already spilled.
|
|
1059 // Only do it for capacity spills not conflict spills.
|
|
1060 if( !deflrg._direct_conflict )
|
|
1061 set_was_spilled(n);
|
|
1062 assert(!n->is_Phi(),"Cannot insert Phi into DEFS list");
|
|
1063 // Grab UP info for DEF
|
|
1064 const RegMask &dmask = n->out_RegMask();
|
|
1065 bool defup = dmask.is_UP();
|
|
1066 // Only split at Def if this is a HRP block or bound (and spilled once)
|
|
1067 if( !n->rematerialize() &&
|
|
1068 (((dmask.is_bound1() || dmask.is_bound2() || dmask.is_misaligned_Pair()) &&
|
|
1069 (deflrg._direct_conflict || deflrg._must_spill)) ||
|
|
1070 // Check for LRG being up in a register and we are inside a high
|
|
1071 // pressure area. Spill it down immediately.
|
|
1072 (defup && is_high_pressure(b,&deflrg,insidx))) ) {
|
|
1073 assert( !n->rematerialize(), "" );
|
|
1074 assert( !n->is_SpillCopy(), "" );
|
|
1075 // Do a split at the def site.
|
|
1076 maxlrg = split_DEF( n, b, insidx, maxlrg, Reachblock, debug_defs, splits, slidx );
|
|
1077 // If it wasn't split bail
|
|
1078 if (!maxlrg) {
|
|
1079 return 0;
|
|
1080 }
|
|
1081 // Split DEF's Down
|
|
1082 UPblock[slidx] = 0;
|
|
1083 #ifndef PRODUCT
|
|
1084 // DEBUG
|
|
1085 if( trace_spilling() ) {
|
|
1086 tty->print("\nNew Split DOWN DEF of Spill Idx ");
|
|
1087 tty->print("%d, UP %d:\n",slidx,false);
|
|
1088 n->dump();
|
|
1089 }
|
|
1090 #endif
|
|
1091 }
|
|
1092 else { // Neither bound nor HRP, must be LRP
|
|
1093 // otherwise, just record the def
|
|
1094 Reachblock[slidx] = n;
|
|
1095 // UP should come from the outRegmask() of the DEF
|
|
1096 UPblock[slidx] = defup;
|
|
1097 // Update debug list of reaching down definitions, kill if DEF is UP
|
|
1098 debug_defs[slidx] = defup ? NULL : n;
|
|
1099 #ifndef PRODUCT
|
|
1100 // DEBUG
|
|
1101 if( trace_spilling() ) {
|
|
1102 tty->print("\nNew DEF of Spill Idx ");
|
|
1103 tty->print("%d, UP %d:\n",slidx,defup);
|
|
1104 n->dump();
|
|
1105 }
|
|
1106 #endif
|
|
1107 } // End else LRP
|
|
1108 } // End if spill def
|
|
1109
|
|
1110 // ********** Split Left Over Mem-Mem Moves **********
|
|
1111 // Check for mem-mem copies and split them now. Do not do this
|
|
1112 // to copies about to be spilled; they will be Split shortly.
|
|
1113 if( copyidx ) {
|
|
1114 Node *use = n->in(copyidx);
|
|
1115 uint useidx = Find_id(use);
|
|
1116 if( useidx < _maxlrg && // This is not a new split
|
|
1117 OptoReg::is_stack(deflrg.reg()) &&
|
|
1118 deflrg.reg() < LRG::SPILL_REG ) { // And DEF is from stack
|
|
1119 LRG &uselrg = lrgs(useidx);
|
|
1120 if( OptoReg::is_stack(uselrg.reg()) &&
|
|
1121 uselrg.reg() < LRG::SPILL_REG && // USE is from stack
|
|
1122 deflrg.reg() != uselrg.reg() ) { // Not trivially removed
|
|
1123 uint def_ideal_reg = Matcher::base2reg[n->bottom_type()->base()];
|
|
1124 const RegMask &def_rm = *Matcher::idealreg2regmask[def_ideal_reg];
|
|
1125 const RegMask &use_rm = n->in_RegMask(copyidx);
|
|
1126 if( def_rm.overlap(use_rm) && n->is_SpillCopy() ) { // Bug 4707800, 'n' may be a storeSSL
|
|
1127 if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { // Check when generating nodes
|
|
1128 return 0;
|
|
1129 }
|
|
1130 Node *spill = new (C) MachSpillCopyNode(use,use_rm,def_rm);
|
|
1131 n->set_req(copyidx,spill);
|
|
1132 n->as_MachSpillCopy()->set_in_RegMask(def_rm);
|
|
1133 // Put the spill just before the copy
|
|
1134 insert_proj( b, insidx++, spill, maxlrg++ );
|
|
1135 }
|
|
1136 }
|
|
1137 }
|
|
1138 }
|
|
1139 } // End For All Instructions in Block - Non-PHI Pass
|
|
1140
|
|
1141 // Check if each LRG is live out of this block so as not to propagate
|
|
1142 // beyond the last use of a LRG.
