Mercurial > hg > graal-compiler
annotate src/share/vm/opto/compile.cpp @ 14392:b5c8a61d7fa0
Merge
author | kvn |
---|---|
date | Fri, 21 Jun 2013 15:56:24 -0700 |
parents | d2907f74462e f2110083203d |
children | e2722a66aba7 |
rev | line source |
---|---|
0 | 1 /* |
10405 | 2 * Copyright (c) 1997, 2013, 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:
1397
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1397
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:
1397
diff
changeset
|
21 * questions. |
0 | 22 * |
23 */ | |
24 | |
1972 | 25 #include "precompiled.hpp" |
7199
cd3d6a6b95d9
8003240: x86: move MacroAssembler into separate file
twisti
parents:
7196
diff
changeset
|
26 #include "asm/macroAssembler.hpp" |
cd3d6a6b95d9
8003240: x86: move MacroAssembler into separate file
twisti
parents:
7196
diff
changeset
|
27 #include "asm/macroAssembler.inline.hpp" |
1972 | 28 #include "classfile/systemDictionary.hpp" |
29 #include "code/exceptionHandlerTable.hpp" | |
30 #include "code/nmethod.hpp" | |
31 #include "compiler/compileLog.hpp" | |
7199
cd3d6a6b95d9
8003240: x86: move MacroAssembler into separate file
twisti
parents:
7196
diff
changeset
|
32 #include "compiler/disassembler.hpp" |
1972 | 33 #include "compiler/oopMap.hpp" |
34 #include "opto/addnode.hpp" | |
35 #include "opto/block.hpp" | |
36 #include "opto/c2compiler.hpp" | |
37 #include "opto/callGenerator.hpp" | |
38 #include "opto/callnode.hpp" | |
39 #include "opto/cfgnode.hpp" | |
40 #include "opto/chaitin.hpp" | |
41 #include "opto/compile.hpp" | |
42 #include "opto/connode.hpp" | |
43 #include "opto/divnode.hpp" | |
44 #include "opto/escape.hpp" | |
45 #include "opto/idealGraphPrinter.hpp" | |
46 #include "opto/loopnode.hpp" | |
47 #include "opto/machnode.hpp" | |
48 #include "opto/macro.hpp" | |
49 #include "opto/matcher.hpp" | |
50 #include "opto/memnode.hpp" | |
51 #include "opto/mulnode.hpp" | |
52 #include "opto/node.hpp" | |
53 #include "opto/opcodes.hpp" | |
54 #include "opto/output.hpp" | |
55 #include "opto/parse.hpp" | |
56 #include "opto/phaseX.hpp" | |
57 #include "opto/rootnode.hpp" | |
58 #include "opto/runtime.hpp" | |
59 #include "opto/stringopts.hpp" | |
60 #include "opto/type.hpp" | |
61 #include "opto/vectornode.hpp" | |
62 #include "runtime/arguments.hpp" | |
63 #include "runtime/signature.hpp" | |
64 #include "runtime/stubRoutines.hpp" | |
65 #include "runtime/timer.hpp" | |
10405 | 66 #include "trace/tracing.hpp" |
1972 | 67 #include "utilities/copy.hpp" |
68 #ifdef TARGET_ARCH_MODEL_x86_32 | |
69 # include "adfiles/ad_x86_32.hpp" | |
70 #endif | |
71 #ifdef TARGET_ARCH_MODEL_x86_64 | |
72 # include "adfiles/ad_x86_64.hpp" | |
73 #endif | |
74 #ifdef TARGET_ARCH_MODEL_sparc | |
75 # include "adfiles/ad_sparc.hpp" | |
76 #endif | |
77 #ifdef TARGET_ARCH_MODEL_zero | |
78 # include "adfiles/ad_zero.hpp" | |
79 #endif | |
2192
b92c45f2bc75
7016023: Enable building ARM and PPC from src/closed repository
bobv
parents:
2091
diff
changeset
|
80 #ifdef TARGET_ARCH_MODEL_arm |
b92c45f2bc75
7016023: Enable building ARM and PPC from src/closed repository
bobv
parents:
2091
diff
changeset
|
81 # include "adfiles/ad_arm.hpp" |
b92c45f2bc75
7016023: Enable building ARM and PPC from src/closed repository
bobv
parents:
2091
diff
changeset
|
82 #endif |
14391 | 83 #ifdef TARGET_ARCH_MODEL_ppc_32 |
84 # include "adfiles/ad_ppc_32.hpp" | |
85 #endif | |
86 #ifdef TARGET_ARCH_MODEL_ppc_64 | |
87 # include "adfiles/ad_ppc_64.hpp" | |
2192
b92c45f2bc75
7016023: Enable building ARM and PPC from src/closed repository
bobv
parents:
2091
diff
changeset
|
88 #endif |
0 | 89 |
2008 | 90 |
91 // -------------------- Compile::mach_constant_base_node ----------------------- | |
92 // Constant table base node singleton. | |
93 MachConstantBaseNode* Compile::mach_constant_base_node() { | |
94 if (_mach_constant_base_node == NULL) { | |
95 _mach_constant_base_node = new (C) MachConstantBaseNode(); | |
96 _mach_constant_base_node->add_req(C->root()); | |
97 } | |
98 return _mach_constant_base_node; | |
99 } | |
100 | |
101 | |
0 | 102 /// Support for intrinsics. |
103 | |
104 // Return the index at which m must be inserted (or already exists). | |
105 // The sort order is by the address of the ciMethod, with is_virtual as minor key. | |
106 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { | |
107 #ifdef ASSERT | |
108 for (int i = 1; i < _intrinsics->length(); i++) { | |
109 CallGenerator* cg1 = _intrinsics->at(i-1); | |
110 CallGenerator* cg2 = _intrinsics->at(i); | |
111 assert(cg1->method() != cg2->method() | |
112 ? cg1->method() < cg2->method() | |
113 : cg1->is_virtual() < cg2->is_virtual(), | |
114 "compiler intrinsics list must stay sorted"); | |
115 } | |
116 #endif | |
117 // Binary search sorted list, in decreasing intervals [lo, hi]. | |
118 int lo = 0, hi = _intrinsics->length()-1; | |
119 while (lo <= hi) { | |
120 int mid = (uint)(hi + lo) / 2; | |
121 ciMethod* mid_m = _intrinsics->at(mid)->method(); | |
122 if (m < mid_m) { | |
123 hi = mid-1; | |
124 } else if (m > mid_m) { | |
125 lo = mid+1; | |
126 } else { | |
127 // look at minor sort key | |
128 bool mid_virt = _intrinsics->at(mid)->is_virtual(); | |
129 if (is_virtual < mid_virt) { | |
130 hi = mid-1; | |
131 } else if (is_virtual > mid_virt) { | |
132 lo = mid+1; | |
133 } else { | |
134 return mid; // exact match | |
135 } | |
136 } | |
137 } | |
138 return lo; // inexact match | |
139 } | |
140 | |
141 void Compile::register_intrinsic(CallGenerator* cg) { | |
142 if (_intrinsics == NULL) { | |
7473 | 143 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); |
0 | 144 } |
145 // This code is stolen from ciObjectFactory::insert. | |
146 // Really, GrowableArray should have methods for | |
147 // insert_at, remove_at, and binary_search. | |
148 int len = _intrinsics->length(); | |
149 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); | |
150 if (index == len) { | |
151 _intrinsics->append(cg); | |
152 } else { | |
153 #ifdef ASSERT | |
154 CallGenerator* oldcg = _intrinsics->at(index); | |
155 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); | |
156 #endif | |
157 _intrinsics->append(_intrinsics->at(len-1)); | |
158 int pos; | |
159 for (pos = len-2; pos >= index; pos--) { | |
160 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); | |
161 } | |
162 _intrinsics->at_put(index, cg); | |
163 } | |
164 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); | |
165 } | |
166 | |
167 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { | |
168 assert(m->is_loaded(), "don't try this on unloaded methods"); | |
169 if (_intrinsics != NULL) { | |
170 int index = intrinsic_insertion_index(m, is_virtual); | |
171 if (index < _intrinsics->length() | |
172 && _intrinsics->at(index)->method() == m | |
173 && _intrinsics->at(index)->is_virtual() == is_virtual) { | |
174 return _intrinsics->at(index); | |
175 } | |
176 } | |
177 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. | |
856
75596850f863
6862576: vmIntrinsics needs cleanup in order to support JSR 292 intrinsics
jrose
parents:
647
diff
changeset
|
178 if (m->intrinsic_id() != vmIntrinsics::_none && |
75596850f863
6862576: vmIntrinsics needs cleanup in order to support JSR 292 intrinsics
jrose
parents:
647
diff
changeset
|
179 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { |
0 | 180 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); |
181 if (cg != NULL) { | |
182 // Save it for next time: | |
183 register_intrinsic(cg); | |
184 return cg; | |
185 } else { | |
186 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); | |
187 } | |
188 } | |
189 return NULL; | |
190 } | |
191 | |
192 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined | |
193 // in library_call.cpp. | |
194 | |
195 | |
196 #ifndef PRODUCT | |
197 // statistics gathering... | |
198 | |
199 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; | |
200 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; | |
201 | |
202 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { | |
203 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); | |
204 int oflags = _intrinsic_hist_flags[id]; | |
205 assert(flags != 0, "what happened?"); | |
206 if (is_virtual) { | |
207 flags |= _intrinsic_virtual; | |
208 } | |
209 bool changed = (flags != oflags); | |
210 if ((flags & _intrinsic_worked) != 0) { | |
211 juint count = (_intrinsic_hist_count[id] += 1); | |
212 if (count == 1) { | |
213 changed = true; // first time | |
214 } | |
215 // increment the overall count also: | |
216 _intrinsic_hist_count[vmIntrinsics::_none] += 1; | |
217 } | |
218 if (changed) { | |
219 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { | |
220 // Something changed about the intrinsic's virtuality. | |
221 if ((flags & _intrinsic_virtual) != 0) { | |
222 // This is the first use of this intrinsic as a virtual call. | |
223 if (oflags != 0) { | |
224 // We already saw it as a non-virtual, so note both cases. | |
225 flags |= _intrinsic_both; | |
226 } | |
227 } else if ((oflags & _intrinsic_both) == 0) { | |
228 // This is the first use of this intrinsic as a non-virtual | |
229 flags |= _intrinsic_both; | |
230 } | |
231 } | |
232 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); | |
233 } | |
234 // update the overall flags also: | |
235 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; | |
236 return changed; | |
237 } | |
238 | |
239 static char* format_flags(int flags, char* buf) { | |
240 buf[0] = 0; | |
241 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); | |
242 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); | |
243 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); | |
244 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); | |
245 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); | |
246 if (buf[0] == 0) strcat(buf, ","); | |
247 assert(buf[0] == ',', "must be"); | |
248 return &buf[1]; | |
249 } | |
250 | |
251 void Compile::print_intrinsic_statistics() { | |
252 char flagsbuf[100]; | |
253 ttyLocker ttyl; | |
254 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); | |
255 tty->print_cr("Compiler intrinsic usage:"); | |
256 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; | |
257 if (total == 0) total = 1; // avoid div0 in case of no successes | |
258 #define PRINT_STAT_LINE(name, c, f) \ | |
259 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); | |
260 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { | |
261 vmIntrinsics::ID id = (vmIntrinsics::ID) index; | |
262 int flags = _intrinsic_hist_flags[id]; | |
263 juint count = _intrinsic_hist_count[id]; | |
264 if ((flags | count) != 0) { | |
265 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); | |
266 } | |
267 } | |
268 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); | |
269 if (xtty != NULL) xtty->tail("statistics"); | |
270 } | |
271 | |
272 void Compile::print_statistics() { | |
273 { ttyLocker ttyl; | |
274 if (xtty != NULL) xtty->head("statistics type='opto'"); | |
275 Parse::print_statistics(); | |
276 PhaseCCP::print_statistics(); | |
277 PhaseRegAlloc::print_statistics(); | |
278 Scheduling::print_statistics(); | |
279 PhasePeephole::print_statistics(); | |
280 PhaseIdealLoop::print_statistics(); | |
281 if (xtty != NULL) xtty->tail("statistics"); | |
282 } | |
283 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { | |
284 // put this under its own <statistics> element. | |
285 print_intrinsic_statistics(); | |
286 } | |
287 } | |
288 #endif //PRODUCT | |
289 | |
290 // Support for bundling info | |
291 Bundle* Compile::node_bundling(const Node *n) { | |
292 assert(valid_bundle_info(n), "oob"); | |
293 return &_node_bundling_base[n->_idx]; | |
294 } | |
295 | |
296 bool Compile::valid_bundle_info(const Node *n) { | |
297 return (_node_bundling_limit > n->_idx); | |
298 } | |
299 | |
300 | |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
301 void Compile::gvn_replace_by(Node* n, Node* nn) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
302 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
303 Node* use = n->last_out(i); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
304 bool is_in_table = initial_gvn()->hash_delete(use); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
305 uint uses_found = 0; |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
306 for (uint j = 0; j < use->len(); j++) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
307 if (use->in(j) == n) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
308 if (j < use->req()) |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
309 use->set_req(j, nn); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
310 else |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
311 use->set_prec(j, nn); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
312 uses_found++; |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
313 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
314 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
315 if (is_in_table) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
316 // reinsert into table |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
317 initial_gvn()->hash_find_insert(use); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
318 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
319 record_for_igvn(use); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
320 i -= uses_found; // we deleted 1 or more copies of this edge |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
321 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
322 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
323 |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
324 |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
325 static inline bool not_a_node(const Node* n) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
326 if (n == NULL) return true; |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
327 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
328 if (*(address*)n == badAddress) return true; // kill by Node::destruct |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
329 return false; |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
330 } |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
331 |
0 | 332 // Identify all nodes that are reachable from below, useful. |
333 // Use breadth-first pass that records state in a Unique_Node_List, | |
334 // recursive traversal is slower. | |
335 void Compile::identify_useful_nodes(Unique_Node_List &useful) { | |
336 int estimated_worklist_size = unique(); | |
337 useful.map( estimated_worklist_size, NULL ); // preallocate space | |
338 | |
339 // Initialize worklist | |
340 if (root() != NULL) { useful.push(root()); } | |
341 // If 'top' is cached, declare it useful to preserve cached node | |
342 if( cached_top_node() ) { useful.push(cached_top_node()); } | |
343 | |
344 // Push all useful nodes onto the list, breadthfirst | |
345 for( uint next = 0; next < useful.size(); ++next ) { | |
346 assert( next < unique(), "Unique useful nodes < total nodes"); | |
347 Node *n = useful.at(next); | |
348 uint max = n->len(); | |
349 for( uint i = 0; i < max; ++i ) { | |
350 Node *m = n->in(i); | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
351 if (not_a_node(m)) continue; |
0 | 352 useful.push(m); |
353 } | |
354 } | |
355 } | |
356 | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
357 // Update dead_node_list with any missing dead nodes using useful |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
358 // list. Consider all non-useful nodes to be useless i.e., dead nodes. |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
359 void Compile::update_dead_node_list(Unique_Node_List &useful) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
360 uint max_idx = unique(); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
361 VectorSet& useful_node_set = useful.member_set(); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
362 |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
363 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
364 // If node with index node_idx is not in useful set, |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
365 // mark it as dead in dead node list. |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
366 if (! useful_node_set.test(node_idx) ) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
367 record_dead_node(node_idx); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
368 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
369 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
370 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
371 |
7473 | 372 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { |
373 int shift = 0; | |
374 for (int i = 0; i < inlines->length(); i++) { | |
375 CallGenerator* cg = inlines->at(i); | |
376 CallNode* call = cg->call_node(); | |
377 if (shift > 0) { | |
378 inlines->at_put(i-shift, cg); | |
379 } | |
380 if (!useful.member(call)) { | |
381 shift++; | |
382 } | |
383 } | |
384 inlines->trunc_to(inlines->length()-shift); | |
385 } | |
386 | |
0 | 387 // Disconnect all useless nodes by disconnecting those at the boundary. |
388 void Compile::remove_useless_nodes(Unique_Node_List &useful) { | |
389 uint next = 0; | |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
390 while (next < useful.size()) { |
0 | 391 Node *n = useful.at(next++); |
392 // Use raw traversal of out edges since this code removes out edges | |
393 int max = n->outcnt(); | |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
394 for (int j = 0; j < max; ++j) { |
0 | 395 Node* child = n->raw_out(j); |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
396 if (! useful.member(child)) { |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
397 assert(!child->is_top() || child != top(), |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
398 "If top is cached in Compile object it is in useful list"); |
0 | 399 // Only need to remove this out-edge to the useless node |
400 n->raw_del_out(j); | |
401 --j; | |
402 --max; | |
403 } | |
404 } | |
405 if (n->outcnt() == 1 && n->has_special_unique_user()) { | |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
406 record_for_igvn(n->unique_out()); |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
407 } |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
408 } |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
409 // Remove useless macro and predicate opaq nodes |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
410 for (int i = C->macro_count()-1; i >= 0; i--) { |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
411 Node* n = C->macro_node(i); |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
412 if (!useful.member(n)) { |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
413 remove_macro_node(n); |
0 | 414 } |
415 } | |
8048
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
416 // Remove useless expensive node |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
417 for (int i = C->expensive_count()-1; i >= 0; i--) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
418 Node* n = C->expensive_node(i); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
419 if (!useful.member(n)) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
420 remove_expensive_node(n); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
421 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
422 } |
7473 | 423 // clean up the late inline lists |
424 remove_useless_late_inlines(&_string_late_inlines, useful); | |
10278 | 425 remove_useless_late_inlines(&_boxing_late_inlines, useful); |
7473 | 426 remove_useless_late_inlines(&_late_inlines, useful); |
0 | 427 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) |
428 } | |
429 | |
430 //------------------------------frame_size_in_words----------------------------- | |
431 // frame_slots in units of words | |
432 int Compile::frame_size_in_words() const { | |
433 // shift is 0 in LP32 and 1 in LP64 | |
434 const int shift = (LogBytesPerWord - LogBytesPerInt); | |
435 int words = _frame_slots >> shift; | |
436 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); | |
437 return words; | |
438 } | |
439 | |
440 // ============================================================================ | |
441 //------------------------------CompileWrapper--------------------------------- | |
442 class CompileWrapper : public StackObj { | |
443 Compile *const _compile; | |
444 public: | |
445 CompileWrapper(Compile* compile); | |
446 | |
447 ~CompileWrapper(); | |
448 }; | |
449 | |
450 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { | |
451 // the Compile* pointer is stored in the current ciEnv: | |
452 ciEnv* env = compile->env(); | |
453 assert(env == ciEnv::current(), "must already be a ciEnv active"); | |
454 assert(env->compiler_data() == NULL, "compile already active?"); | |
455 env->set_compiler_data(compile); | |
456 assert(compile == Compile::current(), "sanity"); | |
457 | |
458 compile->set_type_dict(NULL); | |
459 compile->set_type_hwm(NULL); | |
460 compile->set_type_last_size(0); | |
461 compile->set_last_tf(NULL, NULL); | |
462 compile->set_indexSet_arena(NULL); | |
463 compile->set_indexSet_free_block_list(NULL); | |
464 compile->init_type_arena(); | |
465 Type::Initialize(compile); | |
466 _compile->set_scratch_buffer_blob(NULL); | |
467 _compile->begin_method(); | |
468 } | |
469 CompileWrapper::~CompileWrapper() { | |
470 _compile->end_method(); | |
471 if (_compile->scratch_buffer_blob() != NULL) | |
472 BufferBlob::free(_compile->scratch_buffer_blob()); | |
473 _compile->env()->set_compiler_data(NULL); | |
474 } | |
475 | |
476 | |
477 //----------------------------print_compile_messages--------------------------- | |
478 void Compile::print_compile_messages() { | |
479 #ifndef PRODUCT | |
480 // Check if recompiling | |
481 if (_subsume_loads == false && PrintOpto) { | |
482 // Recompiling without allowing machine instructions to subsume loads | |
