Mercurial > hg > truffle
annotate src/share/vm/opto/superword.cpp @ 6144:5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
Summary: replace frequent C2 optimizer code patterns with new methods calls
Reviewed-by: kvn, twisti
Contributed-by: vladimir.x.ivanov@oracle.com
author | kvn |
---|---|
date | Tue, 12 Jun 2012 16:23:31 -0700 |
parents | dc682d9431f3 |
children | 8c92982cbbc4 |
rev | line source |
---|---|
0 | 1 /* |
1585 | 2 * Copyright (c) 2007, 2010, 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:
1058
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1058
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:
1058
diff
changeset
|
21 * questions. |
0 | 22 */ |
23 | |
1972 | 24 #include "precompiled.hpp" |
25 #include "compiler/compileLog.hpp" | |
26 #include "libadt/vectset.hpp" | |
27 #include "memory/allocation.inline.hpp" | |
28 #include "opto/addnode.hpp" | |
29 #include "opto/callnode.hpp" | |
30 #include "opto/divnode.hpp" | |
31 #include "opto/matcher.hpp" | |
32 #include "opto/memnode.hpp" | |
33 #include "opto/mulnode.hpp" | |
34 #include "opto/opcodes.hpp" | |
35 #include "opto/superword.hpp" | |
36 #include "opto/vectornode.hpp" | |
0 | 37 |
38 // | |
39 // S U P E R W O R D T R A N S F O R M | |
40 //============================================================================= | |
41 | |
42 //------------------------------SuperWord--------------------------- | |
43 SuperWord::SuperWord(PhaseIdealLoop* phase) : | |
44 _phase(phase), | |
45 _igvn(phase->_igvn), | |
46 _arena(phase->C->comp_arena()), | |
47 _packset(arena(), 8, 0, NULL), // packs for the current block | |
48 _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb | |
49 _block(arena(), 8, 0, NULL), // nodes in current block | |
50 _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside | |
51 _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads | |
52 _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails | |
53 _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node | |
54 _align_to_ref(NULL), // memory reference to align vectors to | |
55 _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs | |
56 _dg(_arena), // dependence graph | |
57 _visited(arena()), // visited node set | |
58 _post_visited(arena()), // post visited node set | |
59 _n_idx_list(arena(), 8), // scratch list of (node,index) pairs | |
60 _stk(arena(), 8, 0, NULL), // scratch stack of nodes | |
61 _nlist(arena(), 8, 0, NULL), // scratch list of nodes | |
62 _lpt(NULL), // loop tree node | |
63 _lp(NULL), // LoopNode | |
64 _bb(NULL), // basic block | |
65 _iv(NULL) // induction var | |
66 {} | |
67 | |
68 //------------------------------transform_loop--------------------------- | |
69 void SuperWord::transform_loop(IdealLoopTree* lpt) { | |
70 assert(lpt->_head->is_CountedLoop(), "must be"); | |
71 CountedLoopNode *cl = lpt->_head->as_CountedLoop(); | |
72 | |
3850
6987871cfb9b
7077439: Possible reference through NULL in loopPredicate.cpp:726
kvn
parents:
3842
diff
changeset
|
73 if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop |
6987871cfb9b
7077439: Possible reference through NULL in loopPredicate.cpp:726
kvn
parents:
3842
diff
changeset
|
74 |
0 | 75 if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops |
76 | |
77 // Check for no control flow in body (other than exit) | |
78 Node *cl_exit = cl->loopexit(); | |
79 if (cl_exit->in(0) != lpt->_head) return; | |
80 | |
105
a7d0f95410bd
6646020: assert(in_bb(n),"must be in block") in -Xcomp mode
never
parents:
72
diff
changeset
|
81 // Make sure the are no extra control users of the loop backedge |
a7d0f95410bd
6646020: assert(in_bb(n),"must be in block") in -Xcomp mode
never
parents:
72
diff
changeset
|
82 if (cl->back_control()->outcnt() != 1) { |
a7d0f95410bd
6646020: assert(in_bb(n),"must be in block") in -Xcomp mode
never
parents:
72
diff
changeset
|
83 return; |
a7d0f95410bd
6646020: assert(in_bb(n),"must be in block") in -Xcomp mode
never
parents:
72
diff
changeset
|
84 } |
a7d0f95410bd
6646020: assert(in_bb(n),"must be in block") in -Xcomp mode
never
parents:
72
diff
changeset
|
85 |
0 | 86 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit)))) |
87 CountedLoopEndNode* pre_end = get_pre_loop_end(cl); | |
88 if (pre_end == NULL) return; | |
89 Node *pre_opaq1 = pre_end->limit(); | |
90 if (pre_opaq1->Opcode() != Op_Opaque1) return; | |
91 | |
92 // Do vectors exist on this architecture? | |
93 if (vector_width_in_bytes() == 0) return; | |
94 | |
95 init(); // initialize data structures | |
96 | |
97 set_lpt(lpt); | |
98 set_lp(cl); | |
99 | |
100 // For now, define one block which is the entire loop body | |
101 set_bb(cl); | |
102 | |
103 assert(_packset.length() == 0, "packset must be empty"); | |
104 SLP_extract(); | |
105 } | |
106 | |
107 //------------------------------SLP_extract--------------------------- | |
108 // Extract the superword level parallelism | |
109 // | |
110 // 1) A reverse post-order of nodes in the block is constructed. By scanning | |
111 // this list from first to last, all definitions are visited before their uses. | |
112 // | |
113 // 2) A point-to-point dependence graph is constructed between memory references. | |
114 // This simplies the upcoming "independence" checker. | |
115 // | |
116 // 3) The maximum depth in the node graph from the beginning of the block | |
117 // to each node is computed. This is used to prune the graph search | |
118 // in the independence checker. | |
119 // | |
120 // 4) For integer types, the necessary bit width is propagated backwards | |
121 // from stores to allow packed operations on byte, char, and short | |
122 // integers. This reverses the promotion to type "int" that javac | |
123 // did for operations like: char c1,c2,c3; c1 = c2 + c3. | |
124 // | |
125 // 5) One of the memory references is picked to be an aligned vector reference. | |
126 // The pre-loop trip count is adjusted to align this reference in the | |
127 // unrolled body. | |
128 // | |
129 // 6) The initial set of pack pairs is seeded with memory references. | |
130 // | |
131 // 7) The set of pack pairs is extended by following use->def and def->use links. | |
132 // | |
133 // 8) The pairs are combined into vector sized packs. | |
134 // | |
135 // 9) Reorder the memory slices to co-locate members of the memory packs. | |
136 // | |
137 // 10) Generate ideal vector nodes for the final set of packs and where necessary, | |
138 // inserting scalar promotion, vector creation from multiple scalars, and | |
139 // extraction of scalar values from vectors. | |
140 // | |
141 void SuperWord::SLP_extract() { | |
142 | |
143 // Ready the block | |
144 | |
145 construct_bb(); | |
146 | |
147 dependence_graph(); | |
148 | |
149 compute_max_depth(); | |
150 | |
151 compute_vector_element_type(); | |
152 | |
153 // Attempt vectorization | |
154 | |
155 find_adjacent_refs(); | |
156 | |
157 extend_packlist(); | |
158 | |
159 combine_packs(); | |
160 | |
161 construct_my_pack_map(); | |
162 | |
163 filter_packs(); | |
164 | |
165 schedule(); | |
166 | |
167 output(); | |
168 } | |
169 | |
170 //------------------------------find_adjacent_refs--------------------------- | |
171 // Find the adjacent memory references and create pack pairs for them. | |
172 // This is the initial set of packs that will then be extended by | |
173 // following use->def and def->use links. The align positions are | |
174 // assigned relative to the reference "align_to_ref" | |
175 void SuperWord::find_adjacent_refs() { | |
176 // Get list of memory operations | |
177 Node_List memops; | |
178 for (int i = 0; i < _block.length(); i++) { | |
179 Node* n = _block.at(i); | |
29
d5fc211aea19
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
0
diff
changeset
|
180 if (n->is_Mem() && in_bb(n) && |
d5fc211aea19
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
0
diff
changeset
|
181 is_java_primitive(n->as_Mem()->memory_type())) { |
0 | 182 int align = memory_alignment(n->as_Mem(), 0); |
183 if (align != bottom_align) { | |
184 memops.push(n); | |
185 } | |
186 } | |
187 } | |
188 if (memops.size() == 0) return; | |
189 | |
190 // Find a memory reference to align to. The pre-loop trip count | |
191 // is modified to align this reference to a vector-aligned address | |
192 find_align_to_ref(memops); | |
193 if (align_to_ref() == NULL) return; | |
194 | |
195 SWPointer align_to_ref_p(align_to_ref(), this); | |
196 int offset = align_to_ref_p.offset_in_bytes(); | |
197 int scale = align_to_ref_p.scale_in_bytes(); | |
198 int vw = vector_width_in_bytes(); | |
199 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1; | |
200 int iv_adjustment = (stride_sign * vw - (offset % vw)) % vw; | |
201 | |
202 #ifndef PRODUCT | |
203 if (TraceSuperWord) | |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
204 tty->print_cr("\noffset = %d iv_adjustment = %d elt_align = %d scale = %d iv_stride = %d", |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
205 offset, iv_adjustment, align_to_ref_p.memory_size(), align_to_ref_p.scale_in_bytes(), iv_stride()); |
0 | 206 #endif |
207 | |
208 // Set alignment relative to "align_to_ref" | |
209 for (int i = memops.size() - 1; i >= 0; i--) { | |
210 MemNode* s = memops.at(i)->as_Mem(); | |
211 SWPointer p2(s, this); | |
212 if (p2.comparable(align_to_ref_p)) { | |
213 int align = memory_alignment(s, iv_adjustment); | |
214 set_alignment(s, align); | |
215 } else { | |
216 memops.remove(i); | |
217 } | |
218 } | |
219 | |
220 // Create initial pack pairs of memory operations | |
221 for (uint i = 0; i < memops.size(); i++) { | |
222 Node* s1 = memops.at(i); | |
223 for (uint j = 0; j < memops.size(); j++) { | |
224 Node* s2 = memops.at(j); | |
225 if (s1 != s2 && are_adjacent_refs(s1, s2)) { | |
226 int align = alignment(s1); | |
227 if (stmts_can_pack(s1, s2, align)) { | |
228 Node_List* pair = new Node_List(); | |
229 pair->push(s1); | |
230 pair->push(s2); | |
231 _packset.append(pair); | |
232 } | |
233 } | |
234 } | |
235 } | |
236 | |
237 #ifndef PRODUCT | |
238 if (TraceSuperWord) { | |
239 tty->print_cr("\nAfter find_adjacent_refs"); | |
240 print_packset(); | |
241 } | |
242 #endif | |
243 } | |
244 | |
245 //------------------------------find_align_to_ref--------------------------- | |
246 // Find a memory reference to align the loop induction variable to. | |
247 // Looks first at stores then at loads, looking for a memory reference | |
248 // with the largest number of references similar to it. | |
249 void SuperWord::find_align_to_ref(Node_List &memops) { | |
250 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0); | |
251 | |
252 // Count number of comparable memory ops | |
253 for (uint i = 0; i < memops.size(); i++) { | |
254 MemNode* s1 = memops.at(i)->as_Mem(); | |
255 SWPointer p1(s1, this); | |
256 // Discard if pre loop can't align this reference | |
257 if (!ref_is_alignable(p1)) { | |
258 *cmp_ct.adr_at(i) = 0; | |
259 continue; | |
260 } | |
261 for (uint j = i+1; j < memops.size(); j++) { | |
262 MemNode* s2 = memops.at(j)->as_Mem(); | |
263 if (isomorphic(s1, s2)) { | |
264 SWPointer p2(s2, this); | |
265 if (p1.comparable(p2)) { | |
266 (*cmp_ct.adr_at(i))++; | |
267 (*cmp_ct.adr_at(j))++; | |
268 } | |
269 } | |
270 } | |
271 } | |
272 | |
273 // Find Store (or Load) with the greatest number of "comparable" references | |
274 int max_ct = 0; | |
