1 | /* |
2 | * Copyright (c) 2000, 2018, 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 | */ |
24 | |
25 | #include "precompiled.hpp" |
26 | #include "libadt/vectset.hpp" |
27 | #include "memory/allocation.inline.hpp" |
28 | #include "memory/resourceArea.inline.hpp" |
29 | #include "opto/addnode.hpp" |
30 | #include "opto/c2compiler.hpp" |
31 | #include "opto/callnode.hpp" |
32 | #include "opto/cfgnode.hpp" |
33 | #include "opto/chaitin.hpp" |
34 | #include "opto/loopnode.hpp" |
35 | #include "opto/machnode.hpp" |
36 | |
37 | //------------------------------Split-------------------------------------- |
38 | // Walk the graph in RPO and for each lrg which spills, propagate reaching |
39 | // definitions. During propagation, split the live range around regions of |
40 | // High Register Pressure (HRP). If a Def is in a region of Low Register |
41 | // Pressure (LRP), it will not get spilled until we encounter a region of |
42 | // HRP between it and one of its uses. We will spill at the transition |
43 | // point between LRP and HRP. Uses in the HRP region will use the spilled |
44 | // Def. The first Use outside the HRP region will generate a SpillCopy to |
45 | // hoist the live range back up into a register, and all subsequent uses |
46 | // will use that new Def until another HRP region is encountered. Defs in |
47 | // HRP regions will get trailing SpillCopies to push the LRG down into the |
48 | // stack immediately. |
49 | // |
50 | // As a side effect, unlink from (hence make dead) coalesced copies. |
51 | // |
52 | |
53 | static const char out_of_nodes[] = "out of nodes during split" ; |
54 | |
55 | //------------------------------get_spillcopy_wide----------------------------- |
56 | // Get a SpillCopy node with wide-enough masks. Use the 'wide-mask', the |
57 | // wide ideal-register spill-mask if possible. If the 'wide-mask' does |
58 | // not cover the input (or output), use the input (or output) mask instead. |
59 | Node *PhaseChaitin::get_spillcopy_wide(MachSpillCopyNode::SpillType spill_type, Node *def, Node *use, uint uidx) { |
60 | // If ideal reg doesn't exist we've got a bad schedule happening |
61 | // that is forcing us to spill something that isn't spillable. |
62 | // Bail rather than abort |
63 | uint ireg = def->ideal_reg(); |
64 | if (ireg == 0 || ireg == Op_RegFlags) { |
65 | assert(false, "attempted to spill a non-spillable item: %d: %s <- %d: %s, ireg = %u, spill_type: %s" , |
66 | def->_idx, def->Name(), use->_idx, use->Name(), ireg, |
67 | MachSpillCopyNode::spill_type(spill_type)); |
68 | C->record_method_not_compilable("attempted to spill a non-spillable item" ); |
69 | return NULL; |
70 | } |
71 | if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { |
72 | return NULL; |
73 | } |
74 | const RegMask *i_mask = &def->out_RegMask(); |
75 | const RegMask *w_mask = C->matcher()->idealreg2spillmask[ireg]; |
76 | const RegMask *o_mask = use ? &use->in_RegMask(uidx) : w_mask; |
77 | const RegMask *w_i_mask = w_mask->overlap( *i_mask ) ? w_mask : i_mask; |
78 | const RegMask *w_o_mask; |
79 | |
80 | int num_regs = RegMask::num_registers(ireg); |
81 | bool is_vect = RegMask::is_vector(ireg); |
82 | if( w_mask->overlap( *o_mask ) && // Overlap AND |
83 | (num_regs == 1 // Single use or aligned |
84 | || is_vect // or vector |
85 | || (!is_vect && o_mask->is_aligned_pairs())) ) { |
86 | assert(!is_vect || o_mask->is_aligned_sets(num_regs), "vectors are aligned" ); |
87 | // Don't come here for mis-aligned doubles |
88 | w_o_mask = w_mask; |
89 | } else { // wide ideal mask does not overlap with o_mask |
90 | // Mis-aligned doubles come here and XMM->FPR moves on x86. |
91 | w_o_mask = o_mask; // Must target desired registers |
92 | // Does the ideal-reg-mask overlap with o_mask? I.e., can I use |
93 | // a reg-reg move or do I need a trip across register classes |
94 | // (and thus through memory)? |
95 | if( !C->matcher()->idealreg2regmask[ireg]->overlap( *o_mask) && o_mask->is_UP() ) |
96 | // Here we assume a trip through memory is required. |
97 | w_i_mask = &C->FIRST_STACK_mask(); |
98 | } |
99 | return new MachSpillCopyNode(spill_type, def, *w_i_mask, *w_o_mask ); |
100 | } |
101 | |
102 | //------------------------------insert_proj------------------------------------ |
103 | // Insert the spill at chosen location. Skip over any intervening Proj's or |
104 | // Phis. Skip over a CatchNode and projs, inserting in the fall-through block |
105 | // instead. Update high-pressure indices. Create a new live range. |
106 | void PhaseChaitin::insert_proj( Block *b, uint i, Node *spill, uint maxlrg ) { |
107 | // Skip intervening ProjNodes. Do not insert between a ProjNode and |
108 | // its definer. |
109 | while( i < b->number_of_nodes() && |
110 | (b->get_node(i)->is_Proj() || |
111 | b->get_node(i)->is_Phi() ) ) |
112 | i++; |
113 | |
114 | // Do not insert between a call and his Catch |
115 | if( b->get_node(i)->is_Catch() ) { |
116 | // Put the instruction at the top of the fall-thru block. |
117 | // Find the fall-thru projection |
118 | while( 1 ) { |
119 | const CatchProjNode *cp = b->get_node(++i)->as_CatchProj(); |
120 | if( cp->_con == CatchProjNode::fall_through_index ) |
121 | break; |
122 | } |
123 | int sidx = i - b->end_idx()-1; |
124 | b = b->_succs[sidx]; // Switch to successor block |
125 | i = 1; // Right at start of block |
126 | } |
127 | |
128 | b->insert_node(spill, i); // Insert node in block |
129 | _cfg.map_node_to_block(spill, b); // Update node->block mapping to reflect |
130 | // Adjust the point where we go hi-pressure |
131 | if( i <= b->_ihrp_index ) b->_ihrp_index++; |
132 | if( i <= b->_fhrp_index ) b->_fhrp_index++; |
133 | |
134 | // Assign a new Live Range Number to the SpillCopy and grow |
135 | // the node->live range mapping. |
136 | new_lrg(spill,maxlrg); |
137 | } |
138 | |
139 | //------------------------------split_DEF-------------------------------------- |
140 | // There are four categories of Split; UP/DOWN x DEF/USE |
141 | // Only three of these really occur as DOWN/USE will always color |
142 | // Any Split with a DEF cannot CISC-Spill now. Thus we need |
143 | // two helper routines, one for Split DEFS (insert after instruction), |
144 | // one for Split USES (insert before instruction). DEF insertion |
145 | // happens inside Split, where the Leaveblock array is updated. |
146 | uint PhaseChaitin::split_DEF( Node *def, Block *b, int loc, uint maxlrg, Node **Reachblock, Node **debug_defs, GrowableArray<uint> splits, int slidx ) { |
147 | #ifdef ASSERT |
148 | // Increment the counter for this lrg |
149 | splits.at_put(slidx, splits.at(slidx)+1); |
150 | #endif |
151 | // If we are spilling the memory op for an implicit null check, at the |
152 | // null check location (ie - null check is in HRP block) we need to do |
153 | // the null-check first, then spill-down in the following block. |
154 | // (The implicit_null_check function ensures the use is also dominated |
155 | // by the branch-not-taken block.) |
156 | Node *be = b->end(); |
157 | if( be->is_MachNullCheck() && be->in(1) == def && def == b->get_node(loc)) { |
158 | // Spill goes in the branch-not-taken block |
159 | b = b->_succs[b->get_node(b->end_idx()+1)->Opcode() == Op_IfTrue]; |
160 | loc = 0; // Just past the Region |
161 | } |
162 | assert( loc >= 0, "must insert past block head" ); |
163 | |
164 | // Get a def-side SpillCopy |
165 | Node *spill = get_spillcopy_wide(MachSpillCopyNode::Definition, def, NULL, 0); |
166 | // Did we fail to split?, then bail |
167 | if (!spill) { |
168 | return 0; |
169 | } |
