Tue, 02 Sep 2014 12:48:45 -0700
8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
Summary: Add new C2 intrinsic for BigInteger::multiplyToLen() on x86 in 64-bit VM.
Reviewed-by: roland
1 /*
2 * Copyright (c) 2000, 2013, 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->number_of_nodes() &&
107 (b->get_node(i)->is_Proj() ||
108 b->get_node(i)->is_Phi() ) )
109 i++;
111 // Do not insert between a call and his Catch
112 if( b->get_node(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->get_node(++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->insert_node(spill, i); // Insert node in block
126 _cfg.map_node_to_block(spill, 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->get_node(loc)) {
155 // Spill goes in the branch-not-taken block
156 b = b->_succs[b->get_node(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.get_block_for_node(use);
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.get_block_for_node(b->pred(useidx));
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 uint lidx = _lrg_map.live_range_id(in);
321 // We do not need this for live ranges that are only defined once.
322 // However, this is not true for spill copies that are added in this
323 // Split() pass, since they might get coalesced later on in this pass.
324 if (lidx < _lrg_map.max_lrg_id() && lrgs(lidx).is_singledef()) {
325 continue;
326 }
328 Block *b_def = _cfg.get_block_for_node(def);
329 int idx_def = b_def->find_node(def);
330 Node *in_spill = get_spillcopy_wide( in, def, i );
331 if( !in_spill ) return 0; // Bailed out
332 insert_proj(b_def,idx_def,in_spill,maxlrg++);
333 if( b_def == b )
334 insidx++;
335 def->set_req(i,in_spill);
336 }
337 }
339 Node *spill = clone_node(def, b, C);
340 if (spill == NULL || C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) {
341 // Check when generating nodes
342 return 0;
343 }
345 // See if any inputs are currently being spilled, and take the
346 // latest copy of spilled inputs.
347 if( spill->req() > 1 ) {
348 for( uint i = 1; i < spill->req(); i++ ) {
349 Node *in = spill->in(i);
350 uint lidx = _lrg_map.find_id(in);
352 // Walk backwards thru spill copy node intermediates
353 if (walkThru) {
354 while (in->is_SpillCopy() && lidx >= _lrg_map.max_lrg_id()) {
355 in = in->in(1);
356 lidx = _lrg_map.find_id(in);
357 }
359 if (lidx < _lrg_map.max_lrg_id() && lrgs(lidx).is_multidef()) {
360 // walkThru found a multidef LRG, which is unsafe to use, so
361 // just keep the original def used in the clone.
362 in = spill->in(i);
363 lidx = _lrg_map.find_id(in);
364 }
365 }
367 if (lidx < _lrg_map.max_lrg_id() && lrgs(lidx).reg() >= LRG::SPILL_REG) {
368 Node *rdef = Reachblock[lrg2reach[lidx]];
369 if (rdef) {
370 spill->set_req(i, rdef);
371 }
372 }
373 }
374 }
377 assert( spill->out_RegMask().is_UP(), "rematerialize to a reg" );
378 // Rematerialized op is def->spilled+1
379 set_was_spilled(spill);
380 if( _spilled_once.test(def->_idx) )
381 set_was_spilled(spill);
383 insert_proj( b, insidx, spill, maxlrg++ );
384 #ifdef ASSERT
385 // Increment the counter for this lrg
386 splits.at_put(slidx, splits.at(slidx)+1);
387 #endif
388 // See if the cloned def kills any flags, and copy those kills as well
389 uint i = insidx+1;
390 int found_projs = clone_projs( b, i, def, spill, maxlrg);
391 if (found_projs > 0) {
392 // Adjust the point where we go hi-pressure
393 if (i <= b->_ihrp_index) {
394 b->_ihrp_index += found_projs;
395 }
396 if (i <= b->_fhrp_index) {
397 b->_fhrp_index += found_projs;
398 }
399 }
401 return spill;
402 }
404 //------------------------------is_high_pressure-------------------------------
405 // Function to compute whether or not this live range is "high pressure"
406 // in this block - whether it spills eagerly or not.
407 bool PhaseChaitin::is_high_pressure( Block *b, LRG *lrg, uint insidx ) {
408 if( lrg->_was_spilled1 ) return true;
409 // Forced spilling due to conflict? Then split only at binding uses
410 // or defs, not for supposed capacity problems.
411 // CNC - Turned off 7/8/99, causes too much spilling
412 // if( lrg->_is_bound ) return false;
414 // Use float pressure numbers for vectors.
415 bool is_float_or_vector = lrg->_is_float || lrg->_is_vector;
416 // Not yet reached the high-pressure cutoff point, so low pressure
417 uint hrp_idx = is_float_or_vector ? b->_fhrp_index : b->_ihrp_index;
418 if( insidx < hrp_idx ) return false;
419 // Register pressure for the block as a whole depends on reg class
420 int block_pres = is_float_or_vector ? b->_freg_pressure : b->_reg_pressure;
421 // Bound live ranges will split at the binding points first;
422 // Intermediate splits should assume the live range's register set
423 // got "freed up" and that num_regs will become INT_PRESSURE.
424 int bound_pres = is_float_or_vector ? FLOATPRESSURE : INTPRESSURE;
425 // Effective register pressure limit.
426 int lrg_pres = (lrg->get_invalid_mask_size() > lrg->num_regs())
427 ? (lrg->get_invalid_mask_size() >> (lrg->num_regs()-1)) : bound_pres;
428 // High pressure if block pressure requires more register freedom
429 // than live range has.
430 return block_pres >= lrg_pres;
431 }
434 //------------------------------prompt_use---------------------------------
435 // True if lidx is used before any real register is def'd in the block
436 bool PhaseChaitin::prompt_use( Block *b, uint lidx ) {
437 if (lrgs(lidx)._was_spilled2) {
438 return false;
439 }
441 // Scan block for 1st use.
442 for( uint i = 1; i <= b->end_idx(); i++ ) {
443 Node *n = b->get_node(i);
444 // Ignore PHI use, these can be up or down
445 if (n->is_Phi()) {
446 continue;
447 }
448 for (uint j = 1; j < n->req(); j++) {
449 if (_lrg_map.find_id(n->in(j)) == lidx) {
450 return true; // Found 1st use!