275 int max_idx = -1; | |
276 int min_size = max_jint; | |
277 int min_iv_offset = max_jint; | |
278 for (uint j = 0; j < memops.size(); j++) { | |
279 MemNode* s = memops.at(j)->as_Mem(); | |
280 if (s->is_Store()) { | |
281 SWPointer p(s, this); | |
282 if (cmp_ct.at(j) > max_ct || | |
283 cmp_ct.at(j) == max_ct && (data_size(s) < min_size || | |
284 data_size(s) == min_size && | |
285 p.offset_in_bytes() < min_iv_offset)) { | |
286 max_ct = cmp_ct.at(j); | |
287 max_idx = j; | |
288 min_size = data_size(s); | |
289 min_iv_offset = p.offset_in_bytes(); | |
290 } | |
291 } | |
292 } | |
293 // If no stores, look at loads | |
294 if (max_ct == 0) { | |
295 for (uint j = 0; j < memops.size(); j++) { | |
296 MemNode* s = memops.at(j)->as_Mem(); | |
297 if (s->is_Load()) { | |
298 SWPointer p(s, this); | |
299 if (cmp_ct.at(j) > max_ct || | |
300 cmp_ct.at(j) == max_ct && (data_size(s) < min_size || | |
301 data_size(s) == min_size && | |
302 p.offset_in_bytes() < min_iv_offset)) { | |
303 max_ct = cmp_ct.at(j); | |
304 max_idx = j; | |
305 min_size = data_size(s); | |
306 min_iv_offset = p.offset_in_bytes(); | |
307 } | |
308 } | |
309 } | |
310 } | |
311 | |
312 if (max_ct > 0) | |
313 set_align_to_ref(memops.at(max_idx)->as_Mem()); | |
314 | |
315 #ifndef PRODUCT | |
316 if (TraceSuperWord && Verbose) { | |
317 tty->print_cr("\nVector memops after find_align_to_refs"); | |
318 for (uint i = 0; i < memops.size(); i++) { | |
319 MemNode* s = memops.at(i)->as_Mem(); | |
320 s->dump(); | |
321 } | |
322 } | |
323 #endif | |
324 } | |
325 | |
326 //------------------------------ref_is_alignable--------------------------- | |
327 // Can the preloop align the reference to position zero in the vector? | |
328 bool SuperWord::ref_is_alignable(SWPointer& p) { | |
329 if (!p.has_iv()) { | |
330 return true; // no induction variable | |
331 } | |
332 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop()); | |
333 assert(pre_end->stride_is_con(), "pre loop stride is constant"); | |
334 int preloop_stride = pre_end->stride_con(); | |
335 | |
336 int span = preloop_stride * p.scale_in_bytes(); | |
337 | |
338 // Stride one accesses are alignable. | |
339 if (ABS(span) == p.memory_size()) | |
340 return true; | |
341 | |
342 // If initial offset from start of object is computable, | |
343 // compute alignment within the vector. | |
344 int vw = vector_width_in_bytes(); | |
345 if (vw % span == 0) { | |
346 Node* init_nd = pre_end->init_trip(); | |
347 if (init_nd->is_Con() && p.invar() == NULL) { | |
348 int init = init_nd->bottom_type()->is_int()->get_con(); | |
349 | |
350 int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes(); | |
351 assert(init_offset >= 0, "positive offset from object start"); | |
352 | |
353 if (span > 0) { | |
354 return (vw - (init_offset % vw)) % span == 0; | |
355 } else { | |
356 assert(span < 0, "nonzero stride * scale"); | |
357 return (init_offset % vw) % -span == 0; | |
358 } | |
359 } | |
360 } | |
361 return false; | |
362 } | |
363 | |
364 //---------------------------dependence_graph--------------------------- | |
365 // Construct dependency graph. | |
366 // Add dependence edges to load/store nodes for memory dependence | |
367 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x) | |
368 void SuperWord::dependence_graph() { | |
369 // First, assign a dependence node to each memory node | |
370 for (int i = 0; i < _block.length(); i++ ) { | |
371 Node *n = _block.at(i); | |
372 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) { | |
373 _dg.make_node(n); | |
374 } | |
375 } | |
376 | |
377 // For each memory slice, create the dependences | |
378 for (int i = 0; i < _mem_slice_head.length(); i++) { | |
379 Node* n = _mem_slice_head.at(i); | |
380 Node* n_tail = _mem_slice_tail.at(i); | |
381 | |
382 // Get slice in predecessor order (last is first) | |
383 mem_slice_preds(n_tail, n, _nlist); | |
384 | |
385 // Make the slice dependent on the root | |
386 DepMem* slice = _dg.dep(n); | |
387 _dg.make_edge(_dg.root(), slice); | |
388 | |
389 // Create a sink for the slice | |
390 DepMem* slice_sink = _dg.make_node(NULL); | |
391 _dg.make_edge(slice_sink, _dg.tail()); | |
392 | |
393 // Now visit each pair of memory ops, creating the edges | |
394 for (int j = _nlist.length() - 1; j >= 0 ; j--) { | |
395 Node* s1 = _nlist.at(j); | |
396 | |
397 // If no dependency yet, use slice | |
398 if (_dg.dep(s1)->in_cnt() == 0) { | |
399 _dg.make_edge(slice, s1); | |
400 } | |
401 SWPointer p1(s1->as_Mem(), this); | |
402 bool sink_dependent = true; | |
403 for (int k = j - 1; k >= 0; k--) { | |
404 Node* s2 = _nlist.at(k); | |
405 if (s1->is_Load() && s2->is_Load()) | |
406 continue; | |
407 SWPointer p2(s2->as_Mem(), this); | |
408 | |
409 int cmp = p1.cmp(p2); | |
410 if (SuperWordRTDepCheck && | |
411 p1.base() != p2.base() && p1.valid() && p2.valid()) { | |
412 // Create a runtime check to disambiguate | |
413 OrderedPair pp(p1.base(), p2.base()); | |
414 _disjoint_ptrs.append_if_missing(pp); | |
415 } else if (!SWPointer::not_equal(cmp)) { | |
416 // Possibly same address | |
417 _dg.make_edge(s1, s2); | |
418 sink_dependent = false; | |
419 } | |
420 } | |
421 if (sink_dependent) { | |
422 _dg.make_edge(s1, slice_sink); | |
423 } | |
424 } | |
425 #ifndef PRODUCT | |
426 if (TraceSuperWord) { | |
427 tty->print_cr("\nDependence graph for slice: %d", n->_idx); | |
428 for (int q = 0; q < _nlist.length(); q++) { | |
429 _dg.print(_nlist.at(q)); | |
430 } | |
431 tty->cr(); | |
432 } | |
433 #endif | |
434 _nlist.clear(); | |
435 } | |
436 | |
437 #ifndef PRODUCT | |
438 if (TraceSuperWord) { | |
439 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE"); | |
440 for (int r = 0; r < _disjoint_ptrs.length(); r++) { | |
441 _disjoint_ptrs.at(r).print(); | |
442 tty->cr(); | |
443 } | |
444 tty->cr(); | |
445 } | |
446 #endif | |
447 } | |
448 | |
449 //---------------------------mem_slice_preds--------------------------- | |
450 // Return a memory slice (node list) in predecessor order starting at "start" | |
451 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) { | |
452 assert(preds.length() == 0, "start empty"); | |
453 Node* n = start; | |
454 Node* prev = NULL; | |
455 while (true) { | |
456 assert(in_bb(n), "must be in block"); | |
457 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
458 Node* out = n->fast_out(i); | |
459 if (out->is_Load()) { | |
460 if (in_bb(out)) { | |
461 preds.push(out); | |
462 } | |
463 } else { | |
464 // FIXME | |
465 if (out->is_MergeMem() && !in_bb(out)) { | |
466 // Either unrolling is causing a memory edge not to disappear, | |
467 // or need to run igvn.optimize() again before SLP | |
468 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) { | |
469 // Ditto. Not sure what else to check further. | |
667 | 470 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) { |
0 | 471 // StoreCM has an input edge used as a precedence edge. |
472 // Maybe an issue when oop stores are vectorized. | |
473 } else { | |
474 assert(out == prev || prev == NULL, "no branches off of store slice"); | |
475 } | |
476 } | |
477 } | |
478 if (n == stop) break; | |
479 preds.push(n); | |
480 prev = n; | |
481 n = n->in(MemNode::Memory); | |
482 } | |
483 } | |
484 | |
485 //------------------------------stmts_can_pack--------------------------- | |
605 | 486 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and |
0 | 487 // s1 aligned at "align" |
488 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) { | |
987
00977607da34
6879921: CTW failure jdk6_18/hotspot/src/share/vm/utilities/globalDefinitions.cpp:268
cfang
parents:
985
diff
changeset
|
489 |
00977607da34
6879921: CTW failure jdk6_18/hotspot/src/share/vm/utilities/globalDefinitions.cpp:268
cfang
parents:
985
diff
changeset
|
490 // Do not use superword for non-primitives |
00977607da34
6879921: CTW failure jdk6_18/hotspot/src/share/vm/utilities/globalDefinitions.cpp:268
cfang
parents:
985
diff
changeset
|
491 if((s1->is_Mem() && !is_java_primitive(s1->as_Mem()->memory_type())) || |
00977607da34
6879921: CTW failure jdk6_18/hotspot/src/share/vm/utilities/globalDefinitions.cpp:268
cfang
parents:
985
diff
changeset
|
492 (s2->is_Mem() && !is_java_primitive(s2->as_Mem()->memory_type()))) |
00977607da34
6879921: CTW failure jdk6_18/hotspot/src/share/vm/utilities/globalDefinitions.cpp:268
cfang
parents:
985
diff
changeset
|
493 return false; |
00977607da34
6879921: CTW failure jdk6_18/hotspot/src/share/vm/utilities/globalDefinitions.cpp:268
cfang
parents:
985
diff
changeset
|
494 |
0 | 495 if (isomorphic(s1, s2)) { |
496 if (independent(s1, s2)) { | |
497 if (!exists_at(s1, 0) && !exists_at(s2, 1)) { | |
498 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) { | |
499 int s1_align = alignment(s1); | |
500 int s2_align = alignment(s2); | |
501 if (s1_align == top_align || s1_align == align) { | |
502 if (s2_align == top_align || s2_align == align + data_size(s1)) { | |
503 return true; | |
504 } | |
505 } | |
506 } | |
507 } | |
508 } | |
509 } | |
510 return false; | |
511 } | |
512 | |
513 //------------------------------exists_at--------------------------- | |
514 // Does s exist in a pack at position pos? | |
515 bool SuperWord::exists_at(Node* s, uint pos) { | |
516 for (int i = 0; i < _packset.length(); i++) { | |
517 Node_List* p = _packset.at(i); | |
518 if (p->at(pos) == s) { | |
519 return true; | |
520 } | |
521 } | |
522 return false; | |
523 } | |
524 | |
525 //------------------------------are_adjacent_refs--------------------------- | |
526 // Is s1 immediately before s2 in memory? | |
527 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) { | |
528 if (!s1->is_Mem() || !s2->is_Mem()) return false; | |
529 if (!in_bb(s1) || !in_bb(s2)) return false; | |
1585 | 530 |
531 // Do not use superword for non-primitives | |
532 if (!is_java_primitive(s1->as_Mem()->memory_type()) || | |
533 !is_java_primitive(s2->as_Mem()->memory_type())) { | |
534 return false; | |
535 } | |
536 | |
0 | 537 // FIXME - co_locate_pack fails on Stores in different mem-slices, so |
538 // only pack memops that are in the same alias set until that's fixed. | |
539 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) != | |
540 _phase->C->get_alias_index(s2->as_Mem()->adr_type())) | |
541 return false; | |
542 SWPointer p1(s1->as_Mem(), this); | |
543 SWPointer p2(s2->as_Mem(), this); | |
544 if (p1.base() != p2.base() || !p1.comparable(p2)) return false; | |
545 int diff = p2.offset_in_bytes() - p1.offset_in_bytes(); | |
546 return diff == data_size(s1); | |
547 } | |
548 | |
549 //------------------------------isomorphic--------------------------- | |
550 // Are s1 and s2 similar? | |
551 bool SuperWord::isomorphic(Node* s1, Node* s2) { | |
552 if (s1->Opcode() != s2->Opcode()) return false; | |
553 if (s1->req() != s2->req()) return false; | |
554 if (s1->in(0) != s2->in(0)) return false; | |
555 if (velt_type(s1) != velt_type(s2)) return false; | |
556 return true; | |
557 } | |
558 | |
559 //------------------------------independent--------------------------- | |
560 // Is there no data path from s1 to s2 or s2 to s1? | |
561 bool SuperWord::independent(Node* s1, Node* s2) { | |
562 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first"); | |
563 int d1 = depth(s1); | |
564 int d2 = depth(s2); | |
565 if (d1 == d2) return s1 != s2; | |