483 tty->print_cr("*********************************************************"); | |
484 tty->print_cr("** Bailout: Recompile without subsuming loads **"); | |
485 tty->print_cr("*********************************************************"); | |
486 } | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
487 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
488 // Recompiling without escape analysis |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
489 tty->print_cr("*********************************************************"); |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
490 tty->print_cr("** Bailout: Recompile without escape analysis **"); |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
491 tty->print_cr("*********************************************************"); |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
492 } |
10278 | 493 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { |
494 // Recompiling without boxing elimination | |
495 tty->print_cr("*********************************************************"); | |
496 tty->print_cr("** Bailout: Recompile without boxing elimination **"); | |
497 tty->print_cr("*********************************************************"); | |
498 } | |
0 | 499 if (env()->break_at_compile()) { |
605 | 500 // Open the debugger when compiling this method. |
0 | 501 tty->print("### Breaking when compiling: "); |
502 method()->print_short_name(); | |
503 tty->cr(); | |
504 BREAKPOINT; | |
505 } | |
506 | |
507 if( PrintOpto ) { | |
508 if (is_osr_compilation()) { | |
509 tty->print("[OSR]%3d", _compile_id); | |
510 } else { | |
511 tty->print("%3d", _compile_id); | |
512 } | |
513 } | |
514 #endif | |
515 } | |
516 | |
517 | |
2091
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
518 //-----------------------init_scratch_buffer_blob------------------------------ |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
519 // Construct a temporary BufferBlob and cache it for this compile. |
2008 | 520 void Compile::init_scratch_buffer_blob(int const_size) { |
2091
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
521 // If there is already a scratch buffer blob allocated and the |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
522 // constant section is big enough, use it. Otherwise free the |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
523 // current and allocate a new one. |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
524 BufferBlob* blob = scratch_buffer_blob(); |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
525 if ((blob != NULL) && (const_size <= _scratch_const_size)) { |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
526 // Use the current blob. |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
527 } else { |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
528 if (blob != NULL) { |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
529 BufferBlob::free(blob); |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
530 } |
0 | 531 |
2091
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
532 ResourceMark rm; |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
533 _scratch_const_size = const_size; |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
534 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
535 blob = BufferBlob::create("Compile::scratch_buffer", size); |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
536 // Record the buffer blob for next time. |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
537 set_scratch_buffer_blob(blob); |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
538 // Have we run out of code space? |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
539 if (scratch_buffer_blob() == NULL) { |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
540 // Let CompilerBroker disable further compilations. |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
541 record_failure("Not enough space for scratch buffer in CodeCache"); |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
542 return; |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
543 } |
163 | 544 } |
0 | 545 |
546 // Initialize the relocation buffers | |
1748 | 547 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; |
0 | 548 set_scratch_locs_memory(locs_buf); |
549 } | |
550 | |
551 | |
552 //-----------------------scratch_emit_size------------------------------------- | |
553 // Helper function that computes size by emitting code | |
554 uint Compile::scratch_emit_size(const Node* n) { | |
2008 | 555 // Start scratch_emit_size section. |
556 set_in_scratch_emit_size(true); | |
557 | |
0 | 558 // Emit into a trash buffer and count bytes emitted. |
559 // This is a pretty expensive way to compute a size, | |
560 // but it works well enough if seldom used. | |
561 // All common fixed-size instructions are given a size | |
562 // method by the AD file. | |
563 // Note that the scratch buffer blob and locs memory are | |
564 // allocated at the beginning of the compile task, and | |
565 // may be shared by several calls to scratch_emit_size. | |
566 // The allocation of the scratch buffer blob is particularly | |
567 // expensive, since it has to grab the code cache lock. | |
568 BufferBlob* blob = this->scratch_buffer_blob(); | |
569 assert(blob != NULL, "Initialize BufferBlob at start"); | |
570 assert(blob->size() > MAX_inst_size, "sanity"); | |
571 relocInfo* locs_buf = scratch_locs_memory(); | |
1748 | 572 address blob_begin = blob->content_begin(); |
0 | 573 address blob_end = (address)locs_buf; |
1748 | 574 assert(blob->content_contains(blob_end), "sanity"); |
0 | 575 CodeBuffer buf(blob_begin, blob_end - blob_begin); |
2008 | 576 buf.initialize_consts_size(_scratch_const_size); |
0 | 577 buf.initialize_stubs_size(MAX_stubs_size); |
578 assert(locs_buf != NULL, "sanity"); | |
2008 | 579 int lsize = MAX_locs_size / 3; |
580 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); | |
581 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); | |
582 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); | |
583 | |
584 // Do the emission. | |
3839 | 585 |
586 Label fakeL; // Fake label for branch instructions. | |
3853
11211f7cb5a0
7079317: Incorrect branch's destination block in PrintoOptoAssembly output
kvn
parents:
3839
diff
changeset
|
587 Label* saveL = NULL; |
11211f7cb5a0
7079317: Incorrect branch's destination block in PrintoOptoAssembly output
kvn
parents:
3839
diff
changeset
|
588 uint save_bnum = 0; |
11211f7cb5a0
7079317: Incorrect branch's destination block in PrintoOptoAssembly output
kvn
parents:
3839
diff
changeset
|
589 bool is_branch = n->is_MachBranch(); |
3839 | 590 if (is_branch) { |
591 MacroAssembler masm(&buf); | |
592 masm.bind(fakeL); | |
3853
11211f7cb5a0
7079317: Incorrect branch's destination block in PrintoOptoAssembly output
kvn
parents:
3839
diff
changeset
|
593 n->as_MachBranch()->save_label(&saveL, &save_bnum); |
11211f7cb5a0
7079317: Incorrect branch's destination block in PrintoOptoAssembly output
kvn
parents:
3839
diff
changeset
|
594 n->as_MachBranch()->label_set(&fakeL, 0); |
3839 | 595 } |
0 | 596 n->emit(buf, this->regalloc()); |
3853
11211f7cb5a0
7079317: Incorrect branch's destination block in PrintoOptoAssembly output
kvn
parents:
3839
diff
changeset
|
597 if (is_branch) // Restore label. |
11211f7cb5a0
7079317: Incorrect branch's destination block in PrintoOptoAssembly output
kvn
parents:
3839
diff
changeset
|
598 n->as_MachBranch()->label_set(saveL, save_bnum); |
2008 | 599 |
600 // End scratch_emit_size section. | |
601 set_in_scratch_emit_size(false); | |
602 | |
1748 | 603 return buf.insts_size(); |
0 | 604 } |
605 | |
606 | |
607 // ============================================================================ | |
608 //------------------------------Compile standard------------------------------- | |
609 debug_only( int Compile::_debug_idx = 100000; ) | |
610 | |
611 // Compile a method. entry_bci is -1 for normal compilations and indicates | |
612 // the continuation bci for on stack replacement. | |
613 | |
614 | |
10278 | 615 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, |
616 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) | |
0 | 617 : Phase(Compiler), |
618 _env(ci_env), | |
619 _log(ci_env->log()), | |
620 _compile_id(ci_env->compile_id()), | |
621 _save_argument_registers(false), | |
622 _stub_name(NULL), | |
623 _stub_function(NULL), | |
624 _stub_entry_point(NULL), | |
625 _method(target), | |
626 _entry_bci(osr_bci), | |
627 _initial_gvn(NULL), | |
628 _for_igvn(NULL), | |
629 _warm_calls(NULL), | |
630 _subsume_loads(subsume_loads), | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
631 _do_escape_analysis(do_escape_analysis), |
10278 | 632 _eliminate_boxing(eliminate_boxing), |
0 | 633 _failure_reason(NULL), |
634 _code_buffer("Compile::Fill_buffer"), | |
635 _orig_pc_slot(0), | |
636 _orig_pc_slot_offset_in_bytes(0), | |
1265 | 637 _has_method_handle_invokes(false), |
2008 | 638 _mach_constant_base_node(NULL), |
0 | 639 _node_bundling_limit(0), |
640 _node_bundling_base(NULL), | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
641 _java_calls(0), |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
642 _inner_loops(0), |
2008 | 643 _scratch_const_size(-1), |
644 _in_scratch_emit_size(false), | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
645 _dead_node_list(comp_arena()), |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
646 _dead_node_count(0), |
0 | 647 #ifndef PRODUCT |
648 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), | |
649 _printer(IdealGraphPrinter::printer()), | |
650 #endif | |
7421
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
651 _congraph(NULL), |
7473 | 652 _late_inlines(comp_arena(), 2, 0, NULL), |
653 _string_late_inlines(comp_arena(), 2, 0, NULL), | |
10278 | 654 _boxing_late_inlines(comp_arena(), 2, 0, NULL), |
7473 | 655 _late_inlines_pos(0), |
656 _number_of_mh_late_inlines(0), | |
657 _inlining_progress(false), | |
658 _inlining_incrementally(false), | |
7421
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
659 _print_inlining_list(NULL), |
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
660 _print_inlining(0) { |
0 | 661 C = this; |
662 | |
663 CompileWrapper cw(this); | |
664 #ifndef PRODUCT | |
665 if (TimeCompiler2) { | |
666 tty->print(" "); | |
667 target->holder()->name()->print(); | |
668 tty->print("."); | |
669 target->print_short_name(); | |
670 tty->print(" "); | |
671 } | |
672 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); | |
673 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); | |
100
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
674 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
675 if (!print_opto_assembly) { |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
676 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
677 if (print_assembly && !Disassembler::can_decode()) { |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
678 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
679 print_opto_assembly = true; |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
680 } |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
681 } |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
682 set_print_assembly(print_opto_assembly); |
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
333
diff
changeset
|
683 set_parsed_irreducible_loop(false); |
0 | 684 #endif |
685 | |
686 if (ProfileTraps) { | |
687 // Make sure the method being compiled gets its own MDO, | |
688 // so we can at least track the decompile_count(). | |
2007
5ddfcf4b079e
7003554: (tiered) assert(is_null_object() || handle() != NULL) failed: cannot embed null pointer
iveresov
parents:
1972
diff
changeset
|
689 method()->ensure_method_data(); |
0 | 690 } |
691 | |
692 Init(::AliasLevel); | |
693 | |
694 | |
695 print_compile_messages(); | |
696 | |
697 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) | |
698 _ilt = InlineTree::build_inline_tree_root(); | |
699 else | |
700 _ilt = NULL; | |
701 | |
702 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice | |
703 assert(num_alias_types() >= AliasIdxRaw, ""); | |
704 | |
705 #define MINIMUM_NODE_HASH 1023 | |
706 // Node list that Iterative GVN will start with | |
707 Unique_Node_List for_igvn(comp_arena()); | |
708 set_for_igvn(&for_igvn); | |
709 | |
710 // GVN that will be run immediately on new nodes | |
711 uint estimated_size = method()->code_size()*4+64; | |
712 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); | |
713 PhaseGVN gvn(node_arena(), estimated_size); | |
714 set_initial_gvn(&gvn); | |
715 | |
7605 | 716 if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) { |
7421
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
717 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); |
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
718 } |
0 | 719 { // Scope for timing the parser |
720 TracePhase t3("parse", &_t_parser, true); | |
721 | |
722 // Put top into the hash table ASAP. | |
723 initial_gvn()->transform_no_reclaim(top()); | |
724 | |
725 // Set up tf(), start(), and find a CallGenerator. | |
3249
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
726 CallGenerator* cg = NULL; |
0 | 727 if (is_osr_compilation()) { |
728 const TypeTuple *domain = StartOSRNode::osr_domain(); | |
729 const TypeTuple *range = TypeTuple::make_range(method()->signature()); | |
730 init_tf(TypeFunc::make(domain, range)); | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
731 StartNode* s = new (this) StartOSRNode(root(), domain); |
0 | 732 initial_gvn()->set_type_bottom(s); |
733 init_start(s); | |
734 cg = CallGenerator::for_osr(method(), entry_bci()); | |
735 } else { | |
736 // Normal case. | |
737 init_tf(TypeFunc::make(method())); | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
738 StartNode* s = new (this) StartNode(root(), tf()->domain()); |
0 | 739 initial_gvn()->set_type_bottom(s); |
740 init_start(s); | |
3249
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
741 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
742 // With java.lang.ref.reference.get() we must go through the |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
743 // intrinsic when G1 is enabled - even when get() is the root |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
744 // method of the compile - so that, if necessary, the value in |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
745 // the referent field of the reference object gets recorded by |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
746 // the pre-barrier code. |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
747 // Specifically, if G1 is enabled, the value in the referent |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
748 // field is recorded by the G1 SATB pre barrier. This will |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
749 // result in the referent being marked live and the reference |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
750 // object removed from the list of discovered references during |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
751 // reference processing. |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
752 cg = find_intrinsic(method(), false); |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
753 } |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
754 if (cg == NULL) { |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
755 float past_uses = method()->interpreter_invocation_count(); |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
756 float expected_uses = past_uses; |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
757 cg = CallGenerator::for_inline(method(), expected_uses); |
e1162778c1c8
7009266: G1: assert(obj->is_oop_or_null(true )) failed: Error
johnc
parents:
3248
diff
changeset
|
758 } |
0 | 759 } |
760 if (failing()) return; | |
761 if (cg == NULL) { | |
762 record_method_not_compilable_all_tiers("cannot parse method"); | |
763 return; | |
764 } | |
765 JVMState* jvms = build_start_state(start(), tf()); | |
766 if ((jvms = cg->generate(jvms)) == NULL) { | |
767 record_method_not_compilable("method parse failed"); | |
768 return; | |
769 } | |
770 GraphKit kit(jvms); | |
771 | |
772 if (!kit.stopped()) { | |
773 // Accept return values, and transfer control we know not where. | |
774 // This is done by a special, unique ReturnNode bound to root. | |
775 return_values(kit.jvms()); | |
776 } | |
777 | |
778 if (kit.has_exceptions()) { | |
779 // Any exceptions that escape from this call must be rethrown | |
780 // to whatever caller is dynamically above us on the stack. | |
781 // This is done by a special, unique RethrowNode bound to root. | |
782 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); | |
783 } | |
784 | |
7473 | 785 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); |
786 | |
787 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { | |
788 inline_string_calls(true); | |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
789 } |
7473 | 790 |
791 if (failing()) return; | |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
792 |
10405 | 793 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); |
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
333
diff
changeset
|
794 |
0 | 795 // Remove clutter produced by parsing. |
796 if (!failing()) { | |
797 ResourceMark rm; | |
798 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); | |
799 } | |
800 } | |
801 | |
802 // Note: Large methods are capped off in do_one_bytecode(). | |
803 if (failing()) return; | |
804 | |
805 // After parsing, node notes are no longer automagic. | |
806 // They must be propagated by register_new_node_with_optimizer(), | |
807 // clone(), or the like. | |
808 set_default_node_notes(NULL); | |
809 | |
810 for (;;) { | |
811 int successes = Inline_Warm(); | |
812 if (failing()) return; | |
813 if (successes == 0) break; | |
814 } | |
815 | |
816 // Drain the list. | |
817 Finish_Warm(); | |
818 #ifndef PRODUCT | |
819 if (_printer) { | |
820 _printer->print_inlining(this); | |
821 } | |
822 #endif | |
823 | |
824 if (failing()) return; | |
825 NOT_PRODUCT( verify_graph_edges(); ) | |
826 | |
827 // Now optimize | |
828 Optimize(); | |
829 if (failing()) return; | |
830 NOT_PRODUCT( verify_graph_edges(); ) | |
831 | |
832 #ifndef PRODUCT | |
833 if (PrintIdeal) { | |
834 ttyLocker ttyl; // keep the following output all in one block | |
835 // This output goes directly to the tty, not the compiler log. | |
836 // To enable tools to match it up with the compilation activity, | |
837 // be sure to tag this tty output with the compile ID. | |
838 if (xtty != NULL) { | |
839 xtty->head("ideal compile_id='%d'%s", compile_id(), | |
840 is_osr_compilation() ? " compile_kind='osr'" : | |
841 ""); | |
842 } | |
843 root()->dump(9999); | |
844 if (xtty != NULL) { | |
845 xtty->tail("ideal"); | |
846 } | |
847 } | |
848 #endif | |
849 | |
850 // Now that we know the size of all the monitors we can add a fixed slot | |
851 // for the original deopt pc. | |
852 | |
853 _orig_pc_slot = fixed_slots(); | |
854 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); | |
855 set_fixed_slots(next_slot); | |
856 | |
857 // Now generate code | |
858 Code_Gen(); | |
859 if (failing()) return; | |
860 | |
861 // Check if we want to skip execution of all compiled code. | |
862 { | |
863 #ifndef PRODUCT | |
864 if (OptoNoExecute) { | |
865 record_method_not_compilable("+OptoNoExecute"); // Flag as failed | |
866 return; | |
867 } | |
868 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); | |
869 #endif | |
870 | |
871 if (is_osr_compilation()) { | |
872 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); | |
873 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); | |
874 } else { | |
875 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); | |
876 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); | |
877 } | |
878 | |
879 env()->register_method(_method, _entry_bci, | |
880 &_code_offsets, | |
881 _orig_pc_slot_offset_in_bytes, | |
882 code_buffer(), | |
883 frame_size_in_words(), _oop_map_set, | |
884 &_handler_table, &_inc_table, | |
885 compiler, | |
886 env()->comp_level(), | |
6792
137868b7aa6f
7196199: java/text/Bidi/Bug6665028.java failed: Bidi run count incorrect
kvn
parents:
6725
diff
changeset
|
887 has_unsafe_access(), |
137868b7aa6f
7196199: java/text/Bidi/Bug6665028.java failed: Bidi run count incorrect
kvn
parents:
6725
diff
changeset
|
888 SharedRuntime::is_wide_vector(max_vector_size()) |
0 | 889 ); |
6843 | 890 |
891 if (log() != NULL) // Print code cache state into compiler log | |
892 log()->code_cache_state(); | |
0 | 893 } |
894 } | |
895 | |
896 //------------------------------Compile---------------------------------------- | |
897 // Compile a runtime stub | |
898 Compile::Compile( ciEnv* ci_env, | |
899 TypeFunc_generator generator, | |
900 address stub_function, | |
901 const char *stub_name, | |
902 int is_fancy_jump, | |
903 bool pass_tls, | |
904 bool save_arg_registers, | |
905 bool return_pc ) | |
906 : Phase(Compiler), | |
907 _env(ci_env), | |
908 _log(ci_env->log()), | |
8730
b7c2c5b2572c
8005772: Stubs report compile id -1 in phase events
neliasso
parents:
8694
diff
changeset
|
909 _compile_id(0), |
0 | 910 _save_argument_registers(save_arg_registers), |
911 _method(NULL), | |
912 _stub_name(stub_name), | |
913 _stub_function(stub_function), | |
914 _stub_entry_point(NULL), | |
915 _entry_bci(InvocationEntryBci), | |
916 _initial_gvn(NULL), | |
917 _for_igvn(NULL), | |
918 _warm_calls(NULL), | |
919 _orig_pc_slot(0), | |
920 _orig_pc_slot_offset_in_bytes(0), | |
921 _subsume_loads(true), | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
922 _do_escape_analysis(false), |
10278 | 923 _eliminate_boxing(false), |
0 | 924 _failure_reason(NULL), |
925 _code_buffer("Compile::Fill_buffer"), | |
1265 | 926 _has_method_handle_invokes(false), |
