Mon, 20 Aug 2012 09:07:21 -0700
6340864: Implement vectorization optimizations in hotspot-server
Summary: Added asm encoding and mach nodes for vector arithmetic instructions on x86.
Reviewed-by: roland
1 /*
2 * Copyright (c) 2000, 2012, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "libadt/vectset.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/c2compiler.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/chaitin.hpp"
33 #include "opto/loopnode.hpp"
34 #include "opto/machnode.hpp"
36 //------------------------------Split--------------------------------------
37 // Walk the graph in RPO and for each lrg which spills, propagate reaching
38 // definitions. During propagation, split the live range around regions of
39 // High Register Pressure (HRP). If a Def is in a region of Low Register
40 // Pressure (LRP), it will not get spilled until we encounter a region of
41 // HRP between it and one of its uses. We will spill at the transition
42 // point between LRP and HRP. Uses in the HRP region will use the spilled
43 // Def. The first Use outside the HRP region will generate a SpillCopy to
44 // hoist the live range back up into a register, and all subsequent uses
45 // will use that new Def until another HRP region is encountered. Defs in
46 // HRP regions will get trailing SpillCopies to push the LRG down into the
47 // stack immediately.
48 //
49 // As a side effect, unlink from (hence make dead) coalesced copies.
50 //
52 static const char out_of_nodes[] = "out of nodes during split";
54 //------------------------------get_spillcopy_wide-----------------------------
55 // Get a SpillCopy node with wide-enough masks. Use the 'wide-mask', the
56 // wide ideal-register spill-mask if possible. If the 'wide-mask' does
57 // not cover the input (or output), use the input (or output) mask instead.
58 Node *PhaseChaitin::get_spillcopy_wide( Node *def, Node *use, uint uidx ) {
59 // If ideal reg doesn't exist we've got a bad schedule happening
60 // that is forcing us to spill something that isn't spillable.
61 // Bail rather than abort
62 int ireg = def->ideal_reg();
63 if( ireg == 0 || ireg == Op_RegFlags ) {
64 assert(false, "attempted to spill a non-spillable item");
65 C->record_method_not_compilable("attempted to spill a non-spillable item");
66 return NULL;
67 }
68 if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) {
69 return NULL;
70 }
71 const RegMask *i_mask = &def->out_RegMask();
72 const RegMask *w_mask = C->matcher()->idealreg2spillmask[ireg];
73 const RegMask *o_mask = use ? &use->in_RegMask(uidx) : w_mask;
74 const RegMask *w_i_mask = w_mask->overlap( *i_mask ) ? w_mask : i_mask;
75 const RegMask *w_o_mask;
77 int num_regs = RegMask::num_registers(ireg);
78 bool is_vect = RegMask::is_vector(ireg);
79 if( w_mask->overlap( *o_mask ) && // Overlap AND
80 ((num_regs == 1) // Single use or aligned
81 || is_vect // or vector
82 || !is_vect && o_mask->is_aligned_pairs()) ) {
83 assert(!is_vect || o_mask->is_aligned_sets(num_regs), "vectors are aligned");
84 // Don't come here for mis-aligned doubles
85 w_o_mask = w_mask;
86 } else { // wide ideal mask does not overlap with o_mask
87 // Mis-aligned doubles come here and XMM->FPR moves on x86.
88 w_o_mask = o_mask; // Must target desired registers
89 // Does the ideal-reg-mask overlap with o_mask? I.e., can I use
90 // a reg-reg move or do I need a trip across register classes
91 // (and thus through memory)?
92 if( !C->matcher()->idealreg2regmask[ireg]->overlap( *o_mask) && o_mask->is_UP() )
93 // Here we assume a trip through memory is required.
94 w_i_mask = &C->FIRST_STACK_mask();
95 }
96 return new (C) MachSpillCopyNode( def, *w_i_mask, *w_o_mask );
97 }
99 //------------------------------insert_proj------------------------------------
100 // Insert the spill at chosen location. Skip over any intervening Proj's or
101 // Phis. Skip over a CatchNode and projs, inserting in the fall-through block
102 // instead. Update high-pressure indices. Create a new live range.
103 void PhaseChaitin::insert_proj( Block *b, uint i, Node *spill, uint maxlrg ) {
104 // Skip intervening ProjNodes. Do not insert between a ProjNode and
105 // its definer.
106 while( i < b->_nodes.size() &&
107 (b->_nodes[i]->is_Proj() ||
108 b->_nodes[i]->is_Phi() ) )
109 i++;
111 // Do not insert between a call and his Catch
112 if( b->_nodes[i]->is_Catch() ) {
113 // Put the instruction at the top of the fall-thru block.
114 // Find the fall-thru projection
115 while( 1 ) {
116 const CatchProjNode *cp = b->_nodes[++i]->as_CatchProj();
117 if( cp->_con == CatchProjNode::fall_through_index )
118 break;
119 }
120 int sidx = i - b->end_idx()-1;
121 b = b->_succs[sidx]; // Switch to successor block
122 i = 1; // Right at start of block
123 }
125 b->_nodes.insert(i,spill); // Insert node in block
126 _cfg._bbs.map(spill->_idx,b); // Update node->block mapping to reflect
127 // Adjust the point where we go hi-pressure
128 if( i <= b->_ihrp_index ) b->_ihrp_index++;
129 if( i <= b->_fhrp_index ) b->_fhrp_index++;
131 // Assign a new Live Range Number to the SpillCopy and grow
132 // the node->live range mapping.
133 new_lrg(spill,maxlrg);
134 }
136 //------------------------------split_DEF--------------------------------------
137 // There are four categories of Split; UP/DOWN x DEF/USE
138 // Only three of these really occur as DOWN/USE will always color
139 // Any Split with a DEF cannot CISC-Spill now. Thus we need
140 // two helper routines, one for Split DEFS (insert after instruction),
141 // one for Split USES (insert before instruction). DEF insertion
142 // happens inside Split, where the Leaveblock array is updated.
143 uint PhaseChaitin::split_DEF( Node *def, Block *b, int loc, uint maxlrg, Node **Reachblock, Node **debug_defs, GrowableArray<uint> splits, int slidx ) {
144 #ifdef ASSERT
145 // Increment the counter for this lrg
146 splits.at_put(slidx, splits.at(slidx)+1);
147 #endif
148 // If we are spilling the memory op for an implicit null check, at the
149 // null check location (ie - null check is in HRP block) we need to do
150 // the null-check first, then spill-down in the following block.
151 // (The implicit_null_check function ensures the use is also dominated
152 // by the branch-not-taken block.)
153 Node *be = b->end();
154 if( be->is_MachNullCheck() && be->in(1) == def && def == b->_nodes[loc] ) {
155 // Spill goes in the branch-not-taken block
156 b = b->_succs[b->_nodes[b->end_idx()+1]->Opcode() == Op_IfTrue];
157 loc = 0; // Just past the Region
158 }
159 assert( loc >= 0, "must insert past block head" );
161 // Get a def-side SpillCopy
162 Node *spill = get_spillcopy_wide(def,NULL,0);
163 // Did we fail to split?, then bail
164 if (!spill) {
165 return 0;
166 }
168 // Insert the spill at chosen location
169 insert_proj( b, loc+1, spill, maxlrg++);
171 // Insert new node into Reaches array
172 Reachblock[slidx] = spill;
173 // Update debug list of reaching down definitions by adding this one
174 debug_defs[slidx] = spill;
176 // return updated count of live ranges
177 return maxlrg;
178 }
180 //------------------------------split_USE--------------------------------------
181 // Splits at uses can involve redeffing the LRG, so no CISC Spilling there.