566 Node* deep = d1 > d2 ? s1 : s2; | |
567 Node* shallow = d1 > d2 ? s2 : s1; | |
568 | |
569 visited_clear(); | |
570 | |
571 return independent_path(shallow, deep); | |
572 } | |
573 | |
574 //------------------------------independent_path------------------------------ | |
575 // Helper for independent | |
576 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) { | |
577 if (dp >= 1000) return false; // stop deep recursion | |
578 visited_set(deep); | |
579 int shal_depth = depth(shallow); | |
580 assert(shal_depth <= depth(deep), "must be"); | |
581 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) { | |
582 Node* pred = preds.current(); | |
583 if (in_bb(pred) && !visited_test(pred)) { | |
584 if (shallow == pred) { | |
585 return false; | |
586 } | |
587 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) { | |
588 return false; | |
589 } | |
590 } | |
591 } | |
592 return true; | |
593 } | |
594 | |
595 //------------------------------set_alignment--------------------------- | |
596 void SuperWord::set_alignment(Node* s1, Node* s2, int align) { | |
597 set_alignment(s1, align); | |
598 set_alignment(s2, align + data_size(s1)); | |
599 } | |
600 | |
601 //------------------------------data_size--------------------------- | |
602 int SuperWord::data_size(Node* s) { | |
603 const Type* t = velt_type(s); | |
604 BasicType bt = t->array_element_basic_type(); | |
29
d5fc211aea19
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
0
diff
changeset
|
605 int bsize = type2aelembytes(bt); |
0 | 606 assert(bsize != 0, "valid size"); |
607 return bsize; | |
608 } | |
609 | |
610 //------------------------------extend_packlist--------------------------- | |
611 // Extend packset by following use->def and def->use links from pack members. | |
612 void SuperWord::extend_packlist() { | |
613 bool changed; | |
614 do { | |
615 changed = false; | |
616 for (int i = 0; i < _packset.length(); i++) { | |
617 Node_List* p = _packset.at(i); | |
618 changed |= follow_use_defs(p); | |
619 changed |= follow_def_uses(p); | |
620 } | |
621 } while (changed); | |
622 | |
623 #ifndef PRODUCT | |
624 if (TraceSuperWord) { | |
625 tty->print_cr("\nAfter extend_packlist"); | |
626 print_packset(); | |
627 } | |
628 #endif | |
629 } | |
630 | |
631 //------------------------------follow_use_defs--------------------------- | |
632 // Extend the packset by visiting operand definitions of nodes in pack p | |
633 bool SuperWord::follow_use_defs(Node_List* p) { | |
634 Node* s1 = p->at(0); | |
635 Node* s2 = p->at(1); | |
636 assert(p->size() == 2, "just checking"); | |
637 assert(s1->req() == s2->req(), "just checking"); | |
638 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking"); | |
639 | |
640 if (s1->is_Load()) return false; | |
641 | |
642 int align = alignment(s1); | |
643 bool changed = false; | |
644 int start = s1->is_Store() ? MemNode::ValueIn : 1; | |
645 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req(); | |
646 for (int j = start; j < end; j++) { | |
647 Node* t1 = s1->in(j); | |
648 Node* t2 = s2->in(j); | |
649 if (!in_bb(t1) || !in_bb(t2)) | |
650 continue; | |
651 if (stmts_can_pack(t1, t2, align)) { | |
652 if (est_savings(t1, t2) >= 0) { | |
653 Node_List* pair = new Node_List(); | |
654 pair->push(t1); | |
655 pair->push(t2); | |
656 _packset.append(pair); | |
657 set_alignment(t1, t2, align); | |
658 changed = true; | |
659 } | |
660 } | |
661 } | |
662 return changed; | |
663 } | |
664 | |
665 //------------------------------follow_def_uses--------------------------- | |
666 // Extend the packset by visiting uses of nodes in pack p | |
667 bool SuperWord::follow_def_uses(Node_List* p) { | |
668 bool changed = false; | |
669 Node* s1 = p->at(0); | |
670 Node* s2 = p->at(1); | |
671 assert(p->size() == 2, "just checking"); | |
672 assert(s1->req() == s2->req(), "just checking"); | |
673 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking"); | |
674 | |
675 if (s1->is_Store()) return false; | |
676 | |
677 int align = alignment(s1); | |
678 int savings = -1; | |
679 Node* u1 = NULL; | |
680 Node* u2 = NULL; | |
681 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) { | |
682 Node* t1 = s1->fast_out(i); | |
683 if (!in_bb(t1)) continue; | |
684 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) { | |
685 Node* t2 = s2->fast_out(j); | |
686 if (!in_bb(t2)) continue; | |
687 if (!opnd_positions_match(s1, t1, s2, t2)) | |
688 continue; | |
689 if (stmts_can_pack(t1, t2, align)) { | |
690 int my_savings = est_savings(t1, t2); | |
691 if (my_savings > savings) { | |
692 savings = my_savings; | |
693 u1 = t1; | |
694 u2 = t2; | |
695 } | |
696 } | |
697 } | |
698 } | |
699 if (savings >= 0) { | |
700 Node_List* pair = new Node_List(); | |
701 pair->push(u1); | |
702 pair->push(u2); | |
703 _packset.append(pair); | |
704 set_alignment(u1, u2, align); | |
705 changed = true; | |
706 } | |
707 return changed; | |
708 } | |
709 | |
710 //---------------------------opnd_positions_match------------------------- | |
711 // Is the use of d1 in u1 at the same operand position as d2 in u2? | |
712 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) { | |
713 uint ct = u1->req(); | |
714 if (ct != u2->req()) return false; | |
715 uint i1 = 0; | |
716 uint i2 = 0; | |
717 do { | |
718 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break; | |
719 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break; | |
720 if (i1 != i2) { | |
721 return false; | |
722 } | |
723 } while (i1 < ct); | |
724 return true; | |
725 } | |
726 | |
727 //------------------------------est_savings--------------------------- | |
728 // Estimate the savings from executing s1 and s2 as a pack | |
729 int SuperWord::est_savings(Node* s1, Node* s2) { | |
730 int save = 2 - 1; // 2 operations per instruction in packed form | |
731 | |
732 // inputs | |
733 for (uint i = 1; i < s1->req(); i++) { | |
734 Node* x1 = s1->in(i); | |
735 Node* x2 = s2->in(i); | |
736 if (x1 != x2) { | |
737 if (are_adjacent_refs(x1, x2)) { | |
738 save += adjacent_profit(x1, x2); | |
739 } else if (!in_packset(x1, x2)) { | |
740 save -= pack_cost(2); | |
741 } else { | |
742 save += unpack_cost(2); | |
743 } | |
744 } | |
745 } | |
746 | |
747 // uses of result | |
748 uint ct = 0; | |
749 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) { | |
750 Node* s1_use = s1->fast_out(i); | |
751 for (int j = 0; j < _packset.length(); j++) { | |
752 Node_List* p = _packset.at(j); | |
753 if (p->at(0) == s1_use) { | |
754 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) { | |
755 Node* s2_use = s2->fast_out(k); | |
756 if (p->at(p->size()-1) == s2_use) { | |
757 ct++; | |
758 if (are_adjacent_refs(s1_use, s2_use)) { | |
759 save += adjacent_profit(s1_use, s2_use); | |
760 } | |
761 } | |
762 } | |
763 } | |
764 } | |
765 } | |
766 | |
767 if (ct < s1->outcnt()) save += unpack_cost(1); | |
768 if (ct < s2->outcnt()) save += unpack_cost(1); | |
769 | |
770 return save; | |
771 } | |
772 | |
773 //------------------------------costs--------------------------- | |
774 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; } | |
775 int SuperWord::pack_cost(int ct) { return ct; } | |
776 int SuperWord::unpack_cost(int ct) { return ct; } | |
777 | |
778 //------------------------------combine_packs--------------------------- | |
779 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last | |
780 void SuperWord::combine_packs() { | |
781 bool changed; | |
782 do { | |
783 changed = false; | |
784 for (int i = 0; i < _packset.length(); i++) { | |
785 Node_List* p1 = _packset.at(i); | |
786 if (p1 == NULL) continue; | |
787 for (int j = 0; j < _packset.length(); j++) { | |
788 Node_List* p2 = _packset.at(j); | |
789 if (p2 == NULL) continue; | |
790 if (p1->at(p1->size()-1) == p2->at(0)) { | |
791 for (uint k = 1; k < p2->size(); k++) { | |
792 p1->push(p2->at(k)); | |
793 } | |
794 _packset.at_put(j, NULL); | |
795 changed = true; | |
796 } | |
797 } | |
798 } | |
799 } while (changed); | |
800 | |
801 for (int i = _packset.length() - 1; i >= 0; i--) { | |
802 Node_List* p1 = _packset.at(i); | |
803 if (p1 == NULL) { | |
804 _packset.remove_at(i); | |
805 } | |
806 } | |
807 | |
808 #ifndef PRODUCT | |
809 if (TraceSuperWord) { | |
810 tty->print_cr("\nAfter combine_packs"); | |
811 print_packset(); | |
812 } | |
813 #endif | |
814 } | |
815 | |
816 //-----------------------------construct_my_pack_map-------------------------- | |
817 // Construct the map from nodes to packs. Only valid after the | |
818 // point where a node is only in one pack (after combine_packs). | |
819 void SuperWord::construct_my_pack_map() { | |
820 Node_List* rslt = NULL; | |
821 for (int i = 0; i < _packset.length(); i++) { | |
822 Node_List* p = _packset.at(i); | |
823 for (uint j = 0; j < p->size(); j++) { | |
824 Node* s = p->at(j); | |
825 assert(my_pack(s) == NULL, "only in one pack"); | |
826 set_my_pack(s, p); | |
827 } | |
828 } | |
829 } | |
830 | |
831 //------------------------------filter_packs--------------------------- | |
832 // Remove packs that are not implemented or not profitable. | |
833 void SuperWord::filter_packs() { | |
834 | |
835 // Remove packs that are not implemented | |
836 for (int i = _packset.length() - 1; i >= 0; i--) { | |
837 Node_List* pk = _packset.at(i); | |
838 bool impl = implemented(pk); | |
839 if (!impl) { | |
840 #ifndef PRODUCT | |
841 if (TraceSuperWord && Verbose) { | |
842 tty->print_cr("Unimplemented"); | |
843 pk->at(0)->dump(); | |
844 } | |
845 #endif | |
846 remove_pack_at(i); | |
847 } | |
848 } | |
849 | |
850 // Remove packs that are not profitable | |
851 bool changed; | |
852 do { | |
853 changed = false; | |
854 for (int i = _packset.length() - 1; i >= 0; i--) { | |
855 Node_List* pk = _packset.at(i); | |
856 bool prof = profitable(pk); | |
857 if (!prof) { | |
858 #ifndef PRODUCT | |
859 if (TraceSuperWord && Verbose) { | |
860 tty->print_cr("Unprofitable"); | |
861 pk->at(0)->dump(); | |
862 } | |
863 #endif | |
864 remove_pack_at(i); | |
865 changed = true; | |
866 } | |
867 } | |
868 } while (changed); | |
869 | |
870 #ifndef PRODUCT | |
871 if (TraceSuperWord) { | |
872 tty->print_cr("\nAfter filter_packs"); | |
873 print_packset(); | |
874 tty->cr(); | |
875 } | |
876 #endif | |
877 } | |
878 | |
879 //------------------------------implemented--------------------------- | |
880 // Can code be generated for pack p? | |
881 bool SuperWord::implemented(Node_List* p) { | |
882 Node* p0 = p->at(0); | |
883 int vopc = VectorNode::opcode(p0->Opcode(), p->size(), velt_type(p0)); | |
884 return vopc > 0 && Matcher::has_match_rule(vopc); | |
885 } | |
886 | |
887 //------------------------------profitable--------------------------- | |
888 // For pack p, are all operands and all uses (with in the block) vector? | |
889 bool SuperWord::profitable(Node_List* p) { | |
890 Node* p0 = p->at(0); | |
891 uint start, end; | |
892 vector_opd_range(p0, &start, &end); | |
893 | |
894 // Return false if some input is not vector and inside block | |
895 for (uint i = start; i < end; i++) { | |
896 if (!is_vector_use(p0, i)) { | |
897 // For now, return false if not scalar promotion case (inputs are the same.) | |
605 | 898 // Later, implement PackNode and allow differing, non-vector inputs |
0 | 899 // (maybe just the ones from outside the block.) |
900 Node* p0_def = p0->in(i); | |