170 | |
171 | // Insert the spill at chosen location |
172 | insert_proj( b, loc+1, spill, maxlrg++); |
173 | |
174 | // Insert new node into Reaches array |
175 | Reachblock[slidx] = spill; |
176 | // Update debug list of reaching down definitions by adding this one |
177 | debug_defs[slidx] = spill; |
178 | |
179 | // return updated count of live ranges |
180 | return maxlrg; |
181 | } |
182 | |
183 | //------------------------------split_USE-------------------------------------- |
184 | // Splits at uses can involve redeffing the LRG, so no CISC Spilling there. |
185 | // Debug uses want to know if def is already stack enabled. |
186 | uint PhaseChaitin::split_USE(MachSpillCopyNode::SpillType spill_type, Node *def, Block *b, Node *use, uint useidx, uint maxlrg, bool def_down, bool cisc_sp, GrowableArray<uint> splits, int slidx ) { |
187 | #ifdef ASSERT |
188 | // Increment the counter for this lrg |
189 | splits.at_put(slidx, splits.at(slidx)+1); |
190 | #endif |
191 | |
192 | // Some setup stuff for handling debug node uses |
193 | JVMState* jvms = use->jvms(); |
194 | uint debug_start = jvms ? jvms->debug_start() : 999999; |
195 | uint debug_end = jvms ? jvms->debug_end() : 999999; |
196 | |
197 | //------------------------------------------- |
198 | // Check for use of debug info |
199 | if (useidx >= debug_start && useidx < debug_end) { |
200 | // Actually it's perfectly legal for constant debug info to appear |
201 | // just unlikely. In this case the optimizer left a ConI of a 4 |
202 | // as both inputs to a Phi with only a debug use. It's a single-def |
203 | // live range of a rematerializable value. The live range spills, |
204 | // rematerializes and now the ConI directly feeds into the debug info. |
205 | // assert(!def->is_Con(), "constant debug info already constructed directly"); |
206 | |
207 | // Special split handling for Debug Info |
208 | // If DEF is DOWN, just hook the edge and return |
209 | // If DEF is UP, Split it DOWN for this USE. |
210 | if( def->is_Mach() ) { |
211 | if( def_down ) { |
212 | // DEF is DOWN, so connect USE directly to the DEF |
213 | use->set_req(useidx, def); |
214 | } else { |
215 | // Block and index where the use occurs. |
216 | Block *b = _cfg.get_block_for_node(use); |
217 | // Put the clone just prior to use |
218 | int bindex = b->find_node(use); |
219 | // DEF is UP, so must copy it DOWN and hook in USE |
220 | // Insert SpillCopy before the USE, which uses DEF as its input, |
221 | // and defs a new live range, which is used by this node. |
222 | Node *spill = get_spillcopy_wide(spill_type, def,use,useidx); |
223 | // did we fail to split? |
224 | if (!spill) { |
225 | // Bail |
226 | return 0; |
227 | } |
228 | // insert into basic block |
229 | insert_proj( b, bindex, spill, maxlrg++ ); |
230 | // Use the new split |
231 | use->set_req(useidx,spill); |
232 | } |
233 | // No further split handling needed for this use |
234 | return maxlrg; |
235 | } // End special splitting for debug info live range |
236 | } // If debug info |
237 | |
238 | // CISC-SPILLING |
239 | // Finally, check to see if USE is CISC-Spillable, and if so, |
240 | // gather_lrg_masks will add the flags bit to its mask, and |
241 | // no use side copy is needed. This frees up the live range |
242 | // register choices without causing copy coalescing, etc. |
243 | if( UseCISCSpill && cisc_sp ) { |
244 | int inp = use->cisc_operand(); |
245 | if( inp != AdlcVMDeps::Not_cisc_spillable ) |
246 | // Convert operand number to edge index number |
247 | inp = use->as_Mach()->operand_index(inp); |
248 | if( inp == (int)useidx ) { |
249 | use->set_req(useidx, def); |
250 | #ifndef PRODUCT |
251 | if( TraceCISCSpill ) { |
252 | tty->print(" set_split: " ); |
253 | use->dump(); |
254 | } |
255 | #endif |
256 | return maxlrg; |
257 | } |
258 | } |
259 | |
260 | //------------------------------------------- |
261 | // Insert a Copy before the use |
262 | |
263 | // Block and index where the use occurs. |
264 | int bindex; |
265 | // Phi input spill-copys belong at the end of the prior block |
266 | if( use->is_Phi() ) { |
267 | b = _cfg.get_block_for_node(b->pred(useidx)); |
268 | bindex = b->end_idx(); |
269 | } else { |
270 | // Put the clone just prior to use |
271 | bindex = b->find_node(use); |
272 | } |
273 | |
274 | Node *spill = get_spillcopy_wide(spill_type, def, use, useidx ); |
275 | if( !spill ) return 0; // Bailed out |
276 | // Insert SpillCopy before the USE, which uses the reaching DEF as |
277 | // its input, and defs a new live range, which is used by this node. |
278 | insert_proj( b, bindex, spill, maxlrg++ ); |
279 | // Use the spill/clone |
280 | use->set_req(useidx,spill); |
281 | |
282 | // return updated live range count |
283 | return maxlrg; |
284 | } |
285 | |
286 | //------------------------------clone_node---------------------------- |
287 | // Clone node with anti dependence check. |
288 | Node* clone_node(Node* def, Block *b, Compile* C) { |
289 | if (def->needs_anti_dependence_check()) { |
290 | #ifdef ASSERT |
291 | if (PrintOpto && WizardMode) { |
292 | tty->print_cr("RA attempts to clone node with anti_dependence:" ); |
293 | def->dump(-1); tty->cr(); |
294 | tty->print_cr("into block:" ); |
295 | b->dump(); |
296 | } |
297 | #endif |
298 | if (C->subsume_loads() == true && !C->failing()) { |
299 | // Retry with subsume_loads == false |
300 | // If this is the first failure, the sentinel string will "stick" |
301 | // to the Compile object, and the C2Compiler will see it and retry. |
302 | C->record_failure(C2Compiler::retry_no_subsuming_loads()); |
303 | } else { |
304 | // Bailout without retry |
305 | C->record_method_not_compilable("RA Split failed: attempt to clone node with anti_dependence" ); |
306 | } |
307 | return 0; |
308 | } |
309 | return def->clone(); |
310 | } |
311 | |
312 | //------------------------------split_Rematerialize---------------------------- |
313 | // Clone a local copy of the def. |
314 | Node *PhaseChaitin::split_Rematerialize(Node *def, Block *b, uint insidx, uint &maxlrg, |
315 | GrowableArray<uint> splits, int slidx, uint *lrg2reach, |
316 | Node **Reachblock, bool walkThru) { |
317 | // The input live ranges will be stretched to the site of the new |
318 | // instruction. They might be stretched past a def and will thus |
319 | // have the old and new values of the same live range alive at the |
320 | // same time - a definite no-no. Split out private copies of |
321 | // the inputs. |
322 | if (def->req() > 1) { |
323 | for (uint i = 1; i < def->req(); i++) { |
324 | Node *in = def->in(i); |
325 | uint lidx = _lrg_map.live_range_id(in); |
326 | // We do not need this for live ranges that are only defined once. |
327 | // However, this is not true for spill copies that are added in this |
328 | // Split() pass, since they might get coalesced later on in this pass. |
329 | if (lidx < _lrg_map.max_lrg_id() && lrgs(lidx).is_singledef()) { |
330 | continue; |
331 | } |
332 | |
333 | Block *b_def = _cfg.get_block_for_node(def); |
334 | int idx_def = b_def->find_node(def); |
335 | // Cannot spill Op_RegFlags. |
336 | Node *in_spill; |
337 | if (in->ideal_reg() != Op_RegFlags) { |
338 | in_spill = get_spillcopy_wide(MachSpillCopyNode::InputToRematerialization, in, def, i); |
339 | if (!in_spill) { return 0; } // Bailed out |
340 | insert_proj(b_def, idx_def, in_spill, maxlrg++); |
341 | if (b_def == b) { |
342 | insidx++; |
343 | } |
344 | def->set_req(i, in_spill); |
345 | } else { |
346 | // The 'in' defines a flag register. Flag registers can not be spilled. |