182 // Debug uses want to know if def is already stack enabled.
183 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 ) {
184 #ifdef ASSERT
185 // Increment the counter for this lrg
186 splits.at_put(slidx, splits.at(slidx)+1);
187 #endif
189 // Some setup stuff for handling debug node uses
190 JVMState* jvms = use->jvms();
191 uint debug_start = jvms ? jvms->debug_start() : 999999;
192 uint debug_end = jvms ? jvms->debug_end() : 999999;
194 //-------------------------------------------
195 // Check for use of debug info
196 if (useidx >= debug_start && useidx < debug_end) {
197 // Actually it's perfectly legal for constant debug info to appear
198 // just unlikely. In this case the optimizer left a ConI of a 4
199 // as both inputs to a Phi with only a debug use. It's a single-def
200 // live range of a rematerializable value. The live range spills,
201 // rematerializes and now the ConI directly feeds into the debug info.
202 // assert(!def->is_Con(), "constant debug info already constructed directly");
204 // Special split handling for Debug Info
205 // If DEF is DOWN, just hook the edge and return
206 // If DEF is UP, Split it DOWN for this USE.
207 if( def->is_Mach() ) {
208 if( def_down ) {
209 // DEF is DOWN, so connect USE directly to the DEF
210 use->set_req(useidx, def);
211 } else {
212 // Block and index where the use occurs.
213 Block *b = _cfg._bbs[use->_idx];
214 // Put the clone just prior to use
215 int bindex = b->find_node(use);
216 // DEF is UP, so must copy it DOWN and hook in USE
217 // Insert SpillCopy before the USE, which uses DEF as its input,
218 // and defs a new live range, which is used by this node.
219 Node *spill = get_spillcopy_wide(def,use,useidx);
220 // did we fail to split?
221 if (!spill) {
222 // Bail
223 return 0;
224 }
225 // insert into basic block
226 insert_proj( b, bindex, spill, maxlrg++ );
227 // Use the new split
228 use->set_req(useidx,spill);
229 }
230 // No further split handling needed for this use
231 return maxlrg;
232 } // End special splitting for debug info live range
233 } // If debug info
235 // CISC-SPILLING
236 // Finally, check to see if USE is CISC-Spillable, and if so,
237 // gather_lrg_masks will add the flags bit to its mask, and
238 // no use side copy is needed. This frees up the live range
239 // register choices without causing copy coalescing, etc.
240 if( UseCISCSpill && cisc_sp ) {
241 int inp = use->cisc_operand();
242 if( inp != AdlcVMDeps::Not_cisc_spillable )
243 // Convert operand number to edge index number
244 inp = use->as_Mach()->operand_index(inp);
245 if( inp == (int)useidx ) {
246 use->set_req(useidx, def);
247 #ifndef PRODUCT
248 if( TraceCISCSpill ) {
249 tty->print(" set_split: ");
250 use->dump();
251 }
252 #endif
253 return maxlrg;
254 }
255 }
257 //-------------------------------------------
258 // Insert a Copy before the use
260 // Block and index where the use occurs.
261 int bindex;
262 // Phi input spill-copys belong at the end of the prior block
263 if( use->is_Phi() ) {
264 b = _cfg._bbs[b->pred(useidx)->_idx];
265 bindex = b->end_idx();
266 } else {
267 // Put the clone just prior to use
268 bindex = b->find_node(use);
269 }
271 Node *spill = get_spillcopy_wide( def, use, useidx );
272 if( !spill ) return 0; // Bailed out
273 // Insert SpillCopy before the USE, which uses the reaching DEF as
274 // its input, and defs a new live range, which is used by this node.
275 insert_proj( b, bindex, spill, maxlrg++ );
276 // Use the spill/clone
277 use->set_req(useidx,spill);
279 // return updated live range count
280 return maxlrg;
281 }
283 //------------------------------clone_node----------------------------
284 // Clone node with anti dependence check.
285 Node* clone_node(Node* def, Block *b, Compile* C) {
286 if (def->needs_anti_dependence_check()) {
287 #ifdef ASSERT
288 if (Verbose) {
289 tty->print_cr("RA attempts to clone node with anti_dependence:");
290 def->dump(-1); tty->cr();
291 tty->print_cr("into block:");
292 b->dump();
293 }
294 #endif
295 if (C->subsume_loads() == true && !C->failing()) {
296 // Retry with subsume_loads == false
297 // If this is the first failure, the sentinel string will "stick"
298 // to the Compile object, and the C2Compiler will see it and retry.
299 C->record_failure(C2Compiler::retry_no_subsuming_loads());
300 } else {
301 // Bailout without retry
302 C->record_method_not_compilable("RA Split failed: attempt to clone node with anti_dependence");
303 }
304 return 0;
305 }
306 return def->clone();
307 }
309 //------------------------------split_Rematerialize----------------------------
310 // Clone a local copy of the def.
311 Node *PhaseChaitin::split_Rematerialize( Node *def, Block *b, uint insidx, uint &maxlrg, GrowableArray<uint> splits, int slidx, uint *lrg2reach, Node **Reachblock, bool walkThru ) {
312 // The input live ranges will be stretched to the site of the new
313 // instruction. They might be stretched past a def and will thus
314 // have the old and new values of the same live range alive at the
315 // same time - a definite no-no. Split out private copies of
316 // the inputs.
317 if( def->req() > 1 ) {
318 for( uint i = 1; i < def->req(); i++ ) {
319 Node *in = def->in(i);
320 // Check for single-def (LRG cannot redefined)
321 uint lidx = n2lidx(in);
322 if( lidx >= _maxlrg ) continue; // Value is a recent spill-copy
323 if (lrgs(lidx).is_singledef()) continue;
325 Block *b_def = _cfg._bbs[def->_idx];
326 int idx_def = b_def->find_node(def);
327 Node *in_spill = get_spillcopy_wide( in, def, i );
328 if( !in_spill ) return 0; // Bailed out
329 insert_proj(b_def,idx_def,in_spill,maxlrg++);
330 if( b_def == b )
331 insidx++;
332 def->set_req(i,in_spill);
333 }
334 }
336 Node *spill = clone_node(def, b, C);
337 if (spill == NULL || C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) {
338 // Check when generating nodes
339 return 0;
340 }
342 // See if any inputs are currently being spilled, and take the
343 // latest copy of spilled inputs.
344 if( spill->req() > 1 ) {
345 for( uint i = 1; i < spill->req(); i++ ) {
346 Node *in = spill->in(i);
347 uint lidx = Find_id(in);
349 // Walk backwards thru spill copy node intermediates
350 if (walkThru) {
351 while ( in->is_SpillCopy() && lidx >= _maxlrg ) {
352 in = in->in(1);
353 lidx = Find_id(in);
354 }
356 if (lidx < _maxlrg && lrgs(lidx).is_multidef()) {
357 // walkThru found a multidef LRG, which is unsafe to use, so
358 // just keep the original def used in the clone.