901 for (uint j = 1; j < p->size(); j++) { | |
902 Node* use = p->at(j); | |
903 Node* def = use->in(i); | |
904 if (p0_def != def) | |
905 return false; | |
906 } | |
907 } | |
908 } | |
909 if (!p0->is_Store()) { | |
910 // For now, return false if not all uses are vector. | |
911 // Later, implement ExtractNode and allow non-vector uses (maybe | |
912 // just the ones outside the block.) | |
913 for (uint i = 0; i < p->size(); i++) { | |
914 Node* def = p->at(i); | |
915 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { | |
916 Node* use = def->fast_out(j); | |
917 for (uint k = 0; k < use->req(); k++) { | |
918 Node* n = use->in(k); | |
919 if (def == n) { | |
920 if (!is_vector_use(use, k)) { | |
921 return false; | |
922 } | |
923 } | |
924 } | |
925 } | |
926 } | |
927 } | |
928 return true; | |
929 } | |
930 | |
931 //------------------------------schedule--------------------------- | |
932 // Adjust the memory graph for the packed operations | |
933 void SuperWord::schedule() { | |
934 | |
935 // Co-locate in the memory graph the members of each memory pack | |
936 for (int i = 0; i < _packset.length(); i++) { | |
937 co_locate_pack(_packset.at(i)); | |
938 } | |
939 } | |
940 | |
667 | 941 //-------------------------------remove_and_insert------------------- |
942 //remove "current" from its current position in the memory graph and insert | |
943 //it after the appropriate insertion point (lip or uip) | |
944 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip, | |
945 Node *uip, Unique_Node_List &sched_before) { | |
946 Node* my_mem = current->in(MemNode::Memory); | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
6045
diff
changeset
|
947 _igvn.rehash_node_delayed(current); |
667 | 948 _igvn.hash_delete(my_mem); |
949 | |
950 //remove current_store from its current position in the memmory graph | |
951 for (DUIterator i = current->outs(); current->has_out(i); i++) { | |
952 Node* use = current->out(i); | |
953 if (use->is_Mem()) { | |
954 assert(use->in(MemNode::Memory) == current, "must be"); | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
6045
diff
changeset
|
955 _igvn.rehash_node_delayed(use); |
667 | 956 if (use == prev) { // connect prev to my_mem |
957 use->set_req(MemNode::Memory, my_mem); | |
958 } else if (sched_before.member(use)) { | |
959 _igvn.hash_delete(uip); | |
960 use->set_req(MemNode::Memory, uip); | |
961 } else { | |
962 _igvn.hash_delete(lip); | |
963 use->set_req(MemNode::Memory, lip); | |
964 } | |
965 --i; //deleted this edge; rescan position | |
966 } | |
967 } | |
968 | |
969 bool sched_up = sched_before.member(current); | |
970 Node *insert_pt = sched_up ? uip : lip; | |
971 _igvn.hash_delete(insert_pt); | |
972 | |
973 // all uses of insert_pt's memory state should use current's instead | |
974 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) { | |
975 Node* use = insert_pt->out(i); | |
976 if (use->is_Mem()) { | |
977 assert(use->in(MemNode::Memory) == insert_pt, "must be"); | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
6045
diff
changeset
|
978 _igvn.replace_input_of(use, MemNode::Memory, current); |
667 | 979 --i; //deleted this edge; rescan position |
980 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) { | |
981 uint pos; //lip (lower insert point) must be the last one in the memory slice | |
982 for (pos=1; pos < use->req(); pos++) { | |
983 if (use->in(pos) == insert_pt) break; | |
984 } | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
6045
diff
changeset
|
985 _igvn.replace_input_of(use, pos, current); |
667 | 986 --i; |
987 } | |
988 } | |
989 | |
990 //connect current to insert_pt | |
991 current->set_req(MemNode::Memory, insert_pt); | |
992 } | |
993 | |
994 //------------------------------co_locate_pack---------------------------------- | |
995 // To schedule a store pack, we need to move any sandwiched memory ops either before | |
996 // or after the pack, based upon dependence information: | |
997 // (1) If any store in the pack depends on the sandwiched memory op, the | |
998 // sandwiched memory op must be scheduled BEFORE the pack; | |
999 // (2) If a sandwiched memory op depends on any store in the pack, the | |
1000 // sandwiched memory op must be scheduled AFTER the pack; | |
1001 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched | |
1002 // memory op (say memB), memB must be scheduled before memA. So, if memA is | |
1003 // scheduled before the pack, memB must also be scheduled before the pack; | |
1004 // (4) If there is no dependence restriction for a sandwiched memory op, we simply | |
1005 // schedule this store AFTER the pack | |
1006 // (5) We know there is no dependence cycle, so there in no other case; | |
1007 // (6) Finally, all memory ops in another single pack should be moved in the same direction. | |
1008 // | |
952 | 1009 // To schedule a load pack, we use the memory state of either the first or the last load in |
1010 // the pack, based on the dependence constraint. | |
0 | 1011 void SuperWord::co_locate_pack(Node_List* pk) { |
1012 if (pk->at(0)->is_Store()) { | |
1013 MemNode* first = executed_first(pk)->as_Mem(); | |
1014 MemNode* last = executed_last(pk)->as_Mem(); | |
667 | 1015 Unique_Node_List schedule_before_pack; |
1016 Unique_Node_List memops; | |
1017 | |
0 | 1018 MemNode* current = last->in(MemNode::Memory)->as_Mem(); |
667 | 1019 MemNode* previous = last; |
0 | 1020 while (true) { |
1021 assert(in_bb(current), "stay in block"); | |
667 | 1022 memops.push(previous); |
1023 for (DUIterator i = current->outs(); current->has_out(i); i++) { | |
1024 Node* use = current->out(i); | |
1025 if (use->is_Mem() && use != previous) | |
1026 memops.push(use); | |
1027 } | |
1028 if(current == first) break; | |
1029 previous = current; | |
1030 current = current->in(MemNode::Memory)->as_Mem(); | |
1031 } | |
1032 | |
1033 // determine which memory operations should be scheduled before the pack | |
1034 for (uint i = 1; i < memops.size(); i++) { | |
1035 Node *s1 = memops.at(i); | |
1036 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) { | |
1037 for (uint j = 0; j< i; j++) { | |
1038 Node *s2 = memops.at(j); | |
1039 if (!independent(s1, s2)) { | |
1040 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) { | |
1041 schedule_before_pack.push(s1); //s1 must be scheduled before | |
1042 Node_List* mem_pk = my_pack(s1); | |
1043 if (mem_pk != NULL) { | |
1044 for (uint ii = 0; ii < mem_pk->size(); ii++) { | |
1045 Node* s = mem_pk->at(ii); // follow partner | |
1046 if (memops.member(s) && !schedule_before_pack.member(s)) | |
1047 schedule_before_pack.push(s); | |
1048 } | |
1049 } | |
1050 } | |
1051 } | |
1052 } | |
1053 } | |
1054 } | |
1055 | |
1056 MemNode* lower_insert_pt = last; | |
1057 Node* upper_insert_pt = first->in(MemNode::Memory); | |
1058 previous = last; //previous store in pk | |
1059 current = last->in(MemNode::Memory)->as_Mem(); | |
1060 | |
1061 //start scheduling from "last" to "first" | |
1062 while (true) { | |
1063 assert(in_bb(current), "stay in block"); | |
1064 assert(in_pack(previous, pk), "previous stays in pack"); | |
0 | 1065 Node* my_mem = current->in(MemNode::Memory); |
667 | 1066 |
0 | 1067 if (in_pack(current, pk)) { |
667 | 1068 // Forward users of my memory state (except "previous) to my input memory state |
0 | 1069 _igvn.hash_delete(current); |
1070 for (DUIterator i = current->outs(); current->has_out(i); i++) { | |
1071 Node* use = current->out(i); | |
667 | 1072 if (use->is_Mem() && use != previous) { |
0 | 1073 assert(use->in(MemNode::Memory) == current, "must be"); |
667 | 1074 if (schedule_before_pack.member(use)) { |
1075 _igvn.hash_delete(upper_insert_pt); | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
6045
diff
changeset
|
1076 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt); |
667 | 1077 } else { |
1078 _igvn.hash_delete(lower_insert_pt); | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
6045
diff
changeset
|
1079 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt); |
667 | 1080 } |
0 | 1081 --i; // deleted this edge; rescan position |
1082 } | |
1083 } | |
667 | 1084 previous = current; |
1085 } else { // !in_pack(current, pk) ==> a sandwiched store | |
1086 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack); | |
0 | 1087 } |
667 | 1088 |
0 | 1089 if (current == first) break; |
1090 current = my_mem->as_Mem(); | |
667 | 1091 } // end while |
1092 } else if (pk->at(0)->is_Load()) { //load | |
952 | 1093 // all loads in the pack should have the same memory state. By default, |
1094 // we use the memory state of the last load. However, if any load could | |
1095 // not be moved down due to the dependence constraint, we use the memory | |
1096 // state of the first load. | |
1097 Node* last_mem = executed_last(pk)->in(MemNode::Memory); | |
1098 Node* first_mem = executed_first(pk)->in(MemNode::Memory); | |
1099 bool schedule_last = true; | |
1100 for (uint i = 0; i < pk->size(); i++) { | |
1101 Node* ld = pk->at(i); | |
1102 for (Node* current = last_mem; current != ld->in(MemNode::Memory); | |
1103 current=current->in(MemNode::Memory)) { | |
1104 assert(current != first_mem, "corrupted memory graph"); | |
1105 if(current->is_Mem() && !independent(current, ld)){ | |
1106 schedule_last = false; // a later store depends on this load | |
1107 break; | |
1108 } | |
1109 } | |
1110 } | |
1111 | |
1112 Node* mem_input = schedule_last ? last_mem : first_mem; | |
1113 _igvn.hash_delete(mem_input); | |
1114 // Give each load the same memory state | |
0 | 1115 for (uint i = 0; i < pk->size(); i++) { |
1116 LoadNode* ld = pk->at(i)->as_Load(); | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
6045
diff
changeset
|
1117 _igvn.replace_input_of(ld, MemNode::Memory, mem_input); |
0 | 1118 } |
1119 } | |
1120 } | |
1121 | |
1122 //------------------------------output--------------------------- | |
1123 // Convert packs into vector node operations | |
1124 void SuperWord::output() { | |
1125 if (_packset.length() == 0) return; | |
1126 | |
2445 | 1127 #ifndef PRODUCT |
1128 if (TraceLoopOpts) { | |
1129 tty->print("SuperWord "); | |
1130 lpt()->dump_head(); | |
1131 } | |
1132 #endif | |
1133 | |
0 | 1134 // MUST ENSURE main loop's initial value is properly aligned: |
1135 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0 | |
1136 | |
1137 align_initial_loop_index(align_to_ref()); | |
1138 | |
1139 // Insert extract (unpack) operations for scalar uses | |
1140 for (int i = 0; i < _packset.length(); i++) { | |
1141 insert_extracts(_packset.at(i)); | |
1142 } | |
1143 | |
1144 for (int i = 0; i < _block.length(); i++) { | |
1145 Node* n = _block.at(i); | |
1146 Node_List* p = my_pack(n); | |
1147 if (p && n == executed_last(p)) { | |
1148 uint vlen = p->size(); | |
1149 Node* vn = NULL; | |
1150 Node* low_adr = p->at(0); | |
1151 Node* first = executed_first(p); | |
1152 if (n->is_Load()) { | |
1153 int opc = n->Opcode(); | |
1154 Node* ctl = n->in(MemNode::Control); | |
1155 Node* mem = first->in(MemNode::Memory); | |
1156 Node* adr = low_adr->in(MemNode::Address); | |
1157 const TypePtr* atyp = n->adr_type(); | |
1158 vn = VectorLoadNode::make(_phase->C, opc, ctl, mem, adr, atyp, vlen); | |
1159 | |
1160 } else if (n->is_Store()) { | |
1161 // Promote value to be stored to vector | |
3842 | 1162 Node* val = vector_opd(p, MemNode::ValueIn); |
0 | 1163 |