2008 | 927 _mach_constant_base_node(NULL), |
0 | 928 _node_bundling_limit(0), |
929 _node_bundling_base(NULL), | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
930 _java_calls(0), |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
931 _inner_loops(0), |
0 | 932 #ifndef PRODUCT |
933 _trace_opto_output(TraceOptoOutput), | |
934 _printer(NULL), | |
935 #endif | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
936 _dead_node_list(comp_arena()), |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
937 _dead_node_count(0), |
7421
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
938 _congraph(NULL), |
7473 | 939 _number_of_mh_late_inlines(0), |
940 _inlining_progress(false), | |
941 _inlining_incrementally(false), | |
7421
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
942 _print_inlining_list(NULL), |
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
943 _print_inlining(0) { |
0 | 944 C = this; |
945 | |
946 #ifndef PRODUCT | |
947 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); | |
948 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); | |
949 set_print_assembly(PrintFrameConverterAssembly); | |
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
333
diff
changeset
|
950 set_parsed_irreducible_loop(false); |
0 | 951 #endif |
952 CompileWrapper cw(this); | |
953 Init(/*AliasLevel=*/ 0); | |
954 init_tf((*generator)()); | |
955 | |
956 { | |
957 // The following is a dummy for the sake of GraphKit::gen_stub | |
958 Unique_Node_List for_igvn(comp_arena()); | |
959 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this | |
960 PhaseGVN gvn(Thread::current()->resource_area(),255); | |
961 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively | |
962 gvn.transform_no_reclaim(top()); | |
963 | |
964 GraphKit kit; | |
965 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); | |
966 } | |
967 | |
968 NOT_PRODUCT( verify_graph_edges(); ) | |
969 Code_Gen(); | |
970 if (failing()) return; | |
971 | |
972 | |
973 // Entry point will be accessed using compile->stub_entry_point(); | |
974 if (code_buffer() == NULL) { | |
975 Matcher::soft_match_failure(); | |
976 } else { | |
977 if (PrintAssembly && (WizardMode || Verbose)) | |
978 tty->print_cr("### Stub::%s", stub_name); | |
979 | |
980 if (!failing()) { | |
981 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); | |
982 | |
983 // Make the NMethod | |
984 // For now we mark the frame as never safe for profile stackwalking | |
985 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, | |
986 code_buffer(), | |
987 CodeOffsets::frame_never_safe, | |
988 // _code_offsets.value(CodeOffsets::Frame_Complete), | |
989 frame_size_in_words(), | |
990 _oop_map_set, | |
991 save_arg_registers); | |
992 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); | |
993 | |
994 _stub_entry_point = rs->entry_point(); | |
995 } | |
996 } | |
997 } | |
998 | |
999 //------------------------------Init------------------------------------------- | |
1000 // Prepare for a single compilation | |
1001 void Compile::Init(int aliaslevel) { | |
1002 _unique = 0; | |
1003 _regalloc = NULL; | |
1004 | |
1005 _tf = NULL; // filled in later | |
1006 _top = NULL; // cached later | |
1007 _matcher = NULL; // filled in later | |
1008 _cfg = NULL; // filled in later | |
1009 | |
1010 set_24_bit_selection_and_mode(Use24BitFP, false); | |
1011 | |
1012 _node_note_array = NULL; | |
1013 _default_node_notes = NULL; | |
1014 | |
1015 _immutable_memory = NULL; // filled in at first inquiry | |
1016 | |
1017 // Globally visible Nodes | |
1018 // First set TOP to NULL to give safe behavior during creation of RootNode | |
1019 set_cached_top_node(NULL); | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
1020 set_root(new (this) RootNode()); |
0 | 1021 // Now that you have a Root to point to, create the real TOP |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
1022 set_cached_top_node( new (this) ConNode(Type::TOP) ); |
0 | 1023 set_recent_alloc(NULL, NULL); |
1024 | |
1025 // Create Debug Information Recorder to record scopes, oopmaps, etc. | |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
1026 env()->set_oop_recorder(new OopRecorder(env()->arena())); |
0 | 1027 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); |
1028 env()->set_dependencies(new Dependencies(env())); | |
1029 | |
1030 _fixed_slots = 0; | |
1031 set_has_split_ifs(false); | |
1032 set_has_loops(has_method() && method()->has_loops()); // first approximation | |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
1033 set_has_stringbuilder(false); |
10278 | 1034 set_has_boxed_value(false); |
0 | 1035 _trap_can_recompile = false; // no traps emitted yet |
1036 _major_progress = true; // start out assuming good things will happen | |
1037 set_has_unsafe_access(false); | |
6792
137868b7aa6f
7196199: java/text/Bidi/Bug6665028.java failed: Bidi run count incorrect
kvn
parents:
6725
diff
changeset
|
1038 set_max_vector_size(0); |
0 | 1039 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); |
1040 set_decompile_count(0); | |
1041 | |
418 | 1042 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); |
1783 | 1043 set_num_loop_opts(LoopOptsCount); |
1044 set_do_inlining(Inline); | |
1045 set_max_inline_size(MaxInlineSize); | |
1046 set_freq_inline_size(FreqInlineSize); | |
1047 set_do_scheduling(OptoScheduling); | |
1048 set_do_count_invocations(false); | |
1049 set_do_method_data_update(false); | |
0 | 1050 |
1051 if (debug_info()->recording_non_safepoints()) { | |
1052 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> | |
1053 (comp_arena(), 8, 0, NULL)); | |
1054 set_default_node_notes(Node_Notes::make(this)); | |
1055 } | |
1056 | |
1057 // // -- Initialize types before each compile -- | |
1058 // // Update cached type information | |
1059 // if( _method && _method->constants() ) | |
1060 // Type::update_loaded_types(_method, _method->constants()); | |
1061 | |
1062 // Init alias_type map. | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
1063 if (!_do_escape_analysis && aliaslevel == 3) |
0 | 1064 aliaslevel = 2; // No unique types without escape analysis |
1065 _AliasLevel = aliaslevel; | |
1066 const int grow_ats = 16; | |
1067 _max_alias_types = grow_ats; | |
1068 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); | |
1069 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); | |
1070 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); | |
1071 { | |
1072 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; | |
1073 } | |
1074 // Initialize the first few types. | |
1075 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); | |
1076 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); | |
1077 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); | |
1078 _num_alias_types = AliasIdxRaw+1; | |
1079 // Zero out the alias type cache. | |
1080 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); | |
1081 // A NULL adr_type hits in the cache right away. Preload the right answer. | |
1082 probe_alias_cache(NULL)->_index = AliasIdxTop; | |
1083 | |
1084 _intrinsics = NULL; | |
1685 | 1085 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); |
1086 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); | |
8048
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
1087 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); |
0 | 1088 register_library_intrinsics(); |
1089 } | |
1090 | |
1091 //---------------------------init_start---------------------------------------- | |
1092 // Install the StartNode on this compile object. | |
1093 void Compile::init_start(StartNode* s) { | |
1094 if (failing()) | |
1095 return; // already failing | |
1096 assert(s == start(), ""); | |
1097 } | |
1098 | |
1099 StartNode* Compile::start() const { | |
1100 assert(!failing(), ""); | |
1101 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { | |
1102 Node* start = root()->fast_out(i); | |
1103 if( start->is_Start() ) | |
1104 return start->as_Start(); | |
1105 } | |
1106 ShouldNotReachHere(); | |
1107 return NULL; | |
1108 } | |
1109 | |
1110 //-------------------------------immutable_memory------------------------------------- | |
1111 // Access immutable memory | |
1112 Node* Compile::immutable_memory() { | |
1113 if (_immutable_memory != NULL) { | |
1114 return _immutable_memory; | |
1115 } | |
1116 StartNode* s = start(); | |
1117 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { | |
1118 Node *p = s->fast_out(i); | |
1119 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { | |
1120 _immutable_memory = p; | |
1121 return _immutable_memory; | |
1122 } | |
1123 } | |
1124 ShouldNotReachHere(); | |
1125 return NULL; | |
1126 } | |
1127 | |
1128 //----------------------set_cached_top_node------------------------------------ | |
1129 // Install the cached top node, and make sure Node::is_top works correctly. | |
1130 void Compile::set_cached_top_node(Node* tn) { | |
1131 if (tn != NULL) verify_top(tn); | |
1132 Node* old_top = _top; | |
1133 _top = tn; | |
1134 // Calling Node::setup_is_top allows the nodes the chance to adjust | |
1135 // their _out arrays. | |
1136 if (_top != NULL) _top->setup_is_top(); | |
1137 if (old_top != NULL) old_top->setup_is_top(); | |
1138 assert(_top == NULL || top()->is_top(), ""); | |
1139 } | |
1140 | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1141 #ifdef ASSERT |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1142 uint Compile::count_live_nodes_by_graph_walk() { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1143 Unique_Node_List useful(comp_arena()); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1144 // Get useful node list by walking the graph. |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1145 identify_useful_nodes(useful); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1146 return useful.size(); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1147 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1148 |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1149 void Compile::print_missing_nodes() { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1150 |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1151 // Return if CompileLog is NULL and PrintIdealNodeCount is false. |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1152 if ((_log == NULL) && (! PrintIdealNodeCount)) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1153 return; |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1154 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1155 |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1156 // This is an expensive function. It is executed only when the user |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1157 // specifies VerifyIdealNodeCount option or otherwise knows the |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1158 // additional work that needs to be done to identify reachable nodes |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1159 // by walking the flow graph and find the missing ones using |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1160 // _dead_node_list. |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1161 |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1162 Unique_Node_List useful(comp_arena()); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1163 // Get useful node list by walking the graph. |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1164 identify_useful_nodes(useful); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1165 |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1166 uint l_nodes = C->live_nodes(); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1167 uint l_nodes_by_walk = useful.size(); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1168 |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1169 if (l_nodes != l_nodes_by_walk) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1170 if (_log != NULL) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1171 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1172 _log->stamp(); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1173 _log->end_head(); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1174 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1175 VectorSet& useful_member_set = useful.member_set(); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1176 int last_idx = l_nodes_by_walk; |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1177 for (int i = 0; i < last_idx; i++) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1178 if (useful_member_set.test(i)) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1179 if (_dead_node_list.test(i)) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1180 if (_log != NULL) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1181 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1182 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1183 if (PrintIdealNodeCount) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1184 // Print the log message to tty |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1185 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1186 useful.at(i)->dump(); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1187 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1188 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1189 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1190 else if (! _dead_node_list.test(i)) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1191 if (_log != NULL) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1192 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1193 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1194 if (PrintIdealNodeCount) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1195 // Print the log message to tty |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1196 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1197 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1198 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1199 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1200 if (_log != NULL) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1201 _log->tail("mismatched_nodes"); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1202 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1203 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1204 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1205 #endif |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
1206 |
0 | 1207 #ifndef PRODUCT |
1208 void Compile::verify_top(Node* tn) const { | |
1209 if (tn != NULL) { | |
1210 assert(tn->is_Con(), "top node must be a constant"); | |
1211 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); | |
1212 assert(tn->in(0) != NULL, "must have live top node"); | |
1213 } | |
1214 } | |
1215 #endif | |
1216 | |
1217 | |
1218 ///-------------------Managing Per-Node Debug & Profile Info------------------- | |
1219 | |
1220 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { | |
1221 guarantee(arr != NULL, ""); | |
1222 int num_blocks = arr->length(); | |
1223 if (grow_by < num_blocks) grow_by = num_blocks; | |
1224 int num_notes = grow_by * _node_notes_block_size; | |
1225 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); | |
1226 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); | |
1227 while (num_notes > 0) { | |
1228 arr->append(notes); | |
1229 notes += _node_notes_block_size; | |
1230 num_notes -= _node_notes_block_size; | |
1231 } | |
1232 assert(num_notes == 0, "exact multiple, please"); | |
1233 } | |
1234 | |
1235 bool Compile::copy_node_notes_to(Node* dest, Node* source) { | |
1236 if (source == NULL || dest == NULL) return false; | |
1237 | |
1238 if (dest->is_Con()) | |
1239 return false; // Do not push debug info onto constants. | |
1240 | |
1241 #ifdef ASSERT | |
1242 // Leave a bread crumb trail pointing to the original node: | |
1243 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { | |
1244 dest->set_debug_orig(source); | |
1245 } | |
1246 #endif | |
1247 | |
1248 if (node_note_array() == NULL) | |
1249 return false; // Not collecting any notes now. | |
1250 | |
1251 // This is a copy onto a pre-existing node, which may already have notes. | |
1252 // If both nodes have notes, do not overwrite any pre-existing notes. | |
1253 Node_Notes* source_notes = node_notes_at(source->_idx); | |
1254 if (source_notes == NULL || source_notes->is_clear()) return false; | |
1255 Node_Notes* dest_notes = node_notes_at(dest->_idx); | |
1256 if (dest_notes == NULL || dest_notes->is_clear()) { | |
1257 return set_node_notes_at(dest->_idx, source_notes); | |
1258 } | |
1259 | |
1260 Node_Notes merged_notes = (*source_notes); | |
1261 // The order of operations here ensures that dest notes will win... | |
1262 merged_notes.update_from(dest_notes); | |
1263 return set_node_notes_at(dest->_idx, &merged_notes); | |
1264 } | |
1265 | |
1266 | |
1267 //--------------------------allow_range_check_smearing------------------------- | |
1268 // Gating condition for coalescing similar range checks. | |
1269 // Sometimes we try 'speculatively' replacing a series of a range checks by a | |
1270 // single covering check that is at least as strong as any of them. | |
1271 // If the optimization succeeds, the simplified (strengthened) range check | |
1272 // will always succeed. If it fails, we will deopt, and then give up | |
1273 // on the optimization. | |
1274 bool Compile::allow_range_check_smearing() const { | |
1275 // If this method has already thrown a range-check, | |
1276 // assume it was because we already tried range smearing | |
1277 // and it failed. | |
1278 uint already_trapped = trap_count(Deoptimization::Reason_range_check); | |
1279 return !already_trapped; | |
1280 } | |
1281 | |
1282 | |
1283 //------------------------------flatten_alias_type----------------------------- | |
1284 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { | |
1285 int offset = tj->offset(); | |
1286 TypePtr::PTR ptr = tj->ptr(); | |
1287 | |
247 | 1288 // Known instance (scalarizable allocation) alias only with itself. |
1289 bool is_known_inst = tj->isa_oopptr() != NULL && | |
1290 tj->is_oopptr()->is_known_instance(); | |
1291 | |
0 | 1292 // Process weird unsafe references. |
1293 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { | |
1294 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); | |
247 | 1295 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); |
0 | 1296 tj = TypeOopPtr::BOTTOM; |
1297 ptr = tj->ptr(); | |
1298 offset = tj->offset(); | |
1299 } | |
1300 | |
1301 // Array pointers need some flattening | |
1302 const TypeAryPtr *ta = tj->isa_aryptr(); | |
247 | 1303 if( ta && is_known_inst ) { |
1304 if ( offset != Type::OffsetBot && | |
1305 offset > arrayOopDesc::length_offset_in_bytes() ) { | |
1306 offset = Type::OffsetBot; // Flatten constant access into array body only | |
1307 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); | |
1308 } | |
1309 } else if( ta && _AliasLevel >= 2 ) { | |
0 | 1310 // For arrays indexed by constant indices, we flatten the alias |
1311 // space to include all of the array body. Only the header, klass | |
1312 // and array length can be accessed un-aliased. | |
1313 if( offset != Type::OffsetBot ) { | |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
1314 if( ta->const_oop() ) { // MethodData* or Method* |
0 | 1315 offset = Type::OffsetBot; // Flatten constant access into array body |
247 | 1316 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); |
0 | 1317 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { |
1318 // range is OK as-is. | |
1319 tj = ta = TypeAryPtr::RANGE; | |
1320 } else if( offset == oopDesc::klass_offset_in_bytes() ) { | |
1321 tj = TypeInstPtr::KLASS; // all klass loads look alike | |
1322 ta = TypeAryPtr::RANGE; // generic ignored junk | |
1323 ptr = TypePtr::BotPTR; | |
1324 } else if( offset == oopDesc::mark_offset_in_bytes() ) { | |
1325 tj = TypeInstPtr::MARK; | |
1326 ta = TypeAryPtr::RANGE; // generic ignored junk | |
1327 ptr = TypePtr::BotPTR; | |
1328 } else { // Random constant offset into array body | |
1329 offset = Type::OffsetBot; // Flatten constant access into array body | |
247 | 1330 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); |
0 | 1331 } |
1332 } | |
1333 // Arrays of fixed size alias with arrays of unknown size. | |
1334 if (ta->size() != TypeInt::POS) { | |
1335 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); | |
247 | 1336 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); |
0 | 1337 } |
1338 // Arrays of known objects become arrays of unknown objects. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1339 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1340 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); |
247 | 1341 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1342 } |
0 | 1343 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { |
1344 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); | |
247 | 1345 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); |
0 | 1346 } |
1347 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so | |
1348 // cannot be distinguished by bytecode alone. | |
1349 if (ta->elem() == TypeInt::BOOL) { | |
1350 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); | |
1351 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); | |
247 | 1352 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); |
0 | 1353 } |
1354 // During the 2nd round of IterGVN, NotNull castings are removed. | |
1355 // Make sure the Bottom and NotNull variants alias the same. | |
1356 // Also, make sure exact and non-exact variants alias the same. | |
1357 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { | |
3789 | 1358 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); |
0 | 1359 } |
1360 } | |
1361 | |
1362 // Oop pointers need some flattening | |
1363 const TypeInstPtr *to = tj->isa_instptr(); | |
1364 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { | |
2376
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1365 ciInstanceKlass *k = to->klass()->as_instance_klass(); |
0 | 1366 if( ptr == TypePtr::Constant ) { |
2376
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1367 if (to->klass() != ciEnv::current()->Class_klass() || |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1368 offset < k->size_helper() * wordSize) { |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1369 // No constant oop pointers (such as Strings); they alias with |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1370 // unknown strings. |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1371 assert(!is_known_inst, "not scalarizable allocation"); |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1372 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1373 } |
247 | 1374 } else if( is_known_inst ) { |
163 | 1375 tj = to; // Keep NotNull and klass_is_exact for instance type |
0 | 1376 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { |
1377 // During the 2nd round of IterGVN, NotNull castings are removed. | |
1378 // Make sure the Bottom and NotNull variants alias the same. | |
1379 // Also, make sure exact and non-exact variants alias the same. | |
247 | 1380 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); |
0 | 1381 } |
1382 // Canonicalize the holder of this field | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1383 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { |
0 | 1384 // First handle header references such as a LoadKlassNode, even if the |
1385 // object's klass is unloaded at compile time (4965979). | |
247 | 1386 if (!is_known_inst) { // Do it only for non-instance types |
1387 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); | |
1388 } | |
0 | 1389 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { |
2376
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1390 // Static fields are in the space above the normal instance |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1391 // fields in the java.lang.Class instance. |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1392 if (to->klass() != ciEnv::current()->Class_klass()) { |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1393 to = NULL; |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1394 tj = TypeOopPtr::BOTTOM; |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1395 offset = tj->offset(); |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1396 } |
0 | 1397 } else { |
1398 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); | |
1399 if (!k->equals(canonical_holder) || tj->offset() != offset) { | |
247 | 1400 if( is_known_inst ) { |
1401 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); | |
1402 } else { | |
1403 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); | |
1404 } | |
0 | 1405 } |
1406 } | |
1407 } | |
1408 | |
1409 // Klass pointers to object array klasses need some flattening | |
1410 const TypeKlassPtr *tk = tj->isa_klassptr(); | |
1411 if( tk ) { | |
1412 // If we are referencing a field within a Klass, we need | |
1413 // to assume the worst case of an Object. Both exact and | |
4760
669f6a7d5b70
7121073: secondary_super_cache memory slice has incorrect bounds in flatten_alias_type
never
parents:
4115
diff
changeset
|
1414 // inexact types must flatten to the same alias class so |
669f6a7d5b70
7121073: secondary_super_cache memory slice has incorrect bounds in flatten_alias_type
never
parents:
4115
diff
changeset
|
1415 // use NotNull as the PTR. |
0 | 1416 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { |
1417 | |
4760
669f6a7d5b70
7121073: secondary_super_cache memory slice has incorrect bounds in flatten_alias_type
never
parents:
4115
diff
changeset
|
1418 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, |
0 | 1419 TypeKlassPtr::OBJECT->klass(), |
1420 offset); | |
1421 } | |
1422 | |
1423 ciKlass* klass = tk->klass(); | |
1424 if( klass->is_obj_array_klass() ) { | |
1425 ciKlass* k = TypeAryPtr::OOPS->klass(); | |
1426 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs | |
1427 k = TypeInstPtr::BOTTOM->klass(); | |
1428 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); | |
1429 } | |
1430 | |
1431 // Check for precise loads from the primary supertype array and force them | |
1432 // to the supertype cache alias index. Check for generic array loads from | |
1433 // the primary supertype array and also force them to the supertype cache | |
1434 // alias index. Since the same load can reach both, we need to merge | |
1435 // these 2 disparate memories into the same alias class. Since the | |
1436 // primary supertype array is read-only, there's no chance of confusion | |
1437 // where we bypass an array load and an array store. | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4760
diff
changeset
|
1438 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); |
4760
669f6a7d5b70
7121073: secondary_super_cache memory slice has incorrect bounds in flatten_alias_type
never
parents:
4115
diff
changeset
|
1439 if (offset == Type::OffsetBot || |
669f6a7d5b70
7121073: secondary_super_cache memory slice has incorrect bounds in flatten_alias_type
never
parents:
4115
diff
changeset
|
1440 (offset >= primary_supers_offset && |
669f6a7d5b70
7121073: secondary_super_cache memory slice has incorrect bounds in flatten_alias_type
never
parents:
4115
diff
changeset
|
1441 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4760
diff
changeset
|
1442 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { |
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4760
diff
changeset
|
1443 offset = in_bytes(Klass::secondary_super_cache_offset()); |
0 | 1444 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); |
1445 } | |
1446 } | |
1447 | |
1448 // Flatten all Raw pointers together. | |
1449 if (tj->base() == Type::RawPtr) | |
1450 tj = TypeRawPtr::BOTTOM; | |
1451 | |
1452 if (tj->base() == Type::AnyPtr) | |
1453 tj = TypePtr::BOTTOM; // An error, which the caller must check for. | |
1454 | |
1455 // Flatten all to bottom for now | |
1456 switch( _AliasLevel ) { | |
1457 case 0: | |
1458 tj = TypePtr::BOTTOM; | |
1459 break; | |
1460 case 1: // Flatten to: oop, static, field or array | |
1461 switch (tj->base()) { | |
1462 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; | |
1463 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; | |
1464 case Type::AryPtr: // do not distinguish arrays at all | |
1465 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; | |
1466 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; | |
1467 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it | |
1468 default: ShouldNotReachHere(); | |
1469 } | |
1470 break; | |
605 | 1471 case 2: // No collapsing at level 2; keep all splits |
1472 case 3: // No collapsing at level 3; keep all splits | |
0 | 1473 break; |
1474 default: | |
1475 Unimplemented(); | |
1476 } | |
1477 | |
1478 offset = tj->offset(); | |
1479 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); | |
1480 | |
1481 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || | |
1482 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || | |
1483 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || | |
1484 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || | |
1485 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || | |
1486 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || | |
1487 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , | |
1488 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); | |
1489 assert( tj->ptr() != TypePtr::TopPTR && | |
1490 tj->ptr() != TypePtr::AnyNull && | |
1491 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); | |
1492 // assert( tj->ptr() != TypePtr::Constant || | |
1493 // tj->base() == Type::RawPtr || | |
1494 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); | |
1495 | |
1496 return tj; | |
1497 } | |
1498 | |
1499 void Compile::AliasType::Init(int i, const TypePtr* at) { | |
1500 _index = i; | |
1501 _adr_type = at; | |
1502 _field = NULL; | |
1503 _is_rewritable = true; // default | |
1504 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; | |
223 | 1505 if (atoop != NULL && atoop->is_known_instance()) { |
1506 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); | |
0 | 1507 _general_index = Compile::current()->get_alias_index(gt); |
1508 } else { | |
1509 _general_index = 0; | |
1510 } | |
1511 } | |
1512 | |
1513 //---------------------------------print_on------------------------------------ | |
1514 #ifndef PRODUCT | |
1515 void Compile::AliasType::print_on(outputStream* st) { | |
1516 if (index() < 10) | |
1517 st->print("@ <%d> ", index()); | |
1518 else st->print("@ <%d>", index()); | |
1519 st->print(is_rewritable() ? " " : " RO"); | |
1520 int offset = adr_type()->offset(); | |
1521 if (offset == Type::OffsetBot) | |
1522 st->print(" +any"); | |
1523 else st->print(" +%-3d", offset); | |
1524 st->print(" in "); | |
1525 adr_type()->dump_on(st); | |
1526 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); | |
1527 if (field() != NULL && tjp) { | |
1528 if (tjp->klass() != field()->holder() || | |
1529 tjp->offset() != field()->offset_in_bytes()) { | |
1530 st->print(" != "); | |
1531 field()->print(); | |
1532 st->print(" ***"); | |
1533 } | |
1534 } | |
1535 } | |
1536 | |
1537 void print_alias_types() { | |
1538 Compile* C = Compile::current(); | |
1539 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); | |
1540 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { | |
1541 C->alias_type(idx)->print_on(tty); | |
1542 tty->cr(); | |
1543 } | |
1544 } | |
1545 #endif | |
1546 | |
1547 | |
1548 //----------------------------probe_alias_cache-------------------------------- | |
1549 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { | |
1550 intptr_t key = (intptr_t) adr_type; | |
1551 key ^= key >> logAliasCacheSize; | |
1552 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; | |
1553 } | |
1554 | |
1555 | |
1556 //-----------------------------grow_alias_types-------------------------------- | |
1557 void Compile::grow_alias_types() { | |
1558 const int old_ats = _max_alias_types; // how many before? | |
1559 const int new_ats = old_ats; // how many more? | |
1560 const int grow_ats = old_ats+new_ats; // how many now? | |
1561 _max_alias_types = grow_ats; | |
1562 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); | |
1563 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); | |
1564 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); | |
1565 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; | |
1566 } | |
1567 | |
1568 | |
1569 //--------------------------------find_alias_type------------------------------ | |
2376
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1570 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { |
0 | 1571 if (_AliasLevel == 0) |
1572 return alias_type(AliasIdxBot); | |
1573 | |
1574 AliasCacheEntry* ace = probe_alias_cache(adr_type); | |
1575 if (ace->_adr_type == adr_type) { | |
1576 return alias_type(ace->_index); | |
1577 } | |
1578 | |
1579 // Handle special cases. | |
1580 if (adr_type == NULL) return alias_type(AliasIdxTop); | |
1581 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); | |
1582 | |
1583 // Do it the slow way. | |
1584 const TypePtr* flat = flatten_alias_type(adr_type); | |
1585 | |
1586 #ifdef ASSERT | |
1587 assert(flat == flatten_alias_type(flat), "idempotent"); | |
1588 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); | |
1589 if (flat->isa_oopptr() && !flat->isa_klassptr()) { | |
1590 const TypeOopPtr* foop = flat->is_oopptr(); | |
247 | 1591 // Scalarizable allocations have exact klass always. |
1592 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); | |
1593 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); | |
0 | 1594 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); |
1595 } | |
1596 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); | |
1597 #endif | |
1598 | |
1599 int idx = AliasIdxTop; | |
1600 for (int i = 0; i < num_alias_types(); i++) { | |
1601 if (alias_type(i)->adr_type() == flat) { | |
1602 idx = i; | |
1603 break; | |
1604 } | |
1605 } | |
1606 | |
1607 if (idx == AliasIdxTop) { | |
1608 if (no_create) return NULL; | |
1609 // Grow the array if necessary. | |
1610 if (_num_alias_types == _max_alias_types) grow_alias_types(); | |
1611 // Add a new alias type. | |
1612 idx = _num_alias_types++; | |
1613 _alias_types[idx]->Init(idx, flat); | |
1614 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); | |
1615 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); | |
1616 if (flat->isa_instptr()) { | |
1617 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() | |
1618 && flat->is_instptr()->klass() == env()->Class_klass()) | |
1619 alias_type(idx)->set_rewritable(false); | |
1620 } | |
1621 if (flat->isa_klassptr()) { | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4760
diff
changeset
|
1622 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) |
0 | 1623 alias_type(idx)->set_rewritable(false); |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4760
diff
changeset
|
1624 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) |
0 | 1625 alias_type(idx)->set_rewritable(false); |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4760
diff
changeset
|
1626 if (flat->offset() == in_bytes(Klass::access_flags_offset())) |
0 | 1627 alias_type(idx)->set_rewritable(false); |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4760
diff
changeset
|
1628 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) |
0 | 1629 alias_type(idx)->set_rewritable(false); |
1630 } | |
1631 // %%% (We would like to finalize JavaThread::threadObj_offset(), | |
1632 // but the base pointer type is not distinctive enough to identify | |
1633 // references into JavaThread.) | |
1634 | |
2376
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1635 // Check for final fields. |
0 | 1636 const TypeInstPtr* tinst = flat->isa_instptr(); |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1637 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { |
2376
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1638 ciField* field; |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1639 if (tinst->const_oop() != NULL && |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1640 tinst->klass() == ciEnv::current()->Class_klass() && |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1641 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1642 // static field |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1643 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1644 field = k->get_field_by_offset(tinst->offset(), true); |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1645 } else { |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1646 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1647 field = k->get_field_by_offset(tinst->offset(), false); |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1648 } |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1649 assert(field == NULL || |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1650 original_field == NULL || |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1651 (field->holder() == original_field->holder() && |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1652 field->offset() == original_field->offset() && |
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1653 field->is_static() == original_field->is_static()), "wrong field?"); |
0 | 1654 // Set field() and is_rewritable() attributes. |
1655 if (field != NULL) alias_type(idx)->set_field(field); | |
1656 } | |
1657 } | |
1658 | |
1659 // Fill the cache for next time. | |
1660 ace->_adr_type = adr_type; | |
1661 ace->_index = idx; | |
1662 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); | |
1663 | |
1664 // Might as well try to fill the cache for the flattened version, too. | |
1665 AliasCacheEntry* face = probe_alias_cache(flat); | |
1666 if (face->_adr_type == NULL) { | |
1667 face->_adr_type = flat; | |
1668 face->_index = idx; | |
1669 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); | |
1670 } | |
1671 | |
1672 return alias_type(idx); | |
1673 } | |
1674 | |
1675 | |
1676 Compile::AliasType* Compile::alias_type(ciField* field) { | |
1677 const TypeOopPtr* t; | |
1678 if (field->is_static()) | |
2376
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1679 t = TypeInstPtr::make(field->holder()->java_mirror()); |
0 | 1680 else |
1681 t = TypeOopPtr::make_from_klass_raw(field->holder()); | |
2376
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1682 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); |
0 | 1683 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); |
1684 return atp; | |
1685 } | |
1686 | |
1687 | |
1688 //------------------------------have_alias_type-------------------------------- | |
1689 bool Compile::have_alias_type(const TypePtr* adr_type) { | |
1690 AliasCacheEntry* ace = probe_alias_cache(adr_type); | |
1691 if (ace->_adr_type == adr_type) { | |
1692 return true; | |
1693 } | |
1694 | |
1695 // Handle special cases. | |
1696 if (adr_type == NULL) return true; | |
1697 if (adr_type == TypePtr::BOTTOM) return true; | |
1698 | |
2376
c7f3d0b4570f
7017732: move static fields into Class to prepare for perm gen removal
never
parents:
2192
diff
changeset
|
1699 return find_alias_type(adr_type, true, NULL) != NULL; |
0 | 1700 } |
1701 | |
1702 //-----------------------------must_alias-------------------------------------- | |
1703 // True if all values of the given address type are in the given alias category. | |
1704 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { | |
1705 if (alias_idx == AliasIdxBot) return true; // the universal category | |
1706 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP | |
1707 if (alias_idx == AliasIdxTop) return false; // the empty category | |
1708 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins | |
1709 | |
1710 // the only remaining possible overlap is identity | |
1711 int adr_idx = get_alias_index(adr_type); | |
1712 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); | |
1713 assert(adr_idx == alias_idx || | |
1714 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM | |
1715 && adr_type != TypeOopPtr::BOTTOM), | |
1716 "should not be testing for overlap with an unsafe pointer"); | |
1717 return adr_idx == alias_idx; | |
1718 } | |
1719 | |
1720 //------------------------------can_alias-------------------------------------- | |
1721 // True if any values of the given address type are in the given alias category. | |
1722 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { | |
1723 if (alias_idx == AliasIdxTop) return false; // the empty category | |
1724 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP | |
1725 if (alias_idx == AliasIdxBot) return true; // the universal category | |
1726 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins | |
1727 | |
1728 // the only remaining possible overlap is identity | |
1729 int adr_idx = get_alias_index(adr_type); | |
1730 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); | |
1731 return adr_idx == alias_idx; | |
1732 } | |
1733 | |
1734 | |
1735 | |
1736 //---------------------------pop_warm_call------------------------------------- | |
1737 WarmCallInfo* Compile::pop_warm_call() { | |
1738 WarmCallInfo* wci = _warm_calls; | |
1739 if (wci != NULL) _warm_calls = wci->remove_from(wci); | |
1740 return wci; | |
1741 } | |
1742 | |
1743 //----------------------------Inline_Warm-------------------------------------- | |
1744 int Compile::Inline_Warm() { | |
1745 // If there is room, try to inline some more warm call sites. | |
1746 // %%% Do a graph index compaction pass when we think we're out of space? | |
1747 if (!InlineWarmCalls) return 0; | |
1748 | |
1749 int calls_made_hot = 0; | |
1750 int room_to_grow = NodeCountInliningCutoff - unique(); | |
1751 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); | |
1752 int amount_grown = 0; | |
1753 WarmCallInfo* call; | |
1754 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { | |
1755 int est_size = (int)call->size(); | |
1756 if (est_size > (room_to_grow - amount_grown)) { | |
1757 // This one won't fit anyway. Get rid of it. | |
1758 call->make_cold(); | |
1759 continue; | |
1760 } | |
1761 call->make_hot(); | |
1762 calls_made_hot++; | |
1763 amount_grown += est_size; | |
1764 amount_to_grow -= est_size; | |
1765 } | |
1766 | |
1767 if (calls_made_hot > 0) set_major_progress(); | |
1768 return calls_made_hot; | |
1769 } | |
1770 | |
1771 | |
1772 //----------------------------Finish_Warm-------------------------------------- | |
1773 void Compile::Finish_Warm() { | |
1774 if (!InlineWarmCalls) return; | |
1775 if (failing()) return; | |
1776 if (warm_calls() == NULL) return; | |
1777 | |
1778 // Clean up loose ends, if we are out of space for inlining. | |
1779 WarmCallInfo* call; | |
1780 while ((call = pop_warm_call()) != NULL) { | |
1781 call->make_cold(); | |
1782 } | |
1783 } | |
1784 | |
1172 | 1785 //---------------------cleanup_loop_predicates----------------------- |
1786 // Remove the opaque nodes that protect the predicates so that all unused | |
1787 // checks and uncommon_traps will be eliminated from the ideal graph | |
1788 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { | |
1789 if (predicate_count()==0) return; | |
1790 for (int i = predicate_count(); i > 0; i--) { | |
1791 Node * n = predicate_opaque1_node(i-1); | |
1792 assert(n->Opcode() == Op_Opaque1, "must be"); | |
1793 igvn.replace_node(n, n->in(1)); | |
1794 } | |
1795 assert(predicate_count()==0, "should be clean!"); | |
1796 } | |
0 | 1797 |
7473 | 1798 // StringOpts and late inlining of string methods |
1799 void Compile::inline_string_calls(bool parse_time) { | |
1800 { | |
1801 // remove useless nodes to make the usage analysis simpler | |
1802 ResourceMark rm; | |