359 in = spill->in(i);
360 lidx = Find_id(in);
361 }
362 }
364 if( lidx < _maxlrg && lrgs(lidx).reg() >= LRG::SPILL_REG ) {
365 Node *rdef = Reachblock[lrg2reach[lidx]];
366 if( rdef ) spill->set_req(i,rdef);
367 }
368 }
369 }
372 assert( spill->out_RegMask().is_UP(), "rematerialize to a reg" );
373 // Rematerialized op is def->spilled+1
374 set_was_spilled(spill);
375 if( _spilled_once.test(def->_idx) )
376 set_was_spilled(spill);
378 insert_proj( b, insidx, spill, maxlrg++ );
379 #ifdef ASSERT
380 // Increment the counter for this lrg
381 splits.at_put(slidx, splits.at(slidx)+1);
382 #endif
383 // See if the cloned def kills any flags, and copy those kills as well
384 uint i = insidx+1;
385 if( clone_projs( b, i, def, spill, maxlrg ) ) {
386 // Adjust the point where we go hi-pressure
387 if( i <= b->_ihrp_index ) b->_ihrp_index++;
388 if( i <= b->_fhrp_index ) b->_fhrp_index++;
389 }
391 return spill;
392 }
394 //------------------------------is_high_pressure-------------------------------
395 // Function to compute whether or not this live range is "high pressure"
396 // in this block - whether it spills eagerly or not.
397 bool PhaseChaitin::is_high_pressure( Block *b, LRG *lrg, uint insidx ) {
398 if( lrg->_was_spilled1 ) return true;
399 // Forced spilling due to conflict? Then split only at binding uses
400 // or defs, not for supposed capacity problems.
401 // CNC - Turned off 7/8/99, causes too much spilling
402 // if( lrg->_is_bound ) return false;
404 // Use float pressure numbers for vectors.
405 bool is_float_or_vector = lrg->_is_float || lrg->_is_vector;
406 // Not yet reached the high-pressure cutoff point, so low pressure
407 uint hrp_idx = is_float_or_vector ? b->_fhrp_index : b->_ihrp_index;
408 if( insidx < hrp_idx ) return false;
409 // Register pressure for the block as a whole depends on reg class
410 int block_pres = is_float_or_vector ? b->_freg_pressure : b->_reg_pressure;
411 // Bound live ranges will split at the binding points first;
412 // Intermediate splits should assume the live range's register set
413 // got "freed up" and that num_regs will become INT_PRESSURE.
414 int bound_pres = is_float_or_vector ? FLOATPRESSURE : INTPRESSURE;
415 // Effective register pressure limit.
416 int lrg_pres = (lrg->get_invalid_mask_size() > lrg->num_regs())
417 ? (lrg->get_invalid_mask_size() >> (lrg->num_regs()-1)) : bound_pres;
418 // High pressure if block pressure requires more register freedom
419 // than live range has.
420 return block_pres >= lrg_pres;
421 }
424 //------------------------------prompt_use---------------------------------
425 // True if lidx is used before any real register is def'd in the block
426 bool PhaseChaitin::prompt_use( Block *b, uint lidx ) {
427 if( lrgs(lidx)._was_spilled2 ) return false;
429 // Scan block for 1st use.
430 for( uint i = 1; i <= b->end_idx(); i++ ) {
431 Node *n = b->_nodes[i];
432 // Ignore PHI use, these can be up or down
433 if( n->is_Phi() ) continue;
434 for( uint j = 1; j < n->req(); j++ )
435 if( Find_id(n->in(j)) == lidx )
436 return true; // Found 1st use!
437 if( n->out_RegMask().is_NotEmpty() ) return false;
438 }
439 return false;
440 }
442 //------------------------------Split--------------------------------------
443 //----------Split Routine----------
444 // ***** NEW SPLITTING HEURISTIC *****
445 // DEFS: If the DEF is in a High Register Pressure(HRP) Block, split there.
446 // Else, no split unless there is a HRP block between a DEF and
447 // one of its uses, and then split at the HRP block.
448 //
449 // USES: If USE is in HRP, split at use to leave main LRG on stack.
450 // Else, hoist LRG back up to register only (ie - split is also DEF)
451 // We will compute a new maxlrg as we go
452 uint PhaseChaitin::Split( uint maxlrg ) {
453 NOT_PRODUCT( Compile::TracePhase t3("regAllocSplit", &_t_regAllocSplit, TimeCompiler); )
455 uint bidx, pidx, slidx, insidx, inpidx, twoidx;
456 uint non_phi = 1, spill_cnt = 0;
457 Node **Reachblock;
458 Node *n1, *n2, *n3;
459 Node_List *defs,*phis;
460 bool *UPblock;
461 bool u1, u2, u3;
462 Block *b, *pred;
463 PhiNode *phi;
464 GrowableArray<uint> lidxs;
466 // Array of counters to count splits per live range
467 GrowableArray<uint> splits;
469 //----------Setup Code----------
470 // Create a convenient mapping from lrg numbers to reaches/leaves indices
471 uint *lrg2reach = NEW_RESOURCE_ARRAY( uint, _maxlrg );
472 // Keep track of DEFS & Phis for later passes
473 defs = new Node_List();
474 phis = new Node_List();
475 // Gather info on which LRG's are spilling, and build maps
476 for( bidx = 1; bidx < _maxlrg; bidx++ ) {
477 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) {
478 assert(!lrgs(bidx).mask().is_AllStack(),"AllStack should color");
479 lrg2reach[bidx] = spill_cnt;
480 spill_cnt++;
481 lidxs.append(bidx);
482 #ifdef ASSERT
483 // Initialize the split counts to zero
484 splits.append(0);
485 #endif
486 #ifndef PRODUCT
487 if( PrintOpto && WizardMode && lrgs(bidx)._was_spilled1 )
488 tty->print_cr("Warning, 2nd spill of L%d",bidx);
489 #endif
490 }
491 }
493 // Create side arrays for propagating reaching defs info.
494 // Each block needs a node pointer for each spilling live range for the
495 // Def which is live into the block. Phi nodes handle multiple input
496 // Defs by querying the output of their predecessor blocks and resolving
497 // them to a single Def at the phi. The pointer is updated for each
498 // Def in the block, and then becomes the output for the block when
499 // processing of the block is complete. We also need to track whether
500 // a Def is UP or DOWN. UP means that it should get a register (ie -
501 // it is always in LRP regions), and DOWN means that it is probably
502 // on the stack (ie - it crosses HRP regions).
503 Node ***Reaches = NEW_RESOURCE_ARRAY( Node**, _cfg._num_blocks+1 );
504 bool **UP = NEW_RESOURCE_ARRAY( bool*, _cfg._num_blocks+1 );
505 Node **debug_defs = NEW_RESOURCE_ARRAY( Node*, spill_cnt );
506 VectorSet **UP_entry= NEW_RESOURCE_ARRAY( VectorSet*, spill_cnt );
508 // Initialize Reaches & UP
509 for( bidx = 0; bidx < _cfg._num_blocks+1; bidx++ ) {
510 Reaches[bidx] = NEW_RESOURCE_ARRAY( Node*, spill_cnt );
511 UP[bidx] = NEW_RESOURCE_ARRAY( bool, spill_cnt );
512 Node **Reachblock = Reaches[bidx];
513 bool *UPblock = UP[bidx];
514 for( slidx = 0; slidx < spill_cnt; slidx++ ) {
515 UPblock[slidx] = true; // Assume they start in registers
516 Reachblock[slidx] = NULL; // Assume that no def is present
517 }
518 }
520 // Initialize to array of empty vectorsets
521 for( slidx = 0; slidx < spill_cnt; slidx++ )
522 UP_entry[slidx] = new VectorSet(Thread::current()->resource_area());
524 //----------PASS 1----------
525 //----------Propagation & Node Insertion Code----------
526 // Walk the Blocks in RPO for DEF & USE info
527 for( bidx = 0; bidx < _cfg._num_blocks; bidx++ ) {
529 if (C->check_node_count(spill_cnt, out_of_nodes)) {
530 return 0;
531 }
533 b = _cfg._blocks[bidx];
534 // Reaches & UP arrays for this block
535 Reachblock = Reaches[b->_pre_order];
536 UPblock = UP[b->_pre_order];
537 // Reset counter of start of non-Phi nodes in block
538 non_phi = 1;
539 //----------Block Entry Handling----------
540 // Check for need to insert a new phi
541 // Cycle through this block's predecessors, collecting Reaches
542 // info for each spilled LRG. If they are identical, no phi is
543 // needed. If they differ, check for a phi, and insert if missing,
544 // or update edges if present. Set current block's Reaches set to
545 // be either the phi's or the reaching def, as appropriate.