1164 int opc = n->Opcode(); | |
1165 Node* ctl = n->in(MemNode::Control); | |
1166 Node* mem = first->in(MemNode::Memory); | |
1167 Node* adr = low_adr->in(MemNode::Address); | |
1168 const TypePtr* atyp = n->adr_type(); | |
1169 vn = VectorStoreNode::make(_phase->C, opc, ctl, mem, adr, atyp, val, vlen); | |
1170 | |
1171 } else if (n->req() == 3) { | |
1172 // Promote operands to vector | |
1173 Node* in1 = vector_opd(p, 1); | |
1174 Node* in2 = vector_opd(p, 2); | |
1175 vn = VectorNode::make(_phase->C, n->Opcode(), in1, in2, vlen, velt_type(n)); | |
1176 | |
1177 } else { | |
1178 ShouldNotReachHere(); | |
1179 } | |
1180 | |
1181 _phase->_igvn.register_new_node_with_optimizer(vn); | |
1182 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0))); | |
1183 for (uint j = 0; j < p->size(); j++) { | |
1184 Node* pm = p->at(j); | |
1621
6027dddc26c6
6677629: PhaseIterGVN::subsume_node() should call hash_delete() and add_users_to_worklist()
kvn
parents:
1585
diff
changeset
|
1185 _igvn.replace_node(pm, vn); |
0 | 1186 } |
1187 _igvn._worklist.push(vn); | |
1188 } | |
1189 } | |
1190 } | |
1191 | |
1192 //------------------------------vector_opd--------------------------- | |
1193 // Create a vector operand for the nodes in pack p for operand: in(opd_idx) | |
3842 | 1194 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) { |
0 | 1195 Node* p0 = p->at(0); |
1196 uint vlen = p->size(); | |
1197 Node* opd = p0->in(opd_idx); | |
1198 | |
1199 bool same_opd = true; | |
1200 for (uint i = 1; i < vlen; i++) { | |
1201 Node* pi = p->at(i); | |
1202 Node* in = pi->in(opd_idx); | |
1203 if (opd != in) { | |
1204 same_opd = false; | |
1205 break; | |
1206 } | |
1207 } | |
1208 | |
1209 if (same_opd) { | |
3842 | 1210 if (opd->is_Vector() || opd->is_VectorLoad()) { |
1211 return opd; // input is matching vector | |
0 | 1212 } |
3842 | 1213 assert(!opd->is_VectorStore(), "such vector is not expected here"); |
6045 | 1214 // Convert scalar input to vector with the same number of elements as |
1215 // p0's vector. Use p0's type because size of operand's container in | |
1216 // vector should match p0's size regardless operand's size. | |
1217 const Type* p0_t = velt_type(p0); | |
1218 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, p0_t); | |
0 | 1219 |
1220 _phase->_igvn.register_new_node_with_optimizer(vn); | |
1221 _phase->set_ctrl(vn, _phase->get_ctrl(opd)); | |
1222 return vn; | |
1223 } | |
1224 | |
1225 // Insert pack operation | |
6045 | 1226 const Type* p0_t = velt_type(p0); |
1227 PackNode* pk = PackNode::make(_phase->C, opd, p0_t); | |
1228 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); ) | |
0 | 1229 |
1230 for (uint i = 1; i < vlen; i++) { | |
1231 Node* pi = p->at(i); | |
1232 Node* in = pi->in(opd_idx); | |
1233 assert(my_pack(in) == NULL, "Should already have been unpacked"); | |
6045 | 1234 assert(opd_bt == in->bottom_type()->basic_type(), "all same type"); |
0 | 1235 pk->add_opd(in); |
1236 } | |
1237 _phase->_igvn.register_new_node_with_optimizer(pk); | |
1238 _phase->set_ctrl(pk, _phase->get_ctrl(opd)); | |
1239 return pk; | |
1240 } | |
1241 | |
1242 //------------------------------insert_extracts--------------------------- | |
1243 // If a use of pack p is not a vector use, then replace the | |
1244 // use with an extract operation. | |
1245 void SuperWord::insert_extracts(Node_List* p) { | |
1246 if (p->at(0)->is_Store()) return; | |
1247 assert(_n_idx_list.is_empty(), "empty (node,index) list"); | |
1248 | |
1249 // Inspect each use of each pack member. For each use that is | |
1250 // not a vector use, replace the use with an extract operation. | |
1251 | |
1252 for (uint i = 0; i < p->size(); i++) { | |
1253 Node* def = p->at(i); | |
1254 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { | |
1255 Node* use = def->fast_out(j); | |
1256 for (uint k = 0; k < use->req(); k++) { | |
1257 Node* n = use->in(k); | |
1258 if (def == n) { | |
1259 if (!is_vector_use(use, k)) { | |
1260 _n_idx_list.push(use, k); | |
1261 } | |
1262 } | |
1263 } | |
1264 } | |
1265 } | |
1266 | |
1267 while (_n_idx_list.is_nonempty()) { | |
1268 Node* use = _n_idx_list.node(); | |
1269 int idx = _n_idx_list.index(); | |
1270 _n_idx_list.pop(); | |
1271 Node* def = use->in(idx); | |
1272 | |
1273 // Insert extract operation | |
1274 _igvn.hash_delete(def); | |
1275 int def_pos = alignment(def) / data_size(def); | |
1276 const Type* def_t = velt_type(def); | |
1277 | |
1278 Node* ex = ExtractNode::make(_phase->C, def, def_pos, def_t); | |
1279 _phase->_igvn.register_new_node_with_optimizer(ex); | |
1280 _phase->set_ctrl(ex, _phase->get_ctrl(def)); | |
6144
5e990493719e
7173340: C2: code cleanup: use PhaseIterGVN::replace_edge(Node*, int, Node*) where applicable
kvn
parents:
6045
diff
changeset
|
1281 _igvn.replace_input_of(use, idx, ex); |
0 | 1282 _igvn._worklist.push(def); |
1283 | |
1284 bb_insert_after(ex, bb_idx(def)); | |
1285 set_velt_type(ex, def_t); | |
1286 } | |
1287 } | |
1288 | |
1289 //------------------------------is_vector_use--------------------------- | |
1290 // Is use->in(u_idx) a vector use? | |
1291 bool SuperWord::is_vector_use(Node* use, int u_idx) { | |
1292 Node_List* u_pk = my_pack(use); | |
1293 if (u_pk == NULL) return false; | |
1294 Node* def = use->in(u_idx); | |
1295 Node_List* d_pk = my_pack(def); | |
1296 if (d_pk == NULL) { | |
1297 // check for scalar promotion | |
1298 Node* n = u_pk->at(0)->in(u_idx); | |
1299 for (uint i = 1; i < u_pk->size(); i++) { | |
1300 if (u_pk->at(i)->in(u_idx) != n) return false; | |
1301 } | |
1302 return true; | |
1303 } | |
1304 if (u_pk->size() != d_pk->size()) | |
1305 return false; | |
1306 for (uint i = 0; i < u_pk->size(); i++) { | |
1307 Node* ui = u_pk->at(i); | |
1308 Node* di = d_pk->at(i); | |
1309 if (ui->in(u_idx) != di || alignment(ui) != alignment(di)) | |
1310 return false; | |
1311 } | |
1312 return true; | |
1313 } | |
1314 | |
1315 //------------------------------construct_bb--------------------------- | |
1316 // Construct reverse postorder list of block members | |
1317 void SuperWord::construct_bb() { | |
1318 Node* entry = bb(); | |
1319 | |
1320 assert(_stk.length() == 0, "stk is empty"); | |
1321 assert(_block.length() == 0, "block is empty"); | |
1322 assert(_data_entry.length() == 0, "data_entry is empty"); | |
1323 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty"); | |
1324 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty"); | |
1325 | |
1326 // Find non-control nodes with no inputs from within block, | |
1327 // create a temporary map from node _idx to bb_idx for use | |
1328 // by the visited and post_visited sets, | |
1329 // and count number of nodes in block. | |
1330 int bb_ct = 0; | |
1331 for (uint i = 0; i < lpt()->_body.size(); i++ ) { | |
1332 Node *n = lpt()->_body.at(i); | |
1333 set_bb_idx(n, i); // Create a temporary map | |
1334 if (in_bb(n)) { | |
1335 bb_ct++; | |
1336 if (!n->is_CFG()) { | |
1337 bool found = false; | |
1338 for (uint j = 0; j < n->req(); j++) { | |
1339 Node* def = n->in(j); | |
1340 if (def && in_bb(def)) { | |
1341 found = true; | |
1342 break; | |
1343 } | |
1344 } | |
1345 if (!found) { | |
1346 assert(n != entry, "can't be entry"); | |
1347 _data_entry.push(n); | |
1348 } | |
1349 } | |
1350 } | |
1351 } | |
1352 | |
1353 // Find memory slices (head and tail) | |
1354 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) { | |
1355 Node *n = lp()->fast_out(i); | |
1356 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) { | |
1357 Node* n_tail = n->in(LoopNode::LoopBackControl); | |
253
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
235
diff
changeset
|
1358 if (n_tail != n->in(LoopNode::EntryControl)) { |
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
235
diff
changeset
|
1359 _mem_slice_head.push(n); |
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
235
diff
changeset
|
1360 _mem_slice_tail.push(n_tail); |
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
235
diff
changeset
|
1361 } |
0 | 1362 } |
1363 } | |
1364 | |
1365 // Create an RPO list of nodes in block | |
1366 | |
1367 visited_clear(); | |
1368 post_visited_clear(); | |
1369 | |
1370 // Push all non-control nodes with no inputs from within block, then control entry | |
1371 for (int j = 0; j < _data_entry.length(); j++) { | |
1372 Node* n = _data_entry.at(j); | |
1373 visited_set(n); | |
1374 _stk.push(n); | |
1375 } | |
1376 visited_set(entry); | |
1377 _stk.push(entry); | |
1378 | |
1379 // Do a depth first walk over out edges | |
1380 int rpo_idx = bb_ct - 1; | |
1381 int size; | |
1382 while ((size = _stk.length()) > 0) { | |
1383 Node* n = _stk.top(); // Leave node on stack | |
1384 if (!visited_test_set(n)) { | |
1385 // forward arc in graph | |
1386 } else if (!post_visited_test(n)) { | |
1387 // cross or back arc | |
1388 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { | |
1389 Node *use = n->fast_out(i); | |
1390 if (in_bb(use) && !visited_test(use) && | |
1391 // Don't go around backedge | |
1392 (!use->is_Phi() || n == entry)) { | |
1393 _stk.push(use); | |
1394 } | |
1395 } | |
1396 if (_stk.length() == size) { | |
1397 // There were no additional uses, post visit node now | |
1398 _stk.pop(); // Remove node from stack | |
1399 assert(rpo_idx >= 0, ""); | |
1400 _block.at_put_grow(rpo_idx, n); | |
1401 rpo_idx--; | |
1402 post_visited_set(n); | |
1403 assert(rpo_idx >= 0 || _stk.is_empty(), ""); | |
1404 } | |
1405 } else { | |
1406 _stk.pop(); // Remove post-visited node from stack | |
1407 } | |
1408 } | |
1409 | |
1410 // Create real map of block indices for nodes | |
1411 for (int j = 0; j < _block.length(); j++) { | |
1412 Node* n = _block.at(j); | |
1413 set_bb_idx(n, j); | |
1414 } | |
1415 | |
1416 initialize_bb(); // Ensure extra info is allocated. | |
1417 | |
1418 #ifndef PRODUCT | |
1419 if (TraceSuperWord) { | |
1420 print_bb(); | |
1421 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE"); | |
1422 for (int m = 0; m < _data_entry.length(); m++) { | |
1423 tty->print("%3d ", m); | |
1424 _data_entry.at(m)->dump(); | |
1425 } | |
1426 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE"); | |
1427 for (int m = 0; m < _mem_slice_head.length(); m++) { | |
1428 tty->print("%3d ", m); _mem_slice_head.at(m)->dump(); | |
1429 tty->print(" "); _mem_slice_tail.at(m)->dump(); | |
1430 } | |
1431 } | |
1432 #endif | |
1433 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found"); | |
1434 } | |
1435 | |
1436 //------------------------------initialize_bb--------------------------- | |
1437 // Initialize per node info | |
1438 void SuperWord::initialize_bb() { | |
1439 Node* last = _block.at(_block.length() - 1); | |
1440 grow_node_info(bb_idx(last)); | |
1441 } | |
1442 | |
1443 //------------------------------bb_insert_after--------------------------- | |
1444 // Insert n into block after pos | |
1445 void SuperWord::bb_insert_after(Node* n, int pos) { | |
1446 int n_pos = pos + 1; | |
1447 // Make room | |
1448 for (int i = _block.length() - 1; i >= n_pos; i--) { | |
1449 _block.at_put_grow(i+1, _block.at(i)); | |
1450 } | |
1451 for (int j = _node_info.length() - 1; j >= n_pos; j--) { | |
1452 _node_info.at_put_grow(j+1, _node_info.at(j)); | |
1453 } | |
1454 // Set value | |
1455 _block.at_put_grow(n_pos, n); | |
1456 _node_info.at_put_grow(n_pos, SWNodeInfo::initial); | |