451 }
452 }
453 if (n->out_RegMask().is_NotEmpty()) {
454 return false;
455 }
456 }
457 return false;
458 }
460 //------------------------------Split--------------------------------------
461 //----------Split Routine----------
462 // ***** NEW SPLITTING HEURISTIC *****
463 // DEFS: If the DEF is in a High Register Pressure(HRP) Block, split there.
464 // Else, no split unless there is a HRP block between a DEF and
465 // one of its uses, and then split at the HRP block.
466 //
467 // USES: If USE is in HRP, split at use to leave main LRG on stack.
468 // Else, hoist LRG back up to register only (ie - split is also DEF)
469 // We will compute a new maxlrg as we go
470 uint PhaseChaitin::Split(uint maxlrg, ResourceArea* split_arena) {
471 NOT_PRODUCT( Compile::TracePhase t3("regAllocSplit", &_t_regAllocSplit, TimeCompiler); )
473 // Free thread local resources used by this method on exit.
474 ResourceMark rm(split_arena);
476 uint bidx, pidx, slidx, insidx, inpidx, twoidx;
477 uint non_phi = 1, spill_cnt = 0;
478 Node *n1, *n2, *n3;
479 Node_List *defs,*phis;
480 bool *UPblock;
481 bool u1, u2, u3;
482 Block *b, *pred;
483 PhiNode *phi;
484 GrowableArray<uint> lidxs(split_arena, maxlrg, 0, 0);
486 // Array of counters to count splits per live range
487 GrowableArray<uint> splits(split_arena, maxlrg, 0, 0);
489 #define NEW_SPLIT_ARRAY(type, size)\
490 (type*) split_arena->allocate_bytes((size) * sizeof(type))
492 //----------Setup Code----------
493 // Create a convenient mapping from lrg numbers to reaches/leaves indices
494 uint *lrg2reach = NEW_SPLIT_ARRAY(uint, maxlrg);
495 // Keep track of DEFS & Phis for later passes
496 defs = new Node_List();
497 phis = new Node_List();
498 // Gather info on which LRG's are spilling, and build maps
499 for (bidx = 1; bidx < maxlrg; bidx++) {
500 if (lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG) {
501 assert(!lrgs(bidx).mask().is_AllStack(),"AllStack should color");
502 lrg2reach[bidx] = spill_cnt;
503 spill_cnt++;
504 lidxs.append(bidx);
505 #ifdef ASSERT
506 // Initialize the split counts to zero
507 splits.append(0);
508 #endif
509 #ifndef PRODUCT
510 if( PrintOpto && WizardMode && lrgs(bidx)._was_spilled1 )
511 tty->print_cr("Warning, 2nd spill of L%d",bidx);
512 #endif
513 }
514 }
516 // Create side arrays for propagating reaching defs info.
517 // Each block needs a node pointer for each spilling live range for the
518 // Def which is live into the block. Phi nodes handle multiple input
519 // Defs by querying the output of their predecessor blocks and resolving
520 // them to a single Def at the phi. The pointer is updated for each
521 // Def in the block, and then becomes the output for the block when
522 // processing of the block is complete. We also need to track whether
523 // a Def is UP or DOWN. UP means that it should get a register (ie -
524 // it is always in LRP regions), and DOWN means that it is probably
525 // on the stack (ie - it crosses HRP regions).
526 Node ***Reaches = NEW_SPLIT_ARRAY( Node**, _cfg.number_of_blocks() + 1);
527 bool **UP = NEW_SPLIT_ARRAY( bool*, _cfg.number_of_blocks() + 1);
528 Node **debug_defs = NEW_SPLIT_ARRAY( Node*, spill_cnt );
529 VectorSet **UP_entry= NEW_SPLIT_ARRAY( VectorSet*, spill_cnt );
531 // Initialize Reaches & UP
532 for (bidx = 0; bidx < _cfg.number_of_blocks() + 1; bidx++) {
533 Reaches[bidx] = NEW_SPLIT_ARRAY( Node*, spill_cnt );
534 UP[bidx] = NEW_SPLIT_ARRAY( bool, spill_cnt );
535 Node **Reachblock = Reaches[bidx];
536 bool *UPblock = UP[bidx];
537 for( slidx = 0; slidx < spill_cnt; slidx++ ) {
538 UPblock[slidx] = true; // Assume they start in registers
539 Reachblock[slidx] = NULL; // Assume that no def is present
540 }
541 }
543 #undef NEW_SPLIT_ARRAY
545 // Initialize to array of empty vectorsets
546 for( slidx = 0; slidx < spill_cnt; slidx++ )
547 UP_entry[slidx] = new VectorSet(split_arena);
549 //----------PASS 1----------
550 //----------Propagation & Node Insertion Code----------
551 // Walk the Blocks in RPO for DEF & USE info
552 for( bidx = 0; bidx < _cfg.number_of_blocks(); bidx++ ) {
554 if (C->check_node_count(spill_cnt, out_of_nodes)) {
555 return 0;
556 }
558 b = _cfg.get_block(bidx);
559 // Reaches & UP arrays for this block
560 Node** Reachblock = Reaches[b->_pre_order];
561 UPblock = UP[b->_pre_order];
562 // Reset counter of start of non-Phi nodes in block
563 non_phi = 1;
564 //----------Block Entry Handling----------
565 // Check for need to insert a new phi
566 // Cycle through this block's predecessors, collecting Reaches
567 // info for each spilled LRG. If they are identical, no phi is
568 // needed. If they differ, check for a phi, and insert if missing,
569 // or update edges if present. Set current block's Reaches set to
570 // be either the phi's or the reaching def, as appropriate.
571 // If no Phi is needed, check if the LRG needs to spill on entry
572 // to the block due to HRP.
573 for( slidx = 0; slidx < spill_cnt; slidx++ ) {
574 // Grab the live range number
575 uint lidx = lidxs.at(slidx);
576 // Do not bother splitting or putting in Phis for single-def
577 // rematerialized live ranges. This happens alot to constants
578 // with long live ranges.