347 | // Register allocation handles live ranges with flag registers |
348 | // by rematerializing the def (in this case 'in'). Thus, this is not |
349 | // critical if the input can be rematerialized, too. |
350 | if (!in->rematerialize()) { |
351 | assert(false, "Can not rematerialize %d: %s. Prolongs RegFlags live" |
352 | " range and defining node %d: %s may not be rematerialized." , |
353 | def->_idx, def->Name(), in->_idx, in->Name()); |
354 | C->record_method_not_compilable("attempted to spill a non-spillable item with RegFlags input" ); |
355 | return 0; // Bailed out |
356 | } |
357 | } |
358 | } |
359 | } |
360 | |
361 | Node *spill = clone_node(def, b, C); |
362 | if (spill == NULL || C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { |
363 | // Check when generating nodes |
364 | return 0; |
365 | } |
366 | |
367 | // See if any inputs are currently being spilled, and take the |
368 | // latest copy of spilled inputs. |
369 | if( spill->req() > 1 ) { |
370 | for( uint i = 1; i < spill->req(); i++ ) { |
371 | Node *in = spill->in(i); |
372 | uint lidx = _lrg_map.find_id(in); |
373 | |
374 | // Walk backwards thru spill copy node intermediates |
375 | if (walkThru) { |
376 | while (in->is_SpillCopy() && lidx >= _lrg_map.max_lrg_id()) { |
377 | in = in->in(1); |
378 | lidx = _lrg_map.find_id(in); |
379 | } |
380 | |
381 | if (lidx < _lrg_map.max_lrg_id() && lrgs(lidx).is_multidef()) { |
382 | // walkThru found a multidef LRG, which is unsafe to use, so |
383 | // just keep the original def used in the clone. |
384 | in = spill->in(i); |
385 | lidx = _lrg_map.find_id(in); |
386 | } |
387 | } |
388 | |
389 | if (lidx < _lrg_map.max_lrg_id() && lrgs(lidx).reg() >= LRG::SPILL_REG) { |
390 | Node *rdef = Reachblock[lrg2reach[lidx]]; |
391 | if (rdef) { |
392 | spill->set_req(i, rdef); |
393 | } |
394 | } |
395 | } |
396 | } |
397 | |
398 | |
399 | assert( spill->out_RegMask().is_UP(), "rematerialize to a reg" ); |
400 | // Rematerialized op is def->spilled+1 |
401 | set_was_spilled(spill); |
402 | if( _spilled_once.test(def->_idx) ) |
403 | set_was_spilled(spill); |
404 | |
405 | insert_proj( b, insidx, spill, maxlrg++ ); |
406 | #ifdef ASSERT |
407 | // Increment the counter for this lrg |
408 | splits.at_put(slidx, splits.at(slidx)+1); |
409 | #endif |
410 | // See if the cloned def kills any flags, and copy those kills as well |
411 | uint i = insidx+1; |
412 | int found_projs = clone_projs( b, i, def, spill, maxlrg); |
413 | if (found_projs > 0) { |
414 | // Adjust the point where we go hi-pressure |
415 | if (i <= b->_ihrp_index) { |
416 | b->_ihrp_index += found_projs; |
417 | } |
418 | if (i <= b->_fhrp_index) { |
419 | b->_fhrp_index += found_projs; |
420 | } |
421 | } |
422 | |
423 | return spill; |
424 | } |
425 | |
426 | //------------------------------is_high_pressure------------------------------- |
427 | // Function to compute whether or not this live range is "high pressure" |
428 | // in this block - whether it spills eagerly or not. |
429 | bool PhaseChaitin::is_high_pressure( Block *b, LRG *lrg, uint insidx ) { |
430 | if( lrg->_was_spilled1 ) return true; |
431 | // Forced spilling due to conflict? Then split only at binding uses |
432 | // or defs, not for supposed capacity problems. |
433 | // CNC - Turned off 7/8/99, causes too much spilling |
434 | // if( lrg->_is_bound ) return false; |
435 | |
436 | // Use float pressure numbers for vectors. |
437 | bool is_float_or_vector = lrg->_is_float || lrg->_is_vector; |
438 | // Not yet reached the high-pressure cutoff point, so low pressure |
439 | uint hrp_idx = is_float_or_vector ? b->_fhrp_index : b->_ihrp_index; |
440 | if( insidx < hrp_idx ) return false; |
441 | // Register pressure for the block as a whole depends on reg class |
442 | int block_pres = is_float_or_vector ? b->_freg_pressure : b->_reg_pressure; |
443 | // Bound live ranges will split at the binding points first; |
444 | // Intermediate splits should assume the live range's register set |
445 | // got "freed up" and that num_regs will become INT_PRESSURE. |
446 | int bound_pres = is_float_or_vector ? FLOATPRESSURE : INTPRESSURE; |
447 | // Effective register pressure limit. |
448 | int lrg_pres = (lrg->get_invalid_mask_size() > lrg->num_regs()) |
449 | ? (lrg->get_invalid_mask_size() >> (lrg->num_regs()-1)) : bound_pres; |
450 | // High pressure if block pressure requires more register freedom |
451 | // than live range has. |
452 | return block_pres >= lrg_pres; |
453 | } |
454 | |
455 | |
456 | //------------------------------prompt_use--------------------------------- |
457 | // True if lidx is used before any real register is def'd in the block |
458 | bool PhaseChaitin::prompt_use( Block *b, uint lidx ) { |
459 | if (lrgs(lidx)._was_spilled2) { |
460 | return false; |
461 | } |
462 | |
463 | // Scan block for 1st use. |
464 | for( uint i = 1; i <= b->end_idx(); i++ ) { |
465 | Node *n = b->get_node(i); |
466 | // Ignore PHI use, these can be up or down |
467 | if (n->is_Phi()) { |
468 | continue; |
469 | } |
470 | for (uint j = 1; j < n->req(); j++) { |
471 | if (_lrg_map.find_id(n->in(j)) == lidx) { |
472 | return true; // Found 1st use! |
473 | } |
474 | } |
475 | if (n->out_RegMask().is_NotEmpty()) { |
476 | return false; |
477 | } |
478 | } |
479 | return false; |
480 | } |
481 | |
482 | //------------------------------Split-------------------------------------- |
483 | //----------Split Routine---------- |
484 | // ***** NEW SPLITTING HEURISTIC ***** |
485 | // DEFS: If the DEF is in a High Register Pressure(HRP) Block, split there. |
486 | // Else, no split unless there is a HRP block between a DEF and |
487 | // one of its uses, and then split at the HRP block. |
488 | // |
489 | // USES: If USE is in HRP, split at use to leave main LRG on stack. |
490 | // Else, hoist LRG back up to register only (ie - split is also DEF) |
491 | // We will compute a new maxlrg as we go |
492 | uint PhaseChaitin::Split(uint maxlrg, ResourceArea* split_arena) { |
493 | Compile::TracePhase tp("regAllocSplit" , &timers[_t_regAllocSplit]); |
494 | |
495 | // Free thread local resources used by this method on exit. |
496 | ResourceMark rm(split_arena); |
497 | |
498 | uint bidx, pidx, slidx, insidx, inpidx, twoidx; |
499 | uint non_phi = 1, spill_cnt = 0; |
500 | Node *n1, *n2, *n3; |
501 | Node_List *defs,*phis; |
502 | bool *UPblock; |
503 | bool u1, u2, u3; |
504 | Block *b, *pred; |
505 | PhiNode *phi; |
506 | GrowableArray<uint> lidxs(split_arena, maxlrg, 0, 0); |
507 | |
508 | // Array of counters to count splits per live range |
509 | GrowableArray<uint> splits(split_arena, maxlrg, 0, 0); |
510 | |
511 | #define NEW_SPLIT_ARRAY(type, size)\ |
512 | (type*) split_arena->allocate_bytes((size) * sizeof(type)) |
513 | |
514 | //----------Setup Code---------- |
515 | // Create a convenient mapping from lrg numbers to reaches/leaves indices |
516 | uint *lrg2reach = NEW_SPLIT_ARRAY(uint, maxlrg); |
517 | // Keep track of DEFS & Phis for later passes |
518 | defs = new Node_List(); |
519 | phis = new Node_List(); |
520 | // Gather info on which LRG's are spilling, and build maps |
521 | for (bidx = 1; bidx < maxlrg; bidx++) { |
522 | if (lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG) { |
523 | assert(!lrgs(bidx).mask().is_AllStack(),"AllStack should color" ); |
524 | lrg2reach[bidx] = spill_cnt; |
525 | spill_cnt++; |
526 | lidxs.append(bidx); |
527 | #ifdef ASSERT |
528 | // Initialize the split counts to zero |
529 | splits.append(0); |
530 | #endif |
531 | if (PrintOpto && WizardMode && lrgs(bidx)._was_spilled1) { |