1457 // Adjust map from node->_idx to _block index | |
1458 for (int i = n_pos; i < _block.length(); i++) { | |
1459 set_bb_idx(_block.at(i), i); | |
1460 } | |
1461 } | |
1462 | |
1463 //------------------------------compute_max_depth--------------------------- | |
1464 // Compute max depth for expressions from beginning of block | |
1465 // Use to prune search paths during test for independence. | |
1466 void SuperWord::compute_max_depth() { | |
1467 int ct = 0; | |
1468 bool again; | |
1469 do { | |
1470 again = false; | |
1471 for (int i = 0; i < _block.length(); i++) { | |
1472 Node* n = _block.at(i); | |
1473 if (!n->is_Phi()) { | |
1474 int d_orig = depth(n); | |
1475 int d_in = 0; | |
1476 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) { | |
1477 Node* pred = preds.current(); | |
1478 if (in_bb(pred)) { | |
1479 d_in = MAX2(d_in, depth(pred)); | |
1480 } | |
1481 } | |
1482 if (d_in + 1 != d_orig) { | |
1483 set_depth(n, d_in + 1); | |
1484 again = true; | |
1485 } | |
1486 } | |
1487 } | |
1488 ct++; | |
1489 } while (again); | |
1490 #ifndef PRODUCT | |
1491 if (TraceSuperWord && Verbose) | |
1492 tty->print_cr("compute_max_depth iterated: %d times", ct); | |
1493 #endif | |
1494 } | |
1495 | |
1496 //-------------------------compute_vector_element_type----------------------- | |
1497 // Compute necessary vector element type for expressions | |
1498 // This propagates backwards a narrower integer type when the | |
1499 // upper bits of the value are not needed. | |
1500 // Example: char a,b,c; a = b + c; | |
1501 // Normally the type of the add is integer, but for packed character | |
1502 // operations the type of the add needs to be char. | |
1503 void SuperWord::compute_vector_element_type() { | |
1504 #ifndef PRODUCT | |
1505 if (TraceSuperWord && Verbose) | |
1506 tty->print_cr("\ncompute_velt_type:"); | |
1507 #endif | |
1508 | |
1509 // Initial type | |
1510 for (int i = 0; i < _block.length(); i++) { | |
1511 Node* n = _block.at(i); | |
1512 const Type* t = n->is_Mem() ? Type::get_const_basic_type(n->as_Mem()->memory_type()) | |
1513 : _igvn.type(n); | |
1514 const Type* vt = container_type(t); | |
1515 set_velt_type(n, vt); | |
1516 } | |
1517 | |
1518 // Propagate narrowed type backwards through operations | |
1519 // that don't depend on higher order bits | |
1520 for (int i = _block.length() - 1; i >= 0; i--) { | |
1521 Node* n = _block.at(i); | |
1522 // Only integer types need be examined | |
1523 if (n->bottom_type()->isa_int()) { | |
1524 uint start, end; | |
1525 vector_opd_range(n, &start, &end); | |
1526 const Type* vt = velt_type(n); | |
1527 | |
1528 for (uint j = start; j < end; j++) { | |
1529 Node* in = n->in(j); | |
1530 // Don't propagate through a type conversion | |
1531 if (n->bottom_type() != in->bottom_type()) | |
1532 continue; | |
1533 switch(in->Opcode()) { | |
1534 case Op_AddI: case Op_AddL: | |
1535 case Op_SubI: case Op_SubL: | |
1536 case Op_MulI: case Op_MulL: | |
1537 case Op_AndI: case Op_AndL: | |
1538 case Op_OrI: case Op_OrL: | |
1539 case Op_XorI: case Op_XorL: | |
1540 case Op_LShiftI: case Op_LShiftL: | |
1541 case Op_CMoveI: case Op_CMoveL: | |
1542 if (in_bb(in)) { | |
1543 bool same_type = true; | |
1544 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) { | |
1545 Node *use = in->fast_out(k); | |
1546 if (!in_bb(use) || velt_type(use) != vt) { | |
1547 same_type = false; | |
1548 break; | |
1549 } | |
1550 } | |
1551 if (same_type) { | |
1552 set_velt_type(in, vt); | |
1553 } | |
1554 } | |
1555 } | |
1556 } | |
1557 } | |
1558 } | |
1559 #ifndef PRODUCT | |
1560 if (TraceSuperWord && Verbose) { | |
1561 for (int i = 0; i < _block.length(); i++) { | |
1562 Node* n = _block.at(i); | |
1563 velt_type(n)->dump(); | |
1564 tty->print("\t"); | |
1565 n->dump(); | |
1566 } | |
1567 } | |
1568 #endif | |
1569 } | |
1570 | |
1571 //------------------------------memory_alignment--------------------------- | |
1572 // Alignment within a vector memory reference | |
1573 int SuperWord::memory_alignment(MemNode* s, int iv_adjust_in_bytes) { | |
1574 SWPointer p(s, this); | |
1575 if (!p.valid()) { | |
1576 return bottom_align; | |
1577 } | |
1578 int offset = p.offset_in_bytes(); | |
1579 offset += iv_adjust_in_bytes; | |
1580 int off_rem = offset % vector_width_in_bytes(); | |
1581 int off_mod = off_rem >= 0 ? off_rem : off_rem + vector_width_in_bytes(); | |
1582 return off_mod; | |
1583 } | |
1584 | |
1585 //---------------------------container_type--------------------------- | |
1586 // Smallest type containing range of values | |
1587 const Type* SuperWord::container_type(const Type* t) { | |
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
113
diff
changeset
|
1588 const Type* tp = t->make_ptr(); |
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
113
diff
changeset
|
1589 if (tp && tp->isa_aryptr()) { |
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
113
diff
changeset
|
1590 t = tp->is_aryptr()->elem(); |
0 | 1591 } |
1592 if (t->basic_type() == T_INT) { | |
1593 if (t->higher_equal(TypeInt::BOOL)) return TypeInt::BOOL; | |
1594 if (t->higher_equal(TypeInt::BYTE)) return TypeInt::BYTE; | |
1595 if (t->higher_equal(TypeInt::CHAR)) return TypeInt::CHAR; | |
1596 if (t->higher_equal(TypeInt::SHORT)) return TypeInt::SHORT; | |
1597 return TypeInt::INT; | |
1598 } | |
1599 return t; | |
1600 } | |
1601 | |
1602 //-------------------------vector_opd_range----------------------- | |
1603 // (Start, end] half-open range defining which operands are vector | |
1604 void SuperWord::vector_opd_range(Node* n, uint* start, uint* end) { | |
1605 switch (n->Opcode()) { | |
558
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
253
diff
changeset
|
1606 case Op_LoadB: case Op_LoadUS: |
0 | 1607 case Op_LoadI: case Op_LoadL: |
1608 case Op_LoadF: case Op_LoadD: | |
1609 case Op_LoadP: | |
1610 *start = 0; | |
1611 *end = 0; | |
1612 return; | |
1613 case Op_StoreB: case Op_StoreC: | |
1614 case Op_StoreI: case Op_StoreL: | |
1615 case Op_StoreF: case Op_StoreD: | |
1616 case Op_StoreP: | |
1617 *start = MemNode::ValueIn; | |
1618 *end = *start + 1; | |
1619 return; | |
1620 case Op_LShiftI: case Op_LShiftL: | |
1621 *start = 1; | |
1622 *end = 2; | |
1623 return; | |
1624 case Op_CMoveI: case Op_CMoveL: case Op_CMoveF: case Op_CMoveD: | |
1625 *start = 2; | |
1626 *end = n->req(); | |
1627 return; | |
1628 } | |
1629 *start = 1; | |
1630 *end = n->req(); // default is all operands | |
1631 } | |
1632 | |
1633 //------------------------------in_packset--------------------------- | |
1634 // Are s1 and s2 in a pack pair and ordered as s1,s2? | |
1635 bool SuperWord::in_packset(Node* s1, Node* s2) { | |
1636 for (int i = 0; i < _packset.length(); i++) { | |
1637 Node_List* p = _packset.at(i); | |
1638 assert(p->size() == 2, "must be"); | |
1639 if (p->at(0) == s1 && p->at(p->size()-1) == s2) { | |
1640 return true; | |
1641 } | |
1642 } | |
1643 return false; | |
1644 } | |
1645 | |
1646 //------------------------------in_pack--------------------------- | |
1647 // Is s in pack p? | |
1648 Node_List* SuperWord::in_pack(Node* s, Node_List* p) { | |
1649 for (uint i = 0; i < p->size(); i++) { | |
1650 if (p->at(i) == s) { | |
1651 return p; | |
1652 } | |
1653 } | |
1654 return NULL; | |
1655 } | |
1656 | |
1657 //------------------------------remove_pack_at--------------------------- | |
1658 // Remove the pack at position pos in the packset | |
1659 void SuperWord::remove_pack_at(int pos) { | |
1660 Node_List* p = _packset.at(pos); | |
1661 for (uint i = 0; i < p->size(); i++) { | |
1662 Node* s = p->at(i); | |
1663 set_my_pack(s, NULL); | |
1664 } | |
1665 _packset.remove_at(pos); | |
1666 } | |
1667 | |
1668 //------------------------------executed_first--------------------------- | |
1669 // Return the node executed first in pack p. Uses the RPO block list | |
1670 // to determine order. | |
1671 Node* SuperWord::executed_first(Node_List* p) { | |
1672 Node* n = p->at(0); | |
1673 int n_rpo = bb_idx(n); | |
1674 for (uint i = 1; i < p->size(); i++) { | |
1675 Node* s = p->at(i); | |
1676 int s_rpo = bb_idx(s); | |
1677 if (s_rpo < n_rpo) { | |
1678 n = s; | |
1679 n_rpo = s_rpo; | |
1680 } | |
1681 } | |
1682 return n; | |
1683 } | |
1684 | |
1685 //------------------------------executed_last--------------------------- | |
1686 // Return the node executed last in pack p. | |
1687 Node* SuperWord::executed_last(Node_List* p) { | |
1688 Node* n = p->at(0); | |
1689 int n_rpo = bb_idx(n); | |
1690 for (uint i = 1; i < p->size(); i++) { | |
1691 Node* s = p->at(i); | |
1692 int s_rpo = bb_idx(s); | |
1693 if (s_rpo > n_rpo) { | |
1694 n = s; | |
1695 n_rpo = s_rpo; | |
1696 } | |
1697 } | |
1698 return n; | |
1699 } | |
1700 | |
1701 //----------------------------align_initial_loop_index--------------------------- | |
1702 // Adjust pre-loop limit so that in main loop, a load/store reference | |
1703 // to align_to_ref will be a position zero in the vector. | |
1704 // (iv + k) mod vector_align == 0 | |
1705 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) { | |
1706 CountedLoopNode *main_head = lp()->as_CountedLoop(); | |
1707 assert(main_head->is_main_loop(), ""); | |
1708 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head); | |
1709 assert(pre_end != NULL, ""); | |
1710 Node *pre_opaq1 = pre_end->limit(); | |
1711 assert(pre_opaq1->Opcode() == Op_Opaque1, ""); | |
1712 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; | |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1713 Node *lim0 = pre_opaq->in(1); |
0 | 1714 |
1715 // Where we put new limit calculations | |
1716 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); | |
1717 | |
1718 // Ensure the original loop limit is available from the | |
1719 // pre-loop Opaque1 node. | |
1720 Node *orig_limit = pre_opaq->original_loop_limit(); | |
1721 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, ""); | |
1722 | |
1723 SWPointer align_to_ref_p(align_to_ref, this); | |
1724 | |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1725 // Given: |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1726 // lim0 == original pre loop limit |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1727 // V == v_align (power of 2) |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1728 // invar == extra invariant piece of the address expression |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1729 // e == k [ +/- invar ] |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1730 // |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1731 // When reassociating expressions involving '%' the basic rules are: |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1732 // (a - b) % k == 0 => a % k == b % k |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1733 // and: |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1734 // (a + b) % k == 0 => a % k == (k - b) % k |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1735 // |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1736 // For stride > 0 && scale > 0, |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1737 // Derive the new pre-loop limit "lim" such that the two constraints: |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1738 // (1) lim = lim0 + N (where N is some positive integer < V) |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1739 // (2) (e + lim) % V == 0 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1740 // are true. |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1741 // |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1742 // Substituting (1) into (2), |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1743 // (e + lim0 + N) % V == 0 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1744 // solve for N: |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1745 // N = (V - (e + lim0)) % V |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1746 // substitute back into (1), so that new limit |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1747 // lim = lim0 + (V - (e + lim0)) % V |