579 if( lrgs(lidx).is_singledef() &&
580 lrgs(lidx)._def->rematerialize() ) {
581 // reset the Reaches & UP entries
582 Reachblock[slidx] = lrgs(lidx)._def;
583 UPblock[slidx] = true;
584 // Record following instruction in case 'n' rematerializes and
585 // kills flags
586 Block *pred1 = _cfg.get_block_for_node(b->pred(1));
587 continue;
588 }
590 // Initialize needs_phi and needs_split
591 bool needs_phi = false;
592 bool needs_split = false;
593 bool has_phi = false;
594 // Walk the predecessor blocks to check inputs for that live range
595 // Grab predecessor block header
596 n1 = b->pred(1);
597 // Grab the appropriate reaching def info for inpidx
598 pred = _cfg.get_block_for_node(n1);
599 pidx = pred->_pre_order;
600 Node **Ltmp = Reaches[pidx];
601 bool *Utmp = UP[pidx];
602 n1 = Ltmp[slidx];
603 u1 = Utmp[slidx];
604 // Initialize node for saving type info
605 n3 = n1;
606 u3 = u1;
608 // Compare inputs to see if a Phi is needed
609 for( inpidx = 2; inpidx < b->num_preds(); inpidx++ ) {
610 // Grab predecessor block headers
611 n2 = b->pred(inpidx);
612 // Grab the appropriate reaching def info for inpidx
613 pred = _cfg.get_block_for_node(n2);
614 pidx = pred->_pre_order;
615 Ltmp = Reaches[pidx];
616 Utmp = UP[pidx];
617 n2 = Ltmp[slidx];
618 u2 = Utmp[slidx];
619 // For each LRG, decide if a phi is necessary
620 if( n1 != n2 ) {
621 needs_phi = true;
622 }
623 // See if the phi has mismatched inputs, UP vs. DOWN
624 if( n1 && n2 && (u1 != u2) ) {
625 needs_split = true;
626 }
627 // Move n2/u2 to n1/u1 for next iteration
628 n1 = n2;
629 u1 = u2;
630 // Preserve a non-NULL predecessor for later type referencing
631 if( (n3 == NULL) && (n2 != NULL) ){
632 n3 = n2;
633 u3 = u2;
634 }
635 } // End for all potential Phi inputs
637 // check block for appropriate phinode & update edges
638 for( insidx = 1; insidx <= b->end_idx(); insidx++ ) {
639 n1 = b->get_node(insidx);
640 // bail if this is not a phi
641 phi = n1->is_Phi() ? n1->as_Phi() : NULL;
642 if( phi == NULL ) {
643 // Keep track of index of first non-PhiNode instruction in block
644 non_phi = insidx;
645 // break out of the for loop as we have handled all phi nodes
646 break;
647 }
648 // must be looking at a phi
649 if (_lrg_map.find_id(n1) == lidxs.at(slidx)) {
650 // found the necessary phi
651 needs_phi = false;
652 has_phi = true;
653 // initialize the Reaches entry for this LRG
654 Reachblock[slidx] = phi;
655 break;
656 } // end if found correct phi
657 } // end for all phi's
659 // If a phi is needed or exist, check for it
660 if( needs_phi || has_phi ) {
661 // add new phinode if one not already found
662 if( needs_phi ) {
663 // create a new phi node and insert it into the block
664 // type is taken from left over pointer to a predecessor
665 assert(n3,"No non-NULL reaching DEF for a Phi");
666 phi = new (C) PhiNode(b->head(), n3->bottom_type());
667 // initialize the Reaches entry for this LRG
668 Reachblock[slidx] = phi;
670 // add node to block & node_to_block mapping
671 insert_proj(b, insidx++, phi, maxlrg++);
672 non_phi++;
673 // Reset new phi's mapping to be the spilling live range
674 _lrg_map.map(phi->_idx, lidx);
675 assert(_lrg_map.find_id(phi) == lidx, "Bad update on Union-Find mapping");
676 } // end if not found correct phi
677 // Here you have either found or created the Phi, so record it
678 assert(phi != NULL,"Must have a Phi Node here");
679 phis->push(phi);
680 // PhiNodes should either force the LRG UP or DOWN depending
681 // on its inputs and the register pressure in the Phi's block.
682 UPblock[slidx] = true; // Assume new DEF is UP
683 // If entering a high-pressure area with no immediate use,
684 // assume Phi is DOWN
685 if( is_high_pressure( b, &lrgs(lidx), b->end_idx()) && !prompt_use(b,lidx) )
686 UPblock[slidx] = false;
687 // If we are not split up/down and all inputs are down, then we
688 // are down
689 if( !needs_split && !u3 )
690 UPblock[slidx] = false;
691 } // end if phi is needed
693 // Do not need a phi, so grab the reaching DEF
694 else {
695 // Grab predecessor block header
696 n1 = b->pred(1);
697 // Grab the appropriate reaching def info for k
698 pred = _cfg.get_block_for_node(n1);
699 pidx = pred->_pre_order;
700 Node **Ltmp = Reaches[pidx];
701 bool *Utmp = UP[pidx];
702 // reset the Reaches & UP entries
703 Reachblock[slidx] = Ltmp[slidx];
704 UPblock[slidx] = Utmp[slidx];
705 } // end else no Phi is needed
706 } // end for all spilling live ranges
707 // DEBUG
708 #ifndef PRODUCT
709 if(trace_spilling()) {
710 tty->print("/`\nBlock %d: ", b->_pre_order);
711 tty->print("Reaching Definitions after Phi handling\n");
712 for( uint x = 0; x < spill_cnt; x++ ) {
713 tty->print("Spill Idx %d: UP %d: Node\n",x,UPblock[x]);
714 if( Reachblock[x] )
715 Reachblock[x]->dump();
716 else
717 tty->print("Undefined\n");
718 }
719 }
720 #endif
722 //----------Non-Phi Node Splitting----------
723 // Since phi-nodes have now been handled, the Reachblock array for this
724 // block is initialized with the correct starting value for the defs which
725 // reach non-phi instructions in this block. Thus, process non-phi
726 // instructions normally, inserting SpillCopy nodes for all spill
727 // locations.
