Mercurial > hg > truffle
annotate src/share/vm/opto/ifg.cpp @ 9691:bd5c6b3dedc5
implement inlining support for JSR 292
author | twisti |
---|---|
date | Tue, 14 May 2013 11:27:09 -0700 |
parents | b808febcad9a |
children | 8373c19be854 |
rev | line source |
---|---|
0 | 1 /* |
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
2 * Copyright (c) 1998, 2012, Oracle and/or its affiliates. All rights reserved. |
0 | 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 * | |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1212
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1212
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1212
diff
changeset
|
21 * questions. |
0 | 22 * |
23 */ | |
24 | |
1972 | 25 #include "precompiled.hpp" |
26 #include "compiler/oopMap.hpp" | |
27 #include "memory/allocation.inline.hpp" | |
28 #include "opto/addnode.hpp" | |
29 #include "opto/block.hpp" | |
30 #include "opto/callnode.hpp" | |
31 #include "opto/cfgnode.hpp" | |
32 #include "opto/chaitin.hpp" | |
33 #include "opto/coalesce.hpp" | |
34 #include "opto/connode.hpp" | |
35 #include "opto/indexSet.hpp" | |
36 #include "opto/machnode.hpp" | |
37 #include "opto/memnode.hpp" | |
38 #include "opto/opcodes.hpp" | |
0 | 39 |
40 //============================================================================= | |
41 //------------------------------IFG-------------------------------------------- | |
42 PhaseIFG::PhaseIFG( Arena *arena ) : Phase(Interference_Graph), _arena(arena) { | |
43 } | |
44 | |
45 //------------------------------init------------------------------------------- | |
46 void PhaseIFG::init( uint maxlrg ) { | |
47 _maxlrg = maxlrg; | |
48 _yanked = new (_arena) VectorSet(_arena); | |
49 _is_square = false; | |
50 // Make uninitialized adjacency lists | |
51 _adjs = (IndexSet*)_arena->Amalloc(sizeof(IndexSet)*maxlrg); | |
52 // Also make empty live range structures | |
53 _lrgs = (LRG *)_arena->Amalloc( maxlrg * sizeof(LRG) ); | |
54 memset(_lrgs,0,sizeof(LRG)*maxlrg); | |
55 // Init all to empty | |
56 for( uint i = 0; i < maxlrg; i++ ) { | |
57 _adjs[i].initialize(maxlrg); | |
58 _lrgs[i].Set_All(); | |
59 } | |
60 } | |
61 | |
62 //------------------------------add-------------------------------------------- | |
63 // Add edge between vertices a & b. These are sorted (triangular matrix), | |
64 // then the smaller number is inserted in the larger numbered array. | |
65 int PhaseIFG::add_edge( uint a, uint b ) { | |
66 lrgs(a).invalid_degree(); | |
67 lrgs(b).invalid_degree(); | |
68 // Sort a and b, so that a is bigger | |
69 assert( !_is_square, "only on triangular" ); | |
70 if( a < b ) { uint tmp = a; a = b; b = tmp; } | |
71 return _adjs[a].insert( b ); | |
72 } | |
73 | |
74 //------------------------------add_vector------------------------------------- | |
75 // Add an edge between 'a' and everything in the vector. | |
76 void PhaseIFG::add_vector( uint a, IndexSet *vec ) { | |
77 // IFG is triangular, so do the inserts where 'a' < 'b'. | |
78 assert( !_is_square, "only on triangular" ); | |
79 IndexSet *adjs_a = &_adjs[a]; | |
80 if( !vec->count() ) return; | |
81 | |
82 IndexSetIterator elements(vec); | |
83 uint neighbor; | |
84 while ((neighbor = elements.next()) != 0) { | |
85 add_edge( a, neighbor ); | |
86 } | |
87 } | |
88 | |
89 //------------------------------test------------------------------------------- | |
90 // Is there an edge between a and b? | |
91 int PhaseIFG::test_edge( uint a, uint b ) const { | |
92 // Sort a and b, so that a is larger | |
93 assert( !_is_square, "only on triangular" ); | |
94 if( a < b ) { uint tmp = a; a = b; b = tmp; } | |
95 return _adjs[a].member(b); | |
96 } | |
97 | |
98 //------------------------------SquareUp--------------------------------------- | |
99 // Convert triangular matrix to square matrix | |
100 void PhaseIFG::SquareUp() { | |
101 assert( !_is_square, "only on triangular" ); | |
102 | |
103 // Simple transpose | |
104 for( uint i = 0; i < _maxlrg; i++ ) { | |
105 IndexSetIterator elements(&_adjs[i]); | |
106 uint datum; | |
107 while ((datum = elements.next()) != 0) { | |
108 _adjs[datum].insert( i ); | |
109 } | |
110 } | |
111 _is_square = true; | |
112 } | |
113 | |
114 //------------------------------Compute_Effective_Degree----------------------- | |
115 // Compute effective degree in bulk | |
116 void PhaseIFG::Compute_Effective_Degree() { | |
117 assert( _is_square, "only on square" ); | |
118 | |
119 for( uint i = 0; i < _maxlrg; i++ ) | |
120 lrgs(i).set_degree(effective_degree(i)); | |
121 } | |
122 | |