532 | tty->print_cr("Warning, 2nd spill of L%d" ,bidx); |
533 | } |
534 | } |
535 | } |
536 | |
537 | // Create side arrays for propagating reaching defs info. |
538 | // Each block needs a node pointer for each spilling live range for the |
539 | // Def which is live into the block. Phi nodes handle multiple input |
540 | // Defs by querying the output of their predecessor blocks and resolving |
541 | // them to a single Def at the phi. The pointer is updated for each |
542 | // Def in the block, and then becomes the output for the block when |
543 | // processing of the block is complete. We also need to track whether |
544 | // a Def is UP or DOWN. UP means that it should get a register (ie - |
545 | // it is always in LRP regions), and DOWN means that it is probably |
546 | // on the stack (ie - it crosses HRP regions). |
547 | Node ***Reaches = NEW_SPLIT_ARRAY( Node**, _cfg.number_of_blocks() + 1); |
548 | bool **UP = NEW_SPLIT_ARRAY( bool*, _cfg.number_of_blocks() + 1); |
549 | Node **debug_defs = NEW_SPLIT_ARRAY( Node*, spill_cnt ); |
550 | VectorSet **UP_entry= NEW_SPLIT_ARRAY( VectorSet*, spill_cnt ); |
551 | |
552 | // Initialize Reaches & UP |
553 | for (bidx = 0; bidx < _cfg.number_of_blocks() + 1; bidx++) { |
554 | Reaches[bidx] = NEW_SPLIT_ARRAY( Node*, spill_cnt ); |
555 | UP[bidx] = NEW_SPLIT_ARRAY( bool, spill_cnt ); |
556 | Node **Reachblock = Reaches[bidx]; |
557 | bool *UPblock = UP[bidx]; |
558 | for( slidx = 0; slidx < spill_cnt; slidx++ ) { |
559 | UPblock[slidx] = true; // Assume they start in registers |
560 | Reachblock[slidx] = NULL; // Assume that no def is present |
561 | } |
562 | } |
563 | |
564 | #undef NEW_SPLIT_ARRAY |
565 | |
566 | // Initialize to array of empty vectorsets |
567 | for( slidx = 0; slidx < spill_cnt; slidx++ ) |
568 | UP_entry[slidx] = new VectorSet(split_arena); |
569 | |
570 | //----------PASS 1---------- |
571 | //----------Propagation & Node Insertion Code---------- |
572 | // Walk the Blocks in RPO for DEF & USE info |
573 | for( bidx = 0; bidx < _cfg.number_of_blocks(); bidx++ ) { |
574 | |
575 | if (C->check_node_count(spill_cnt, out_of_nodes)) { |
576 | return 0; |
577 | } |
578 | |
579 | b = _cfg.get_block(bidx); |
580 | // Reaches & UP arrays for this block |
581 | Node** Reachblock = Reaches[b->_pre_order]; |
582 | UPblock = UP[b->_pre_order]; |
583 | // Reset counter of start of non-Phi nodes in block |
584 | non_phi = 1; |
585 | //----------Block Entry Handling---------- |
586 | // Check for need to insert a new phi |
587 | // Cycle through this block's predecessors, collecting Reaches |
588 | // info for each spilled LRG. If they are identical, no phi is |
589 | // needed. If they differ, check for a phi, and insert if missing, |
590 | // or update edges if present. Set current block's Reaches set to |
591 | // be either the phi's or the reaching def, as appropriate. |
592 | // If no Phi is needed, check if the LRG needs to spill on entry |
593 | // to the block due to HRP. |
594 | for( slidx = 0; slidx < spill_cnt; slidx++ ) { |
595 | // Grab the live range number |
596 | uint lidx = lidxs.at(slidx); |
597 | // Do not bother splitting or putting in Phis for single-def |
598 | // rematerialized live ranges. This happens alot to constants |
599 | // with long live ranges. |
600 | if( lrgs(lidx).is_singledef() && |
601 | lrgs(lidx)._def->rematerialize() ) { |
602 | // reset the Reaches & UP entries |
603 | Reachblock[slidx] = lrgs(lidx)._def; |
604 | UPblock[slidx] = true; |
605 | // Record following instruction in case 'n' rematerializes and |
606 | // kills flags |
607 | Block *pred1 = _cfg.get_block_for_node(b->pred(1)); |
608 | continue; |
609 | } |
610 | |
611 | // Initialize needs_phi and needs_split |
612 | bool needs_phi = false; |
613 | bool needs_split = false; |
614 | bool has_phi = false; |
615 | // Walk the predecessor blocks to check inputs for that live range |
616 | // Grab predecessor block header |
617 | n1 = b->pred(1); |
618 | // Grab the appropriate reaching def info for inpidx |
619 | pred = _cfg.get_block_for_node(n1); |
620 | pidx = pred->_pre_order; |
621 | Node **Ltmp = Reaches[pidx]; |
622 | bool *Utmp = UP[pidx]; |
623 | n1 = Ltmp[slidx]; |
624 | u1 = Utmp[slidx]; |
625 | // Initialize node for saving type info |
626 | n3 = n1; |
627 | u3 = u1; |
628 | |
629 | // Compare inputs to see if a Phi is needed |
630 | for( inpidx = 2; inpidx < b->num_preds(); inpidx++ ) { |
631 | // Grab predecessor block headers |
632 | n2 = b->pred(inpidx); |
633 | // Grab the appropriate reaching def info for inpidx |
634 | pred = _cfg.get_block_for_node(n2); |
635 | pidx = pred->_pre_order; |
636 | Ltmp = Reaches[pidx]; |
637 | Utmp = UP[pidx]; |
638 | n2 = Ltmp[slidx]; |
639 | u2 = Utmp[slidx]; |
640 | // For each LRG, decide if a phi is necessary |
641 | if( n1 != n2 ) { |
642 | needs_phi = true; |
643 | } |
644 | // See if the phi has mismatched inputs, UP vs. DOWN |
645 | if( n1 && n2 && (u1 != u2) ) { |
646 | needs_split = true; |
647 | } |
648 | // Move n2/u2 to n1/u1 for next iteration |
649 | n1 = n2; |
650 | u1 = u2; |
651 | // Preserve a non-NULL predecessor for later type referencing |
652 | if( (n3 == NULL) && (n2 != NULL) ){ |
653 | n3 = n2; |
654 | u3 = u2; |
655 | } |
656 | } // End for all potential Phi inputs |
657 | |
658 | // check block for appropriate phinode & update edges |
659 | for( insidx = 1; insidx <= b->end_idx(); insidx++ ) { |
660 | n1 = b->get_node(insidx); |
661 | // bail if this is not a phi |
662 | phi = n1->is_Phi() ? n1->as_Phi() : NULL; |
663 | if( phi == NULL ) { |
664 | // Keep track of index of first non-PhiNode instruction in block |
665 | non_phi = insidx; |
666 | // break out of the for loop as we have handled all phi nodes |
667 | break; |
668 | } |
669 | // must be looking at a phi |
670 | if (_lrg_map.find_id(n1) == lidxs.at(slidx)) { |
671 | // found the necessary phi |
672 | needs_phi = false; |
673 | has_phi = true; |
674 | // initialize the Reaches entry for this LRG |
675 | Reachblock[slidx] = phi; |
676 | break; |
677 | } // end if found correct phi |
678 | } // end for all phi's |
679 | |
680 | // If a phi is needed or exist, check for it |
681 | if( needs_phi || has_phi ) { |
682 | // add new phinode if one not already found |
683 | if( needs_phi ) { |
684 | // create a new phi node and insert it into the block |
685 | // type is taken from left over pointer to a predecessor |
686 | guarantee(n3, "No non-NULL reaching DEF for a Phi" ); |
687 | phi = new PhiNode(b->head(), n3->bottom_type()); |
688 | // initialize the Reaches entry for this LRG |
689 | Reachblock[slidx] = phi; |
690 | |
691 | // add node to block & node_to_block mapping |
692 | insert_proj(b, insidx++, phi, maxlrg++); |
693 | non_phi++; |
694 | // Reset new phi's mapping to be the spilling live range |
695 | _lrg_map.map(phi->_idx, lidx); |
696 | assert(_lrg_map.find_id(phi) == lidx, "Bad update on Union-Find mapping" ); |
697 | } // end if not found correct phi |
698 | // Here you have either found or created the Phi, so record it |
699 | assert(phi != NULL,"Must have a Phi Node here" ); |
700 | phis->push(phi); |
701 | // PhiNodes should either force the LRG UP or DOWN depending |
702 | // on its inputs and the register pressure in the Phi's block. |
703 | UPblock[slidx] = true; // Assume new DEF is UP |
704 | // If entering a high-pressure area with no immediate use, |
705 | // assume Phi is DOWN |
706 | if( is_high_pressure( b, &lrgs(lidx), b->end_idx()) && !prompt_use(b,lidx) ) |
707 | UPblock[slidx] = false; |
708 | // If we are not split up/down and all inputs are down, then we |