729 // Memoize any DOWN reaching definitions for use as DEBUG info
730 for( insidx = 0; insidx < spill_cnt; insidx++ ) {
731 debug_defs[insidx] = (UPblock[insidx]) ? NULL : Reachblock[insidx];
732 if( UPblock[insidx] ) // Memoize UP decision at block start
733 UP_entry[insidx]->set( b->_pre_order );
734 }
736 //----------Walk Instructions in the Block and Split----------
737 // For all non-phi instructions in the block
738 for( insidx = 1; insidx <= b->end_idx(); insidx++ ) {
739 Node *n = b->get_node(insidx);
740 // Find the defining Node's live range index
741 uint defidx = _lrg_map.find_id(n);
742 uint cnt = n->req();
744 if (n->is_Phi()) {
745 // Skip phi nodes after removing dead copies.
746 if (defidx < _lrg_map.max_lrg_id()) {
747 // Check for useless Phis. These appear if we spill, then
748 // coalesce away copies. Dont touch Phis in spilling live
749 // ranges; they are busy getting modifed in this pass.
750 if( lrgs(defidx).reg() < LRG::SPILL_REG ) {
751 uint i;
752 Node *u = NULL;
753 // Look for the Phi merging 2 unique inputs
754 for( i = 1; i < cnt; i++ ) {
755 // Ignore repeats and self
756 if( n->in(i) != u && n->in(i) != n ) {
757 // Found a unique input
758 if( u != NULL ) // If it's the 2nd, bail out
759 break;
760 u = n->in(i); // Else record it
761 }
762 }
763 assert( u, "at least 1 valid input expected" );
764 if (i >= cnt) { // Found one unique input
765 assert(_lrg_map.find_id(n) == _lrg_map.find_id(u), "should be the same lrg");
766 n->replace_by(u); // Then replace with unique input
767 n->disconnect_inputs(NULL, C);
768 b->remove_node(insidx);
769 insidx--;
770 b->_ihrp_index--;
771 b->_fhrp_index--;
772 }
773 }
774 }
775 continue;
776 }
777 assert( insidx > b->_ihrp_index ||
778 (b->_reg_pressure < (uint)INTPRESSURE) ||
779 b->_ihrp_index > 4000000 ||
780 b->_ihrp_index >= b->end_idx() ||
781 !b->get_node(b->_ihrp_index)->is_Proj(), "" );
782 assert( insidx > b->_fhrp_index ||
783 (b->_freg_pressure < (uint)FLOATPRESSURE) ||
784 b->_fhrp_index > 4000000 ||
785 b->_fhrp_index >= b->end_idx() ||
786 !b->get_node(b->_fhrp_index)->is_Proj(), "" );
788 // ********** Handle Crossing HRP Boundry **********
789 if( (insidx == b->_ihrp_index) || (insidx == b->_fhrp_index) ) {
790 for( slidx = 0; slidx < spill_cnt; slidx++ ) {
791 // Check for need to split at HRP boundary - split if UP
792 n1 = Reachblock[slidx];
793 // bail out if no reaching DEF
794 if( n1 == NULL ) continue;
795 // bail out if live range is 'isolated' around inner loop
796 uint lidx = lidxs.at(slidx);
797 // If live range is currently UP
798 if( UPblock[slidx] ) {
799 // set location to insert spills at
800 // SPLIT DOWN HERE - NO CISC SPILL
801 if( is_high_pressure( b, &lrgs(lidx), insidx ) &&
802 !n1->rematerialize() ) {
803 // If there is already a valid stack definition available, use it
804 if( debug_defs[slidx] != NULL ) {
805 Reachblock[slidx] = debug_defs[slidx];
806 }
807 else {
808 // Insert point is just past last use or def in the block
809 int insert_point = insidx-1;
810 while( insert_point > 0 ) {
811 Node *n = b->get_node(insert_point);
812 // Hit top of block? Quit going backwards
813 if (n->is_Phi()) {
814 break;
815 }
816 // Found a def? Better split after it.
817 if (_lrg_map.live_range_id(n) == lidx) {
818 break;
819 }
820 // Look for a use
821 uint i;
822 for( i = 1; i < n->req(); i++ ) {
823 if (_lrg_map.live_range_id(n->in(i)) == lidx) {
824 break;
825 }
826 }
827 // Found a use? Better split after it.
828 if (i < n->req()) {
829 break;
830 }
831 insert_point--;
832 }
833 uint orig_eidx = b->end_idx();
834 maxlrg = split_DEF( n1, b, insert_point, maxlrg, Reachblock, debug_defs, splits, slidx);
835 // If it wasn't split bail
836 if (!maxlrg) {
837 return 0;
838 }
839 // Spill of NULL check mem op goes into the following block.
840 if (b->end_idx() > orig_eidx) {
841 insidx++;
842 }
843 }
844 // This is a new DEF, so update UP
845 UPblock[slidx] = false;
846 #ifndef PRODUCT
847 // DEBUG
848 if( trace_spilling() ) {
849 tty->print("\nNew Split DOWN DEF of Spill Idx ");
850 tty->print("%d, UP %d:\n",slidx,false);
851 n1->dump();
852 }
853 #endif
854 }
855 } // end if LRG is UP
856 } // end for all spilling live ranges
857 assert( b->get_node(insidx) == n, "got insidx set incorrectly" );
858 } // end if crossing HRP Boundry
860 // If the LRG index is oob, then this is a new spillcopy, skip it.
861 if (defidx >= _lrg_map.max_lrg_id()) {
862 continue;
863 }
864 LRG &deflrg = lrgs(defidx);
865 uint copyidx = n->is_Copy();
866 // Remove coalesced copy from CFG
867 if (copyidx && defidx == _lrg_map.live_range_id(n->in(copyidx))) {
868 n->replace_by( n->in(copyidx) );
869 n->set_req( copyidx, NULL );
870 b->remove_node(insidx--);
871 b->_ihrp_index--; // Adjust the point where we go hi-pressure
872 b->_fhrp_index--;
873 continue;
874 }
876 #define DERIVED 0
878 // ********** Handle USES **********
879 bool nullcheck = false;
880 // Implicit null checks never use the spilled value
881 if( n->is_MachNullCheck() )
882 nullcheck = true;
883 if( !nullcheck ) {
884 // Search all inputs for a Spill-USE
885 JVMState* jvms = n->jvms();
886 uint oopoff = jvms ? jvms->oopoff() : cnt;
887 uint old_last = cnt - 1;
888 for( inpidx = 1; inpidx < cnt; inpidx++ ) {
889 // Derived/base pairs may be added to our inputs during this loop.