546 // If no Phi is needed, check if the LRG needs to spill on entry
547 // to the block due to HRP.
548 for( slidx = 0; slidx < spill_cnt; slidx++ ) {
549 // Grab the live range number
550 uint lidx = lidxs.at(slidx);
551 // Do not bother splitting or putting in Phis for single-def
552 // rematerialized live ranges. This happens alot to constants
553 // with long live ranges.
554 if( lrgs(lidx).is_singledef() &&
555 lrgs(lidx)._def->rematerialize() ) {
556 // reset the Reaches & UP entries
557 Reachblock[slidx] = lrgs(lidx)._def;
558 UPblock[slidx] = true;
559 // Record following instruction in case 'n' rematerializes and
560 // kills flags
561 Block *pred1 = _cfg._bbs[b->pred(1)->_idx];
562 continue;
563 }
565 // Initialize needs_phi and needs_split
566 bool needs_phi = false;
567 bool needs_split = false;
568 bool has_phi = false;
569 // Walk the predecessor blocks to check inputs for that live range
570 // Grab predecessor block header
571 n1 = b->pred(1);
572 // Grab the appropriate reaching def info for inpidx
573 pred = _cfg._bbs[n1->_idx];
574 pidx = pred->_pre_order;
575 Node **Ltmp = Reaches[pidx];
576 bool *Utmp = UP[pidx];
577 n1 = Ltmp[slidx];
578 u1 = Utmp[slidx];
579 // Initialize node for saving type info
580 n3 = n1;
581 u3 = u1;
583 // Compare inputs to see if a Phi is needed
584 for( inpidx = 2; inpidx < b->num_preds(); inpidx++ ) {
585 // Grab predecessor block headers
586 n2 = b->pred(inpidx);
587 // Grab the appropriate reaching def info for inpidx
588 pred = _cfg._bbs[n2->_idx];
589 pidx = pred->_pre_order;
590 Ltmp = Reaches[pidx];
591 Utmp = UP[pidx];
592 n2 = Ltmp[slidx];
593 u2 = Utmp[slidx];
594 // For each LRG, decide if a phi is necessary
595 if( n1 != n2 ) {
596 needs_phi = true;
597 }
598 // See if the phi has mismatched inputs, UP vs. DOWN
599 if( n1 && n2 && (u1 != u2) ) {
600 needs_split = true;
601 }
602 // Move n2/u2 to n1/u1 for next iteration
603 n1 = n2;
604 u1 = u2;
605 // Preserve a non-NULL predecessor for later type referencing
606 if( (n3 == NULL) && (n2 != NULL) ){
607 n3 = n2;
608 u3 = u2;
609 }
610 } // End for all potential Phi inputs
612 // check block for appropriate phinode & update edges
613 for( insidx = 1; insidx <= b->end_idx(); insidx++ ) {
614 n1 = b->_nodes[insidx];
615 // bail if this is not a phi
616 phi = n1->is_Phi() ? n1->as_Phi() : NULL;
617 if( phi == NULL ) {
618 // Keep track of index of first non-PhiNode instruction in block
619 non_phi = insidx;
620 // break out of the for loop as we have handled all phi nodes
621 break;
622 }
623 // must be looking at a phi
624 if( Find_id(n1) == lidxs.at(slidx) ) {
625 // found the necessary phi
626 needs_phi = false;
627 has_phi = true;
628 // initialize the Reaches entry for this LRG
629 Reachblock[slidx] = phi;
630 break;
631 } // end if found correct phi
632 } // end for all phi's
634 // If a phi is needed or exist, check for it
635 if( needs_phi || has_phi ) {
636 // add new phinode if one not already found
637 if( needs_phi ) {
638 // create a new phi node and insert it into the block
639 // type is taken from left over pointer to a predecessor
640 assert(n3,"No non-NULL reaching DEF for a Phi");
641 phi = new (C, b->num_preds()) PhiNode(b->head(), n3->bottom_type());
642 // initialize the Reaches entry for this LRG
643 Reachblock[slidx] = phi;
645 // add node to block & node_to_block mapping
646 insert_proj( b, insidx++, phi, maxlrg++ );
647 non_phi++;
648 // Reset new phi's mapping to be the spilling live range
649 _names.map(phi->_idx, lidx);
650 assert(Find_id(phi) == lidx,"Bad update on Union-Find mapping");
651 } // end if not found correct phi
652 // Here you have either found or created the Phi, so record it
653 assert(phi != NULL,"Must have a Phi Node here");
654 phis->push(phi);
655 // PhiNodes should either force the LRG UP or DOWN depending
656 // on its inputs and the register pressure in the Phi's block.
657 UPblock[slidx] = true; // Assume new DEF is UP
658 // If entering a high-pressure area with no immediate use,
659 // assume Phi is DOWN
660 if( is_high_pressure( b, &lrgs(lidx), b->end_idx()) && !prompt_use(b,lidx) )
661 UPblock[slidx] = false;
662 // If we are not split up/down and all inputs are down, then we
663 // are down
664 if( !needs_split && !u3 )
665 UPblock[slidx] = false;
666 } // end if phi is needed
668 // Do not need a phi, so grab the reaching DEF
669 else {
670 // Grab predecessor block header
671 n1 = b->pred(1);
672 // Grab the appropriate reaching def info for k
673 pred = _cfg._bbs[n1->_idx];
674 pidx = pred->_pre_order;
675 Node **Ltmp = Reaches[pidx];
676 bool *Utmp = UP[pidx];
677 // reset the Reaches & UP entries
678 Reachblock[slidx] = Ltmp[slidx];
679 UPblock[slidx] = Utmp[slidx];
680 } // end else no Phi is needed
681 } // end for all spilling live ranges
682 // DEBUG
683 #ifndef PRODUCT
684 if(trace_spilling()) {
685 tty->print("/`\nBlock %d: ", b->_pre_order);
686 tty->print("Reaching Definitions after Phi handling\n");
687 for( uint x = 0; x < spill_cnt; x++ ) {
688 tty->print("Spill Idx %d: UP %d: Node\n",x,UPblock[x]);
689 if( Reachblock[x] )
690 Reachblock[x]->dump();
691 else
692 tty->print("Undefined\n");
693 }
694 }
695 #endif
697 //----------Non-Phi Node Splitting----------
698 // Since phi-nodes have now been handled, the Reachblock array for this
699 // block is initialized with the correct starting value for the defs which
700 // reach non-phi instructions in this block. Thus, process non-phi
701 // instructions normally, inserting SpillCopy nodes for all spill
702 // locations.