0 | 1748 // |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1749 // For stride > 0 && scale < 0 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1750 // Constraints: |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1751 // lim = lim0 + N |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1752 // (e - lim) % V == 0 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1753 // Solving for lim: |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1754 // (e - lim0 - N) % V == 0 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1755 // N = (e - lim0) % V |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1756 // lim = lim0 + (e - lim0) % V |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1757 // |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1758 // For stride < 0 && scale > 0 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1759 // Constraints: |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1760 // lim = lim0 - N |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1761 // (e + lim) % V == 0 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1762 // Solving for lim: |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1763 // (e + lim0 - N) % V == 0 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1764 // N = (e + lim0) % V |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1765 // lim = lim0 - (e + lim0) % V |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1766 // |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1767 // For stride < 0 && scale < 0 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1768 // Constraints: |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1769 // lim = lim0 - N |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1770 // (e - lim) % V == 0 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1771 // Solving for lim: |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1772 // (e - lim0 + N) % V == 0 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1773 // N = (V - (e - lim0)) % V |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1774 // lim = lim0 - (V - (e - lim0)) % V |
0 | 1775 |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1776 int stride = iv_stride(); |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1777 int scale = align_to_ref_p.scale_in_bytes(); |
0 | 1778 int elt_size = align_to_ref_p.memory_size(); |
1779 int v_align = vector_width_in_bytes() / elt_size; | |
1780 int k = align_to_ref_p.offset_in_bytes() / elt_size; | |
1781 | |
1782 Node *kn = _igvn.intcon(k); | |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1783 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1784 Node *e = kn; |
0 | 1785 if (align_to_ref_p.invar() != NULL) { |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1786 // incorporate any extra invariant piece producing k +/- invar >>> log2(elt) |
0 | 1787 Node* log2_elt = _igvn.intcon(exact_log2(elt_size)); |
1788 Node* aref = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt); | |
1789 _phase->_igvn.register_new_node_with_optimizer(aref); | |
1790 _phase->set_ctrl(aref, pre_ctrl); | |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1791 if (align_to_ref_p.negate_invar()) { |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1792 e = new (_phase->C, 3) SubINode(e, aref); |
0 | 1793 } else { |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1794 e = new (_phase->C, 3) AddINode(e, aref); |
0 | 1795 } |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1796 _phase->_igvn.register_new_node_with_optimizer(e); |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1797 _phase->set_ctrl(e, pre_ctrl); |
0 | 1798 } |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1799 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1800 // compute e +/- lim0 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1801 if (scale < 0) { |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1802 e = new (_phase->C, 3) SubINode(e, lim0); |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1803 } else { |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1804 e = new (_phase->C, 3) AddINode(e, lim0); |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1805 } |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1806 _phase->_igvn.register_new_node_with_optimizer(e); |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1807 _phase->set_ctrl(e, pre_ctrl); |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1808 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1809 if (stride * scale > 0) { |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1810 // compute V - (e +/- lim0) |
0 | 1811 Node* va = _igvn.intcon(v_align); |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1812 e = new (_phase->C, 3) SubINode(va, e); |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1813 _phase->_igvn.register_new_node_with_optimizer(e); |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1814 _phase->set_ctrl(e, pre_ctrl); |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1815 } |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1816 // compute N = (exp) % V |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1817 Node* va_msk = _igvn.intcon(v_align - 1); |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1818 Node* N = new (_phase->C, 3) AndINode(e, va_msk); |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1819 _phase->_igvn.register_new_node_with_optimizer(N); |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1820 _phase->set_ctrl(N, pre_ctrl); |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1821 |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1822 // substitute back into (1), so that new limit |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1823 // lim = lim0 + N |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1824 Node* lim; |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1825 if (stride < 0) { |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1826 lim = new (_phase->C, 3) SubINode(lim0, N); |
0 | 1827 } else { |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1828 lim = new (_phase->C, 3) AddINode(lim0, N); |
0 | 1829 } |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1830 _phase->_igvn.register_new_node_with_optimizer(lim); |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1831 _phase->set_ctrl(lim, pre_ctrl); |
0 | 1832 Node* constrained = |
72
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1833 (stride > 0) ? (Node*) new (_phase->C,3) MinINode(lim, orig_limit) |
f705f25597eb
6663621: JVM crashes while trying to execute api/java_security/Signature/SignatureTests.html#initSign tests.
never
parents:
29
diff
changeset
|
1834 : (Node*) new (_phase->C,3) MaxINode(lim, orig_limit); |
0 | 1835 _phase->_igvn.register_new_node_with_optimizer(constrained); |
1836 _phase->set_ctrl(constrained, pre_ctrl); | |
1837 _igvn.hash_delete(pre_opaq); | |
1838 pre_opaq->set_req(1, constrained); | |
1839 } | |
1840 | |
1841 //----------------------------get_pre_loop_end--------------------------- | |
1842 // Find pre loop end from main loop. Returns null if none. | |
1843 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) { | |
1844 Node *ctrl = cl->in(LoopNode::EntryControl); | |
1845 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL; | |
1846 Node *iffm = ctrl->in(0); | |
1847 if (!iffm->is_If()) return NULL; | |
1848 Node *p_f = iffm->in(0); | |
1849 if (!p_f->is_IfFalse()) return NULL; | |
1850 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL; | |
1851 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); | |
1852 if (!pre_end->loopnode()->is_pre_loop()) return NULL; | |
1853 return pre_end; | |
1854 } | |
1855 | |
1856 | |
1857 //------------------------------init--------------------------- | |
1858 void SuperWord::init() { | |
1859 _dg.init(); | |
1860 _packset.clear(); | |
1861 _disjoint_ptrs.clear(); | |
1862 _block.clear(); | |
1863 _data_entry.clear(); | |
1864 _mem_slice_head.clear(); | |
1865 _mem_slice_tail.clear(); | |
1866 _node_info.clear(); | |
1867 _align_to_ref = NULL; | |
1868 _lpt = NULL; | |
1869 _lp = NULL; | |
1870 _bb = NULL; | |
1871 _iv = NULL; | |
1872 } | |
1873 | |
1874 //------------------------------print_packset--------------------------- | |
1875 void SuperWord::print_packset() { | |
1876 #ifndef PRODUCT | |
1877 tty->print_cr("packset"); | |
1878 for (int i = 0; i < _packset.length(); i++) { | |
1879 tty->print_cr("Pack: %d", i); | |
1880 Node_List* p = _packset.at(i); | |
1881 print_pack(p); | |
1882 } | |
1883 #endif | |
1884 } | |
1885 | |
1886 //------------------------------print_pack--------------------------- | |
1887 void SuperWord::print_pack(Node_List* p) { | |
1888 for (uint i = 0; i < p->size(); i++) { | |
1889 print_stmt(p->at(i)); | |
1890 } | |
1891 } | |
1892 | |
1893 //------------------------------print_bb--------------------------- | |
1894 void SuperWord::print_bb() { | |
1895 #ifndef PRODUCT | |
1896 tty->print_cr("\nBlock"); | |
1897 for (int i = 0; i < _block.length(); i++) { | |
1898 Node* n = _block.at(i); | |
1899 tty->print("%d ", i); | |
1900 if (n) { | |
1901 n->dump(); | |
1902 } | |
1903 } | |
1904 #endif | |
1905 } | |
1906 | |
1907 //------------------------------print_stmt--------------------------- | |
1908 void SuperWord::print_stmt(Node* s) { | |
1909 #ifndef PRODUCT | |
1910 tty->print(" align: %d \t", alignment(s)); | |
1911 s->dump(); | |
1912 #endif | |
1913 } | |
1914 | |
1915 //------------------------------blank--------------------------- | |
1916 char* SuperWord::blank(uint depth) { | |
1917 static char blanks[101]; | |
1918 assert(depth < 101, "too deep"); | |
1919 for (uint i = 0; i < depth; i++) blanks[i] = ' '; | |
1920 blanks[depth] = '\0'; | |
1921 return blanks; | |
1922 } | |
1923 | |
1924 | |
1925 //==============================SWPointer=========================== | |
1926 | |
1927 //----------------------------SWPointer------------------------ | |
1928 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) : | |
1929 _mem(mem), _slp(slp), _base(NULL), _adr(NULL), | |
1930 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) { | |
1931 | |
1932 Node* adr = mem->in(MemNode::Address); | |
1933 if (!adr->is_AddP()) { | |
1934 assert(!valid(), "too complex"); | |
1935 return; | |
1936 } | |
1937 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant) | |
1938 Node* base = adr->in(AddPNode::Base); | |
1058
73a726751507
6852078: HSX 14/16 in jdk 5.0: api/javax_management api/org_omg jck tests crashes or make tnameserv crash
cfang
parents:
987
diff
changeset
|
1939 //unsafe reference could not be aligned appropriately without runtime checking |
73a726751507