709 | // are down |
710 | if( !needs_split && !u3 ) |
711 | UPblock[slidx] = false; |
712 | } // end if phi is needed |
713 | |
714 | // Do not need a phi, so grab the reaching DEF |
715 | else { |
716 | // Grab predecessor block header |
717 | n1 = b->pred(1); |
718 | // Grab the appropriate reaching def info for k |
719 | pred = _cfg.get_block_for_node(n1); |
720 | pidx = pred->_pre_order; |
721 | Node **Ltmp = Reaches[pidx]; |
722 | bool *Utmp = UP[pidx]; |
723 | // reset the Reaches & UP entries |
724 | Reachblock[slidx] = Ltmp[slidx]; |
725 | UPblock[slidx] = Utmp[slidx]; |
726 | } // end else no Phi is needed |
727 | } // end for all spilling live ranges |
728 | // DEBUG |
729 | #ifndef PRODUCT |
730 | if(trace_spilling()) { |
731 | tty->print("/`\nBlock %d: " , b->_pre_order); |
732 | tty->print("Reaching Definitions after Phi handling\n" ); |
733 | for( uint x = 0; x < spill_cnt; x++ ) { |
734 | tty->print("Spill Idx %d: UP %d: Node\n" ,x,UPblock[x]); |
735 | if( Reachblock[x] ) |
736 | Reachblock[x]->dump(); |
737 | else |
738 | tty->print("Undefined\n" ); |
739 | } |
740 | } |
741 | #endif |
742 | |
743 | //----------Non-Phi Node Splitting---------- |
744 | // Since phi-nodes have now been handled, the Reachblock array for this |
745 | // block is initialized with the correct starting value for the defs which |
746 | // reach non-phi instructions in this block. Thus, process non-phi |
747 | // instructions normally, inserting SpillCopy nodes for all spill |
748 | // locations. |
749 | |
750 | // Memoize any DOWN reaching definitions for use as DEBUG info |
751 | for( insidx = 0; insidx < spill_cnt; insidx++ ) { |
752 | debug_defs[insidx] = (UPblock[insidx]) ? NULL : Reachblock[insidx]; |
753 | if( UPblock[insidx] ) // Memoize UP decision at block start |
754 | UP_entry[insidx]->set( b->_pre_order ); |
755 | } |
756 | |
757 | //----------Walk Instructions in the Block and Split---------- |
758 | // For all non-phi instructions in the block |
759 | for( insidx = 1; insidx <= b->end_idx(); insidx++ ) { |
760 | Node *n = b->get_node(insidx); |
761 | // Find the defining Node's live range index |
762 | uint defidx = _lrg_map.find_id(n); |
763 | uint cnt = n->req(); |
764 | |
765 | if (n->is_Phi()) { |
766 | // Skip phi nodes after removing dead copies. |
767 | if (defidx < _lrg_map.max_lrg_id()) { |
768 | // Check for useless Phis. These appear if we spill, then |
769 | // coalesce away copies. Dont touch Phis in spilling live |
770 | // ranges; they are busy getting modifed in this pass. |
771 | if( lrgs(defidx).reg() < LRG::SPILL_REG ) { |
772 | uint i; |
773 | Node *u = NULL; |
774 | // Look for the Phi merging 2 unique inputs |
775 | for( i = 1; i < cnt; i++ ) { |
776 | // Ignore repeats and self |
777 | if( n->in(i) != u && n->in(i) != n ) { |
778 | // Found a unique input |
779 | if( u != NULL ) // If it's the 2nd, bail out |
780 | break; |
781 | u = n->in(i); // Else record it |
782 | } |
783 | } |
784 | assert( u, "at least 1 valid input expected" ); |
785 | if (i >= cnt) { // Found one unique input |
786 | assert(_lrg_map.find_id(n) == _lrg_map.find_id(u), "should be the same lrg" ); |
787 | n->replace_by(u); // Then replace with unique input |
788 | n->disconnect_inputs(NULL, C); |
789 | b->remove_node(insidx); |
790 | insidx--; |
791 | b->_ihrp_index--; |
792 | b->_fhrp_index--; |
793 | } |
794 | } |
795 | } |
796 | continue; |
797 | } |
798 | assert( insidx > b->_ihrp_index || |
799 | (b->_reg_pressure < (uint)INTPRESSURE) || |
800 | b->_ihrp_index > 4000000 || |
801 | b->_ihrp_index >= b->end_idx() || |
802 | !b->get_node(b->_ihrp_index)->is_Proj(), "" ); |
803 | assert( insidx > b->_fhrp_index || |
804 | (b->_freg_pressure < (uint)FLOATPRESSURE) || |
805 | b->_fhrp_index > 4000000 || |
806 | b->_fhrp_index >= b->end_idx() || |
807 | !b->get_node(b->_fhrp_index)->is_Proj(), "" ); |
808 | |
809 | // ********** Handle Crossing HRP Boundry ********** |
810 | if( (insidx == b->_ihrp_index) || (insidx == b->_fhrp_index) ) { |
811 | for( slidx = 0; slidx < spill_cnt; slidx++ ) { |
812 | // Check for need to split at HRP boundary - split if UP |
813 | n1 = Reachblock[slidx]; |
814 | // bail out if no reaching DEF |
815 | if( n1 == NULL ) continue; |
816 | // bail out if live range is 'isolated' around inner loop |
817 | uint lidx = lidxs.at(slidx); |
818 | // If live range is currently UP |
819 | if( UPblock[slidx] ) { |
820 | // set location to insert spills at |
821 | // SPLIT DOWN HERE - NO CISC SPILL |
822 | if( is_high_pressure( b, &lrgs(lidx), insidx ) && |
823 | !n1->rematerialize() ) { |
824 | // If there is already a valid stack definition available, use it |
825 | if( debug_defs[slidx] != NULL ) { |
826 | Reachblock[slidx] = debug_defs[slidx]; |
827 | } |
828 | else { |
829 | // Insert point is just past last use or def in the block |
830 | int insert_point = insidx-1; |
831 | while( insert_point > 0 ) { |
832 | Node *n = b->get_node(insert_point); |
833 | // Hit top of block? Quit going backwards |
834 | if (n->is_Phi()) { |
835 | break; |
836 | } |
837 | // Found a def? Better split after it. |
838 | if (_lrg_map.live_range_id(n) == lidx) { |
839 | break; |
840 | } |
841 | // Look for a use |
842 | uint i; |
843 | for( i = 1; i < n->req(); i++ ) { |
844 | if (_lrg_map.live_range_id(n->in(i)) == lidx) { |
845 | break; |
846 | } |
847 | } |
848 | // Found a use? Better split after it. |
849 | if (i < n->req()) { |
850 | break; |
851 | } |
852 | insert_point--; |
853 | } |
854 | uint orig_eidx = b->end_idx(); |
855 | maxlrg = split_DEF( n1, b, insert_point, maxlrg, Reachblock, debug_defs, splits, slidx); |
856 | // If it wasn't split bail |
857 | if (!maxlrg) { |
858 | return 0; |
859 | } |
860 | // Spill of NULL check mem op goes into the following block. |
861 | if (b->end_idx() > orig_eidx) { |
862 | insidx++; |
863 | } |
864 | } |
865 | // This is a new DEF, so update UP |
866 | UPblock[slidx] = false; |
867 | #ifndef PRODUCT |
868 | // DEBUG |
869 | if( trace_spilling() ) { |
870 | tty->print("\nNew Split DOWN DEF of Spill Idx " ); |
871 | tty->print("%d, UP %d:\n" ,slidx,false); |
872 | n1->dump(); |
873 | } |
874 | #endif |
875 | } |
876 | } // end if LRG is UP |
877 | } // end for all spilling live ranges |
878 | assert( b->get_node(insidx) == n, "got insidx set incorrectly" ); |
879 | } // end if crossing HRP Boundry |
880 | |
881 | // If the LRG index is oob, then this is a new spillcopy, skip it. |
882 | if (defidx >= _lrg_map.max_lrg_id()) { |
883 | continue; |
884 | } |
885 | LRG &deflrg = lrgs(defidx); |
886 | uint copyidx = n->is_Copy(); |
887 | // Remove coalesced copy from CFG |
888 | if (copyidx && defidx == _lrg_map.live_range_id(n->in(copyidx))) { |
889 | n->replace_by( n->in(copyidx) ); |
890 | n->set_req( copyidx, NULL ); |
891 | b->remove_node(insidx--); |
892 | b->_ihrp_index--; // Adjust the point where we go hi-pressure |
893 | b->_fhrp_index--; |
894 | continue; |
895 | } |
896 | |
897 | #define DERIVED 0 |
898 | |
899 | // ********** Handle USES ********** |
900 | bool nullcheck = false; |
901 | // Implicit null checks never use the spilled value |
902 | if( n->is_MachNullCheck() ) |
903 | nullcheck = true; |
904 | if( !nullcheck ) { |
905 | // Search all inputs for a Spill-USE |
906 | JVMState* jvms = n->jvms(); |
907 | uint oopoff = jvms ? jvms->oopoff() : cnt; |
908 | uint old_last = cnt - 1; |
909 | for( inpidx = 1; inpidx < cnt; inpidx++ ) { |
910 | // Derived/base pairs may be added to our inputs during this loop. |
911 | // If inpidx > old_last, then one of these new inputs is being |