704 // Memoize any DOWN reaching definitions for use as DEBUG info
705 for( insidx = 0; insidx < spill_cnt; insidx++ ) {
706 debug_defs[insidx] = (UPblock[insidx]) ? NULL : Reachblock[insidx];
707 if( UPblock[insidx] ) // Memoize UP decision at block start
708 UP_entry[insidx]->set( b->_pre_order );
709 }
711 //----------Walk Instructions in the Block and Split----------
712 // For all non-phi instructions in the block
713 for( insidx = 1; insidx <= b->end_idx(); insidx++ ) {
714 Node *n = b->_nodes[insidx];
715 // Find the defining Node's live range index
716 uint defidx = Find_id(n);
717 uint cnt = n->req();
719 if( n->is_Phi() ) {
720 // Skip phi nodes after removing dead copies.
721 if( defidx < _maxlrg ) {
722 // Check for useless Phis. These appear if we spill, then
723 // coalesce away copies. Dont touch Phis in spilling live
724 // ranges; they are busy getting modifed in this pass.
725 if( lrgs(defidx).reg() < LRG::SPILL_REG ) {
726 uint i;
727 Node *u = NULL;
728 // Look for the Phi merging 2 unique inputs
729 for( i = 1; i < cnt; i++ ) {
730 // Ignore repeats and self
731 if( n->in(i) != u && n->in(i) != n ) {
732 // Found a unique input
733 if( u != NULL ) // If it's the 2nd, bail out
734 break;
735 u = n->in(i); // Else record it
736 }
737 }
738 assert( u, "at least 1 valid input expected" );
739 if( i >= cnt ) { // Found one unique input
740 assert(Find_id(n) == Find_id(u), "should be the same lrg");
741 n->replace_by(u); // Then replace with unique input
742 n->disconnect_inputs(NULL);
743 b->_nodes.remove(insidx);
744 insidx--;
745 b->_ihrp_index--;
746 b->_fhrp_index--;
747 }
748 }
749 }
750 continue;
751 }
752 assert( insidx > b->_ihrp_index ||
753 (b->_reg_pressure < (uint)INTPRESSURE) ||
754 b->_ihrp_index > 4000000 ||
755 b->_ihrp_index >= b->end_idx() ||
756 !b->_nodes[b->_ihrp_index]->is_Proj(), "" );
757 assert( insidx > b->_fhrp_index ||
758 (b->_freg_pressure < (uint)FLOATPRESSURE) ||
759 b->_fhrp_index > 4000000 ||
760 b->_fhrp_index >= b->end_idx() ||
761 !b->_nodes[b->_fhrp_index]->is_Proj(), "" );
763 // ********** Handle Crossing HRP Boundry **********
764 if( (insidx == b->_ihrp_index) || (insidx == b->_fhrp_index) ) {
765 for( slidx = 0; slidx < spill_cnt; slidx++ ) {
766 // Check for need to split at HRP boundary - split if UP
767 n1 = Reachblock[slidx];
768 // bail out if no reaching DEF
769 if( n1 == NULL ) continue;
770 // bail out if live range is 'isolated' around inner loop
771 uint lidx = lidxs.at(slidx);
772 // If live range is currently UP
773 if( UPblock[slidx] ) {
774 // set location to insert spills at
775 // SPLIT DOWN HERE - NO CISC SPILL
776 if( is_high_pressure( b, &lrgs(lidx), insidx ) &&
777 !n1->rematerialize() ) {
778 // If there is already a valid stack definition available, use it
779 if( debug_defs[slidx] != NULL ) {
780 Reachblock[slidx] = debug_defs[slidx];
781 }
782 else {
783 // Insert point is just past last use or def in the block
784 int insert_point = insidx-1;
785 while( insert_point > 0 ) {
786 Node *n = b->_nodes[insert_point];
787 // Hit top of block? Quit going backwards
788 if( n->is_Phi() ) break;
789 // Found a def? Better split after it.
790 if( n2lidx(n) == lidx ) break;
791 // Look for a use
792 uint i;
793 for( i = 1; i < n->req(); i++ )
794 if( n2lidx(n->in(i)) == lidx )
795 break;
796 // Found a use? Better split after it.
797 if( i < n->req() ) break;
798 insert_point--;
799 }
800 uint orig_eidx = b->end_idx();
801 maxlrg = split_DEF( n1, b, insert_point, maxlrg, Reachblock, debug_defs, splits, slidx);
802 // If it wasn't split bail
803 if (!maxlrg) {
804 return 0;
805 }
806 // Spill of NULL check mem op goes into the following block.
807 if (b->end_idx() > orig_eidx)
808 insidx++;
809 }
810 // This is a new DEF, so update UP
811 UPblock[slidx] = false;
812 #ifndef PRODUCT
813 // DEBUG
814 if( trace_spilling() ) {
815 tty->print("\nNew Split DOWN DEF of Spill Idx ");
816 tty->print("%d, UP %d:\n",slidx,false);
817 n1->dump();
818 }
819 #endif
820 }
821 } // end if LRG is UP
822 } // end for all spilling live ranges
823 assert( b->_nodes[insidx] == n, "got insidx set incorrectly" );
824 } // end if crossing HRP Boundry
826 // If the LRG index is oob, then this is a new spillcopy, skip it.
827 if( defidx >= _maxlrg ) {
828 continue;
829 }
830 LRG &deflrg = lrgs(defidx);
831 uint copyidx = n->is_Copy();
832 // Remove coalesced copy from CFG
833 if( copyidx && defidx == n2lidx(n->in(copyidx)) ) {
834 n->replace_by( n->in(copyidx) );
835 n->set_req( copyidx, NULL );
836 b->_nodes.remove(insidx--);
837 b->_ihrp_index--; // Adjust the point where we go hi-pressure
838 b->_fhrp_index--;
839 continue;
840 }
842 #define DERIVED 0
844 // ********** Handle USES **********
845 bool nullcheck = false;
846 // Implicit null checks never use the spilled value
847 if( n->is_MachNullCheck() )
848 nullcheck = true;
849 if( !nullcheck ) {
850 // Search all inputs for a Spill-USE
851 JVMState* jvms = n->jvms();
852 uint oopoff = jvms ? jvms->oopoff() : cnt;
853 uint old_last = cnt - 1;
854 for( inpidx = 1; inpidx < cnt; inpidx++ ) {
855 // Derived/base pairs may be added to our inputs during this loop.
856 // If inpidx > old_last, then one of these new inputs is being
857 // handled. Skip the derived part of the pair, but process
858 // the base like any other input.
859 if( inpidx > old_last && ((inpidx - oopoff) & 1) == DERIVED ) {
860 continue; // skip derived_debug added below
861 }
862 // Get lidx of input
863 uint useidx = Find_id(n->in(inpidx));
864 // Not a brand-new split, and it is a spill use
865 if( useidx < _maxlrg && lrgs(useidx).reg() >= LRG::SPILL_REG ) {
866 // Check for valid reaching DEF
867 slidx = lrg2reach[useidx];
868 Node *def = Reachblock[slidx];
869 assert( def != NULL, "Using Undefined Value in Split()\n");
871 // (+++) %%%% remove this in favor of pre-pass in matcher.cpp
872 // monitor references do not care where they live, so just hook
873 if ( jvms && jvms->is_monitor_use(inpidx) ) {
874 // The effect of this clone is to drop the node out of the block,
875 // so that the allocator does not see it anymore, and therefore
876 // does not attempt to assign it a register.