890 // If inpidx > old_last, then one of these new inputs is being
891 // handled. Skip the derived part of the pair, but process
892 // the base like any other input.
893 if (inpidx > old_last && ((inpidx - oopoff) & 1) == DERIVED) {
894 continue; // skip derived_debug added below
895 }
896 // Get lidx of input
897 uint useidx = _lrg_map.find_id(n->in(inpidx));
898 // Not a brand-new split, and it is a spill use
899 if (useidx < _lrg_map.max_lrg_id() && lrgs(useidx).reg() >= LRG::SPILL_REG) {
900 // Check for valid reaching DEF
901 slidx = lrg2reach[useidx];
902 Node *def = Reachblock[slidx];
903 assert( def != NULL, "Using Undefined Value in Split()\n");
905 // (+++) %%%% remove this in favor of pre-pass in matcher.cpp
906 // monitor references do not care where they live, so just hook
907 if ( jvms && jvms->is_monitor_use(inpidx) ) {
908 // The effect of this clone is to drop the node out of the block,
909 // so that the allocator does not see it anymore, and therefore
910 // does not attempt to assign it a register.
911 def = clone_node(def, b, C);
912 if (def == NULL || C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) {
913 return 0;
914 }
915 _lrg_map.extend(def->_idx, 0);
916 _cfg.map_node_to_block(def, b);
917 n->set_req(inpidx, def);
918 continue;
919 }
921 // Rematerializable? Then clone def at use site instead
922 // of store/load
923 if( def->rematerialize() ) {
924 int old_size = b->number_of_nodes();
925 def = split_Rematerialize( def, b, insidx, maxlrg, splits, slidx, lrg2reach, Reachblock, true );
926 if( !def ) return 0; // Bail out
927 insidx += b->number_of_nodes()-old_size;
928 }
930 MachNode *mach = n->is_Mach() ? n->as_Mach() : NULL;
931 // Base pointers and oopmap references do not care where they live.
932 if ((inpidx >= oopoff) ||
933 (mach && mach->ideal_Opcode() == Op_AddP && inpidx == AddPNode::Base)) {
934 if (def->rematerialize() && lrgs(useidx)._was_spilled2) {
935 // This def has been rematerialized a couple of times without
936 // progress. It doesn't care if it lives UP or DOWN, so
937 // spill it down now.
938 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false,splits,slidx);
939 // If it wasn't split bail
940 if (!maxlrg) {
941 return 0;
942 }
943 insidx++; // Reset iterator to skip USE side split
944 } else {
945 // Just hook the def edge
946 n->set_req(inpidx, def);
947 }
949 if (inpidx >= oopoff) {
950 // After oopoff, we have derived/base pairs. We must mention all
951 // derived pointers here as derived/base pairs for GC. If the
952 // derived value is spilling and we have a copy both in Reachblock
953 // (called here 'def') and debug_defs[slidx] we need to mention
954 // both in derived/base pairs or kill one.
955 Node *derived_debug = debug_defs[slidx];
956 if( ((inpidx - oopoff) & 1) == DERIVED && // derived vs base?
957 mach && mach->ideal_Opcode() != Op_Halt &&
958 derived_debug != NULL &&
959 derived_debug != def ) { // Actual 2nd value appears
960 // We have already set 'def' as a derived value.
961 // Also set debug_defs[slidx] as a derived value.
962 uint k;
963 for( k = oopoff; k < cnt; k += 2 )
964 if( n->in(k) == derived_debug )
965 break; // Found an instance of debug derived
966 if( k == cnt ) {// No instance of debug_defs[slidx]
967 // Add a derived/base pair to cover the debug info.
968 // We have to process the added base later since it is not
969 // handled yet at this point but skip derived part.
970 assert(((n->req() - oopoff) & 1) == DERIVED,
971 "must match skip condition above");
972 n->add_req( derived_debug ); // this will be skipped above
973 n->add_req( n->in(inpidx+1) ); // this will be processed
974 // Increment cnt to handle added input edges on
975 // subsequent iterations.
976 cnt += 2;
977 }
978 }
979 }
980 continue;
981 }
982 // Special logic for DEBUG info
983 if( jvms && b->_freq > BLOCK_FREQUENCY(0.5) ) {
984 uint debug_start = jvms->debug_start();
985 // If this is debug info use & there is a reaching DOWN def
986 if ((debug_start <= inpidx) && (debug_defs[slidx] != NULL)) {
987 assert(inpidx < oopoff, "handle only debug info here");
988 // Just hook it in & move on
989 n->set_req(inpidx, debug_defs[slidx]);
990 // (Note that this can make two sides of a split live at the
991 // same time: The debug def on stack, and another def in a
992 // register. The GC needs to know about both of them, but any
993 // derived pointers after oopoff will refer to only one of the
994 // two defs and the GC would therefore miss the other. Thus
995 // this hack is only allowed for debug info which is Java state
996 // and therefore never a derived pointer.)
997 continue;
998 }
999 }
1000 // Grab register mask info
1001 const RegMask &dmask = def->out_RegMask();
1002 const RegMask &umask = n->in_RegMask(inpidx);
1003 bool is_vect = RegMask::is_vector(def->ideal_reg());
1004 assert(inpidx < oopoff, "cannot use-split oop map info");
1006 bool dup = UPblock[slidx];
1007 bool uup = umask.is_UP();
1009 // Need special logic to handle bound USES. Insert a split at this
1010 // bound use if we can't rematerialize the def, or if we need the
1011 // split to form a misaligned pair.