1803 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); | |
1804 } | |
1805 | |
1806 { | |
1807 ResourceMark rm; | |
10405 | 1808 print_method(PHASE_BEFORE_STRINGOPTS, 3); |
7473 | 1809 PhaseStringOpts pso(initial_gvn(), for_igvn()); |
10405 | 1810 print_method(PHASE_AFTER_STRINGOPTS, 3); |
7473 | 1811 } |
1812 | |
1813 // now inline anything that we skipped the first time around | |
1814 if (!parse_time) { | |
1815 _late_inlines_pos = _late_inlines.length(); | |
1816 } | |
1817 | |
1818 while (_string_late_inlines.length() > 0) { | |
1819 CallGenerator* cg = _string_late_inlines.pop(); | |
1820 cg->do_late_inline(); | |
1821 if (failing()) return; | |
1822 } | |
1823 _string_late_inlines.trunc_to(0); | |
1824 } | |
1825 | |
10278 | 1826 // Late inlining of boxing methods |
1827 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { | |
1828 if (_boxing_late_inlines.length() > 0) { | |
1829 assert(has_boxed_value(), "inconsistent"); | |
1830 | |
1831 PhaseGVN* gvn = initial_gvn(); | |
1832 set_inlining_incrementally(true); | |
1833 | |
1834 assert( igvn._worklist.size() == 0, "should be done with igvn" ); | |
1835 for_igvn()->clear(); | |
1836 gvn->replace_with(&igvn); | |
1837 | |
1838 while (_boxing_late_inlines.length() > 0) { | |
1839 CallGenerator* cg = _boxing_late_inlines.pop(); | |
1840 cg->do_late_inline(); | |
1841 if (failing()) return; | |
1842 } | |
1843 _boxing_late_inlines.trunc_to(0); | |
1844 | |
1845 { | |
1846 ResourceMark rm; | |
1847 PhaseRemoveUseless pru(gvn, for_igvn()); | |
1848 } | |
1849 | |
1850 igvn = PhaseIterGVN(gvn); | |
1851 igvn.optimize(); | |
1852 | |
1853 set_inlining_progress(false); | |
1854 set_inlining_incrementally(false); | |
1855 } | |
1856 } | |
1857 | |
7473 | 1858 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { |
1859 assert(IncrementalInline, "incremental inlining should be on"); | |
1860 PhaseGVN* gvn = initial_gvn(); | |
1861 | |
1862 set_inlining_progress(false); | |
1863 for_igvn()->clear(); | |
1864 gvn->replace_with(&igvn); | |
1865 | |
1866 int i = 0; | |
1867 | |
1868 for (; i <_late_inlines.length() && !inlining_progress(); i++) { | |
1869 CallGenerator* cg = _late_inlines.at(i); | |
1870 _late_inlines_pos = i+1; | |
1871 cg->do_late_inline(); | |
1872 if (failing()) return; | |
1873 } | |
1874 int j = 0; | |
1875 for (; i < _late_inlines.length(); i++, j++) { | |
1876 _late_inlines.at_put(j, _late_inlines.at(i)); | |
1877 } | |
1878 _late_inlines.trunc_to(j); | |
1879 | |
1880 { | |
1881 ResourceMark rm; | |
10278 | 1882 PhaseRemoveUseless pru(gvn, for_igvn()); |
7473 | 1883 } |
1884 | |
1885 igvn = PhaseIterGVN(gvn); | |
1886 } | |
1887 | |
1888 // Perform incremental inlining until bound on number of live nodes is reached | |
1889 void Compile::inline_incrementally(PhaseIterGVN& igvn) { | |
1890 PhaseGVN* gvn = initial_gvn(); | |
1891 | |
1892 set_inlining_incrementally(true); | |
1893 set_inlining_progress(true); | |
1894 uint low_live_nodes = 0; | |
1895 | |
1896 while(inlining_progress() && _late_inlines.length() > 0) { | |
1897 | |
1898 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { | |
1899 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { | |
1900 // PhaseIdealLoop is expensive so we only try it once we are | |
1901 // out of loop and we only try it again if the previous helped | |
1902 // got the number of nodes down significantly | |
1903 PhaseIdealLoop ideal_loop( igvn, false, true ); | |
1904 if (failing()) return; | |
1905 low_live_nodes = live_nodes(); | |
1906 _major_progress = true; | |
1907 } | |
1908 | |
1909 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { | |
1910 break; | |
1911 } | |
1912 } | |
1913 | |
1914 inline_incrementally_one(igvn); | |
1915 | |
1916 if (failing()) return; | |
1917 | |
1918 igvn.optimize(); | |
1919 | |
1920 if (failing()) return; | |
1921 } | |
1922 | |
1923 assert( igvn._worklist.size() == 0, "should be done with igvn" ); | |
1924 | |
1925 if (_string_late_inlines.length() > 0) { | |
1926 assert(has_stringbuilder(), "inconsistent"); | |
1927 for_igvn()->clear(); | |
1928 initial_gvn()->replace_with(&igvn); | |
1929 | |
1930 inline_string_calls(false); | |
1931 | |
1932 if (failing()) return; | |
1933 | |
1934 { | |
1935 ResourceMark rm; | |
1936 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); | |
1937 } | |
1938 | |
1939 igvn = PhaseIterGVN(gvn); | |
1940 | |
1941 igvn.optimize(); | |
1942 } | |
1943 | |
1944 set_inlining_incrementally(false); | |
1945 } | |
1946 | |
1947 | |
0 | 1948 //------------------------------Optimize--------------------------------------- |
1949 // Given a graph, optimize it. | |
1950 void Compile::Optimize() { | |
1951 TracePhase t1("optimizer", &_t_optimizer, true); | |
1952 | |
1953 #ifndef PRODUCT | |
1954 if (env()->break_at_compile()) { | |
1955 BREAKPOINT; | |
1956 } | |
1957 | |
1958 #endif | |
1959 | |
1960 ResourceMark rm; | |
1961 int loop_opts_cnt; | |
1962 | |
1963 NOT_PRODUCT( verify_graph_edges(); ) | |
1964 | |
10405 | 1965 print_method(PHASE_AFTER_PARSING); |
0 | 1966 |
1967 { | |
1968 // Iterative Global Value Numbering, including ideal transforms | |
1969 // Initialize IterGVN with types and values from parse-time GVN | |
1970 PhaseIterGVN igvn(initial_gvn()); | |
1971 { | |
1972 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) | |
1973 igvn.optimize(); | |
1974 } | |
1975 | |
10405 | 1976 print_method(PHASE_ITER_GVN1, 2); |
0 | 1977 |
1978 if (failing()) return; | |
1979 | |
10278 | 1980 { |
1981 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) | |
1982 inline_incrementally(igvn); | |
1983 } | |
7473 | 1984 |
10405 | 1985 print_method(PHASE_INCREMENTAL_INLINE, 2); |
7473 | 1986 |
1987 if (failing()) return; | |
1988 | |
10278 | 1989 if (eliminate_boxing()) { |
1990 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) | |
1991 // Inline valueOf() methods now. | |
1992 inline_boxing_calls(igvn); | |
1993 | |
10405 | 1994 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); |
10278 | 1995 |
1996 if (failing()) return; | |
1997 } | |
1998 | |
8048
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
1999 // No more new expensive nodes will be added to the list from here |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
2000 // so keep only the actual candidates for optimizations. |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
2001 cleanup_expensive_nodes(igvn); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
2002 |
1634
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
2003 // Perform escape analysis |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
2004 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
2005 if (has_loops()) { |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
2006 // Cleanup graph (remove dead nodes). |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
2007 TracePhase t2("idealLoop", &_t_idealLoop, true); |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
2008 PhaseIdealLoop ideal_loop( igvn, false, true ); |
10405 | 2009 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
2010 if (failing()) return; |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3899
diff
changeset
|
2011 } |
1634
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
2012 ConnectionGraph::do_analysis(this, &igvn); |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
2013 |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
2014 if (failing()) return; |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
2015 |
4115 | 2016 // Optimize out fields loads from scalar replaceable allocations. |
1634
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
2017 igvn.optimize(); |
10405 | 2018 print_method(PHASE_ITER_GVN_AFTER_EA, 2); |
1634
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
2019 |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
2020 if (failing()) return; |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
2021 |
4115 | 2022 if (congraph() != NULL && macro_count() > 0) { |
5948
ee138854b3a6
7147744: CTW: assert(false) failed: infinite EA connection graph build
kvn
parents:
4767
diff
changeset
|
2023 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) |
4115 | 2024 PhaseMacroExpand mexp(igvn); |
2025 mexp.eliminate_macro_nodes(); | |
2026 igvn.set_delay_transform(false); | |
2027 | |
2028 igvn.optimize(); | |
10405 | 2029 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); |
4115 | 2030 |
2031 if (failing()) return; | |
2032 } | |
1634
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
2033 } |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
2034 |
0 | 2035 // Loop transforms on the ideal graph. Range Check Elimination, |
2036 // peeling, unrolling, etc. | |
2037 | |
2038 // Set loop opts counter | |
2039 loop_opts_cnt = num_loop_opts(); | |
2040 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { | |
2041 { | |
2042 TracePhase t2("idealLoop", &_t_idealLoop, true); | |
2445 | 2043 PhaseIdealLoop ideal_loop( igvn, true ); |
0 | 2044 loop_opts_cnt--; |
10405 | 2045 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); |
0 | 2046 if (failing()) return; |
2047 } | |
2048 // Loop opts pass if partial peeling occurred in previous pass | |
2049 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { | |
2050 TracePhase t3("idealLoop", &_t_idealLoop, true); | |
2445 | 2051 PhaseIdealLoop ideal_loop( igvn, false ); |
0 | 2052 loop_opts_cnt--; |
10405 | 2053 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); |
0 | 2054 if (failing()) return; |
2055 } | |
2056 // Loop opts pass for loop-unrolling before CCP | |
2057 if(major_progress() && (loop_opts_cnt > 0)) { | |
2058 TracePhase t4("idealLoop", &_t_idealLoop, true); | |
2445 | 2059 PhaseIdealLoop ideal_loop( igvn, false ); |
0 | 2060 loop_opts_cnt--; |
10405 | 2061 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); |
0 | 2062 } |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
2063 if (!failing()) { |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
2064 // Verify that last round of loop opts produced a valid graph |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
2065 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
2066 PhaseIdealLoop::verify(igvn); |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
2067 } |
0 | 2068 } |
2069 if (failing()) return; | |
2070 | |
2071 // Conditional Constant Propagation; | |
2072 PhaseCCP ccp( &igvn ); | |
2073 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); | |
2074 { | |
2075 TracePhase t2("ccp", &_t_ccp, true); | |
2076 ccp.do_transform(); | |
2077 } | |
10405 | 2078 print_method(PHASE_CPP1, 2); |
0 | 2079 |
2080 assert( true, "Break here to ccp.dump_old2new_map()"); | |
2081 | |
2082 // Iterative Global Value Numbering, including ideal transforms | |
2083 { | |
2084 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) | |
2085 igvn = ccp; | |
2086 igvn.optimize(); | |
2087 } | |
2088 | |
10405 | 2089 print_method(PHASE_ITER_GVN2, 2); |
0 | 2090 |
2091 if (failing()) return; | |
2092 | |
2093 // Loop transforms on the ideal graph. Range Check Elimination, | |
2094 // peeling, unrolling, etc. | |
2095 if(loop_opts_cnt > 0) { | |
2096 debug_only( int cnt = 0; ); | |
2097 while(major_progress() && (loop_opts_cnt > 0)) { | |
2098 TracePhase t2("idealLoop", &_t_idealLoop, true); | |
2099 assert( cnt++ < 40, "infinite cycle in loop optimization" ); | |
2445 | 2100 PhaseIdealLoop ideal_loop( igvn, true); |
0 | 2101 loop_opts_cnt--; |
10405 | 2102 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); |
0 | 2103 if (failing()) return; |
2104 } | |
2105 } | |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
2106 |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
2107 { |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
2108 // Verify that all previous optimizations produced a valid graph |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
2109 // at least to this point, even if no loop optimizations were done. |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
2110 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
2111 PhaseIdealLoop::verify(igvn); |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
2112 } |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
2113 |
0 | 2114 { |
2115 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) | |
2116 PhaseMacroExpand mex(igvn); | |
2117 if (mex.expand_macro_nodes()) { | |
2118 assert(failing(), "must bail out w/ explicit message"); | |
2119 return; | |
2120 } | |
2121 } | |
2122 | |
2123 } // (End scope of igvn; run destructor if necessary for asserts.) | |
2124 | |
7605 | 2125 dump_inlining(); |
0 | 2126 // A method with only infinite loops has no edges entering loops from root |
2127 { | |
2128 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) | |
2129 if (final_graph_reshaping()) { | |
2130 assert(failing(), "must bail out w/ explicit message"); | |
2131 return; | |
2132 } | |
2133 } | |
2134 | |
10405 | 2135 print_method(PHASE_OPTIMIZE_FINISHED, 2); |
0 | 2136 } |
2137 | |
2138 | |
2139 //------------------------------Code_Gen--------------------------------------- | |
2140 // Given a graph, generate code for it | |
2141 void Compile::Code_Gen() { | |
2142 if (failing()) return; | |
2143 | |
2144 // Perform instruction selection. You might think we could reclaim Matcher | |
2145 // memory PDQ, but actually the Matcher is used in generating spill code. | |
2146 // Internals of the Matcher (including some VectorSets) must remain live | |
2147 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage | |
2148 // set a bit in reclaimed memory. | |
2149 | |
2150 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine | |
2151 // nodes. Mapping is only valid at the root of each matched subtree. | |
2152 NOT_PRODUCT( verify_graph_edges(); ) | |
2153 | |
2154 Node_List proj_list; | |
2155 Matcher m(proj_list); | |
2156 _matcher = &m; | |
2157 { | |
2158 TracePhase t2("matcher", &_t_matcher, true); | |
2159 m.match(); | |
2160 } | |
2161 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine | |
2162 // nodes. Mapping is only valid at the root of each matched subtree. | |
2163 NOT_PRODUCT( verify_graph_edges(); ) | |
2164 | |
2165 // If you have too many nodes, or if matching has failed, bail out | |
2166 check_node_count(0, "out of nodes matching instructions"); | |
2167 if (failing()) return; | |
2168 | |
2169 // Build a proper-looking CFG | |
2170 PhaseCFG cfg(node_arena(), root(), m); | |
2171 _cfg = &cfg; | |
2172 { | |
2173 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) | |
2174 cfg.Dominators(); | |
2175 if (failing()) return; | |
2176 | |
2177 NOT_PRODUCT( verify_graph_edges(); ) | |
2178 | |
2179 cfg.Estimate_Block_Frequency(); | |
2180 cfg.GlobalCodeMotion(m,unique(),proj_list); | |
5951 | 2181 if (failing()) return; |
0 | 2182 |
10405 | 2183 print_method(PHASE_GLOBAL_CODE_MOTION, 2); |
0 | 2184 |
2185 NOT_PRODUCT( verify_graph_edges(); ) | |
2186 | |
2187 debug_only( cfg.verify(); ) | |
2188 } | |
2189 NOT_PRODUCT( verify_graph_edges(); ) | |
2190 | |
10111 | 2191 PhaseChaitin regalloc(unique(), cfg, m); |
0 | 2192 _regalloc = ®alloc; |
2193 { | |
2194 TracePhase t2("regalloc", &_t_registerAllocation, true); | |
2195 // Perform register allocation. After Chaitin, use-def chains are | |
2196 // no longer accurate (at spill code) and so must be ignored. | |
2197 // Node->LRG->reg mappings are still accurate. | |
2198 _regalloc->Register_Allocate(); | |
2199 | |
2200 // Bail out if the allocator builds too many nodes | |
10111 | 2201 if (failing()) { |
2202 return; | |
2203 } | |
0 | 2204 } |
2205 | |
2206 // Prior to register allocation we kept empty basic blocks in case the | |
2207 // the allocator needed a place to spill. After register allocation we | |
2208 // are not adding any new instructions. If any basic block is empty, we | |
2209 // can now safely remove it. | |
2210 { | |
418 | 2211 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) |
2212 cfg.remove_empty(); | |
2213 if (do_freq_based_layout()) { | |
2214 PhaseBlockLayout layout(cfg); | |
2215 } else { | |
2216 cfg.set_loop_alignment(); | |
2217 } | |
2218 cfg.fixup_flow(); | |
0 | 2219 } |
2220 | |
2221 // Apply peephole optimizations | |
2222 if( OptoPeephole ) { | |
2223 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) | |
2224 PhasePeephole peep( _regalloc, cfg); | |
2225 peep.do_transform(); | |
2226 } | |
2227 | |
2228 // Convert Nodes to instruction bits in a buffer | |
2229 { | |
2230 // %%%% workspace merge brought two timers together for one job | |
2231 TracePhase t2a("output", &_t_output, true); | |
2232 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) | |
2233 Output(); | |
2234 } | |
2235 | |
10405 | 2236 print_method(PHASE_FINAL_CODE); |
0 | 2237 |
2238 // He's dead, Jim. | |
2239 _cfg = (PhaseCFG*)0xdeadbeef; | |
2240 _regalloc = (PhaseChaitin*)0xdeadbeef; | |
2241 } | |
2242 | |
2243 | |
2244 //------------------------------dump_asm--------------------------------------- | |
2245 // Dump formatted assembly | |
2246 #ifndef PRODUCT | |
2247 void Compile::dump_asm(int *pcs, uint pc_limit) { | |
2248 bool cut_short = false; | |
2249 tty->print_cr("#"); | |
2250 tty->print("# "); _tf->dump(); tty->cr(); | |
2251 tty->print_cr("#"); | |
2252 | |
2253 // For all blocks | |
2254 int pc = 0x0; // Program counter | |
2255 char starts_bundle = ' '; | |
2256 _regalloc->dump_frame(); | |
2257 | |
2258 Node *n = NULL; | |
2259 for( uint i=0; i<_cfg->_num_blocks; i++ ) { | |
2260 if (VMThread::should_terminate()) { cut_short = true; break; } | |
2261 Block *b = _cfg->_blocks[i]; | |
2262 if (b->is_connector() && !Verbose) continue; | |
2263 n = b->_nodes[0]; | |
2264 if (pcs && n->_idx < pc_limit) | |
2265 tty->print("%3.3x ", pcs[n->_idx]); | |
2266 else | |
2267 tty->print(" "); | |
2268 b->dump_head( &_cfg->_bbs ); | |
2269 if (b->is_connector()) { | |
2270 tty->print_cr(" # Empty connector block"); | |
2271 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { | |
2272 tty->print_cr(" # Block is sole successor of call"); | |
2273 } | |
2274 | |
2275 // For all instructions | |
2276 Node *delay = NULL; | |
2277 for( uint j = 0; j<b->_nodes.size(); j++ ) { | |
2278 if (VMThread::should_terminate()) { cut_short = true; break; } | |
2279 n = b->_nodes[j]; | |
2280 if (valid_bundle_info(n)) { | |
2281 Bundle *bundle = node_bundling(n); | |
2282 if (bundle->used_in_unconditional_delay()) { | |
2283 delay = n; | |
2284 continue; | |
2285 } | |
2286 if (bundle->starts_bundle()) | |
2287 starts_bundle = '+'; | |
2288 } | |
2289 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
2290 if (WizardMode) n->dump(); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
2291 |
0 | 2292 if( !n->is_Region() && // Dont print in the Assembly |
2293 !n->is_Phi() && // a few noisely useless nodes | |
2294 !n->is_Proj() && | |
2295 !n->is_MachTemp() && | |
1100
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2296 !n->is_SafePointScalarObject() && |
0 | 2297 !n->is_Catch() && // Would be nice to print exception table targets |
2298 !n->is_MergeMem() && // Not very interesting | |
2299 !n->is_top() && // Debug info table constants | |
2300 !(n->is_Con() && !n->is_Mach())// Debug info table constants | |
2301 ) { | |
2302 if (pcs && n->_idx < pc_limit) | |
2303 tty->print("%3.3x", pcs[n->_idx]); | |
2304 else | |
2305 tty->print(" "); | |
2306 tty->print(" %c ", starts_bundle); | |
2307 starts_bundle = ' '; | |
2308 tty->print("\t"); | |
2309 n->format(_regalloc, tty); | |
2310 tty->cr(); | |
2311 } | |
2312 | |
2313 // If we have an instruction with a delay slot, and have seen a delay, | |
2314 // then back up and print it | |
2315 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { | |
2316 assert(delay != NULL, "no unconditional delay instruction"); | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
2317 if (WizardMode) delay->dump(); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
2318 |
0 | 2319 if (node_bundling(delay)->starts_bundle()) |
2320 starts_bundle = '+'; | |
2321 if (pcs && n->_idx < pc_limit) | |
2322 tty->print("%3.3x", pcs[n->_idx]); | |
2323 else | |
2324 tty->print(" "); | |
2325 tty->print(" %c ", starts_bundle); | |
2326 starts_bundle = ' '; | |
2327 tty->print("\t"); | |
2328 delay->format(_regalloc, tty); | |
2329 tty->print_cr(""); | |
2330 delay = NULL; | |
2331 } | |
2332 | |
2333 // Dump the exception table as well | |
2334 if( n->is_Catch() && (Verbose || WizardMode) ) { | |
2335 // Print the exception table for this offset | |
2336 _handler_table.print_subtable_for(pc); | |
2337 } | |
2338 } | |
2339 | |
2340 if (pcs && n->_idx < pc_limit) | |
2341 tty->print_cr("%3.3x", pcs[n->_idx]); | |
2342 else | |
2343 tty->print_cr(""); | |
2344 | |
2345 assert(cut_short || delay == NULL, "no unconditional delay branch"); | |
2346 | |
2347 } // End of per-block dump | |
2348 tty->print_cr(""); | |
2349 | |
2350 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); | |
2351 } | |
2352 #endif | |
2353 | |
2354 //------------------------------Final_Reshape_Counts--------------------------- | |
2355 // This class defines counters to help identify when a method | |
2356 // may/must be executed using hardware with only 24-bit precision. | |
2357 struct Final_Reshape_Counts : public StackObj { | |
2358 int _call_count; // count non-inlined 'common' calls | |
2359 int _float_count; // count float ops requiring 24-bit precision | |
2360 int _double_count; // count double ops requiring more precision | |
2361 int _java_call_count; // count non-inlined 'java' calls | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2362 int _inner_loop_count; // count loops which need alignment |
0 | 2363 VectorSet _visited; // Visitation flags |
2364 Node_List _tests; // Set of IfNodes & PCTableNodes | |
2365 | |
2366 Final_Reshape_Counts() : | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2367 _call_count(0), _float_count(0), _double_count(0), |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2368 _java_call_count(0), _inner_loop_count(0), |