877 def = clone_node(def, b, C);
878 if (def == NULL || C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) {
879 return 0;
880 }
881 _names.extend(def->_idx,0);
882 _cfg._bbs.map(def->_idx,b);
883 n->set_req(inpidx, def);
884 continue;
885 }
887 // Rematerializable? Then clone def at use site instead
888 // of store/load
889 if( def->rematerialize() ) {
890 int old_size = b->_nodes.size();
891 def = split_Rematerialize( def, b, insidx, maxlrg, splits, slidx, lrg2reach, Reachblock, true );
892 if( !def ) return 0; // Bail out
893 insidx += b->_nodes.size()-old_size;
894 }
896 MachNode *mach = n->is_Mach() ? n->as_Mach() : NULL;
897 // Base pointers and oopmap references do not care where they live.
898 if ((inpidx >= oopoff) ||
899 (mach && mach->ideal_Opcode() == Op_AddP && inpidx == AddPNode::Base)) {
900 if (def->rematerialize() && lrgs(useidx)._was_spilled2) {
901 // This def has been rematerialized a couple of times without
902 // progress. It doesn't care if it lives UP or DOWN, so
903 // spill it down now.
904 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false,splits,slidx);
905 // If it wasn't split bail
906 if (!maxlrg) {
907 return 0;
908 }
909 insidx++; // Reset iterator to skip USE side split
910 } else {
911 // Just hook the def edge
912 n->set_req(inpidx, def);
913 }
915 if (inpidx >= oopoff) {
916 // After oopoff, we have derived/base pairs. We must mention all
917 // derived pointers here as derived/base pairs for GC. If the
918 // derived value is spilling and we have a copy both in Reachblock
919 // (called here 'def') and debug_defs[slidx] we need to mention
920 // both in derived/base pairs or kill one.
921 Node *derived_debug = debug_defs[slidx];
922 if( ((inpidx - oopoff) & 1) == DERIVED && // derived vs base?
923 mach && mach->ideal_Opcode() != Op_Halt &&
924 derived_debug != NULL &&
925 derived_debug != def ) { // Actual 2nd value appears
926 // We have already set 'def' as a derived value.
927 // Also set debug_defs[slidx] as a derived value.
928 uint k;
929 for( k = oopoff; k < cnt; k += 2 )
930 if( n->in(k) == derived_debug )
931 break; // Found an instance of debug derived
932 if( k == cnt ) {// No instance of debug_defs[slidx]
933 // Add a derived/base pair to cover the debug info.
934 // We have to process the added base later since it is not
935 // handled yet at this point but skip derived part.
936 assert(((n->req() - oopoff) & 1) == DERIVED,
937 "must match skip condition above");
938 n->add_req( derived_debug ); // this will be skipped above
939 n->add_req( n->in(inpidx+1) ); // this will be processed
940 // Increment cnt to handle added input edges on
941 // subsequent iterations.
942 cnt += 2;
943 }
944 }
945 }
946 continue;
947 }
948 // Special logic for DEBUG info
949 if( jvms && b->_freq > BLOCK_FREQUENCY(0.5) ) {
950 uint debug_start = jvms->debug_start();
951 // If this is debug info use & there is a reaching DOWN def
952 if ((debug_start <= inpidx) && (debug_defs[slidx] != NULL)) {
953 assert(inpidx < oopoff, "handle only debug info here");
954 // Just hook it in & move on
955 n->set_req(inpidx, debug_defs[slidx]);
956 // (Note that this can make two sides of a split live at the
957 // same time: The debug def on stack, and another def in a
958 // register. The GC needs to know about both of them, but any
959 // derived pointers after oopoff will refer to only one of the
960 // two defs and the GC would therefore miss the other. Thus
961 // this hack is only allowed for debug info which is Java state
962 // and therefore never a derived pointer.)
963 continue;
964 }
965 }
966 // Grab register mask info
967 const RegMask &dmask = def->out_RegMask();
968 const RegMask &umask = n->in_RegMask(inpidx);
969 bool is_vect = RegMask::is_vector(def->ideal_reg());
970 assert(inpidx < oopoff, "cannot use-split oop map info");
972 bool dup = UPblock[slidx];
973 bool uup = umask.is_UP();
975 // Need special logic to handle bound USES. Insert a split at this
976 // bound use if we can't rematerialize the def, or if we need the
977 // split to form a misaligned pair.
978 if( !umask.is_AllStack() &&
979 (int)umask.Size() <= lrgs(useidx).num_regs() &&
980 (!def->rematerialize() ||
981 !is_vect && umask.is_misaligned_pair())) {
982 // These need a Split regardless of overlap or pressure
983 // SPLIT - NO DEF - NO CISC SPILL
984 maxlrg = split_USE(def,b,n,inpidx,maxlrg,dup,false, splits,slidx);
985 // If it wasn't split bail
986 if (!maxlrg) {
987 return 0;
988 }
989 insidx++; // Reset iterator to skip USE side split
990 continue;
991 }
993 if (UseFPUForSpilling && n->is_MachCall() && !uup && !dup ) {
994 // The use at the call can force the def down so insert
995 // a split before the use to allow the def more freedom.
996 maxlrg = split_USE(def,b,n,inpidx,maxlrg,dup,false, splits,slidx);
997 // If it wasn't split bail
998 if (!maxlrg) {
999 return 0;
1000 }
1001 insidx++; // Reset iterator to skip USE side split
1002 continue;
1003 }
1005 // Here is the logic chart which describes USE Splitting:
1006 // 0 = false or DOWN, 1 = true or UP
1007 //
1008 // Overlap | DEF | USE | Action
1009 //-------------------------------------------------------
1010 // 0 | 0 | 0 | Copy - mem -> mem
1011 // 0 | 0 | 1 | Split-UP - Check HRP
1012 // 0 | 1 | 0 | Split-DOWN - Debug Info?
1013 // 0 | 1 | 1 | Copy - reg -> reg
1014 // 1 | 0 | 0 | Reset Input Edge (no Split)
1015 // 1 | 0 | 1 | Split-UP - Check HRP
1016 // 1 | 1 | 0 | Split-DOWN - Debug Info?
1017 // 1 | 1 | 1 | Reset Input Edge (no Split)
1018 //
1019 // So, if (dup == uup), then overlap test determines action,
1020 // with true being no split, and false being copy. Else,
1021 // if DEF is DOWN, Split-UP, and check HRP to decide on
1022 // resetting DEF. Finally if DEF is UP, Split-DOWN, with
1023 // special handling for Debug Info.
1024 if( dup == uup ) {
1025 if( dmask.overlap(umask) ) {
1026 // Both are either up or down, and there is overlap, No Split
1027 n->set_req(inpidx, def);
1028 }
1029 else { // Both are either up or down, and there is no overlap
1030 if( dup ) { // If UP, reg->reg copy
1031 // COPY ACROSS HERE - NO DEF - NO CISC SPILL
1032 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false, splits,slidx);
1033 // If it wasn't split bail
1034 if (!maxlrg) {
1035 return 0;
1036 }
1037 insidx++; // Reset iterator to skip USE side split
1038 }
1039 else { // DOWN, mem->mem copy
1040 // COPY UP & DOWN HERE - NO DEF - NO CISC SPILL
1041 // First Split-UP to move value into Register
1042 uint def_ideal = def->ideal_reg();
1043 const RegMask* tmp_rm = Matcher::idealreg2regmask[def_ideal];
1044 Node *spill = new (C) MachSpillCopyNode(def, dmask, *tmp_rm);
1045 insert_proj( b, insidx, spill, maxlrg );
1046 // Then Split-DOWN as if previous Split was DEF
1047 maxlrg = split_USE(spill,b,n,inpidx,maxlrg,false,false, splits,slidx);
1048 // If it wasn't split bail
1049 if (!maxlrg) {
1050 return 0;
1051 }
1052 insidx += 2; // Reset iterator to skip USE side splits
1053 }
1054 } // End else no overlap
1055 } // End if dup == uup
1056 // dup != uup, so check dup for direction of Split
1057 else {
1058 if( dup ) { // If UP, Split-DOWN and check Debug Info
1059 // If this node is already a SpillCopy, just patch the edge
1060 // except the case of spilling to stack.