912 | // handled. Skip the derived part of the pair, but process |
913 | // the base like any other input. |
914 | if (inpidx > old_last && ((inpidx - oopoff) & 1) == DERIVED) { |
915 | continue; // skip derived_debug added below |
916 | } |
917 | // Get lidx of input |
918 | uint useidx = _lrg_map.find_id(n->in(inpidx)); |
919 | // Not a brand-new split, and it is a spill use |
920 | if (useidx < _lrg_map.max_lrg_id() && lrgs(useidx).reg() >= LRG::SPILL_REG) { |
921 | // Check for valid reaching DEF |
922 | slidx = lrg2reach[useidx]; |
923 | Node *def = Reachblock[slidx]; |
924 | assert( def != NULL, "Using Undefined Value in Split()\n" ); |
925 | |
926 | // (+++) %%%% remove this in favor of pre-pass in matcher.cpp |
927 | // monitor references do not care where they live, so just hook |
928 | if ( jvms && jvms->is_monitor_use(inpidx) ) { |
929 | // The effect of this clone is to drop the node out of the block, |
930 | // so that the allocator does not see it anymore, and therefore |
931 | // does not attempt to assign it a register. |
932 | def = clone_node(def, b, C); |
933 | if (def == NULL || C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { |
934 | return 0; |
935 | } |
936 | _lrg_map.extend(def->_idx, 0); |
937 | _cfg.map_node_to_block(def, b); |
938 | n->set_req(inpidx, def); |
939 | continue; |
940 | } |
941 | |
942 | // Rematerializable? Then clone def at use site instead |
943 | // of store/load |
944 | if( def->rematerialize() ) { |
945 | int old_size = b->number_of_nodes(); |
946 | def = split_Rematerialize( def, b, insidx, maxlrg, splits, slidx, lrg2reach, Reachblock, true ); |
947 | if( !def ) return 0; // Bail out |
948 | insidx += b->number_of_nodes()-old_size; |
949 | } |
950 | |
951 | MachNode *mach = n->is_Mach() ? n->as_Mach() : NULL; |
952 | // Base pointers and oopmap references do not care where they live. |
953 | if ((inpidx >= oopoff) || |
954 | (mach && mach->ideal_Opcode() == Op_AddP && inpidx == AddPNode::Base)) { |
955 | if (def->rematerialize() && lrgs(useidx)._was_spilled2) { |
956 | // This def has been rematerialized a couple of times without |
957 | // progress. It doesn't care if it lives UP or DOWN, so |
958 | // spill it down now. |
959 | maxlrg = split_USE(MachSpillCopyNode::BasePointerToMem, def,b,n,inpidx,maxlrg,false,false,splits,slidx); |
960 | // If it wasn't split bail |
961 | if (!maxlrg) { |
962 | return 0; |
963 | } |
964 | insidx++; // Reset iterator to skip USE side split |
965 | } else { |
966 | // Just hook the def edge |
967 | n->set_req(inpidx, def); |
968 | } |
969 | |
970 | if (inpidx >= oopoff) { |
971 | // After oopoff, we have derived/base pairs. We must mention all |
972 | // derived pointers here as derived/base pairs for GC. If the |
973 | // derived value is spilling and we have a copy both in Reachblock |
974 | // (called here 'def') and debug_defs[slidx] we need to mention |
975 | // both in derived/base pairs or kill one. |
976 | Node *derived_debug = debug_defs[slidx]; |
977 | if( ((inpidx - oopoff) & 1) == DERIVED && // derived vs base? |
978 | mach && mach->ideal_Opcode() != Op_Halt && |
979 | derived_debug != NULL && |
980 | derived_debug != def ) { // Actual 2nd value appears |
981 | // We have already set 'def' as a derived value. |
982 | // Also set debug_defs[slidx] as a derived value. |
983 | uint k; |
984 | for( k = oopoff; k < cnt; k += 2 ) |
985 | if( n->in(k) == derived_debug ) |
986 | break; // Found an instance of debug derived |
987 | if( k == cnt ) {// No instance of debug_defs[slidx] |
988 | // Add a derived/base pair to cover the debug info. |
989 | // We have to process the added base later since it is not |
990 | // handled yet at this point but skip derived part. |
991 | assert(((n->req() - oopoff) & 1) == DERIVED, |
992 | "must match skip condition above" ); |
993 | n->add_req( derived_debug ); // this will be skipped above |
994 | n->add_req( n->in(inpidx+1) ); // this will be processed |
995 | // Increment cnt to handle added input edges on |
996 | // subsequent iterations. |
997 | cnt += 2; |
998 | } |
999 | } |
1000 | } |
1001 | continue; |
1002 | } |
1003 | // Special logic for DEBUG info |
1004 | if( jvms && b->_freq > BLOCK_FREQUENCY(0.5) ) { |
1005 | uint debug_start = jvms->debug_start(); |
1006 | // If this is debug info use & there is a reaching DOWN def |
1007 | if ((debug_start <= inpidx) && (debug_defs[slidx] != NULL)) { |
1008 | assert(inpidx < oopoff, "handle only debug info here" ); |
1009 | // Just hook it in & move on |
1010 | n->set_req(inpidx, debug_defs[slidx]); |
1011 | // (Note that this can make two sides of a split live at the |
1012 | // same time: The debug def on stack, and another def in a |
1013 | // register. The GC needs to know about both of them, but any |
1014 | // derived pointers after oopoff will refer to only one of the |
1015 | // two defs and the GC would therefore miss the other. Thus |
1016 | // this hack is only allowed for debug info which is Java state |
1017 | // and therefore never a derived pointer.) |
1018 | continue; |
1019 | } |
1020 | } |
1021 | // Grab register mask info |
1022 | const RegMask &dmask = def->out_RegMask(); |
1023 | const RegMask &umask = n->in_RegMask(inpidx); |
1024 | bool is_vect = RegMask::is_vector(def->ideal_reg()); |
1025 | assert(inpidx < oopoff, "cannot use-split oop map info" ); |
1026 | |
1027 | bool dup = UPblock[slidx]; |
1028 | bool uup = umask.is_UP(); |
1029 | |
1030 | // Need special logic to handle bound USES. Insert a split at this |
1031 | // bound use if we can't rematerialize the def, or if we need the |
1032 | // split to form a misaligned pair. |
1033 | if( !umask.is_AllStack() && |
1034 | (int)umask.Size() <= lrgs(useidx).num_regs() && |
1035 | (!def->rematerialize() || |
1036 | (!is_vect && umask.is_misaligned_pair()))) { |
1037 | // These need a Split regardless of overlap or pressure |
1038 | // SPLIT - NO DEF - NO CISC SPILL |
1039 | maxlrg = split_USE(MachSpillCopyNode::Bound, def,b,n,inpidx,maxlrg,dup,false, splits,slidx); |
1040 | // If it wasn't split bail |
1041 | if (!maxlrg) { |
1042 | return 0; |
1043 | } |
1044 | insidx++; // Reset iterator to skip USE side split |
1045 | continue; |
1046 | } |
1047 | |
1048 | if (UseFPUForSpilling && n->is_MachCall() && !uup && !dup ) { |
1049 | // The use at the call can force the def down so insert |
1050 | // a split before the use to allow the def more freedom. |
1051 | maxlrg = split_USE(MachSpillCopyNode::CallUse, def,b,n,inpidx,maxlrg,dup,false, splits,slidx); |
1052 | // If it wasn't split bail |
1053 | if (!maxlrg) { |
1054 | return 0; |
1055 | } |
1056 | insidx++; // Reset iterator to skip USE side split |
1057 | continue; |
1058 | } |
1059 | |
1060 | // Here is the logic chart which describes USE Splitting: |
1061 | // 0 = false or DOWN, 1 = true or UP |
1062 | // |
1063 | // Overlap | DEF | USE | Action |
1064 | //------------------------------------------------------- |
1065 | // 0 | 0 | 0 | Copy - mem -> mem |
1066 | // 0 | 0 | 1 | Split-UP - Check HRP |
1067 | // 0 | 1 | 0 | Split-DOWN - Debug Info? |
1068 | // 0 | 1 | 1 | Copy - reg -> reg |
1069 | // 1 | 0 | 0 | Reset Input Edge (no Split) |
1070 | // 1 | 0 | 1 | Split-UP - Check HRP |
1071 | // 1 | 1 | 0 | Split-DOWN - Debug Info? |
1072 | // 1 | 1 | 1 | Reset Input Edge (no Split) |
1073 | // |
1074 | // So, if (dup == uup), then overlap test determines action, |
1075 | // with true being no split, and false being copy. Else, |
1076 | // if DEF is DOWN, Split-UP, and check HRP to decide on |
1077 | // resetting DEF. Finally if DEF is UP, Split-DOWN, with |