123 //------------------------------test_edge_sq----------------------------------- | |
124 int PhaseIFG::test_edge_sq( uint a, uint b ) const { | |
125 assert( _is_square, "only on square" ); | |
126 // Swap, so that 'a' has the lesser count. Then binary search is on | |
127 // the smaller of a's list and b's list. | |
128 if( neighbor_cnt(a) > neighbor_cnt(b) ) { uint tmp = a; a = b; b = tmp; } | |
129 //return _adjs[a].unordered_member(b); | |
130 return _adjs[a].member(b); | |
131 } | |
132 | |
133 //------------------------------Union------------------------------------------ | |
134 // Union edges of B into A | |
135 void PhaseIFG::Union( uint a, uint b ) { | |
136 assert( _is_square, "only on square" ); | |
137 IndexSet *A = &_adjs[a]; | |
138 IndexSetIterator b_elements(&_adjs[b]); | |
139 uint datum; | |
140 while ((datum = b_elements.next()) != 0) { | |
141 if(A->insert(datum)) { | |
142 _adjs[datum].insert(a); | |
143 lrgs(a).invalid_degree(); | |
144 lrgs(datum).invalid_degree(); | |
145 } | |
146 } | |
147 } | |
148 | |
149 //------------------------------remove_node------------------------------------ | |
150 // Yank a Node and all connected edges from the IFG. Return a | |
151 // list of neighbors (edges) yanked. | |
152 IndexSet *PhaseIFG::remove_node( uint a ) { | |
153 assert( _is_square, "only on square" ); | |
154 assert( !_yanked->test(a), "" ); | |
155 _yanked->set(a); | |
156 | |
157 // I remove the LRG from all neighbors. | |
158 IndexSetIterator elements(&_adjs[a]); | |
159 LRG &lrg_a = lrgs(a); | |
160 uint datum; | |
161 while ((datum = elements.next()) != 0) { | |
162 _adjs[datum].remove(a); | |
163 lrgs(datum).inc_degree( -lrg_a.compute_degree(lrgs(datum)) ); | |
164 } | |
165 return neighbors(a); | |
166 } | |
167 | |
168 //------------------------------re_insert-------------------------------------- | |
169 // Re-insert a yanked Node. | |
170 void PhaseIFG::re_insert( uint a ) { | |
171 assert( _is_square, "only on square" ); | |
172 assert( _yanked->test(a), "" ); | |
173 (*_yanked) >>= a; | |
174 | |
175 IndexSetIterator elements(&_adjs[a]); | |
176 uint datum; | |
177 while ((datum = elements.next()) != 0) { | |
178 _adjs[datum].insert(a); | |
179 lrgs(datum).invalid_degree(); | |
180 } | |
181 } | |
182 | |
183 //------------------------------compute_degree--------------------------------- | |
184 // Compute the degree between 2 live ranges. If both live ranges are | |
185 // aligned-adjacent powers-of-2 then we use the MAX size. If either is | |
186 // mis-aligned (or for Fat-Projections, not-adjacent) then we have to | |
187 // MULTIPLY the sizes. Inspect Brigg's thesis on register pairs to see why | |
188 // this is so. | |
189 int LRG::compute_degree( LRG &l ) const { | |
190 int tmp; | |
191 int num_regs = _num_regs; | |
192 int nregs = l.num_regs(); | |
193 tmp = (_fat_proj || l._fat_proj) // either is a fat-proj? | |
194 ? (num_regs * nregs) // then use product | |
195 : MAX2(num_regs,nregs); // else use max | |
196 return tmp; | |
197 } | |
198 | |
199 //------------------------------effective_degree------------------------------- | |
200 // Compute effective degree for this live range. If both live ranges are | |
201 // aligned-adjacent powers-of-2 then we use the MAX size. If either is | |
202 // mis-aligned (or for Fat-Projections, not-adjacent) then we have to | |
203 // MULTIPLY the sizes. Inspect Brigg's thesis on register pairs to see why | |
204 // this is so. | |
205 int PhaseIFG::effective_degree( uint lidx ) const { | |
206 int eff = 0; | |
207 int num_regs = lrgs(lidx).num_regs(); | |
208 int fat_proj = lrgs(lidx)._fat_proj; | |
209 IndexSet *s = neighbors(lidx); | |
210 IndexSetIterator elements(s); | |
211 uint nidx; | |
212 while((nidx = elements.next()) != 0) { | |
213 LRG &lrgn = lrgs(nidx); | |
214 int nregs = lrgn.num_regs(); | |
215 eff += (fat_proj || lrgn._fat_proj) // either is a fat-proj? | |
216 ? (num_regs * nregs) // then use product | |
217 : MAX2(num_regs,nregs); // else use max | |
218 } | |
219 return eff; | |
220 } | |
221 | |
222 | |
223 #ifndef PRODUCT | |
224 //------------------------------dump------------------------------------------- | |
225 void PhaseIFG::dump() const { | |
226 tty->print_cr("-- Interference Graph --%s--", | |
227 _is_square ? "square" : "triangular" ); | |
228 if( _is_square ) { | |
229 for( uint i = 0; i < _maxlrg; i++ ) { | |
230 tty->print( (*_yanked)[i] ? "XX " : " "); | |
231 tty->print("L%d: { ",i); | |
232 IndexSetIterator elements(&_adjs[i]); | |
233 uint datum; | |
234 while ((datum = elements.next()) != 0) { | |