1012 if( !umask.is_AllStack() &&
1013 (int)umask.Size() <= lrgs(useidx).num_regs() &&
1014 (!def->rematerialize() ||
1015 !is_vect && umask.is_misaligned_pair())) {
1016 // These need a Split regardless of overlap or pressure
1017 // SPLIT - NO DEF - NO CISC SPILL
1018 maxlrg = split_USE(def,b,n,inpidx,maxlrg,dup,false, splits,slidx);
1019 // If it wasn't split bail
1020 if (!maxlrg) {
1021 return 0;
1022 }
1023 insidx++; // Reset iterator to skip USE side split
1024 continue;
1025 }
1027 if (UseFPUForSpilling && n->is_MachCall() && !uup && !dup ) {
1028 // The use at the call can force the def down so insert
1029 // a split before the use to allow the def more freedom.
1030 maxlrg = split_USE(def,b,n,inpidx,maxlrg,dup,false, splits,slidx);
1031 // If it wasn't split bail
1032 if (!maxlrg) {
1033 return 0;
1034 }
1035 insidx++; // Reset iterator to skip USE side split
1036 continue;
1037 }
1039 // Here is the logic chart which describes USE Splitting:
1040 // 0 = false or DOWN, 1 = true or UP
1041 //
1042 // Overlap | DEF | USE | Action
1043 //-------------------------------------------------------
1044 // 0 | 0 | 0 | Copy - mem -> mem
1045 // 0 | 0 | 1 | Split-UP - Check HRP
1046 // 0 | 1 | 0 | Split-DOWN - Debug Info?
1047 // 0 | 1 | 1 | Copy - reg -> reg
1048 // 1 | 0 | 0 | Reset Input Edge (no Split)
1049 // 1 | 0 | 1 | Split-UP - Check HRP
1050 // 1 | 1 | 0 | Split-DOWN - Debug Info?
1051 // 1 | 1 | 1 | Reset Input Edge (no Split)
1052 //
1053 // So, if (dup == uup), then overlap test determines action,
1054 // with true being no split, and false being copy. Else,
1055 // if DEF is DOWN, Split-UP, and check HRP to decide on
1056 // resetting DEF. Finally if DEF is UP, Split-DOWN, with
1057 // special handling for Debug Info.
1058 if( dup == uup ) {
1059 if( dmask.overlap(umask) ) {
1060 // Both are either up or down, and there is overlap, No Split
1061 n->set_req(inpidx, def);
1062 }
1063 else { // Both are either up or down, and there is no overlap
1064 if( dup ) { // If UP, reg->reg copy
1065 // COPY ACROSS HERE - NO DEF - NO CISC SPILL
1066 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false, splits,slidx);
1067 // If it wasn't split bail
1068 if (!maxlrg) {
1069 return 0;
1070 }
1071 insidx++; // Reset iterator to skip USE side split
1072 }
1073 else { // DOWN, mem->mem copy
1074 // COPY UP & DOWN HERE - NO DEF - NO CISC SPILL
1075 // First Split-UP to move value into Register
1076 uint def_ideal = def->ideal_reg();
1077 const RegMask* tmp_rm = Matcher::idealreg2regmask[def_ideal];
1078 Node *spill = new (C) MachSpillCopyNode(def, dmask, *tmp_rm);
1079 insert_proj( b, insidx, spill, maxlrg );
1080 // Then Split-DOWN as if previous Split was DEF
1081 maxlrg = split_USE(spill,b,n,inpidx,maxlrg,false,false, splits,slidx);
1082 // If it wasn't split bail
1083 if (!maxlrg) {
1084 return 0;
1085 }
1086 insidx += 2; // Reset iterator to skip USE side splits
1087 }
1088 } // End else no overlap
1089 } // End if dup == uup
1090 // dup != uup, so check dup for direction of Split
1091 else {
1092 if( dup ) { // If UP, Split-DOWN and check Debug Info
1093 // If this node is already a SpillCopy, just patch the edge
1094 // except the case of spilling to stack.
1095 if( n->is_SpillCopy() ) {
1096 RegMask tmp_rm(umask);
1097 tmp_rm.SUBTRACT(Matcher::STACK_ONLY_mask);
1098 if( dmask.overlap(tmp_rm) ) {
1099 if( def != n->in(inpidx) ) {
1100 n->set_req(inpidx, def);
1101 }
1102 continue;
1103 }
1104 }
1105 // COPY DOWN HERE - NO DEF - NO CISC SPILL
1106 maxlrg = split_USE(def,b,n,inpidx,maxlrg,false,false, splits,slidx);
1107 // If it wasn't split bail
1108 if (!maxlrg) {
1109 return 0;
1110 }
1111 insidx++; // Reset iterator to skip USE side split
1112 // Check for debug-info split. Capture it for later
1113 // debug splits of the same value
1114 if (jvms && jvms->debug_start() <= inpidx && inpidx < oopoff)
1115 debug_defs[slidx] = n->in(inpidx);
1117 }
1118 else { // DOWN, Split-UP and check register pressure
1119 if( is_high_pressure( b, &lrgs(useidx), insidx ) ) {
1120 // COPY UP HERE - NO DEF - CISC SPILL
1121 maxlrg = split_USE(def,b,n,inpidx,maxlrg,true,true, splits,slidx);
1122 // If it wasn't split bail
1123 if (!maxlrg) {
1124 return 0;
1125 }
1126 insidx++; // Reset iterator to skip USE side split
1127 } else { // LRP
1128 // COPY UP HERE - WITH DEF - NO CISC SPILL
1129 maxlrg = split_USE(def,b,n,inpidx,maxlrg,true,false, splits,slidx);
1130 // If it wasn't split bail
1131 if (!maxlrg) {
1132 return 0;
1133 }
1134 // Flag this lift-up in a low-pressure block as
1135 // already-spilled, so if it spills again it will
1136 // spill hard (instead of not spilling hard and
1137 // coalescing away).
1138 set_was_spilled(n->in(inpidx));
1139 // Since this is a new DEF, update Reachblock & UP
1140 Reachblock[slidx] = n->in(inpidx);
1141 UPblock[slidx] = true;
1142 insidx++; // Reset iterator to skip USE side split
1143 }
1144 } // End else DOWN
1145 } // End dup != uup
1146 } // End if Spill USE
1147 } // End For All Inputs
1148 } // End If not nullcheck
1150 // ********** Handle DEFS **********
1151 // DEFS either Split DOWN in HRP regions or when the LRG is bound, or
1152 // just reset the Reaches info in LRP regions. DEFS must always update
1153 // UP info.