0 | 2369 _visited( Thread::current()->resource_area() ) { } |
2370 | |
2371 void inc_call_count () { _call_count ++; } | |
2372 void inc_float_count () { _float_count ++; } | |
2373 void inc_double_count() { _double_count++; } | |
2374 void inc_java_call_count() { _java_call_count++; } | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2375 void inc_inner_loop_count() { _inner_loop_count++; } |
0 | 2376 |
2377 int get_call_count () const { return _call_count ; } | |
2378 int get_float_count () const { return _float_count ; } | |
2379 int get_double_count() const { return _double_count; } | |
2380 int get_java_call_count() const { return _java_call_count; } | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2381 int get_inner_loop_count() const { return _inner_loop_count; } |
0 | 2382 }; |
2383 | |
9060
cc32ccaaf47f
8003310: Enable -Wunused-function when compiling with gcc
mikael
parents:
8730
diff
changeset
|
2384 #ifdef ASSERT |
0 | 2385 static bool oop_offset_is_sane(const TypeInstPtr* tp) { |
2386 ciInstanceKlass *k = tp->klass()->as_instance_klass(); | |
2387 // Make sure the offset goes inside the instance layout. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
2388 return k->contains_field_offset(tp->offset()); |
0 | 2389 // Note that OffsetBot and OffsetTop are very negative. |
2390 } | |
9060
cc32ccaaf47f
8003310: Enable -Wunused-function when compiling with gcc
mikael
parents:
8730
diff
changeset
|
2391 #endif |
0 | 2392 |
3248
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2393 // Eliminate trivially redundant StoreCMs and accumulate their |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2394 // precedence edges. |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2395 void Compile::eliminate_redundant_card_marks(Node* n) { |
3248
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2396 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2397 if (n->in(MemNode::Address)->outcnt() > 1) { |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2398 // There are multiple users of the same address so it might be |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2399 // possible to eliminate some of the StoreCMs |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2400 Node* mem = n->in(MemNode::Memory); |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2401 Node* adr = n->in(MemNode::Address); |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2402 Node* val = n->in(MemNode::ValueIn); |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2403 Node* prev = n; |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2404 bool done = false; |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2405 // Walk the chain of StoreCMs eliminating ones that match. As |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2406 // long as it's a chain of single users then the optimization is |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2407 // safe. Eliminating partially redundant StoreCMs would require |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2408 // cloning copies down the other paths. |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2409 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2410 if (adr == mem->in(MemNode::Address) && |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2411 val == mem->in(MemNode::ValueIn)) { |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2412 // redundant StoreCM |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2413 if (mem->req() > MemNode::OopStore) { |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2414 // Hasn't been processed by this code yet. |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2415 n->add_prec(mem->in(MemNode::OopStore)); |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2416 } else { |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2417 // Already converted to precedence edge |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2418 for (uint i = mem->req(); i < mem->len(); i++) { |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2419 // Accumulate any precedence edges |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2420 if (mem->in(i) != NULL) { |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2421 n->add_prec(mem->in(i)); |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2422 } |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2423 } |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2424 // Everything above this point has been processed. |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2425 done = true; |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2426 } |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2427 // Eliminate the previous StoreCM |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2428 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2429 assert(mem->outcnt() == 0, "should be dead"); |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2430 mem->disconnect_inputs(NULL, this); |
3248
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2431 } else { |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2432 prev = mem; |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2433 } |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2434 mem = prev->in(MemNode::Memory); |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2435 } |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2436 } |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2437 } |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2438 |
0 | 2439 //------------------------------final_graph_reshaping_impl---------------------- |
2440 // Implement items 1-5 from final_graph_reshaping below. | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2441 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { |
0 | 2442 |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2443 if ( n->outcnt() == 0 ) return; // dead node |
0 | 2444 uint nop = n->Opcode(); |
2445 | |
2446 // Check for 2-input instruction with "last use" on right input. | |
2447 // Swap to left input. Implements item (2). | |
2448 if( n->req() == 3 && // two-input instruction | |
2449 n->in(1)->outcnt() > 1 && // left use is NOT a last use | |
2450 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop | |
2451 n->in(2)->outcnt() == 1 &&// right use IS a last use | |
2452 !n->in(2)->is_Con() ) { // right use is not a constant | |
2453 // Check for commutative opcode | |
2454 switch( nop ) { | |
2455 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: | |
2456 case Op_MaxI: case Op_MinI: | |
2457 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: | |
2458 case Op_AndL: case Op_XorL: case Op_OrL: | |
2459 case Op_AndI: case Op_XorI: case Op_OrI: { | |
2460 // Move "last use" input to left by swapping inputs | |
2461 n->swap_edges(1, 2); | |
2462 break; | |
2463 } | |
2464 default: | |
2465 break; | |
2466 } | |
2467 } | |
2468 | |
1609 | 2469 #ifdef ASSERT |
2470 if( n->is_Mem() ) { | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2471 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); |
1609 | 2472 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || |
2473 // oop will be recorded in oop map if load crosses safepoint | |
2474 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || | |
2475 LoadNode::is_immutable_value(n->in(MemNode::Address))), | |
2476 "raw memory operations should have control edge"); | |
2477 } | |
2478 #endif | |
0 | 2479 // Count FPU ops and common calls, implements item (3) |
2480 switch( nop ) { | |
2481 // Count all float operations that may use FPU | |
2482 case Op_AddF: | |
2483 case Op_SubF: | |
2484 case Op_MulF: | |
2485 case Op_DivF: | |
2486 case Op_NegF: | |
2487 case Op_ModF: | |
2488 case Op_ConvI2F: | |
2489 case Op_ConF: | |
2490 case Op_CmpF: | |
2491 case Op_CmpF3: | |
2492 // case Op_ConvL2F: // longs are split into 32-bit halves | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2493 frc.inc_float_count(); |
0 | 2494 break; |
2495 | |
2496 case Op_ConvF2D: | |
2497 case Op_ConvD2F: | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2498 frc.inc_float_count(); |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2499 frc.inc_double_count(); |
0 | 2500 break; |
2501 | |
2502 // Count all double operations that may use FPU | |
2503 case Op_AddD: | |
2504 case Op_SubD: | |
2505 case Op_MulD: | |
2506 case Op_DivD: | |
2507 case Op_NegD: | |
2508 case Op_ModD: | |
2509 case Op_ConvI2D: | |
2510 case Op_ConvD2I: | |
2511 // case Op_ConvL2D: // handled by leaf call | |
2512 // case Op_ConvD2L: // handled by leaf call | |
2513 case Op_ConD: | |
2514 case Op_CmpD: | |
2515 case Op_CmpD3: | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2516 frc.inc_double_count(); |
0 | 2517 break; |
2518 case Op_Opaque1: // Remove Opaque Nodes before matching | |
2519 case Op_Opaque2: // Remove Opaque Nodes before matching | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2520 n->subsume_by(n->in(1), this); |
0 | 2521 break; |
2522 case Op_CallStaticJava: | |
2523 case Op_CallJava: | |
2524 case Op_CallDynamicJava: | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2525 frc.inc_java_call_count(); // Count java call site; |
0 | 2526 case Op_CallRuntime: |
2527 case Op_CallLeaf: | |
2528 case Op_CallLeafNoFP: { | |
2529 assert( n->is_Call(), "" ); | |
2530 CallNode *call = n->as_Call(); | |
2531 // Count call sites where the FP mode bit would have to be flipped. | |
2532 // Do not count uncommon runtime calls: | |
2533 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, | |
2534 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... | |
2535 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2536 frc.inc_call_count(); // Count the call site |
0 | 2537 } else { // See if uncommon argument is shared |
2538 Node *n = call->in(TypeFunc::Parms); | |
2539 int nop = n->Opcode(); | |
2540 // Clone shared simple arguments to uncommon calls, item (1). | |
2541 if( n->outcnt() > 1 && | |
2542 !n->is_Proj() && | |
2543 nop != Op_CreateEx && | |
2544 nop != Op_CheckCastPP && | |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2545 nop != Op_DecodeN && |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2546 nop != Op_DecodeNKlass && |
0 | 2547 !n->is_Mem() ) { |
2548 Node *x = n->clone(); | |
2549 call->set_req( TypeFunc::Parms, x ); | |
2550 } | |
2551 } | |
2552 break; | |
2553 } | |
2554 | |
2555 case Op_StoreD: | |
2556 case Op_LoadD: | |
2557 case Op_LoadD_unaligned: | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2558 frc.inc_double_count(); |
0 | 2559 goto handle_mem; |
2560 case Op_StoreF: | |
2561 case Op_LoadF: | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2562 frc.inc_float_count(); |
0 | 2563 goto handle_mem; |
2564 | |
3248
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2565 case Op_StoreCM: |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2566 { |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2567 // Convert OopStore dependence into precedence edge |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2568 Node* prec = n->in(MemNode::OopStore); |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2569 n->del_req(MemNode::OopStore); |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2570 n->add_prec(prec); |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2571 eliminate_redundant_card_marks(n); |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2572 } |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2573 |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2574 // fall through |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2376
diff
changeset
|
2575 |
0 | 2576 case Op_StoreB: |
2577 case Op_StoreC: | |
2578 case Op_StorePConditional: | |
2579 case Op_StoreI: | |
2580 case Op_StoreL: | |
420
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
418
diff
changeset
|
2581 case Op_StoreIConditional: |
0 | 2582 case Op_StoreLConditional: |
2583 case Op_CompareAndSwapI: | |
2584 case Op_CompareAndSwapL: | |
2585 case Op_CompareAndSwapP: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
2586 case Op_CompareAndSwapN: |
6795
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6792
diff
changeset
|
2587 case Op_GetAndAddI: |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6792
diff
changeset
|
2588 case Op_GetAndAddL: |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6792
diff
changeset
|
2589 case Op_GetAndSetI: |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6792
diff
changeset
|
2590 case Op_GetAndSetL: |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6792
diff
changeset
|
2591 case Op_GetAndSetP: |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6792
diff
changeset
|
2592 case Op_GetAndSetN: |
0 | 2593 case Op_StoreP: |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
2594 case Op_StoreN: |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2595 case Op_StoreNKlass: |
0 | 2596 case Op_LoadB: |
624 | 2597 case Op_LoadUB: |
558
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
492
diff
changeset
|
2598 case Op_LoadUS: |
0 | 2599 case Op_LoadI: |
2600 case Op_LoadKlass: | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2601 case Op_LoadNKlass: |
0 | 2602 case Op_LoadL: |
2603 case Op_LoadL_unaligned: | |
2604 case Op_LoadPLocked: | |
2605 case Op_LoadP: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
2606 case Op_LoadN: |
0 | 2607 case Op_LoadRange: |
2608 case Op_LoadS: { | |
2609 handle_mem: | |
2610 #ifdef ASSERT | |
2611 if( VerifyOptoOopOffsets ) { | |
2612 assert( n->is_Mem(), "" ); | |
2613 MemNode *mem = (MemNode*)n; | |
2614 // Check to see if address types have grounded out somehow. | |
2615 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); | |
2616 assert( !tp || oop_offset_is_sane(tp), "" ); | |
2617 } | |
2618 #endif | |
2619 break; | |
2620 } | |
2621 | |
2622 case Op_AddP: { // Assert sane base pointers | |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2623 Node *addp = n->in(AddPNode::Address); |
0 | 2624 assert( !addp->is_AddP() || |
2625 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation | |
2626 addp->in(AddPNode::Base) == n->in(AddPNode::Base), | |
2627 "Base pointers must match" ); | |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2628 #ifdef _LP64 |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2629 if ((UseCompressedOops || UseCompressedKlassPointers) && |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2630 addp->Opcode() == Op_ConP && |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2631 addp == n->in(AddPNode::Base) && |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2632 n->in(AddPNode::Offset)->is_Con()) { |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2633 // Use addressing with narrow klass to load with offset on x86. |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2634 // On sparc loading 32-bits constant and decoding it have less |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2635 // instructions (4) then load 64-bits constant (7). |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2636 // Do this transformation here since IGVN will convert ConN back to ConP. |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2637 const Type* t = addp->bottom_type(); |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2638 if (t->isa_oopptr() || t->isa_klassptr()) { |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2639 Node* nn = NULL; |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2640 |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2641 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; |
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2642 |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2643 // Look for existing ConN node of the same exact type. |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2644 Node* r = root(); |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2645 uint cnt = r->outcnt(); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2646 for (uint i = 0; i < cnt; i++) { |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2647 Node* m = r->raw_out(i); |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2648 if (m!= NULL && m->Opcode() == op && |
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
182
diff
changeset
|
2649 m->bottom_type()->make_ptr() == t) { |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2650 nn = m; |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2651 break; |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2652 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2653 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2654 if (nn != NULL) { |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2655 // Decode a narrow oop to match address |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2656 // [R12 + narrow_oop_reg<<3 + offset] |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2657 if (t->isa_oopptr()) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2658 nn = new (this) DecodeNNode(nn, t); |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2659 } else { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2660 nn = new (this) DecodeNKlassNode(nn, t); |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2661 } |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2662 n->set_req(AddPNode::Base, nn); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2663 n->set_req(AddPNode::Address, nn); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2664 if (addp->outcnt() == 0) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2665 addp->disconnect_inputs(NULL, this); |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2666 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2667 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2668 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2669 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2670 #endif |
0 | 2671 break; |
2672 } | |
2673 | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2674 #ifdef _LP64 |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2675 case Op_CastPP: |
1575
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2676 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2677 Node* in1 = n->in(1); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2678 const Type* t = n->bottom_type(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2679 Node* new_in1 = in1->clone(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2680 new_in1->as_DecodeN()->set_type(t); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2681 |
1575
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2682 if (!Matcher::narrow_oop_use_complex_address()) { |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2683 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2684 // x86, ARM and friends can handle 2 adds in addressing mode |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2685 // and Matcher can fold a DecodeN node into address by using |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2686 // a narrow oop directly and do implicit NULL check in address: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2687 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2688 // [R12 + narrow_oop_reg<<3 + offset] |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2689 // NullCheck narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2690 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2691 // On other platforms (Sparc) we have to keep new DecodeN node and |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2692 // use it to do implicit NULL check in address: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2693 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2694 // decode_not_null narrow_oop_reg, base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2695 // [base_reg + offset] |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2696 // NullCheck base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2697 // |
605 | 2698 // Pin the new DecodeN node to non-null path on these platform (Sparc) |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2699 // to keep the information to which NULL check the new DecodeN node |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2700 // corresponds to use it as value in implicit_null_check(). |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2701 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2702 new_in1->set_req(0, n->in(0)); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2703 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2704 |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2705 n->subsume_by(new_in1, this); |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2706 if (in1->outcnt() == 0) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2707 in1->disconnect_inputs(NULL, this); |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2708 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2709 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2710 break; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2711 |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2712 case Op_CmpP: |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2713 // Do this transformation here to preserve CmpPNode::sub() and |
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2714 // other TypePtr related Ideal optimizations (for example, ptr nullness). |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2715 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2716 Node* in1 = n->in(1); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2717 Node* in2 = n->in(2); |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2718 if (!in1->is_DecodeNarrowPtr()) { |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2719 in2 = in1; |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2720 in1 = n->in(2); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2721 } |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2722 assert(in1->is_DecodeNarrowPtr(), "sanity"); |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2723 |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2724 Node* new_in2 = NULL; |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2725 if (in2->is_DecodeNarrowPtr()) { |