1078 | // special handling for Debug Info. |
1079 | if( dup == uup ) { |
1080 | if( dmask.overlap(umask) ) { |
1081 | // Both are either up or down, and there is overlap, No Split |
1082 | n->set_req(inpidx, def); |
1083 | } |
1084 | else { // Both are either up or down, and there is no overlap |
1085 | if( dup ) { // If UP, reg->reg copy |
1086 | // COPY ACROSS HERE - NO DEF - NO CISC SPILL |
1087 | maxlrg = split_USE(MachSpillCopyNode::RegToReg, def,b,n,inpidx,maxlrg,false,false, 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 | } |
1094 | else { // DOWN, mem->mem copy |
1095 | // COPY UP & DOWN HERE - NO DEF - NO CISC SPILL |
1096 | // First Split-UP to move value into Register |
1097 | uint def_ideal = def->ideal_reg(); |
1098 | const RegMask* tmp_rm = Matcher::idealreg2regmask[def_ideal]; |
1099 | Node *spill = new MachSpillCopyNode(MachSpillCopyNode::MemToReg, def, dmask, *tmp_rm); |
1100 | insert_proj( b, insidx, spill, maxlrg ); |
1101 | // Then Split-DOWN as if previous Split was DEF |
1102 | maxlrg = split_USE(MachSpillCopyNode::RegToMem, spill,b,n,inpidx,maxlrg,false,false, splits,slidx); |
1103 | // If it wasn't split bail |
1104 | if (!maxlrg) { |
1105 | return 0; |
1106 | } |
1107 | insidx += 2; // Reset iterator to skip USE side splits |
1108 | } |
1109 | } // End else no overlap |
1110 | } // End if dup == uup |
1111 | // dup != uup, so check dup for direction of Split |
1112 | else { |
1113 | if( dup ) { // If UP, Split-DOWN and check Debug Info |
1114 | // If this node is already a SpillCopy, just patch the edge |
1115 | // except the case of spilling to stack. |
1116 | if( n->is_SpillCopy() ) { |
1117 | RegMask tmp_rm(umask); |
1118 | tmp_rm.SUBTRACT(Matcher::STACK_ONLY_mask); |
1119 | if( dmask.overlap(tmp_rm) ) { |
1120 | if( def != n->in(inpidx) ) { |
1121 | n->set_req(inpidx, def); |
1122 | } |
1123 | continue; |
1124 | } |
1125 | } |
1126 | // COPY DOWN HERE - NO DEF - NO CISC SPILL |
1127 | maxlrg = split_USE(MachSpillCopyNode::RegToMem, def,b,n,inpidx,maxlrg,false,false, splits,slidx); |
1128 | // If it wasn't split bail |
1129 | if (!maxlrg) { |
1130 | return 0; |
1131 | } |
1132 | insidx++; // Reset iterator to skip USE side split |
1133 | // Check for debug-info split. Capture it for later |
1134 | // debug splits of the same value |
1135 | if (jvms && jvms->debug_start() <= inpidx && inpidx < oopoff) |
1136 | debug_defs[slidx] = n->in(inpidx); |
1137 | |
1138 | } |
1139 | else { // DOWN, Split-UP and check register pressure |
1140 | if( is_high_pressure( b, &lrgs(useidx), insidx ) ) { |
1141 | // COPY UP HERE - NO DEF - CISC SPILL |
1142 | maxlrg = split_USE(MachSpillCopyNode::MemToReg, def,b,n,inpidx,maxlrg,true,true, splits,slidx); |
1143 | // If it wasn't split bail |
1144 | if (!maxlrg) { |
1145 | return 0; |
1146 | } |
1147 | insidx++; // Reset iterator to skip USE side split |
1148 | } else { // LRP |
1149 | // COPY UP HERE - WITH DEF - NO CISC SPILL |
1150 | maxlrg = split_USE(MachSpillCopyNode::MemToReg, def,b,n,inpidx,maxlrg,true,false, splits,slidx); |
1151 | // If it wasn't split bail |
1152 | if (!maxlrg) { |
1153 | return 0; |
1154 | } |
1155 | // Flag this lift-up in a low-pressure block as |
1156 | // already-spilled, so if it spills again it will |
1157 | // spill hard (instead of not spilling hard and |
1158 | // coalescing away). |
1159 | set_was_spilled(n->in(inpidx)); |
1160 | // Since this is a new DEF, update Reachblock & UP |
1161 | Reachblock[slidx] = n->in(inpidx); |
1162 | UPblock[slidx] = true; |
1163 | insidx++; // Reset iterator to skip USE side split |
1164 | } |
1165 | } // End else DOWN |
1166 | } // End dup != uup |
1167 | } // End if Spill USE |
1168 | } // End For All Inputs |
1169 | } // End If not nullcheck |
1170 | |
1171 | // ********** Handle DEFS ********** |
1172 | // DEFS either Split DOWN in HRP regions or when the LRG is bound, or |
1173 | // just reset the Reaches info in LRP regions. DEFS must always update |
1174 | // UP info. |
1175 | if( deflrg.reg() >= LRG::SPILL_REG ) { // Spilled? |
1176 | uint slidx = lrg2reach[defidx]; |
1177 | // Add to defs list for later assignment of new live range number |
1178 | defs->push(n); |
1179 | // Set a flag on the Node indicating it has already spilled. |
1180 | // Only do it for capacity spills not conflict spills. |
1181 | if( !deflrg._direct_conflict ) |
1182 | set_was_spilled(n); |
1183 | assert(!n->is_Phi(),"Cannot insert Phi into DEFS list" ); |
1184 | // Grab UP info for DEF |
1185 | const RegMask &dmask = n->out_RegMask(); |
1186 | bool defup = dmask.is_UP(); |
1187 | uint ireg = n->ideal_reg(); |
1188 | bool is_vect = RegMask::is_vector(ireg); |
1189 | // Only split at Def if this is a HRP block or bound (and spilled once) |
1190 | if( !n->rematerialize() && |
1191 | (((dmask.is_bound(ireg) || (!is_vect && dmask.is_misaligned_pair())) && |
1192 | (deflrg._direct_conflict || deflrg._must_spill)) || |
1193 | // Check for LRG being up in a register and we are inside a high |
1194 | // pressure area. Spill it down immediately. |
1195 | (defup && is_high_pressure(b,&deflrg,insidx) && !n->is_SpillCopy())) ) { |
1196 | assert( !n->rematerialize(), "" ); |
1197 | // Do a split at the def site. |
1198 | maxlrg = split_DEF( n, b, insidx, maxlrg, Reachblock, debug_defs, splits, slidx ); |
1199 | // If it wasn't split bail |
1200 | if (!maxlrg) { |
1201 | return 0; |
1202 | } |
1203 | // Split DEF's Down |
1204 | UPblock[slidx] = 0; |
1205 | #ifndef PRODUCT |
1206 | // DEBUG |
1207 | if( trace_spilling() ) { |
1208 | tty->print("\nNew Split DOWN DEF of Spill Idx " ); |
1209 | tty->print("%d, UP %d:\n" ,slidx,false); |
1210 | n->dump(); |
1211 | } |
1212 | #endif |
1213 | } |
1214 | else { // Neither bound nor HRP, must be LRP |
1215 | // otherwise, just record the def |
1216 | Reachblock[slidx] = n; |
1217 | // UP should come from the outRegmask() of the DEF |
1218 | UPblock[slidx] = defup; |
1219 | // Update debug list of reaching down definitions, kill if DEF is UP |
1220 | debug_defs[slidx] = defup ? NULL : n; |
1221 | #ifndef PRODUCT |
1222 | // DEBUG |
1223 | if( trace_spilling() ) { |
1224 | tty->print("\nNew DEF of Spill Idx " ); |
1225 | tty->print("%d, UP %d:\n" ,slidx,defup); |
1226 | n->dump(); |
1227 | } |
1228 | #endif |
1229 | } // End else LRP |
1230 | } // End if spill def |
1231 | |
1232 | // ********** Split Left Over Mem-Mem Moves ********** |
1233 | // Check for mem-mem copies and split them now. Do not do this |
1234 | // to copies about to be spilled; they will be Split shortly. |
1235 | if (copyidx) { |
1236 | Node *use = n->in(copyidx); |
1237 | uint useidx = _lrg_map.find_id(use); |
1238 | if (useidx < _lrg_map.max_lrg_id() && // This is not a new split |
1239 | OptoReg::is_stack(deflrg.reg()) && |
1240 | deflrg.reg() < LRG::SPILL_REG ) { // And DEF is from stack |
1241 | LRG &uselrg = lrgs(useidx); |
1242 | if( OptoReg::is_stack(uselrg.reg()) && |
1243 | uselrg.reg() < LRG::SPILL_REG && // USE is from stack |
1244 | deflrg.reg() != uselrg.reg() ) { // Not trivially removed |
1245 | uint def_ideal_reg = n->bottom_type()->ideal_reg(); |
1246 | const RegMask &def_rm = *Matcher::idealreg2regmask[def_ideal_reg]; |
1247 | const RegMask &use_rm = n->in_RegMask(copyidx); |
1248 | if( def_rm.overlap(use_rm) && n->is_SpillCopy() ) { // Bug 4707800, 'n' may be a storeSSL |
1249 | if (C->check_node_count(NodeLimitFudgeFactor, out_of_nodes)) { // Check when generating nodes |
1250 | return 0; |
1251 | } |
1252 | Node *spill = new MachSpillCopyNode(MachSpillCopyNode::MemToReg, use,use_rm,def_rm); |
1253 | n->set_req(copyidx,spill); |