1154 if( deflrg.reg() >= LRG::SPILL_REG ) { // Spilled?
1155 uint slidx = lrg2reach[defidx];
1156 // Add to defs list for later assignment of new live range number
1157 defs->push(n);
1158 // Set a flag on the Node indicating it has already spilled.
1159 // Only do it for capacity spills not conflict spills.
1160 if( !deflrg._direct_conflict )
1161 set_was_spilled(n);
1162 assert(!n->is_Phi(),"Cannot insert Phi into DEFS list");
1163 // Grab UP info for DEF
1164 const RegMask &dmask = n->out_RegMask();
1165 bool defup = dmask.is_UP();
1166 int ireg = n->ideal_reg();
1167 bool is_vect = RegMask::is_vector(ireg);
1168 // Only split at Def if this is a HRP block or bound (and spilled once)
1169 if( !n->rematerialize() &&
1170 (((dmask.is_bound(ireg) || !is_vect && dmask.is_misaligned_pair()) &&
1171 (deflrg._direct_conflict || deflrg._must_spill)) ||
1172 // Check for LRG being up in a register and we are inside a high
1173 // pressure area. Spill it down immediately.
1174 (defup && is_high_pressure(b,&deflrg,insidx))) ) {
1175 assert( !n->rematerialize(), "" );
1176 assert( !n->is_SpillCopy(), "" );
1177 // Do a split at the def site.
1178 maxlrg = split_DEF( n, b, insidx, maxlrg, Reachblock, debug_defs, splits, slidx );
1179 // If it wasn't split bail
1180 if (!maxlrg) {
1181 return 0;
1182 }
1183 // Split DEF's Down
1184 UPblock[slidx] = 0;
1185 #ifndef PRODUCT
1186 // DEBUG
1187 if( trace_spilling() ) {
1188 tty->print("\nNew Split DOWN DEF of Spill Idx ");
1189 tty->print("%d, UP %d:\n",slidx,false);
1190 n->dump();
1191 }
1192 #endif
1193 }
1194 else { // Neither bound nor HRP, must be LRP
1195 // otherwise, just record the def
1196 Reachblock[slidx] = n;
1197 // UP should come from the outRegmask() of the DEF
1198 UPblock[slidx] = defup;
1199 // Update debug list of reaching down definitions, kill if DEF is UP
1200 debug_defs[slidx] = defup ? NULL : n;
1201 #ifndef PRODUCT
1202 // DEBUG
1203 if( trace_spilling() ) {
1204 tty->print("\nNew DEF of Spill Idx ");
1205 tty->print("%d, UP %d:\n",slidx,defup);
1206 n->dump();
1207 }
1208 #endif
1209 } // End else LRP
1210 } // End if spill def
1212 // ********** Split Left Over Mem-Mem Moves **********
1213 // Check for mem-mem copies and split them now. Do not do this
1214 // to copies about to be spilled; they will be Split shortly.
1215 if (copyidx) {
1216 Node *use = n->in(copyidx);
1217 uint useidx = _lrg_map.find_id(use);
1218 if (useidx < _lrg_map.max_lrg_id() && // This is not a new split
1219 OptoReg::is_stack(deflrg.reg()) &&
1220 deflrg.reg() < LRG::SPILL_REG ) { // And DEF is from stack
1221 LRG &uselrg = lrgs(useidx);
1222 if( OptoReg::is_stack(uselrg.reg()) &&
1223 uselrg.reg() < LRG::SPILL_REG && // USE is from stack
1224 deflrg.reg() != uselrg.reg() ) { // Not trivially removed
1225 uint def_ideal_reg = n->bottom_type()->ideal_reg();
1226 const RegMask &def_rm = *Matcher::idealreg2regmask[def_ideal_reg];
1227 const RegMask &use_rm = n->in_RegMask(copyidx);
1228 if( def_rm.overlap(use_rm) && n->is_SpillCopy() ) { // Bug 4707800, 'n' may be a storeSSL
1229 if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { // Check when generating nodes
1230 return 0;
1231 }
1232 Node *spill = new (C) MachSpillCopyNode(use,use_rm,def_rm);
1233 n->set_req(copyidx,spill);
1234 n->as_MachSpillCopy()->set_in_RegMask(def_rm);
1235 // Put the spill just before the copy
1236 insert_proj( b, insidx++, spill, maxlrg++ );
1237 }
1238 }
1239 }
1240 }
1241 } // End For All Instructions in Block - Non-PHI Pass
1243 // Check if each LRG is live out of this block so as not to propagate
1244 // beyond the last use of a LRG.
1245 for( slidx = 0; slidx < spill_cnt; slidx++ ) {
1246 uint defidx = lidxs.at(slidx);
1247 IndexSet *liveout = _live->live(b);
1248 if( !liveout->member(defidx) ) {
1249 #ifdef ASSERT
1250 // The index defidx is not live. Check the liveout array to ensure that
1251 // it contains no members which compress to defidx. Finding such an
1252 // instance may be a case to add liveout adjustment in compress_uf_map().
1253 // See 5063219.
1254 uint member;
1255 IndexSetIterator isi(liveout);
1256 while ((member = isi.next()) != 0) {
1257 assert(defidx != _lrg_map.find_const(member), "Live out member has not been compressed");
1258 }
1259 #endif
1260 Reachblock[slidx] = NULL;
1261 } else {
1262 assert(Reachblock[slidx] != NULL,"No reaching definition for liveout value");
1263 }
1264 }
1265 #ifndef PRODUCT
1266 if( trace_spilling() )
1267 b->dump();
1268 #endif
1269 } // End For All Blocks
1271 //----------PASS 2----------
1272 // Reset all DEF live range numbers here
1273 for( insidx = 0; insidx < defs->size(); insidx++ ) {
1274 // Grab the def
1275 n1 = defs->at(insidx);
1276 // Set new lidx for DEF
1277 new_lrg(n1, maxlrg++);
1278 }
1279 //----------Phi Node Splitting----------
1280 // Clean up a phi here, and assign a new live range number
1281 // Cycle through this block's predecessors, collecting Reaches
1282 // info for each spilled LRG and update edges.