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2726 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2727 new_in2 = in2->in(1); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2728 } else if (in2->Opcode() == Op_ConP) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2729 const Type* t = in2->bottom_type(); |
1575
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2730 if (t == TypePtr::NULL_PTR) { |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2731 assert(in1->is_DecodeN(), "compare klass to null?"); |
1575
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2732 // Don't convert CmpP null check into CmpN if compressed |
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2733 // oops implicit null check is not generated. |
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2734 // This will allow to generate normal oop implicit null check. |
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2735 if (Matcher::gen_narrow_oop_implicit_null_checks()) |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2736 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2737 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2738 // This transformation together with CastPP transformation above |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2739 // will generated code for implicit NULL checks for compressed oops. |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2740 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2741 // The original code after Optimize() |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2742 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2743 // LoadN memory, narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2744 // decode narrow_oop_reg, base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2745 // CmpP base_reg, NULL |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2746 // CastPP base_reg // NotNull |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2747 // Load [base_reg + offset], val_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2748 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2749 // after these transformations will be |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2750 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2751 // LoadN memory, narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2752 // CmpN narrow_oop_reg, NULL |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2753 // decode_not_null narrow_oop_reg, base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2754 // Load [base_reg + offset], val_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2755 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2756 // and the uncommon path (== NULL) will use narrow_oop_reg directly |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2757 // since narrow oops can be used in debug info now (see the code in |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2758 // final_graph_reshaping_walk()). |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2759 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2760 // At the end the code will be matched to |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2761 // on x86: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2762 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2763 // Load_narrow_oop memory, narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2764 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2765 // NullCheck narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2766 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2767 // and on sparc: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2768 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2769 // Load_narrow_oop memory, narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2770 // decode_not_null narrow_oop_reg, base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2771 // Load [base_reg + offset], val_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2772 // NullCheck base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2773 // |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2774 } else if (t->isa_oopptr()) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2775 new_in2 = ConNode::make(this, t->make_narrowoop()); |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2776 } else if (t->isa_klassptr()) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2777 new_in2 = ConNode::make(this, t->make_narrowklass()); |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2778 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2779 } |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2780 if (new_in2 != NULL) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2781 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2782 n->subsume_by(cmpN, this); |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2783 if (in1->outcnt() == 0) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2784 in1->disconnect_inputs(NULL, this); |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2785 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2786 if (in2->outcnt() == 0) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2787 in2->disconnect_inputs(NULL, this); |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2788 } |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2789 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2790 } |
293
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
253
diff
changeset
|
2791 break; |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2792 |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2793 case Op_DecodeN: |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2794 case Op_DecodeNKlass: |
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2795 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); |
1575
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2796 // DecodeN could be pinned when it can't be fold into |
492
5496e074077f
6787050: assert(n->in(0) == 0L,"no control") with UseCompressedOops on sparcv9
kvn
parents:
490
diff
changeset
|
2797 // an address expression, see the code for Op_CastPP above. |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2798 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2799 break; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2800 |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2801 case Op_EncodeP: |
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2802 case Op_EncodePKlass: { |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2803 Node* in1 = n->in(1); |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2804 if (in1->is_DecodeNarrowPtr()) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2805 n->subsume_by(in1->in(1), this); |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2806 } else if (in1->Opcode() == Op_ConP) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2807 const Type* t = in1->bottom_type(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2808 if (t == TypePtr::NULL_PTR) { |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2809 assert(t->isa_oopptr(), "null klass?"); |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2810 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2811 } else if (t->isa_oopptr()) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2812 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2813 } else if (t->isa_klassptr()) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2814 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2815 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2816 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2817 if (in1->outcnt() == 0) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2818 in1->disconnect_inputs(NULL, this); |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2819 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2820 break; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2821 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2822 |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2823 case Op_Proj: { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2824 if (OptimizeStringConcat) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2825 ProjNode* p = n->as_Proj(); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2826 if (p->_is_io_use) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2827 // Separate projections were used for the exception path which |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2828 // are normally removed by a late inline. If it wasn't inlined |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2829 // then they will hang around and should just be replaced with |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2830 // the original one. |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2831 Node* proj = NULL; |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2832 // Replace with just one |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2833 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2834 Node *use = i.get(); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2835 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2836 proj = use; |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2837 break; |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2838 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2839 } |
4767 | 2840 assert(proj != NULL, "must be found"); |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2841 p->subsume_by(proj, this); |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2842 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2843 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2844 break; |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2845 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2846 |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2847 case Op_Phi: |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
2848 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2849 // The EncodeP optimization may create Phi with the same edges |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2850 // for all paths. It is not handled well by Register Allocator. |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2851 Node* unique_in = n->in(1); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2852 assert(unique_in != NULL, ""); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2853 uint cnt = n->req(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2854 for (uint i = 2; i < cnt; i++) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2855 Node* m = n->in(i); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2856 assert(m != NULL, ""); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2857 if (unique_in != m) |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2858 unique_in = NULL; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2859 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2860 if (unique_in != NULL) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2861 n->subsume_by(unique_in, this); |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2862 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2863 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2864 break; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2865 |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2866 #endif |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2867 |
0 | 2868 case Op_ModI: |
2869 if (UseDivMod) { | |
2870 // Check if a%b and a/b both exist | |
2871 Node* d = n->find_similar(Op_DivI); | |
2872 if (d) { | |
2873 // Replace them with a fused divmod if supported | |
2874 if (Matcher::has_match_rule(Op_DivModI)) { | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2875 DivModINode* divmod = DivModINode::make(this, n); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2876 d->subsume_by(divmod->div_proj(), this); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2877 n->subsume_by(divmod->mod_proj(), this); |
0 | 2878 } else { |
2879 // replace a%b with a-((a/b)*b) | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2880 Node* mult = new (this) MulINode(d, d->in(2)); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2881 Node* sub = new (this) SubINode(d->in(1), mult); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2882 n->subsume_by(sub, this); |
0 | 2883 } |
2884 } | |
2885 } | |
2886 break; | |
2887 | |
2888 case Op_ModL: | |
2889 if (UseDivMod) { | |
2890 // Check if a%b and a/b both exist | |
2891 Node* d = n->find_similar(Op_DivL); | |
2892 if (d) { | |
2893 // Replace them with a fused divmod if supported | |
2894 if (Matcher::has_match_rule(Op_DivModL)) { | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2895 DivModLNode* divmod = DivModLNode::make(this, n); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2896 d->subsume_by(divmod->div_proj(), this); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2897 n->subsume_by(divmod->mod_proj(), this); |
0 | 2898 } else { |
2899 // replace a%b with a-((a/b)*b) | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2900 Node* mult = new (this) MulLNode(d, d->in(2)); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2901 Node* sub = new (this) SubLNode(d->in(1), mult); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2902 n->subsume_by(sub, this); |
0 | 2903 } |
2904 } | |
2905 } | |
2906 break; | |
2907 | |
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
6143
diff
changeset
|
2908 case Op_LoadVector: |
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
6143
diff
changeset
|
2909 case Op_StoreVector: |
0 | 2910 break; |
2911 | |
2912 case Op_PackB: | |
2913 case Op_PackS: | |
2914 case Op_PackI: | |
2915 case Op_PackF: | |
2916 case Op_PackL: | |
2917 case Op_PackD: | |
2918 if (n->req()-1 > 2) { | |
2919 // Replace many operand PackNodes with a binary tree for matching | |
2920 PackNode* p = (PackNode*) n; | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2921 Node* btp = p->binary_tree_pack(this, 1, n->req()); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2922 n->subsume_by(btp, this); |
0 | 2923 } |
2924 break; | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2925 case Op_Loop: |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2926 case Op_CountedLoop: |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2927 if (n->as_Loop()->is_inner_loop()) { |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2928 frc.inc_inner_loop_count(); |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2929 } |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2930 break; |
2401
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2931 case Op_LShiftI: |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2932 case Op_RShiftI: |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2933 case Op_URShiftI: |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2934 case Op_LShiftL: |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2935 case Op_RShiftL: |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2936 case Op_URShiftL: |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2937 if (Matcher::need_masked_shift_count) { |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2938 // The cpu's shift instructions don't restrict the count to the |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2939 // lower 5/6 bits. We need to do the masking ourselves. |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2940 Node* in2 = n->in(2); |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2941 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2942 const TypeInt* t = in2->find_int_type(); |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2943 if (t != NULL && t->is_con()) { |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2944 juint shift = t->get_con(); |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2945 if (shift > mask) { // Unsigned cmp |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2946 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); |
2401
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2947 } |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2948 } else { |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2949 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2950 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); |
2401
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2951 n->set_req(2, shift); |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2952 } |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2953 } |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2954 if (in2->outcnt() == 0) { // Remove dead node |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2955 in2->disconnect_inputs(NULL, this); |
2401
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2956 } |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2957 } |
7e88bdae86ec
7029017: Additional architecture support for c2 compiler
roland
parents:
2376
diff
changeset
|
2958 break; |
8694
8651f608fea4
8009460: C2compiler crash in machnode::in_regmask(unsigned int)
roland
parents:
8691
diff
changeset
|
2959 case Op_MemBarStoreStore: |
10278 | 2960 case Op_MemBarRelease: |
8694
8651f608fea4
8009460: C2compiler crash in machnode::in_regmask(unsigned int)
roland
parents:
8691
diff
changeset
|
2961 // Break the link with AllocateNode: it is no longer useful and |
8651f608fea4
8009460: C2compiler crash in machnode::in_regmask(unsigned int)
roland
parents:
8691
diff
changeset
|
2962 // confuses register allocation. |
8651f608fea4
8009460: C2compiler crash in machnode::in_regmask(unsigned int)
roland
parents:
8691
diff
changeset
|
2963 if (n->req() > MemBarNode::Precedent) { |
8651f608fea4
8009460: C2compiler crash in machnode::in_regmask(unsigned int)
roland
parents:
8691
diff
changeset
|
2964 n->set_req(MemBarNode::Precedent, top()); |
8651f608fea4
8009460: C2compiler crash in machnode::in_regmask(unsigned int)
roland
parents:
8691
diff
changeset
|
2965 } |
8651f608fea4
8009460: C2compiler crash in machnode::in_regmask(unsigned int)
roland
parents:
8691
diff
changeset
|
2966 break; |
0 | 2967 default: |
2968 assert( !n->is_Call(), "" ); | |
2969 assert( !n->is_Mem(), "" ); | |
2970 break; | |
2971 } | |
127 | 2972 |
2973 // Collect CFG split points | |
2974 if (n->is_MultiBranch()) | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2975 frc._tests.push(n); |
0 | 2976 } |
2977 | |
2978 //------------------------------final_graph_reshaping_walk--------------------- | |
2979 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), | |
2980 // requires that the walk visits a node's inputs before visiting the node. | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
2981 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2982 ResourceArea *area = Thread::current()->resource_area(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2983 Unique_Node_List sfpt(area); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2984 |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2985 frc._visited.set(root->_idx); // first, mark node as visited |
0 | 2986 uint cnt = root->req(); |
2987 Node *n = root; | |
2988 uint i = 0; | |
2989 while (true) { | |
2990 if (i < cnt) { | |
2991 // Place all non-visited non-null inputs onto stack | |
2992 Node* m = n->in(i); | |
2993 ++i; | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2994 if (m != NULL && !frc._visited.test_set(m->_idx)) { |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2995 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2996 sfpt.push(m); |
0 | 2997 cnt = m->req(); |
2998 nstack.push(n, i); // put on stack parent and next input's index | |
2999 n = m; | |
3000 i = 0; | |
3001 } | |
3002 } else { | |
3003 // Now do post-visit work | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
3004 final_graph_reshaping_impl( n, frc ); |
0 | 3005 if (nstack.is_empty()) |
3006 break; // finished | |
3007 n = nstack.node(); // Get node from stack | |
3008 cnt = n->req(); | |
3009 i = nstack.index(); | |
3010 nstack.pop(); // Shift to the next node on stack | |
3011 } | |
3012 } | |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3013 |
1575
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
3014 // Skip next transformation if compressed oops are not used. |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
3015 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || |
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
3016 (!UseCompressedOops && !UseCompressedKlassPointers)) |
1575
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
3017 return; |
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
3018 |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
3019 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3020 // It could be done for an uncommon traps or any safepoints/calls |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
3021 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3022 while (sfpt.size() > 0) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3023 n = sfpt.pop(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3024 JVMState *jvms = n->as_SafePoint()->jvms(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3025 assert(jvms != NULL, "sanity"); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3026 int start = jvms->debug_start(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3027 int end = n->req(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3028 bool is_uncommon = (n->is_CallStaticJava() && |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3029 n->as_CallStaticJava()->uncommon_trap_request() != 0); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3030 for (int j = start; j < end; j++) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3031 Node* in = n->in(j); |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6843