|
|
1143 for( slidx = 0; slidx < spill_cnt; slidx++ ) {
|
|
1144 uint defidx = lidxs.at(slidx);
|
|
1145 IndexSet *liveout = _live->live(b);
|
|
1146 if( !liveout->member(defidx) ) {
|
|
1147 #ifdef ASSERT
|
|
1148 // The index defidx is not live. Check the liveout array to ensure that
|
|
1149 // it contains no members which compress to defidx. Finding such an
|
|
1150 // instance may be a case to add liveout adjustment in compress_uf_map().
|
|
1151 // See 5063219.
|
|
1152 uint member;
|
|
1153 IndexSetIterator isi(liveout);
|
|
1154 while ((member = isi.next()) != 0) {
|
|
1155 assert(defidx != Find_const(member), "Live out member has not been compressed");
|
|
1156 }
|
|
1157 #endif
|
|
1158 Reachblock[slidx] = NULL;
|
|
1159 } else {
|
|
1160 assert(Reachblock[slidx] != NULL,"No reaching definition for liveout value");
|
|
1161 }
|
|
1162 }
|
|
1163 #ifndef PRODUCT
|
|
1164 if( trace_spilling() )
|
|
1165 b->dump();
|
|
1166 #endif
|
|
1167 } // End For All Blocks
|
|
1168
|
|
1169 //----------PASS 2----------
|
|
1170 // Reset all DEF live range numbers here
|
|
1171 for( insidx = 0; insidx < defs->size(); insidx++ ) {
|
|
1172 // Grab the def
|
|
1173 n1 = defs->at(insidx);
|
|
1174 // Set new lidx for DEF
|
|
1175 new_lrg(n1, maxlrg++);
|
|
1176 }
|
|
1177 //----------Phi Node Splitting----------
|
|
1178 // Clean up a phi here, and assign a new live range number
|
|
1179 // Cycle through this block's predecessors, collecting Reaches
|
|
1180 // info for each spilled LRG and update edges.
|
|
1181 // Walk the phis list to patch inputs, split phis, and name phis
|
|
1182 for( insidx = 0; insidx < phis->size(); insidx++ ) {
|
|
1183 Node *phi = phis->at(insidx);
|
|
1184 assert(phi->is_Phi(),"This list must only contain Phi Nodes");
|
|
1185 Block *b = _cfg._bbs[phi->_idx];
|
|
1186 // Grab the live range number
|
|
1187 uint lidx = Find_id(phi);
|
|
1188 uint slidx = lrg2reach[lidx];
|
|
1189 // Update node to lidx map
|
|
1190 new_lrg(phi, maxlrg++);
|
|
1191 // Get PASS1's up/down decision for the block.
|
|
1192 int phi_up = !!UP_entry[slidx]->test(b->_pre_order);
|
|
1193
|
|
1194 // Force down if double-spilling live range
|
|
1195 if( lrgs(lidx)._was_spilled1 )
|
|
1196 phi_up = false;
|
|
1197
|
|
1198 // When splitting a Phi we an split it normal or "inverted".
|
|
1199 // An inverted split makes the splits target the Phi's UP/DOWN
|
|
1200 // sense inverted; then the Phi is followed by a final def-side
|
|
1201 // split to invert back. It changes which blocks the spill code
|
|
1202 // goes in.
|
|
1203
|
|
1204 // Walk the predecessor blocks and assign the reaching def to the Phi.
|
|
1205 // Split Phi nodes by placing USE side splits wherever the reaching
|
|
1206 // DEF has the wrong UP/DOWN value.