1283 // Walk the phis list to patch inputs, split phis, and name phis
1284 uint lrgs_before_phi_split = maxlrg;
1285 for( insidx = 0; insidx < phis->size(); insidx++ ) {
1286 Node *phi = phis->at(insidx);
1287 assert(phi->is_Phi(),"This list must only contain Phi Nodes");
1288 Block *b = _cfg.get_block_for_node(phi);
1289 // Grab the live range number
1290 uint lidx = _lrg_map.find_id(phi);
1291 uint slidx = lrg2reach[lidx];
1292 // Update node to lidx map
1293 new_lrg(phi, maxlrg++);
1294 // Get PASS1's up/down decision for the block.
1295 int phi_up = !!UP_entry[slidx]->test(b->_pre_order);
1297 // Force down if double-spilling live range
1298 if( lrgs(lidx)._was_spilled1 )
1299 phi_up = false;
1301 // When splitting a Phi we an split it normal or "inverted".
1302 // An inverted split makes the splits target the Phi's UP/DOWN
1303 // sense inverted; then the Phi is followed by a final def-side
1304 // split to invert back. It changes which blocks the spill code
1305 // goes in.
1307 // Walk the predecessor blocks and assign the reaching def to the Phi.
1308 // Split Phi nodes by placing USE side splits wherever the reaching
1309 // DEF has the wrong UP/DOWN value.
1310 for( uint i = 1; i < b->num_preds(); i++ ) {
1311 // Get predecessor block pre-order number
1312 Block *pred = _cfg.get_block_for_node(b->pred(i));
1313 pidx = pred->_pre_order;
1314 // Grab reaching def
1315 Node *def = Reaches[pidx][slidx];
1316 Node** Reachblock = Reaches[pidx];
1317 assert( def, "must have reaching def" );
1318 // If input up/down sense and reg-pressure DISagree
1319 if (def->rematerialize()) {
1320 // Place the rematerialized node above any MSCs created during
1321 // phi node splitting. end_idx points at the insertion point
1322 // so look at the node before it.
1323 int insert = pred->end_idx();
1324 while (insert >= 1 &&
1325 pred->get_node(insert - 1)->is_SpillCopy() &&
1326 _lrg_map.find(pred->get_node(insert - 1)) >= lrgs_before_phi_split) {
1327 insert--;
1328 }
1329 def = split_Rematerialize(def, pred, insert, maxlrg, splits, slidx, lrg2reach, Reachblock, false);
1330 if (!def) {
1331 return 0; // Bail out
1332 }
1333 }
1334 // Update the Phi's input edge array
1335 phi->set_req(i,def);
1336 // Grab the UP/DOWN sense for the input
1337 u1 = UP[pidx][slidx];
1338 if( u1 != (phi_up != 0)) {
1339 maxlrg = split_USE(def, b, phi, i, maxlrg, !u1, false, splits,slidx);
1340 // If it wasn't split bail
1341 if (!maxlrg) {
1342 return 0;
1343 }
1344 }
1345 } // End for all inputs to the Phi
1346 } // End for all Phi Nodes
1347 // Update _maxlrg to save Union asserts
1348 _lrg_map.set_max_lrg_id(maxlrg);
1351 //----------PASS 3----------
1352 // Pass over all Phi's to union the live ranges
1353 for( insidx = 0; insidx < phis->size(); insidx++ ) {
1354 Node *phi = phis->at(insidx);
1355 assert(phi->is_Phi(),"This list must only contain Phi Nodes");
1356 // Walk all inputs to Phi and Union input live range with Phi live range
1357 for( uint i = 1; i < phi->req(); i++ ) {
1358 // Grab the input node
1359 Node *n = phi->in(i);
1360 assert(n, "node should exist");
1361 uint lidx = _lrg_map.find(n);
1362 uint pidx = _lrg_map.find(phi);
1363 if (lidx < pidx) {
1364 Union(n, phi);
1365 }
1366 else if(lidx > pidx) {
1367 Union(phi, n);
1368 }
1369 } // End for all inputs to the Phi Node
1370 } // End for all Phi Nodes
1371 // Now union all two address instructions
1372 for (insidx = 0; insidx < defs->size(); insidx++) {
1373 // Grab the def
1374 n1 = defs->at(insidx);
1375 // Set new lidx for DEF & handle 2-addr instructions
1376 if (n1->is_Mach() && ((twoidx = n1->as_Mach()->two_adr()) != 0)) {
1377 assert(_lrg_map.find(n1->in(twoidx)) < maxlrg,"Assigning bad live range index");
1378 // Union the input and output live ranges
1379 uint lr1 = _lrg_map.find(n1);
1380 uint lr2 = _lrg_map.find(n1->in(twoidx));
1381 if (lr1 < lr2) {
1382 Union(n1, n1->in(twoidx));
1383 }
1384 else if (lr1 > lr2) {
1385 Union(n1->in(twoidx), n1);
1386 }
1387 } // End if two address
1388 } // End for all defs
1389 // DEBUG
1390 #ifdef ASSERT
1391 // Validate all live range index assignments
1392 for (bidx = 0; bidx < _cfg.number_of_blocks(); bidx++) {
1393 b = _cfg.get_block(bidx);
1394 for (insidx = 0; insidx <= b->end_idx(); insidx++) {
1395 Node *n = b->get_node(insidx);
1396 uint defidx = _lrg_map.find(n);
1397 assert(defidx < _lrg_map.max_lrg_id(), "Bad live range index in Split");
1398 assert(defidx < maxlrg,"Bad live range index in Split");
1399 }
1400 }
1401 // Issue a warning if splitting made no progress
1402 int noprogress = 0;
1403 for (slidx = 0; slidx < spill_cnt; slidx++) {
1404 if (PrintOpto && WizardMode && splits.at(slidx) == 0) {
1405 tty->print_cr("Failed to split live range %d", lidxs.at(slidx));
1406 //BREAKPOINT;
1407 }
1408 else {
1409 noprogress++;
1410 }
1411 }
1412 if(!noprogress) {
1413 tty->print_cr("Failed to make progress in Split");
1414 //BREAKPOINT;
1415 }
1416 #endif
1417 // Return updated count of live ranges
1418 return maxlrg;
1419 }