6852078: HSX 14/16 in jdk 5.0: api/javax_management api/org_omg jck tests crashes or make tnameserv crash
cfang
parents:
987
diff
changeset
|
1940 if (base == NULL || base->bottom_type() == Type::TOP) { |
73a726751507
6852078: HSX 14/16 in jdk 5.0: api/javax_management api/org_omg jck tests crashes or make tnameserv crash
cfang
parents:
987
diff
changeset
|
1941 assert(!valid(), "unsafe access"); |
73a726751507
6852078: HSX 14/16 in jdk 5.0: api/javax_management api/org_omg jck tests crashes or make tnameserv crash
cfang
parents:
987
diff
changeset
|
1942 return; |
73a726751507
6852078: HSX 14/16 in jdk 5.0: api/javax_management api/org_omg jck tests crashes or make tnameserv crash
cfang
parents:
987
diff
changeset
|
1943 } |
0 | 1944 for (int i = 0; i < 3; i++) { |
1945 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) { | |
1946 assert(!valid(), "too complex"); | |
1947 return; | |
1948 } | |
1949 adr = adr->in(AddPNode::Address); | |
1950 if (base == adr || !adr->is_AddP()) { | |
1951 break; // stop looking at addp's | |
1952 } | |
1953 } | |
1954 _base = base; | |
1955 _adr = adr; | |
1956 assert(valid(), "Usable"); | |
1957 } | |
1958 | |
1959 // Following is used to create a temporary object during | |
1960 // the pattern match of an address expression. | |
1961 SWPointer::SWPointer(SWPointer* p) : | |
1962 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL), | |
1963 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {} | |
1964 | |
1965 //------------------------scaled_iv_plus_offset-------------------- | |
1966 // Match: k*iv + offset | |
1967 // where: k is a constant that maybe zero, and | |
1968 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional | |
1969 bool SWPointer::scaled_iv_plus_offset(Node* n) { | |
1970 if (scaled_iv(n)) { | |
1971 return true; | |
1972 } | |
1973 if (offset_plus_k(n)) { | |
1974 return true; | |
1975 } | |
1976 int opc = n->Opcode(); | |
1977 if (opc == Op_AddI) { | |
1978 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) { | |
1979 return true; | |
1980 } | |
1981 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) { | |
1982 return true; | |
1983 } | |
1984 } else if (opc == Op_SubI) { | |
1985 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) { | |
1986 return true; | |
1987 } | |
1988 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) { | |
1989 _scale *= -1; | |
1990 return true; | |
1991 } | |
1992 } | |
1993 return false; | |
1994 } | |
1995 | |
1996 //----------------------------scaled_iv------------------------ | |
1997 // Match: k*iv where k is a constant that's not zero | |
1998 bool SWPointer::scaled_iv(Node* n) { | |
1999 if (_scale != 0) { | |
2000 return false; // already found a scale | |
2001 } | |
2002 if (n == iv()) { | |
2003 _scale = 1; | |
2004 return true; | |
2005 } | |
2006 int opc = n->Opcode(); | |
2007 if (opc == Op_MulI) { | |
2008 if (n->in(1) == iv() && n->in(2)->is_Con()) { | |
2009 _scale = n->in(2)->get_int(); | |
2010 return true; | |
2011 } else if (n->in(2) == iv() && n->in(1)->is_Con()) { | |
2012 _scale = n->in(1)->get_int(); | |
2013 return true; | |
2014 } | |
2015 } else if (opc == Op_LShiftI) { | |
2016 if (n->in(1) == iv() && n->in(2)->is_Con()) { | |
2017 _scale = 1 << n->in(2)->get_int(); | |
2018 return true; | |
2019 } | |
2020 } else if (opc == Op_ConvI2L) { | |
2021 if (scaled_iv_plus_offset(n->in(1))) { | |
2022 return true; | |
2023 } | |
2024 } else if (opc == Op_LShiftL) { | |
2025 if (!has_iv() && _invar == NULL) { | |
2026 // Need to preserve the current _offset value, so | |
2027 // create a temporary object for this expression subtree. | |
2028 // Hacky, so should re-engineer the address pattern match. | |
2029 SWPointer tmp(this); | |
2030 if (tmp.scaled_iv_plus_offset(n->in(1))) { | |
2031 if (tmp._invar == NULL) { | |
2032 int mult = 1 << n->in(2)->get_int(); | |
2033 _scale = tmp._scale * mult; | |
2034 _offset += tmp._offset * mult; | |
2035 return true; | |
2036 } | |
2037 } | |
2038 } | |
2039 } | |
2040 return false; | |
2041 } | |
2042 | |
2043 //----------------------------offset_plus_k------------------------ | |
2044 // Match: offset is (k [+/- invariant]) | |
2045 // where k maybe zero and invariant is optional, but not both. | |
2046 bool SWPointer::offset_plus_k(Node* n, bool negate) { | |
2047 int opc = n->Opcode(); | |
2048 if (opc == Op_ConI) { | |
2049 _offset += negate ? -(n->get_int()) : n->get_int(); | |
2050 return true; | |
2051 } else if (opc == Op_ConL) { | |
2052 // Okay if value fits into an int | |
2053 const TypeLong* t = n->find_long_type(); | |
2054 if (t->higher_equal(TypeLong::INT)) { | |
2055 jlong loff = n->get_long(); | |
2056 jint off = (jint)loff; | |
2057 _offset += negate ? -off : loff; | |
2058 return true; | |
2059 } | |
2060 return false; | |
2061 } | |
2062 if (_invar != NULL) return false; // already have an invariant | |
2063 if (opc == Op_AddI) { | |
2064 if (n->in(2)->is_Con() && invariant(n->in(1))) { | |
2065 _negate_invar = negate; | |
2066 _invar = n->in(1); | |
2067 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int(); | |
2068 return true; | |
2069 } else if (n->in(1)->is_Con() && invariant(n->in(2))) { | |
2070 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int(); | |
2071 _negate_invar = negate; | |
2072 _invar = n->in(2); | |
2073 return true; | |
2074 } | |
2075 } | |
2076 if (opc == Op_SubI) { | |
2077 if (n->in(2)->is_Con() && invariant(n->in(1))) { | |
2078 _negate_invar = negate; | |
2079 _invar = n->in(1); | |
2080 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int(); | |
2081 return true; | |
2082 } else if (n->in(1)->is_Con() && invariant(n->in(2))) { | |
2083 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int(); | |
2084 _negate_invar = !negate; | |
2085 _invar = n->in(2); | |
2086 return true; | |
2087 } | |
2088 } | |
2089 if (invariant(n)) { | |
2090 _negate_invar = negate; | |
2091 _invar = n; | |
2092 return true; | |
2093 } | |
2094 return false; | |
2095 } | |
2096 | |
2097 //----------------------------print------------------------ | |
2098 void SWPointer::print() { | |
2099 #ifndef PRODUCT | |
2100 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n", | |
2101 _base != NULL ? _base->_idx : 0, | |
2102 _adr != NULL ? _adr->_idx : 0, | |
2103 _scale, _offset, | |
2104 _negate_invar?'-':'+', | |
2105 _invar != NULL ? _invar->_idx : 0); | |
2106 #endif | |
2107 } | |
2108 | |
2109 // ========================= OrderedPair ===================== | |
2110 | |
2111 const OrderedPair OrderedPair::initial; | |
2112 | |
2113 // ========================= SWNodeInfo ===================== | |
2114 | |
2115 const SWNodeInfo SWNodeInfo::initial; | |
2116 | |
2117 | |
2118 // ============================ DepGraph =========================== | |
2119 | |
2120 //------------------------------make_node--------------------------- | |
2121 // Make a new dependence graph node for an ideal node. | |
2122 DepMem* DepGraph::make_node(Node* node) { | |
2123 DepMem* m = new (_arena) DepMem(node); | |
2124 if (node != NULL) { | |
2125 assert(_map.at_grow(node->_idx) == NULL, "one init only"); | |
2126 _map.at_put_grow(node->_idx, m); | |
2127 } | |
2128 return m; | |
2129 } | |
2130 | |
2131 //------------------------------make_edge--------------------------- | |
2132 // Make a new dependence graph edge from dpred -> dsucc | |
2133 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) { | |
2134 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head()); | |
2135 dpred->set_out_head(e); | |
2136 dsucc->set_in_head(e); | |
2137 return e; | |
2138 } | |
2139 | |
2140 // ========================== DepMem ======================== | |
2141 | |
2142 //------------------------------in_cnt--------------------------- | |
2143 int DepMem::in_cnt() { | |
2144 int ct = 0; | |
2145 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++; | |
2146 return ct; | |
2147 } | |
2148 | |
2149 //------------------------------out_cnt--------------------------- | |
2150 int DepMem::out_cnt() { | |
2151 int ct = 0; | |
2152 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++; | |
2153 return ct; | |
2154 } | |
2155 | |
2156 //------------------------------print----------------------------- | |
2157 void DepMem::print() { | |
2158 #ifndef PRODUCT | |
2159 tty->print(" DepNode %d (", _node->_idx); | |
2160 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) { | |
2161 Node* pred = p->pred()->node(); | |
2162 tty->print(" %d", pred != NULL ? pred->_idx : 0); | |
2163 } | |
2164 tty->print(") ["); | |
2165 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) { | |
2166 Node* succ = s->succ()->node(); | |
2167 tty->print(" %d", succ != NULL ? succ->_idx : 0); | |
2168 } | |
2169 tty->print_cr(" ]"); | |
2170 #endif | |
2171 } | |
2172 | |
2173 // =========================== DepEdge ========================= | |
2174 | |
2175 //------------------------------DepPreds--------------------------- | |
2176 void DepEdge::print() { | |
2177 #ifndef PRODUCT | |
2178 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx); | |
2179 #endif | |
2180 } | |
2181 | |
2182 // =========================== DepPreds ========================= | |
2183 // Iterator over predecessor edges in the dependence graph. | |
2184 | |
2185 //------------------------------DepPreds--------------------------- | |
2186 DepPreds::DepPreds(Node* n, DepGraph& dg) { | |
2187 _n = n; | |
2188 _done = false; | |
2189 if (_n->is_Store() || _n->is_Load()) { | |
2190 _next_idx = MemNode::Address; | |
2191 _end_idx = n->req(); | |
2192 _dep_next = dg.dep(_n)->in_head(); | |
2193 } else if (_n->is_Mem()) { | |
2194 _next_idx = 0; | |
2195 _end_idx = 0; | |
2196 _dep_next = dg.dep(_n)->in_head(); | |
2197 } else { | |
2198 _next_idx = 1; | |
2199 _end_idx = _n->req(); | |
2200 _dep_next = NULL; | |
2201 } | |
2202 next(); | |
2203 } | |
2204 | |
2205 //------------------------------next--------------------------- | |
2206 void DepPreds::next() { | |
2207 if (_dep_next != NULL) { | |
2208 _current = _dep_next->pred()->node(); | |
2209 _dep_next = _dep_next->next_in(); | |
2210 } else if (_next_idx < _end_idx) { | |
2211 _current = _n->in(_next_idx++); | |
2212 } else { | |
2213 _done = true; | |
2214 } | |
2215 } | |
2216 | |
2217 // =========================== DepSuccs ========================= | |
2218 // Iterator over successor edges in the dependence graph. | |
2219 | |
2220 //------------------------------DepSuccs--------------------------- | |
2221 DepSuccs::DepSuccs(Node* n, DepGraph& dg) { | |
2222 _n = n; | |
2223 _done = false; | |
2224 if (_n->is_Load()) { | |
2225 _next_idx = 0; | |
2226 _end_idx = _n->outcnt(); | |
2227 _dep_next = dg.dep(_n)->out_head(); | |
2228 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) { | |
2229 _next_idx = 0; | |
2230 _end_idx = 0; | |
2231 _dep_next = dg.dep(_n)->out_head(); | |
2232 } else { | |
2233 _next_idx = 0; | |
2234 _end_idx = _n->outcnt(); | |
2235 _dep_next = NULL; | |
2236 } | |
2237 next(); | |
2238 } | |
2239 | |
2240 //-------------------------------next--------------------------- | |
2241 void DepSuccs::next() { | |
2242 if (_dep_next != NULL) { | |
2243 _current = _dep_next->succ()->node(); | |
2244 _dep_next = _dep_next->next_out(); | |
2245 } else if (_next_idx < _end_idx) { | |
2246 _current = _n->raw_out(_next_idx++); | |
2247 } else { | |
2248 _done = true; | |
2249 } | |
2250 } |