235 tty->print("L%d ", datum); | |
236 } | |
237 tty->print_cr("}"); | |
238 | |
239 } | |
240 return; | |
241 } | |
242 | |
243 // Triangular | |
244 for( uint i = 0; i < _maxlrg; i++ ) { | |
245 uint j; | |
246 tty->print( (*_yanked)[i] ? "XX " : " "); | |
247 tty->print("L%d: { ",i); | |
248 for( j = _maxlrg; j > i; j-- ) | |
249 if( test_edge(j - 1,i) ) { | |
250 tty->print("L%d ",j - 1); | |
251 } | |
252 tty->print("| "); | |
253 IndexSetIterator elements(&_adjs[i]); | |
254 uint datum; | |
255 while ((datum = elements.next()) != 0) { | |
256 tty->print("L%d ", datum); | |
257 } | |
258 tty->print("}\n"); | |
259 } | |
260 tty->print("\n"); | |
261 } | |
262 | |
263 //------------------------------stats------------------------------------------ | |
264 void PhaseIFG::stats() const { | |
265 ResourceMark rm; | |
266 int *h_cnt = NEW_RESOURCE_ARRAY(int,_maxlrg*2); | |
267 memset( h_cnt, 0, sizeof(int)*_maxlrg*2 ); | |
268 uint i; | |
269 for( i = 0; i < _maxlrg; i++ ) { | |
270 h_cnt[neighbor_cnt(i)]++; | |
271 } | |
272 tty->print_cr("--Histogram of counts--"); | |
273 for( i = 0; i < _maxlrg*2; i++ ) | |
274 if( h_cnt[i] ) | |
275 tty->print("%d/%d ",i,h_cnt[i]); | |
276 tty->print_cr(""); | |
277 } | |
278 | |
279 //------------------------------verify----------------------------------------- | |
280 void PhaseIFG::verify( const PhaseChaitin *pc ) const { | |
281 // IFG is square, sorted and no need for Find | |
282 for( uint i = 0; i < _maxlrg; i++ ) { | |
283 assert(!((*_yanked)[i]) || !neighbor_cnt(i), "Is removed completely" ); | |
284 IndexSet *set = &_adjs[i]; | |
285 IndexSetIterator elements(set); | |
286 uint idx; | |
287 uint last = 0; | |
288 while ((idx = elements.next()) != 0) { | |
289 assert( idx != i, "Must have empty diagonal"); | |
290 assert( pc->Find_const(idx) == idx, "Must not need Find" ); | |
291 assert( _adjs[idx].member(i), "IFG not square" ); | |
292 assert( !(*_yanked)[idx], "No yanked neighbors" ); | |
293 assert( last < idx, "not sorted increasing"); | |
294 last = idx; | |
295 } | |
296 assert( !lrgs(i)._degree_valid || | |
297 effective_degree(i) == lrgs(i).degree(), "degree is valid but wrong" ); | |
298 } | |
299 } | |
300 #endif | |
301 | |
302 //------------------------------interfere_with_live---------------------------- | |
303 // Interfere this register with everything currently live. Use the RegMasks | |
304 // to trim the set of possible interferences. Return a count of register-only | |
605 | 305 // interferences as an estimate of register pressure. |
0 | 306 void PhaseChaitin::interfere_with_live( uint r, IndexSet *liveout ) { |
307 uint retval = 0; | |
308 // Interfere with everything live. | |
309 const RegMask &rm = lrgs(r).mask(); | |
310 // Check for interference by checking overlap of regmasks. | |
311 // Only interfere if acceptable register masks overlap. | |
312 IndexSetIterator elements(liveout); | |
313 uint l; | |
314 while( (l = elements.next()) != 0 ) | |
315 if( rm.overlap( lrgs(l).mask() ) ) | |
316 _ifg->add_edge( r, l ); | |
317 } | |
318 | |
319 //------------------------------build_ifg_virtual------------------------------ | |
320 // Actually build the interference graph. Uses virtual registers only, no | |
321 // physical register masks. This allows me to be very aggressive when | |
322 // coalescing copies. Some of this aggressiveness will have to be undone | |
323 // later, but I'd rather get all the copies I can now (since unremoved copies | |
324 // at this point can end up in bad places). Copies I re-insert later I have | |
325 // more opportunity to insert them in low-frequency locations. | |
326 void PhaseChaitin::build_ifg_virtual( ) { | |
327 | |
328 // For all blocks (in any order) do... | |
329 for( uint i=0; i<_cfg._num_blocks; i++ ) { | |
330 Block *b = _cfg._blocks[i]; | |
331 IndexSet *liveout = _live->live(b); | |
332 | |
333 // The IFG is built by a single reverse pass over each basic block. | |
334 // Starting with the known live-out set, we remove things that get | |
335 // defined and add things that become live (essentially executing one | |
336 // pass of a standard LIVE analysis). Just before a Node defines a value | |
337 // (and removes it from the live-ness set) that value is certainly live. | |
338 // The defined value interferes with everything currently live. The | |
339 // value is then removed from the live-ness set and it's inputs are | |
340 // added to the live-ness set. | |
341 for( uint j = b->end_idx() + 1; j > 1; j-- ) { | |
342 Node *n = b->_nodes[j-1]; | |
343 | |