1061 if( n->is_SpillCopy() ) {
1062 RegMask tmp_rm(umask);
1063 tmp_rm.SUBTRACT(Matcher::STACK_ONLY_mask);
1064 if( dmask.overlap(tmp_rm) ) {
1065 if( def != n->in(inpidx) ) {
1066 n->set_req(inpidx, def);
1067 }
1068 continue;
1069 }
1070 }
1071 // COPY DOWN HERE - NO DEF - NO CISC SPILL
1072 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false, splits,slidx);
1073 // If it wasn't split bail
1074 if (!maxlrg) {
1075 return 0;
1076 }
1077 insidx++; // Reset iterator to skip USE side split
1078 // Check for debug-info split. Capture it for later
1079 // debug splits of the same value
1080 if (jvms && jvms->debug_start() <= inpidx && inpidx < oopoff)
1081 debug_defs[slidx] = n->in(inpidx);
1083 }
1084 else { // DOWN, Split-UP and check register pressure
1085 if( is_high_pressure( b, &lrgs(useidx), insidx ) ) {
1086 // COPY UP HERE - NO DEF - CISC SPILL
1087 maxlrg = split_USE(def,b,n,inpidx,maxlrg,true,true, splits,slidx);
1088 // If it wasn't split bail
1089 if (!maxlrg) {
1090 return 0;
1091 }
1092 insidx++; // Reset iterator to skip USE side split
1093 } else { // LRP
1094 // COPY UP HERE - WITH DEF - NO CISC SPILL
1095 maxlrg = split_USE(def,b,n,inpidx,maxlrg,true,false, splits,slidx);
1096 // If it wasn't split bail
1097 if (!maxlrg) {
1098 return 0;
1099 }
1100 // Flag this lift-up in a low-pressure block as
1101 // already-spilled, so if it spills again it will
1102 // spill hard (instead of not spilling hard and
1103 // coalescing away).
1104 set_was_spilled(n->in(inpidx));
1105 // Since this is a new DEF, update Reachblock & UP
1106 Reachblock[slidx] = n->in(inpidx);
1107 UPblock[slidx] = true;
1108 insidx++; // Reset iterator to skip USE side split
1109 }
1110 } // End else DOWN
1111 } // End dup != uup
1112 } // End if Spill USE
1113 } // End For All Inputs
1114 } // End If not nullcheck
1116 // ********** Handle DEFS **********
1117 // DEFS either Split DOWN in HRP regions or when the LRG is bound, or
1118 // just reset the Reaches info in LRP regions. DEFS must always update
1119 // UP info.
1120 if( deflrg.reg() >= LRG::SPILL_REG ) { // Spilled?
1121 uint slidx = lrg2reach[defidx];
1122 // Add to defs list for later assignment of new live range number
1123 defs->push(n);
1124 // Set a flag on the Node indicating it has already spilled.
1125 // Only do it for capacity spills not conflict spills.
1126 if( !deflrg._direct_conflict )
1127 set_was_spilled(n);
1128 assert(!n->is_Phi(),"Cannot insert Phi into DEFS list");
1129 // Grab UP info for DEF
1130 const RegMask &dmask = n->out_RegMask();
1131 bool defup = dmask.is_UP();
1132 int ireg = n->ideal_reg();
1133 bool is_vect = RegMask::is_vector(ireg);
1134 // Only split at Def if this is a HRP block or bound (and spilled once)
1135 if( !n->rematerialize() &&
1136 (((dmask.is_bound(ireg) || !is_vect && dmask.is_misaligned_pair()) &&
1137 (deflrg._direct_conflict || deflrg._must_spill)) ||
1138 // Check for LRG being up in a register and we are inside a high
1139 // pressure area. Spill it down immediately.
1140 (defup && is_high_pressure(b,&deflrg,insidx))) ) {
1141 assert( !n->rematerialize(), "" );
1142 assert( !n->is_SpillCopy(), "" );
1143 // Do a split at the def site.
1144 maxlrg = split_DEF( n, b, insidx, maxlrg, Reachblock, debug_defs, splits, slidx );
1145 // If it wasn't split bail
1146 if (!maxlrg) {
1147 return 0;
1148 }
1149 // Split DEF's Down
1150 UPblock[slidx] = 0;
1151 #ifndef PRODUCT
1152 // DEBUG
1153 if( trace_spilling() ) {
1154 tty->print("\nNew Split DOWN DEF of Spill Idx ");
1155 tty->print("%d, UP %d:\n",slidx,false);
1156 n->dump();
1157 }
1158 #endif
1159 }
1160 else { // Neither bound nor HRP, must be LRP
1161 // otherwise, just record the def
1162 Reachblock[slidx] = n;
1163 // UP should come from the outRegmask() of the DEF
1164 UPblock[slidx] = defup;
1165 // Update debug list of reaching down definitions, kill if DEF is UP
1166 debug_defs[slidx] = defup ? NULL : n;
1167 #ifndef PRODUCT
1168 // DEBUG
1169 if( trace_spilling() ) {
1170 tty->print("\nNew DEF of Spill Idx ");
1171 tty->print("%d, UP %d:\n",slidx,defup);
1172 n->dump();
1173 }
1174 #endif
1175 } // End else LRP
1176 } // End if spill def
1178 // ********** Split Left Over Mem-Mem Moves **********
1179 // Check for mem-mem copies and split them now. Do not do this
1180 // to copies about to be spilled; they will be Split shortly.
1181 if( copyidx ) {
1182 Node *use = n->in(copyidx);
1183 uint useidx = Find_id(use);
1184 if( useidx < _maxlrg && // This is not a new split
1185 OptoReg::is_stack(deflrg.reg()) &&
1186 deflrg.reg() < LRG::SPILL_REG ) { // And DEF is from stack
1187 LRG &uselrg = lrgs(useidx);
1188 if( OptoReg::is_stack(uselrg.reg()) &&
1189 uselrg.reg() < LRG::SPILL_REG && // USE is from stack
1190 deflrg.reg() != uselrg.reg() ) { // Not trivially removed
1191 uint def_ideal_reg = Matcher::base2reg[n->bottom_type()->base()];
1192 const RegMask &def_rm = *Matcher::idealreg2regmask[def_ideal_reg];
1193 const RegMask &use_rm = n->in_RegMask(copyidx);
1194 if( def_rm.overlap(use_rm) && n->is_SpillCopy() ) { // Bug 4707800, 'n' may be a storeSSL
1195 if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { // Check when generating nodes
1196 return 0;
1197 }
1198 Node *spill = new (C) MachSpillCopyNode(use,use_rm,def_rm);
1199 n->set_req(copyidx,spill);
1200 n->as_MachSpillCopy()->set_in_RegMask(def_rm);
1201 // Put the spill just before the copy
1202 insert_proj( b, insidx++, spill, maxlrg++ );
1203 }
1204 }
1205 }
1206 }
1207 } // End For All Instructions in Block - Non-PHI Pass
1209 // Check if each LRG is live out of this block so as not to propagate
1210 // beyond the last use of a LRG.