1254 | n->as_MachSpillCopy()->set_in_RegMask(def_rm); |
1255 | // Put the spill just before the copy |
1256 | insert_proj( b, insidx++, spill, maxlrg++ ); |
1257 | } |
1258 | } |
1259 | } |
1260 | } |
1261 | } // End For All Instructions in Block - Non-PHI Pass |
1262 | |
1263 | // Check if each LRG is live out of this block so as not to propagate |
1264 | // beyond the last use of a LRG. |
1265 | for( slidx = 0; slidx < spill_cnt; slidx++ ) { |
1266 | uint defidx = lidxs.at(slidx); |
1267 | IndexSet *liveout = _live->live(b); |
1268 | if( !liveout->member(defidx) ) { |
1269 | #ifdef ASSERT |
1270 | // The index defidx is not live. Check the liveout array to ensure that |
1271 | // it contains no members which compress to defidx. Finding such an |
1272 | // instance may be a case to add liveout adjustment in compress_uf_map(). |
1273 | // See 5063219. |
1274 | uint member; |
1275 | IndexSetIterator isi(liveout); |
1276 | while ((member = isi.next()) != 0) { |
1277 | assert(defidx != _lrg_map.find_const(member), "Live out member has not been compressed" ); |
1278 | } |
1279 | #endif |
1280 | Reachblock[slidx] = NULL; |
1281 | } else { |
1282 | assert(Reachblock[slidx] != NULL,"No reaching definition for liveout value" ); |
1283 | } |
1284 | } |
1285 | #ifndef PRODUCT |
1286 | if( trace_spilling() ) |
1287 | b->dump(); |
1288 | #endif |
1289 | } // End For All Blocks |
1290 | |
1291 | //----------PASS 2---------- |
1292 | // Reset all DEF live range numbers here |
1293 | for( insidx = 0; insidx < defs->size(); insidx++ ) { |
1294 | // Grab the def |
1295 | n1 = defs->at(insidx); |
1296 | // Set new lidx for DEF |
1297 | new_lrg(n1, maxlrg++); |
1298 | } |
1299 | //----------Phi Node Splitting---------- |
1300 | // Clean up a phi here, and assign a new live range number |
1301 | // Cycle through this block's predecessors, collecting Reaches |
1302 | // info for each spilled LRG and update edges. |
1303 | // Walk the phis list to patch inputs, split phis, and name phis |
1304 | uint lrgs_before_phi_split = maxlrg; |
1305 | for( insidx = 0; insidx < phis->size(); insidx++ ) { |
1306 | Node *phi = phis->at(insidx); |
1307 | assert(phi->is_Phi(),"This list must only contain Phi Nodes" ); |
1308 | Block *b = _cfg.get_block_for_node(phi); |
1309 | // Grab the live range number |
1310 | uint lidx = _lrg_map.find_id(phi); |
1311 | uint slidx = lrg2reach[lidx]; |
1312 | // Update node to lidx map |
1313 | new_lrg(phi, maxlrg++); |
1314 | // Get PASS1's up/down decision for the block. |
1315 | int phi_up = !!UP_entry[slidx]->test(b->_pre_order); |
1316 | |
1317 | // Force down if double-spilling live range |
1318 | if( lrgs(lidx)._was_spilled1 ) |
1319 | phi_up = false; |
1320 | |
1321 | // When splitting a Phi we an split it normal or "inverted". |
1322 | // An inverted split makes the splits target the Phi's UP/DOWN |
1323 | // sense inverted; then the Phi is followed by a final def-side |
1324 | // split to invert back. It changes which blocks the spill code |
1325 | // goes in. |
1326 | |
1327 | // Walk the predecessor blocks and assign the reaching def to the Phi. |
1328 | // Split Phi nodes by placing USE side splits wherever the reaching |
1329 | // DEF has the wrong UP/DOWN value. |
1330 | for( uint i = 1; i < b->num_preds(); i++ ) { |
1331 | // Get predecessor block pre-order number |
1332 | Block *pred = _cfg.get_block_for_node(b->pred(i)); |
1333 | pidx = pred->_pre_order; |
1334 | // Grab reaching def |
1335 | Node *def = Reaches[pidx][slidx]; |
1336 | Node** Reachblock = Reaches[pidx]; |
1337 | assert( def, "must have reaching def" ); |
1338 | // If input up/down sense and reg-pressure DISagree |
1339 | if (def->rematerialize()) { |
1340 | // Place the rematerialized node above any MSCs created during |
1341 | // phi node splitting. end_idx points at the insertion point |
1342 | // so look at the node before it. |
1343 | int insert = pred->end_idx(); |
1344 | while (insert >= 1 && |
1345 | pred->get_node(insert - 1)->is_SpillCopy() && |
1346 | _lrg_map.find(pred->get_node(insert - 1)) >= lrgs_before_phi_split) { |
1347 | insert--; |
1348 | } |
1349 | def = split_Rematerialize(def, pred, insert, maxlrg, splits, slidx, lrg2reach, Reachblock, false); |
1350 | if (!def) { |
1351 | return 0; // Bail out |
1352 | } |
1353 | } |
1354 | // Update the Phi's input edge array |
1355 | phi->set_req(i,def); |
1356 | // Grab the UP/DOWN sense for the input |
1357 | u1 = UP[pidx][slidx]; |
1358 | if( u1 != (phi_up != 0)) { |
1359 | maxlrg = split_USE(MachSpillCopyNode::PhiLocationDifferToInputLocation, def, b, phi, i, maxlrg, !u1, false, splits,slidx); |
1360 | // If it wasn't split bail |
1361 | if (!maxlrg) { |
1362 | return 0; |
1363 | } |
1364 | } |
1365 | } // End for all inputs to the Phi |
1366 | } // End for all Phi Nodes |
1367 | // Update _maxlrg to save Union asserts |
1368 | _lrg_map.set_max_lrg_id(maxlrg); |
1369 | |
1370 | |
1371 | //----------PASS 3---------- |
1372 | // Pass over all Phi's to union the live ranges |
1373 | for( insidx = 0; insidx < phis->size(); insidx++ ) { |
1374 | Node *phi = phis->at(insidx); |
1375 | assert(phi->is_Phi(),"This list must only contain Phi Nodes" ); |
1376 | // Walk all inputs to Phi and Union input live range with Phi live range |
1377 | for( uint i = 1; i < phi->req(); i++ ) { |
1378 | // Grab the input node |
1379 | Node *n = phi->in(i); |
1380 | assert(n, "node should exist" ); |
1381 | uint lidx = _lrg_map.find(n); |
1382 | uint pidx = _lrg_map.find(phi); |
1383 | if (lidx < pidx) { |
1384 | Union(n, phi); |
1385 | } |
1386 | else if(lidx > pidx) { |
1387 | Union(phi, n); |
1388 | } |
1389 | } // End for all inputs to the Phi Node |
1390 | } // End for all Phi Nodes |
1391 | // Now union all two address instructions |
1392 | for (insidx = 0; insidx < defs->size(); insidx++) { |
1393 | // Grab the def |
1394 | n1 = defs->at(insidx); |
1395 | // Set new lidx for DEF & handle 2-addr instructions |
1396 | if (n1->is_Mach() && ((twoidx = n1->as_Mach()->two_adr()) != 0)) { |
1397 | assert(_lrg_map.find(n1->in(twoidx)) < maxlrg,"Assigning bad live range index" ); |
1398 | // Union the input and output live ranges |
1399 | uint lr1 = _lrg_map.find(n1); |
1400 | uint lr2 = _lrg_map.find(n1->in(twoidx)); |
1401 | if (lr1 < lr2) { |
1402 | Union(n1, n1->in(twoidx)); |
1403 | } |
1404 | else if (lr1 > lr2) { |
1405 | Union(n1->in(twoidx), n1); |
1406 | } |
1407 | } // End if two address |
1408 | } // End for all defs |
1409 | // DEBUG |
1410 | #ifdef ASSERT |
1411 | // Validate all live range index assignments |
1412 | for (bidx = 0; bidx < _cfg.number_of_blocks(); bidx++) { |
1413 | b = _cfg.get_block(bidx); |
1414 | for (insidx = 0; insidx <= b->end_idx(); insidx++) { |
1415 | Node *n = b->get_node(insidx); |
1416 | uint defidx = _lrg_map.find(n); |
1417 | assert(defidx < _lrg_map.max_lrg_id(), "Bad live range index in Split" ); |
1418 | assert(defidx < maxlrg,"Bad live range index in Split" ); |
1419 | } |
1420 | } |
1421 | // Issue a warning if splitting made no progress |
1422 | int noprogress = 0; |
1423 | for (slidx = 0; slidx < spill_cnt; slidx++) { |
1424 | if (PrintOpto && WizardMode && splits.at(slidx) == 0) { |
1425 | tty->print_cr("Failed to split live range %d" , lidxs.at(slidx)); |
1426 | //BREAKPOINT; |
1427 | } |
1428 | else { |
1429 | noprogress++; |
1430 | } |
1431 | } |
1432 | if(!noprogress) { |
1433 | tty->print_cr("Failed to make progress in Split" ); |
1434 | //BREAKPOINT; |
1435 | } |
1436 | #endif |
1437 | // Return updated count of live ranges |
1438 | return maxlrg; |
1439 | } |
1440 | |