344 // Get value being defined | |
345 uint r = n2lidx(n); | |
346 | |
347 // Some special values do not allocate | |
348 if( r ) { | |
349 | |
350 // Remove from live-out set | |
351 liveout->remove(r); | |
352 | |
353 // Copies do not define a new value and so do not interfere. | |
354 // Remove the copies source from the liveout set before interfering. | |
355 uint idx = n->is_Copy(); | |
356 if( idx ) liveout->remove( n2lidx(n->in(idx)) ); | |
357 | |
358 // Interfere with everything live | |
359 interfere_with_live( r, liveout ); | |
360 } | |
361 | |
362 // Make all inputs live | |
363 if( !n->is_Phi() ) { // Phi function uses come from prior block | |
364 for( uint k = 1; k < n->req(); k++ ) | |
365 liveout->insert( n2lidx(n->in(k)) ); | |
366 } | |
367 | |
368 // 2-address instructions always have the defined value live | |
369 // on entry to the instruction, even though it is being defined | |
370 // by the instruction. We pretend a virtual copy sits just prior | |
371 // to the instruction and kills the src-def'd register. | |
372 // In other words, for 2-address instructions the defined value | |
373 // interferes with all inputs. | |
374 uint idx; | |
375 if( n->is_Mach() && (idx = n->as_Mach()->two_adr()) ) { | |
376 const MachNode *mach = n->as_Mach(); | |
377 // Sometimes my 2-address ADDs are commuted in a bad way. | |
378 // We generally want the USE-DEF register to refer to the | |
379 // loop-varying quantity, to avoid a copy. | |
380 uint op = mach->ideal_Opcode(); | |
381 // Check that mach->num_opnds() == 3 to ensure instruction is | |
382 // not subsuming constants, effectively excludes addI_cin_imm | |
383 // Can NOT swap for instructions like addI_cin_imm since it | |
384 // is adding zero to yhi + carry and the second ideal-input | |
385 // points to the result of adding low-halves. | |
386 // Checking req() and num_opnds() does NOT distinguish addI_cout from addI_cout_imm | |
387 if( (op == Op_AddI && mach->req() == 3 && mach->num_opnds() == 3) && | |
388 n->in(1)->bottom_type()->base() == Type::Int && | |
389 // See if the ADD is involved in a tight data loop the wrong way | |
390 n->in(2)->is_Phi() && | |
391 n->in(2)->in(2) == n ) { | |
392 Node *tmp = n->in(1); | |
393 n->set_req( 1, n->in(2) ); | |
394 n->set_req( 2, tmp ); | |
395 } | |
396 // Defined value interferes with all inputs | |
397 uint lidx = n2lidx(n->in(idx)); | |
398 for( uint k = 1; k < n->req(); k++ ) { | |
399 uint kidx = n2lidx(n->in(k)); | |
400 if( kidx != lidx ) | |
401 _ifg->add_edge( r, kidx ); | |
402 } | |
403 } | |
404 } // End of forall instructions in block | |
405 } // End of forall blocks | |
406 } | |
407 | |
408 //------------------------------count_int_pressure----------------------------- | |
409 uint PhaseChaitin::count_int_pressure( IndexSet *liveout ) { | |
410 IndexSetIterator elements(liveout); | |
411 uint lidx; | |
412 uint cnt = 0; | |
413 while ((lidx = elements.next()) != 0) { | |
414 if( lrgs(lidx).mask().is_UP() && | |
415 lrgs(lidx).mask_size() && | |
416 !lrgs(lidx)._is_float && | |
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
417 !lrgs(lidx)._is_vector && |
0 | 418 lrgs(lidx).mask().overlap(*Matcher::idealreg2regmask[Op_RegI]) ) |
419 cnt += lrgs(lidx).reg_pressure(); | |
420 } | |
421 return cnt; | |
422 } | |
423 | |
424 //------------------------------count_float_pressure--------------------------- | |
425 uint PhaseChaitin::count_float_pressure( IndexSet *liveout ) { | |
426 IndexSetIterator elements(liveout); | |
427 uint lidx; | |
428 uint cnt = 0; | |
429 while ((lidx = elements.next()) != 0) { | |
430 if( lrgs(lidx).mask().is_UP() && | |
431 lrgs(lidx).mask_size() && | |
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
432 (lrgs(lidx)._is_float || lrgs(lidx)._is_vector)) |
0 | 433 cnt += lrgs(lidx).reg_pressure(); |
434 } | |
435 return cnt; | |
436 } | |
437 | |
438 //------------------------------lower_pressure--------------------------------- | |
439 // Adjust register pressure down by 1. Capture last hi-to-low transition, | |
440 static void lower_pressure( LRG *lrg, uint where, Block *b, uint *pressure, uint *hrp_index ) { | |
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
441 if (lrg->mask().is_UP() && lrg->mask_size()) { |
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
442 if (lrg->_is_float || lrg->_is_vector) { |
0 | 443 pressure[1] -= lrg->reg_pressure(); |
444 if( pressure[1] == (uint)FLOATPRESSURE ) { | |
445 hrp_index[1] = where; | |
8867 | 446 if( pressure[1] > b->_freg_pressure ) |
447 b->_freg_pressure = pressure[1]+1; | |
0 | 448 } |