1211 for( slidx = 0; slidx < spill_cnt; slidx++ ) {
1212 uint defidx = lidxs.at(slidx);
1213 IndexSet *liveout = _live->live(b);
1214 if( !liveout->member(defidx) ) {
1215 #ifdef ASSERT
1216 // The index defidx is not live. Check the liveout array to ensure that
1217 // it contains no members which compress to defidx. Finding such an
1218 // instance may be a case to add liveout adjustment in compress_uf_map().
1219 // See 5063219.
1220 uint member;
1221 IndexSetIterator isi(liveout);
1222 while ((member = isi.next()) != 0) {
1223 assert(defidx != Find_const(member), "Live out member has not been compressed");
1224 }
1225 #endif
1226 Reachblock[slidx] = NULL;
1227 } else {
1228 assert(Reachblock[slidx] != NULL,"No reaching definition for liveout value");
1229 }
1230 }
1231 #ifndef PRODUCT
1232 if( trace_spilling() )
1233 b->dump();
1234 #endif
1235 } // End For All Blocks
1237 //----------PASS 2----------
1238 // Reset all DEF live range numbers here
1239 for( insidx = 0; insidx < defs->size(); insidx++ ) {
1240 // Grab the def
1241 n1 = defs->at(insidx);
1242 // Set new lidx for DEF
1243 new_lrg(n1, maxlrg++);
1244 }
1245 //----------Phi Node Splitting----------
1246 // Clean up a phi here, and assign a new live range number
1247 // Cycle through this block's predecessors, collecting Reaches
1248 // info for each spilled LRG and update edges.
1249 // Walk the phis list to patch inputs, split phis, and name phis
1250 uint lrgs_before_phi_split = maxlrg;
1251 for( insidx = 0; insidx < phis->size(); insidx++ ) {
1252 Node *phi = phis->at(insidx);
1253 assert(phi->is_Phi(),"This list must only contain Phi Nodes");
1254 Block *b = _cfg._bbs[phi->_idx];
1255 // Grab the live range number
1256 uint lidx = Find_id(phi);
1257 uint slidx = lrg2reach[lidx];
1258 // Update node to lidx map
1259 new_lrg(phi, maxlrg++);
1260 // Get PASS1's up/down decision for the block.
1261 int phi_up = !!UP_entry[slidx]->test(b->_pre_order);
1263 // Force down if double-spilling live range
1264 if( lrgs(lidx)._was_spilled1 )
1265 phi_up = false;
1267 // When splitting a Phi we an split it normal or "inverted".
1268 // An inverted split makes the splits target the Phi's UP/DOWN
1269 // sense inverted; then the Phi is followed by a final def-side
1270 // split to invert back. It changes which blocks the spill code
1271 // goes in.
1273 // Walk the predecessor blocks and assign the reaching def to the Phi.
1274 // Split Phi nodes by placing USE side splits wherever the reaching
1275 // DEF has the wrong UP/DOWN value.
1276 for( uint i = 1; i < b->num_preds(); i++ ) {
1277 // Get predecessor block pre-order number
1278 Block *pred = _cfg._bbs[b->pred(i)->_idx];
1279 pidx = pred->_pre_order;
1280 // Grab reaching def
1281 Node *def = Reaches[pidx][slidx];
1282 assert( def, "must have reaching def" );
1283 // If input up/down sense and reg-pressure DISagree
1284 if( def->rematerialize() ) {
1285 // Place the rematerialized node above any MSCs created during
1286 // phi node splitting. end_idx points at the insertion point
1287 // so look at the node before it.
1288 int insert = pred->end_idx();
1289 while (insert >= 1 &&
1290 pred->_nodes[insert - 1]->is_SpillCopy() &&
1291 Find(pred->_nodes[insert - 1]) >= lrgs_before_phi_split) {
1292 insert--;
1293 }
1294 def = split_Rematerialize( def, pred, insert, maxlrg, splits, slidx, lrg2reach, Reachblock, false );
1295 if( !def ) return 0; // Bail out
1296 }
1297 // Update the Phi's input edge array
1298 phi->set_req(i,def);
1299 // Grab the UP/DOWN sense for the input
1300 u1 = UP[pidx][slidx];
1301 if( u1 != (phi_up != 0)) {
1302 maxlrg = split_USE(def, b, phi, i, maxlrg, !u1, false, splits,slidx);
1303 // If it wasn't split bail
1304 if (!maxlrg) {
1305 return 0;
1306 }
1307 }
1308 } // End for all inputs to the Phi
1309 } // End for all Phi Nodes
1310 // Update _maxlrg to save Union asserts
1311 _maxlrg = maxlrg;
1314 //----------PASS 3----------
1315 // Pass over all Phi's to union the live ranges
1316 for( insidx = 0; insidx < phis->size(); insidx++ ) {
1317 Node *phi = phis->at(insidx);
1318 assert(phi->is_Phi(),"This list must only contain Phi Nodes");
1319 // Walk all inputs to Phi and Union input live range with Phi live range
1320 for( uint i = 1; i < phi->req(); i++ ) {
1321 // Grab the input node
1322 Node *n = phi->in(i);
1323 assert( n, "" );
1324 uint lidx = Find(n);
1325 uint pidx = Find(phi);
1326 if( lidx < pidx )
1327 Union(n, phi);
1328 else if( lidx > pidx )
1329 Union(phi, n);
1330 } // End for all inputs to the Phi Node
1331 } // End for all Phi Nodes
1332 // Now union all two address instructions
1333 for( insidx = 0; insidx < defs->size(); insidx++ ) {
1334 // Grab the def
1335 n1 = defs->at(insidx);
1336 // Set new lidx for DEF & handle 2-addr instructions
1337 if( n1->is_Mach() && ((twoidx = n1->as_Mach()->two_adr()) != 0) ) {
1338 assert( Find(n1->in(twoidx)) < maxlrg,"Assigning bad live range index");
1339 // Union the input and output live ranges
1340 uint lr1 = Find(n1);
1341 uint lr2 = Find(n1->in(twoidx));
1342 if( lr1 < lr2 )
1343 Union(n1, n1->in(twoidx));
1344 else if( lr1 > lr2 )
1345 Union(n1->in(twoidx), n1);
1346 } // End if two address
1347 } // End for all defs
1348 // DEBUG
1349 #ifdef ASSERT
1350 // Validate all live range index assignments
1351 for( bidx = 0; bidx < _cfg._num_blocks; bidx++ ) {
1352 b = _cfg._blocks[bidx];
1353 for( insidx = 0; insidx <= b->end_idx(); insidx++ ) {
1354 Node *n = b->_nodes[insidx];
1355 uint defidx = Find(n);
1356 assert(defidx < _maxlrg,"Bad live range index in Split");
1357 assert(defidx < maxlrg,"Bad live range index in Split");
1358 }
1359 }
1360 // Issue a warning if splitting made no progress
1361 int noprogress = 0;
1362 for( slidx = 0; slidx < spill_cnt; slidx++ ) {
1363 if( PrintOpto && WizardMode && splits.at(slidx) == 0 ) {
1364 tty->print_cr("Failed to split live range %d", lidxs.at(slidx));
1365 //BREAKPOINT;
1366 }
1367 else {
1368 noprogress++;
1369 }
1370 }
1371 if(!noprogress) {
1372 tty->print_cr("Failed to make progress in Split");
1373 //BREAKPOINT;
1374 }
1375 #endif
1376 // Return updated count of live ranges
1377 return maxlrg;
1378 }