|
|
1207 for( uint i = 1; i < b->num_preds(); i++ ) {
|
|
1208 // Get predecessor block pre-order number
|
|
1209 Block *pred = _cfg._bbs[b->pred(i)->_idx];
|
|
1210 pidx = pred->_pre_order;
|
|
1211 // Grab reaching def
|
|
1212 Node *def = Reaches[pidx][slidx];
|
|
1213 assert( def, "must have reaching def" );
|
|
1214 // If input up/down sense and reg-pressure DISagree
|
|
1215 if( def->rematerialize() ) {
|
|
1216 def = split_Rematerialize( def, pred, pred->end_idx(), maxlrg, splits, slidx, lrg2reach, Reachblock, false );
|
|
1217 if( !def ) return 0; // Bail out
|
|
1218 }
|
|
1219 // Update the Phi's input edge array
|
|
1220 phi->set_req(i,def);
|
|
1221 // Grab the UP/DOWN sense for the input
|
|
1222 u1 = UP[pidx][slidx];
|
|
1223 if( u1 != (phi_up != 0)) {
|
|
1224 maxlrg = split_USE(def, b, phi, i, maxlrg, !u1, false, splits,slidx);
|
|
1225 // If it wasn't split bail
|
|
1226 if (!maxlrg) {
|
|
1227 return 0;
|
|
1228 }
|
|
1229 }
|
|
1230 } // End for all inputs to the Phi
|
|
1231 } // End for all Phi Nodes
|
|
1232 // Update _maxlrg to save Union asserts
|
|
1233 _maxlrg = maxlrg;
|
|
1234
|
|
1235
|
|
1236 //----------PASS 3----------
|
|
1237 // Pass over all Phi's to union the live ranges
|
|
1238 for( insidx = 0; insidx < phis->size(); insidx++ ) {
|
|
1239 Node *phi = phis->at(insidx);
|
|
1240 assert(phi->is_Phi(),"This list must only contain Phi Nodes");
|
|
1241 // Walk all inputs to Phi and Union input live range with Phi live range
|
|
1242 for( uint i = 1; i < phi->req(); i++ ) {
|
|
1243 // Grab the input node
|
|
1244 Node *n = phi->in(i);
|
|
1245 assert( n, "" );
|
|
1246 uint lidx = Find(n);
|
|
1247 uint pidx = Find(phi);
|
|
1248 if( lidx < pidx )
|
|
1249 Union(n, phi);
|
|
1250 else if( lidx > pidx )
|
|
1251 Union(phi, n);
|
|
1252 } // End for all inputs to the Phi Node
|
|
1253 } // End for all Phi Nodes
|
|
1254 // Now union all two address instructions
|
|
1255 for( insidx = 0; insidx < defs->size(); insidx++ ) {
|
|
1256 // Grab the def
|
|
1257 n1 = defs->at(insidx);
|
|
1258 // Set new lidx for DEF & handle 2-addr instructions
|
|
1259 if( n1->is_Mach() && ((twoidx = n1->as_Mach()->two_adr()) != 0) ) {
|
|
1260 assert( Find(n1->in(twoidx)) < maxlrg,"Assigning bad live range index");
|
|
1261 // Union the input and output live ranges
|
|
1262 uint lr1 = Find(n1);
|
|
1263 uint lr2 = Find(n1->in(twoidx));
|
|
1264 if( lr1 < lr2 )
|
|
1265 Union(n1, n1->in(twoidx));
|
|
1266 else if( lr1 > lr2 )
|
|
1267 Union(n1->in(twoidx), n1);
|
|
1268 } // End if two address
|
|
1269 } // End for all defs
|
|
1270 // DEBUG
|
|
1271 #ifdef ASSERT
|
|
1272 // Validate all live range index assignments
|
|
1273 for( bidx = 0; bidx < _cfg._num_blocks; bidx++ ) {
|
|
1274 b = _cfg._blocks[bidx];
|
|
1275 for( insidx = 0; insidx <= b->end_idx(); insidx++ ) {
|
|
1276 Node *n = b->_nodes[insidx];
|
|
1277 uint defidx = Find(n);
|
|
1278 assert(defidx < _maxlrg,"Bad live range index in Split");
|
|
1279 assert(defidx < maxlrg,"Bad live range index in Split");
|
|
1280 }
|
|
1281 }
|
|
1282 // Issue a warning if splitting made no progress
|
|
1283 int noprogress = 0;
|
|
1284 for( slidx = 0; slidx < spill_cnt; slidx++ ) {
|
|
1285 if( PrintOpto && WizardMode && splits.at(slidx) == 0 ) {
|
|
1286 tty->print_cr("Failed to split live range %d", lidxs.at(slidx));
|
|
1287 //BREAKPOINT;
|
|
1288 }
|
|
1289 else {
|
|
1290 noprogress++;
|
|
1291 }
|
|
1292 }
|
|
1293 if(!noprogress) {
|
|
1294 tty->print_cr("Failed to make progress in Split");
|
|
1295 //BREAKPOINT;
|
|
1296 }
|
|
1297 #endif
|
|
1298 // Return updated count of live ranges
|
|
1299 return maxlrg;
|
|
1300 }
|