449 } else if( lrg->mask().overlap(*Matcher::idealreg2regmask[Op_RegI]) ) { | |
450 pressure[0] -= lrg->reg_pressure(); | |
451 if( pressure[0] == (uint)INTPRESSURE ) { | |
452 hrp_index[0] = where; | |
8867 | 453 if( pressure[0] > b->_reg_pressure ) |
454 b->_reg_pressure = pressure[0]+1; | |
0 | 455 } |
456 } | |
457 } | |
458 } | |
459 | |
460 //------------------------------build_ifg_physical----------------------------- | |
461 // Build the interference graph using physical registers when available. | |
462 // That is, if 2 live ranges are simultaneously alive but in their acceptable | |
463 // register sets do not overlap, then they do not interfere. | |
464 uint PhaseChaitin::build_ifg_physical( ResourceArea *a ) { | |
465 NOT_PRODUCT( Compile::TracePhase t3("buildIFG", &_t_buildIFGphysical, TimeCompiler); ) | |
466 | |
467 uint spill_reg = LRG::SPILL_REG; | |
468 uint must_spill = 0; | |
469 | |
470 // For all blocks (in any order) do... | |
471 for( uint i = 0; i < _cfg._num_blocks; i++ ) { | |
472 Block *b = _cfg._blocks[i]; | |
473 // Clone (rather than smash in place) the liveout info, so it is alive | |
474 // for the "collect_gc_info" phase later. | |
475 IndexSet liveout(_live->live(b)); | |
476 uint last_inst = b->end_idx(); | |
566
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
477 // Compute first nonphi node index |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
478 uint first_inst; |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
479 for( first_inst = 1; first_inst < last_inst; first_inst++ ) |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
480 if( !b->_nodes[first_inst]->is_Phi() ) |
0 | 481 break; |
482 | |
566
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
483 // Spills could be inserted before CreateEx node which should be |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
484 // first instruction in block after Phis. Move CreateEx up. |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
485 for( uint insidx = first_inst; insidx < last_inst; insidx++ ) { |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
486 Node *ex = b->_nodes[insidx]; |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
487 if( ex->is_SpillCopy() ) continue; |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
488 if( insidx > first_inst && ex->is_Mach() && |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
489 ex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
490 // If the CreateEx isn't above all the MachSpillCopies |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
491 // then move it to the top. |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
492 b->_nodes.remove(insidx); |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
493 b->_nodes.insert(first_inst, ex); |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
494 } |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
495 // Stop once a CreateEx or any other node is found |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
496 break; |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
497 } |
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
498 |
0 | 499 // Reset block's register pressure values for each ifg construction |
500 uint pressure[2], hrp_index[2]; | |
501 pressure[0] = pressure[1] = 0; | |
502 hrp_index[0] = hrp_index[1] = last_inst+1; | |
503 b->_reg_pressure = b->_freg_pressure = 0; | |
504 // Liveout things are presumed live for the whole block. We accumulate | |
505 // 'area' accordingly. If they get killed in the block, we'll subtract | |
506 // the unused part of the block from the area. | |
566
91263420e1c6
6791852: assert(b->_nodes[insidx] == n,"got insidx set incorrectly")
kvn
parents:
380
diff
changeset
|
507 int inst_count = last_inst - first_inst; |
369
5f85534046c2
6750588: assert(lrg._area >= 0,"negative spill area") running NSK stmp0101 test
rasbold
parents:
295
diff
changeset
|
508 double cost = (inst_count <= 0) ? 0.0 : b->_freq * double(inst_count); |
5f85534046c2
6750588: assert(lrg._area >= 0,"negative spill area") running NSK stmp0101 test
rasbold
parents:
295
diff
changeset
|
509 assert(!(cost < 0.0), "negative spill cost" ); |
0 | 510 IndexSetIterator elements(&liveout); |
511 uint lidx; | |
512 while ((lidx = elements.next()) != 0) { | |
513 LRG &lrg = lrgs(lidx); | |
514 lrg._area += cost; | |
515 // Compute initial register pressure | |
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
516 if (lrg.mask().is_UP() && lrg.mask_size()) { |
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
517 if (lrg._is_float || lrg._is_vector) { // Count float pressure |
0 | 518 pressure[1] += lrg.reg_pressure(); |
519 if( pressure[1] > b->_freg_pressure ) | |
520 b->_freg_pressure = pressure[1]; | |
521 // Count int pressure, but do not count the SP, flags | |
522 } else if( lrgs(lidx).mask().overlap(*Matcher::idealreg2regmask[Op_RegI]) ) { | |
523 pressure[0] += lrg.reg_pressure(); | |
524 if( pressure[0] > b->_reg_pressure ) | |
525 b->_reg_pressure = pressure[0]; | |
526 } | |
527 } | |
528 } | |
529 assert( pressure[0] == count_int_pressure (&liveout), "" ); | |
530 assert( pressure[1] == count_float_pressure(&liveout), "" ); | |
531 | |
532 // The IFG is built by a single reverse pass over each basic block. | |
533 // Starting with the known live-out set, we remove things that get | |
534 // defined and add things that become live (essentially executing one | |
535 // pass of a standard LIVE analysis). Just before a Node defines a value | |
536 // (and removes it from the live-ness set) that value is certainly live. | |
537 // The defined value interferes with everything currently live. The | |
538 // value is then removed from the live-ness set and it's inputs are added | |
539 // to the live-ness set. | |
540 uint j; | |
541 for( j = last_inst + 1; j > 1; j-- ) { | |
542 Node *n = b->_nodes[j - 1]; | |
543 | |
544 // Get value being defined | |
545 uint r = n2lidx(n); | |
546 | |
547 // Some special values do not allocate | |
548 if( r ) { | |
549 // A DEF normally costs block frequency; rematerialized values are | |
550 // removed from the DEF sight, so LOWER costs here. | |
551 lrgs(r)._cost += n->rematerialize() ? 0 : b->_freq; | |
552 | |
553 // If it is not live, then this instruction is dead. Probably caused | |
554 // by spilling and rematerialization. Who cares why, yank this baby. | |
555 if( !liveout.member(r) && n->Opcode() != Op_SafePoint ) { | |
556 Node *def = n->in(0); | |
557 if( !n->is_Proj() || | |
558 // Could also be a flags-projection of a dead ADD or such. | |
559 (n2lidx(def) && !liveout.member(n2lidx(def)) ) ) { | |
560 b->_nodes.remove(j - 1); | |
561 if( lrgs(r)._def == n ) lrgs(r)._def = 0; | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6179
diff
changeset
|
562 n->disconnect_inputs(NULL, C); |
0 | 563 _cfg._bbs.map(n->_idx,NULL); |
564 n->replace_by(C->top()); | |
565 // Since yanking a Node from block, high pressure moves up one | |
566 hrp_index[0]--; | |
567 hrp_index[1]--; | |
568 continue; | |
569 } | |
570 | |
571 // Fat-projections kill many registers which cannot be used to | |
572 // hold live ranges. | |
573 if( lrgs(r)._fat_proj ) { | |
574 // Count the int-only registers | |
575 RegMask itmp = lrgs(r).mask(); | |
576 itmp.AND(*Matcher::idealreg2regmask[Op_RegI]); | |
577 int iregs = itmp.Size(); | |
578 if( pressure[0]+iregs > b->_reg_pressure ) | |
579 b->_reg_pressure = pressure[0]+iregs; | |
580 if( pressure[0] <= (uint)INTPRESSURE && | |
581 pressure[0]+iregs > (uint)INTPRESSURE ) { | |
582 hrp_index[0] = j-1; | |
583 } | |
584 // Count the float-only registers | |
585 RegMask ftmp = lrgs(r).mask(); | |
586 ftmp.AND(*Matcher::idealreg2regmask[Op_RegD]); | |
587 int fregs = ftmp.Size(); | |
588 if( pressure[1]+fregs > b->_freg_pressure ) | |
589 b->_freg_pressure = pressure[1]+fregs; | |
590 if( pressure[1] <= (uint)FLOATPRESSURE && | |
591 pressure[1]+fregs > (uint)FLOATPRESSURE ) { | |
592 hrp_index[1] = j-1; | |
593 } | |
594 } | |
595 | |
596 } else { // Else it is live | |
597 // A DEF also ends 'area' partway through the block. | |
598 lrgs(r)._area -= cost; | |
369
5f85534046c2
6750588: assert(lrg._area >= 0,"negative spill area") running NSK stmp0101 test
rasbold
parents:
295
diff
changeset
|
599 assert(!(lrgs(r)._area < 0.0), "negative spill area" ); |
0 | 600 |
601 // Insure high score for immediate-use spill copies so they get a color | |
602 if( n->is_SpillCopy() | |
295
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
603 && lrgs(r).is_singledef() // MultiDef live range can still split |
0 | 604 && n->outcnt() == 1 // and use must be in this block |
605 && _cfg._bbs[n->unique_out()->_idx] == b ) { | |
606 // All single-use MachSpillCopy(s) that immediately precede their | |
607 // use must color early. If a longer live range steals their | |
608 // color, the spill copy will split and may push another spill copy | |
609 // further away resulting in an infinite spill-split-retry cycle. | |
610 // Assigning a zero area results in a high score() and a good | |
611 // location in the simplify list. | |
612 // | |
613 | |
614 Node *single_use = n->unique_out(); | |
615 assert( b->find_node(single_use) >= j, "Use must be later in block"); | |
616 // Use can be earlier in block if it is a Phi, but then I should be a MultiDef | |
617 | |
618 // Find first non SpillCopy 'm' that follows the current instruction | |
619 // (j - 1) is index for current instruction 'n' | |
620 Node *m = n; | |
621 for( uint i = j; i <= last_inst && m->is_SpillCopy(); ++i ) { m = b->_nodes[i]; } | |
622 if( m == single_use ) { | |
623 lrgs(r)._area = 0.0; | |
624 } | |
625 } | |
626 | |
627 // Remove from live-out set | |
628 if( liveout.remove(r) ) { | |
629 // Adjust register pressure. | |
630 // Capture last hi-to-lo pressure transition | |
631 lower_pressure( &lrgs(r), j-1, b, pressure, hrp_index ); | |
632 assert( pressure[0] == count_int_pressure (&liveout), "" ); | |
633 assert( pressure[1] == count_float_pressure(&liveout), "" ); | |
634 } | |
635 | |
636 // Copies do not define a new value and so do not interfere. | |
637 // Remove the copies source from the liveout set before interfering. | |
638 uint idx = n->is_Copy(); | |
639 if( idx ) { | |
640 uint x = n2lidx(n->in(idx)); | |
641 if( liveout.remove( x ) ) { | |
642 lrgs(x)._area -= cost; | |
643 // Adjust register pressure. | |
644 lower_pressure( &lrgs(x), j-1, b, pressure, hrp_index ); | |
645 assert( pressure[0] == count_int_pressure (&liveout), "" ); | |
646 assert( pressure[1] == count_float_pressure(&liveout), "" ); | |
647 } | |
648 } | |
649 } // End of if live or not | |
650 | |
651 // Interfere with everything live. If the defined value must | |
652 // go in a particular register, just remove that register from | |
653 // all conflicting parties and avoid the interference. | |
654 | |
655 // Make exclusions for rematerializable defs. Since rematerializable | |
656 // DEFs are not bound but the live range is, some uses must be bound. | |
657 // If we spill live range 'r', it can rematerialize at each use site | |
658 // according to its bindings. | |
659 const RegMask &rmask = lrgs(r).mask(); | |
660 if( lrgs(r).is_bound() && !(n->rematerialize()) && rmask.is_NotEmpty() ) { | |
661 // Check for common case | |
662 int r_size = lrgs(r).num_regs(); | |
663 OptoReg::Name r_reg = (r_size == 1) ? rmask.find_first_elem() : OptoReg::Physical; | |
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
664 // Smear odd bits |
0 | 665 IndexSetIterator elements(&liveout); |
666 uint l; | |
667 while ((l = elements.next()) != 0) { | |
668 LRG &lrg = lrgs(l); | |
669 // If 'l' must spill already, do not further hack his bits. | |
670 // He'll get some interferences and be forced to spill later. | |
671 if( lrg._must_spill ) continue; | |
672 // Remove bound register(s) from 'l's choices | |
673 RegMask old = lrg.mask(); | |
674 uint old_size = lrg.mask_size(); | |
675 // Remove the bits from LRG 'r' from LRG 'l' so 'l' no | |
676 // longer interferes with 'r'. If 'l' requires aligned | |
677 // adjacent pairs, subtract out bit pairs. | |
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
678 assert(!lrg._is_vector || !lrg._fat_proj, "sanity"); |
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
679 if (lrg.num_regs() > 1 && !lrg._fat_proj) { |
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
680 RegMask r2mask = rmask; |
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
681 // Leave only aligned set of bits. |
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
682 r2mask.smear_to_sets(lrg.num_regs()); |
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
683 // It includes vector case. |
0 | 684 lrg.SUBTRACT( r2mask ); |
685 lrg.compute_set_mask_size(); | |
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
686 } else if( r_size != 1 ) { // fat proj |
0 | 687 lrg.SUBTRACT( rmask ); |
688 lrg.compute_set_mask_size(); | |
689 } else { // Common case: size 1 bound removal | |
690 if( lrg.mask().Member(r_reg) ) { | |
691 lrg.Remove(r_reg); | |
692 lrg.set_mask_size(lrg.mask().is_AllStack() ? 65535:old_size-1); | |
693 } | |
694 } | |
695 // If 'l' goes completely dry, it must spill. | |
696 if( lrg.not_free() ) { | |
697 // Give 'l' some kind of reasonable mask, so he picks up | |
698 // interferences (and will spill later). | |
699 lrg.set_mask( old ); | |
700 lrg.set_mask_size(old_size); | |
701 must_spill++; | |
702 lrg._must_spill = 1; | |
703 lrg.set_reg(OptoReg::Name(LRG::SPILL_REG)); | |
704 } | |
705 } | |
706 } // End of if bound | |
707 | |
708 // Now interference with everything that is live and has | |
709 // compatible register sets. | |
710 interfere_with_live(r,&liveout); | |
711 | |
712 } // End of if normal register-allocated value | |
713 | |
369
5f85534046c2
6750588: assert(lrg._area >= 0,"negative spill area") running NSK stmp0101 test
rasbold
parents:
295
diff
changeset
|
714 // Area remaining in the block |
5f85534046c2
6750588: assert(lrg._area >= 0,"negative spill area") running NSK stmp0101 test
rasbold
parents:
295
diff
changeset
|
715 inst_count--; |
5f85534046c2
6750588: assert(lrg._area >= 0,"negative spill area") running NSK stmp0101 test
rasbold
parents:
295
diff
changeset
|
716 cost = (inst_count <= 0) ? 0.0 : b->_freq * double(inst_count); |
0 | 717 |
718 // Make all inputs live | |
719 if( !n->is_Phi() ) { // Phi function uses come from prior block | |
720 JVMState* jvms = n->jvms(); | |
721 uint debug_start = jvms ? jvms->debug_start() : 999999; | |
722 // Start loop at 1 (skip control edge) for most Nodes. | |
723 // SCMemProj's might be the sole use of a StoreLConditional. | |
724 // While StoreLConditionals set memory (the SCMemProj use) | |
725 // they also def flags; if that flag def is unused the | |
726 // allocator sees a flag-setting instruction with no use of | |
727 // the flags and assumes it's dead. This keeps the (useless) | |
728 // flag-setting behavior alive while also keeping the (useful) | |
729 // memory update effect. | |
1212 | 730 for( uint k = ((n->Opcode() == Op_SCMemProj) ? 0:1); k < n->req(); k++ ) { |
0 | 731 Node *def = n->in(k); |
732 uint x = n2lidx(def); | |
733 if( !x ) continue; | |
734 LRG &lrg = lrgs(x); | |
735 // No use-side cost for spilling debug info | |
736 if( k < debug_start ) | |
737 // A USE costs twice block frequency (once for the Load, once | |
738 // for a Load-delay). Rematerialized uses only cost once. | |
739 lrg._cost += (def->rematerialize() ? b->_freq : (b->_freq + b->_freq)); | |
740 // It is live now | |
741 if( liveout.insert( x ) ) { | |
742 // Newly live things assumed live from here to top of block | |
743 lrg._area += cost; | |
744 // Adjust register pressure | |
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
745 if (lrg.mask().is_UP() && lrg.mask_size()) { |
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
1972
diff
changeset
|
746 if (lrg._is_float || lrg._is_vector) { |
0 | 747 pressure[1] += lrg.reg_pressure(); |
748 if( pressure[1] > b->_freg_pressure ) | |
749 b->_freg_pressure = pressure[1]; | |
750 } else if( lrg.mask().overlap(*Matcher::idealreg2regmask[Op_RegI]) ) { | |
751 pressure[0] += lrg.reg_pressure(); | |
752 if( pressure[0] > b->_reg_pressure ) | |
753 b->_reg_pressure = pressure[0]; | |
754 } | |
755 } | |
756 assert( pressure[0] == count_int_pressure (&liveout), "" ); | |
757 assert( pressure[1] == count_float_pressure(&liveout), "" ); | |
758 } | |
369
5f85534046c2
6750588: assert(lrg._area >= 0,"negative spill area") running NSK stmp0101 test
rasbold
parents:
295
diff
changeset
|
759 assert(!(lrg._area < 0.0), "negative spill area" ); |
0 | 760 } |
761 } | |
762 } // End of reverse pass over all instructions in block | |
763 | |
764 // If we run off the top of the block with high pressure and | |
765 // never see a hi-to-low pressure transition, just record that | |
766 // the whole block is high pressure. | |
767 if( pressure[0] > (uint)INTPRESSURE ) { | |
768 hrp_index[0] = 0; | |
769 if( pressure[0] > b->_reg_pressure ) | |
770 b->_reg_pressure = pressure[0]; | |
771 } | |
772 if( pressure[1] > (uint)FLOATPRESSURE ) { | |
773 hrp_index[1] = 0; | |
774 if( pressure[1] > b->_freg_pressure ) | |
775 b->_freg_pressure = pressure[1]; | |
776 } | |
777 | |
778 // Compute high pressure indice; avoid landing in the middle of projnodes | |
779 j = hrp_index[0]; | |
780 if( j < b->_nodes.size() && j < b->end_idx()+1 ) { | |
781 Node *cur = b->_nodes[j]; | |
782 while( cur->is_Proj() || (cur->is_MachNullCheck()) || cur->is_Catch() ) { | |
783 j--; | |
784 cur = b->_nodes[j]; | |
785 } | |
786 } | |
787 b->_ihrp_index = j; | |
788 j = hrp_index[1]; | |
789 if( j < b->_nodes.size() && j < b->end_idx()+1 ) { | |
790 Node *cur = b->_nodes[j]; | |
791 while( cur->is_Proj() || (cur->is_MachNullCheck()) || cur->is_Catch() ) { | |
792 j--; | |
793 cur = b->_nodes[j]; | |
794 } | |
795 } | |
796 b->_fhrp_index = j; | |
797 | |
798 #ifndef PRODUCT | |
799 // Gather Register Pressure Statistics | |
800 if( PrintOptoStatistics ) { | |
801 if( b->_reg_pressure > (uint)INTPRESSURE || b->_freg_pressure > (uint)FLOATPRESSURE ) | |
802 _high_pressure++; | |
803 else | |
804 _low_pressure++; | |
805 } | |
806 #endif | |
807 } // End of for all blocks | |
808 | |
809 return must_spill; | |
810 } |