diff
changeset
|
3032 if (in->is_DecodeNarrowPtr()) { |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3033 bool safe_to_skip = true; |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3034 if (!is_uncommon ) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3035 // Is it safe to skip? |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3036 for (uint i = 0; i < in->outcnt(); i++) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3037 Node* u = in->raw_out(i); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3038 if (!u->is_SafePoint() || |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3039 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3040 safe_to_skip = false; |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3041 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3042 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3043 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3044 if (safe_to_skip) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3045 n->set_req(j, in->in(1)); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3046 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3047 if (in->outcnt() == 0) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3048 in->disconnect_inputs(NULL, this); |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3049 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3050 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3051 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
3052 } |
0 | 3053 } |
3054 | |
3055 //------------------------------final_graph_reshaping-------------------------- | |
3056 // Final Graph Reshaping. | |
3057 // | |
3058 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late | |
3059 // and not commoned up and forced early. Must come after regular | |
3060 // optimizations to avoid GVN undoing the cloning. Clone constant | |
3061 // inputs to Loop Phis; these will be split by the allocator anyways. | |
3062 // Remove Opaque nodes. | |
3063 // (2) Move last-uses by commutative operations to the left input to encourage | |
3064 // Intel update-in-place two-address operations and better register usage | |
3065 // on RISCs. Must come after regular optimizations to avoid GVN Ideal | |
3066 // calls canonicalizing them back. | |
3067 // (3) Count the number of double-precision FP ops, single-precision FP ops | |
3068 // and call sites. On Intel, we can get correct rounding either by | |
3069 // forcing singles to memory (requires extra stores and loads after each | |
3070 // FP bytecode) or we can set a rounding mode bit (requires setting and | |
3071 // clearing the mode bit around call sites). The mode bit is only used | |
3072 // if the relative frequency of single FP ops to calls is low enough. | |
3073 // This is a key transform for SPEC mpeg_audio. | |
3074 // (4) Detect infinite loops; blobs of code reachable from above but not | |
3075 // below. Several of the Code_Gen algorithms fail on such code shapes, | |
3076 // so we simply bail out. Happens a lot in ZKM.jar, but also happens | |
3077 // from time to time in other codes (such as -Xcomp finalizer loops, etc). | |
3078 // Detection is by looking for IfNodes where only 1 projection is | |
3079 // reachable from below or CatchNodes missing some targets. | |
3080 // (5) Assert for insane oop offsets in debug mode. | |
3081 | |
3082 bool Compile::final_graph_reshaping() { | |
3083 // an infinite loop may have been eliminated by the optimizer, | |
3084 // in which case the graph will be empty. | |
3085 if (root()->req() == 1) { | |
3086 record_method_not_compilable("trivial infinite loop"); | |
3087 return true; | |
3088 } | |
3089 | |
8048
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3090 // Expensive nodes have their control input set to prevent the GVN |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3091 // from freely commoning them. There's no GVN beyond this point so |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3092 // no need to keep the control input. We want the expensive nodes to |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3093 // be freely moved to the least frequent code path by gcm. |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3094 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3095 for (int i = 0; i < expensive_count(); i++) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3096 _expensive_nodes->at(i)->set_req(0, NULL); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3097 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3098 |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
3099 Final_Reshape_Counts frc; |
0 | 3100 |
3101 // Visit everybody reachable! | |
3102 // Allocate stack of size C->unique()/2 to avoid frequent realloc | |
3103 Node_Stack nstack(unique() >> 1); | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
3104 final_graph_reshaping_walk(nstack, root(), frc); |
0 | 3105 |
3106 // Check for unreachable (from below) code (i.e., infinite loops). | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
3107 for( uint i = 0; i < frc._tests.size(); i++ ) { |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
3108 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); |
127 | 3109 // Get number of CFG targets. |
0 | 3110 // Note that PCTables include exception targets after calls. |
127 | 3111 uint required_outcnt = n->required_outcnt(); |
3112 if (n->outcnt() != required_outcnt) { | |
0 | 3113 // Check for a few special cases. Rethrow Nodes never take the |
3114 // 'fall-thru' path, so expected kids is 1 less. | |
3115 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { | |
3116 if (n->in(0)->in(0)->is_Call()) { | |
3117 CallNode *call = n->in(0)->in(0)->as_Call(); | |
3118 if (call->entry_point() == OptoRuntime::rethrow_stub()) { | |
127 | 3119 required_outcnt--; // Rethrow always has 1 less kid |
0 | 3120 } else if (call->req() > TypeFunc::Parms && |
3121 call->is_CallDynamicJava()) { | |
3122 // Check for null receiver. In such case, the optimizer has | |
3123 // detected that the virtual call will always result in a null | |
3124 // pointer exception. The fall-through projection of this CatchNode | |
3125 // will not be populated. | |
3126 Node *arg0 = call->in(TypeFunc::Parms); | |
3127 if (arg0->is_Type() && | |
3128 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { | |
127 | 3129 required_outcnt--; |
0 | 3130 } |
3131 } else if (call->entry_point() == OptoRuntime::new_array_Java() && | |
3132 call->req() > TypeFunc::Parms+1 && | |
3133 call->is_CallStaticJava()) { | |
3134 // Check for negative array length. In such case, the optimizer has | |
3135 // detected that the allocation attempt will always result in an | |
3136 // exception. There is no fall-through projection of this CatchNode . | |
3137 Node *arg1 = call->in(TypeFunc::Parms+1); | |
3138 if (arg1->is_Type() && | |
3139 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { | |
127 | 3140 required_outcnt--; |
0 | 3141 } |
3142 } | |
3143 } | |
3144 } | |
127 | 3145 // Recheck with a better notion of 'required_outcnt' |
3146 if (n->outcnt() != required_outcnt) { | |
0 | 3147 record_method_not_compilable("malformed control flow"); |
3148 return true; // Not all targets reachable! | |
3149 } | |
3150 } | |
3151 // Check that I actually visited all kids. Unreached kids | |
3152 // must be infinite loops. | |
3153 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
3154 if (!frc._visited.test(n->fast_out(j)->_idx)) { |
0 | 3155 record_method_not_compilable("infinite loop"); |
3156 return true; // Found unvisited kid; must be unreach | |
3157 } | |
3158 } | |
3159 | |
3160 // If original bytecodes contained a mixture of floats and doubles | |
3161 // check if the optimizer has made it homogenous, item (3). | |
929
cd18bd5e667c
6873777: FPU control word optimization still performed with SSE
never
parents:
921
diff
changeset
|
3162 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
3163 frc.get_float_count() > 32 && |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
3164 frc.get_double_count() == 0 && |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
3165 (10 * frc.get_call_count() < frc.get_float_count()) ) { |
0 | 3166 set_24_bit_selection_and_mode( false, true ); |
3167 } | |
3168 | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
3169 set_java_calls(frc.get_java_call_count()); |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
3170 set_inner_loops(frc.get_inner_loop_count()); |
0 | 3171 |
3172 // No infinite loops, no reason to bail out. | |
3173 return false; | |
3174 } | |
3175 | |
3176 //-----------------------------too_many_traps---------------------------------- | |
3177 // Report if there are too many traps at the current method and bci. | |
3178 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. | |
3179 bool Compile::too_many_traps(ciMethod* method, | |
3180 int bci, | |
3181 Deoptimization::DeoptReason reason) { | |
3182 ciMethodData* md = method->method_data(); | |
3183 if (md->is_empty()) { | |
3184 // Assume the trap has not occurred, or that it occurred only | |
3185 // because of a transient condition during start-up in the interpreter. | |
3186 return false; | |
3187 } | |
3188 if (md->has_trap_at(bci, reason) != 0) { | |
3189 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. | |
3190 // Also, if there are multiple reasons, or if there is no per-BCI record, | |
3191 // assume the worst. | |
3192 if (log()) | |
3193 log()->elem("observe trap='%s' count='%d'", | |
3194 Deoptimization::trap_reason_name(reason), | |
3195 md->trap_count(reason)); | |
3196 return true; | |
3197 } else { | |
3198 // Ignore method/bci and see if there have been too many globally. | |
3199 return too_many_traps(reason, md); | |
3200 } | |
3201 } | |
3202 | |
3203 // Less-accurate variant which does not require a method and bci. | |
3204 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, | |
3205 ciMethodData* logmd) { | |
3206 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { | |
3207 // Too many traps globally. | |
3208 // Note that we use cumulative trap_count, not just md->trap_count. | |
3209 if (log()) { | |
3210 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); | |
3211 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", | |
3212 Deoptimization::trap_reason_name(reason), | |
3213 mcount, trap_count(reason)); | |
3214 } | |
3215 return true; | |
3216 } else { | |
3217 // The coast is clear. | |
3218 return false; | |
3219 } | |
3220 } | |
3221 | |
3222 //--------------------------too_many_recompiles-------------------------------- | |
3223 // Report if there are too many recompiles at the current method and bci. | |
3224 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. | |
3225 // Is not eager to return true, since this will cause the compiler to use | |
3226 // Action_none for a trap point, to avoid too many recompilations. | |
3227 bool Compile::too_many_recompiles(ciMethod* method, | |
3228 int bci, | |
3229 Deoptimization::DeoptReason reason) { | |
3230 ciMethodData* md = method->method_data(); | |
3231 if (md->is_empty()) { | |
3232 // Assume the trap has not occurred, or that it occurred only | |
3233 // because of a transient condition during start-up in the interpreter. | |
3234 return false; | |
3235 } | |
3236 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. | |
3237 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; | |
3238 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero | |
3239 Deoptimization::DeoptReason per_bc_reason | |
3240 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); | |
3241 if ((per_bc_reason == Deoptimization::Reason_none | |
3242 || md->has_trap_at(bci, reason) != 0) | |
3243 // The trap frequency measure we care about is the recompile count: | |
3244 && md->trap_recompiled_at(bci) | |
3245 && md->overflow_recompile_count() >= bc_cutoff) { | |
3246 // Do not emit a trap here if it has already caused recompilations. | |
3247 // Also, if there are multiple reasons, or if there is no per-BCI record, | |
3248 // assume the worst. | |
3249 if (log()) | |
3250 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", | |
3251 Deoptimization::trap_reason_name(reason), | |
3252 md->trap_count(reason), | |
3253 md->overflow_recompile_count()); | |
3254 return true; | |
3255 } else if (trap_count(reason) != 0 | |
3256 && decompile_count() >= m_cutoff) { | |
3257 // Too many recompiles globally, and we have seen this sort of trap. | |
3258 // Use cumulative decompile_count, not just md->decompile_count. | |
3259 if (log()) | |
3260 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", | |
3261 Deoptimization::trap_reason_name(reason), | |
3262 md->trap_count(reason), trap_count(reason), | |
3263 md->decompile_count(), decompile_count()); | |
3264 return true; | |
3265 } else { | |
3266 // The coast is clear. | |
3267 return false; | |
3268 } | |
3269 } | |
3270 | |
3271 | |
3272 #ifndef PRODUCT | |
3273 //------------------------------verify_graph_edges--------------------------- | |
3274 // Walk the Graph and verify that there is a one-to-one correspondence | |
3275 // between Use-Def edges and Def-Use edges in the graph. | |
3276 void Compile::verify_graph_edges(bool no_dead_code) { | |
3277 if (VerifyGraphEdges) { | |
3278 ResourceArea *area = Thread::current()->resource_area(); | |
3279 Unique_Node_List visited(area); | |
3280 // Call recursive graph walk to check edges | |
3281 _root->verify_edges(visited); | |
3282 if (no_dead_code) { | |
3283 // Now make sure that no visited node is used by an unvisited node. | |
3284 bool dead_nodes = 0; | |
3285 Unique_Node_List checked(area); | |
3286 while (visited.size() > 0) { | |
3287 Node* n = visited.pop(); | |
3288 checked.push(n); | |
3289 for (uint i = 0; i < n->outcnt(); i++) { | |
3290 Node* use = n->raw_out(i); | |
3291 if (checked.member(use)) continue; // already checked | |
3292 if (visited.member(use)) continue; // already in the graph | |
3293 if (use->is_Con()) continue; // a dead ConNode is OK | |
3294 // At this point, we have found a dead node which is DU-reachable. | |
3295 if (dead_nodes++ == 0) | |
3296 tty->print_cr("*** Dead nodes reachable via DU edges:"); | |
3297 use->dump(2); | |
3298 tty->print_cr("---"); | |
3299 checked.push(use); // No repeats; pretend it is now checked. | |
3300 } | |
3301 } | |
3302 assert(dead_nodes == 0, "using nodes must be reachable from root"); | |
3303 } | |
3304 } | |
3305 } | |
3306 #endif | |
3307 | |
3308 // The Compile object keeps track of failure reasons separately from the ciEnv. | |
3309 // This is required because there is not quite a 1-1 relation between the | |
3310 // ciEnv and its compilation task and the Compile object. Note that one | |
3311 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides | |
3312 // to backtrack and retry without subsuming loads. Other than this backtracking | |
3313 // behavior, the Compile's failure reason is quietly copied up to the ciEnv | |
3314 // by the logic in C2Compiler. | |
3315 void Compile::record_failure(const char* reason) { | |
3316 if (log() != NULL) { | |
3317 log()->elem("failure reason='%s' phase='compile'", reason); | |
3318 } | |
3319 if (_failure_reason == NULL) { | |
3320 // Record the first failure reason. | |
3321 _failure_reason = reason; | |
3322 } | |
10405 | 3323 |
3324 EventCompilerFailure event; | |
3325 if (event.should_commit()) { | |
3326 event.set_compileID(Compile::compile_id()); | |
3327 event.set_failure(reason); | |
3328 event.commit(); | |
3329 } | |
3330 | |
222 | 3331 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { |
10405 | 3332 C->print_method(PHASE_FAILURE); |
222 | 3333 } |
0 | 3334 _root = NULL; // flush the graph, too |
3335 } | |
3336 | |
3337 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3338 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3339 _phase_name(name), _dolog(dolog) |
0 | 3340 { |
3341 if (dolog) { | |
3342 C = Compile::current(); | |
3343 _log = C->log(); | |
3344 } else { | |
3345 C = NULL; | |
3346 _log = NULL; | |
3347 } | |
3348 if (_log != NULL) { | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3349 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); |
0 | 3350 _log->stamp(); |
3351 _log->end_head(); | |
3352 } | |
3353 } | |
3354 | |
3355 Compile::TracePhase::~TracePhase() { | |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3356 |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3357 C = Compile::current(); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3358 if (_dolog) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3359 _log = C->log(); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3360 } else { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3361 _log = NULL; |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3362 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3363 |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3364 #ifdef ASSERT |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3365 if (PrintIdealNodeCount) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3366 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3367 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3368 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3369 |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3370 if (VerifyIdealNodeCount) { |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3371 Compile::current()->print_missing_nodes(); |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3372 } |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3373 #endif |
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3374 |
0 | 3375 if (_log != NULL) { |
7196
2aff40cb4703
7092905: C2: Keep track of the number of dead nodes
bharadwaj
parents:
6888
diff
changeset
|
3376 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); |
0 | 3377 } |
3378 } | |
2008 | 3379 |
3380 //============================================================================= | |
3381 // Two Constant's are equal when the type and the value are equal. | |
3382 bool Compile::Constant::operator==(const Constant& other) { | |
3383 if (type() != other.type() ) return false; | |
3384 if (can_be_reused() != other.can_be_reused()) return false; | |
3385 // For floating point values we compare the bit pattern. | |
3386 switch (type()) { | |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3387 case T_FLOAT: return (_v._value.i == other._v._value.i); |
2008 | 3388 case T_LONG: |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3389 case T_DOUBLE: return (_v._value.j == other._v._value.j); |
2008 | 3390 case T_OBJECT: |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3391 case T_ADDRESS: return (_v._value.l == other._v._value.l); |
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3392 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries |
6888
cfe522e6461c
8000623: tools/javac/Diagnostics/6769027/T6769027.java crashes in PSPromotionManager::copy_to_survivor_space
kvn
parents:
6849
diff
changeset
|
3393 case T_METADATA: return (_v._metadata == other._v._metadata); |
2008 | 3394 default: ShouldNotReachHere(); |
3395 } | |
3396 return false; | |
3397 } | |
3398 | |
3399 static int type_to_size_in_bytes(BasicType t) { | |
3400 switch (t) { | |
3401 case T_LONG: return sizeof(jlong ); | |
3402 case T_FLOAT: return sizeof(jfloat ); | |
3403 case T_DOUBLE: return sizeof(jdouble); | |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3404 case T_METADATA: return sizeof(Metadata*); |
2008 | 3405 // We use T_VOID as marker for jump-table entries (labels) which |
4114
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3406 // need an internal word relocation. |
2008 | 3407 case T_VOID: |
3408 case T_ADDRESS: | |
3409 case T_OBJECT: return sizeof(jobject); | |
3410 } | |
3411 | |
3412 ShouldNotReachHere(); | |
3413 return -1; | |
3414 } | |
3415 | |
4114
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3416 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3417 // sort descending |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3418 if (a->freq() > b->freq()) return -1; |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3419 if (a->freq() < b->freq()) return 1; |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3420 return 0; |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3421 } |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3422 |
2008 | 3423 void Compile::ConstantTable::calculate_offsets_and_size() { |
4114
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3424 // First, sort the array by frequencies. |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3425 _constants.sort(qsort_comparator); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3426 |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3427 #ifdef ASSERT |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3428 // Make sure all jump-table entries were sorted to the end of the |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3429 // array (they have a negative frequency). |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3430 bool found_void = false; |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3431 for (int i = 0; i < _constants.length(); i++) { |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3432 Constant con = _constants.at(i); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3433 if (con.type() == T_VOID) |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3434 found_void = true; // jump-tables |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3435 else |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3436 assert(!found_void, "wrong sorting"); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3437 } |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3438 #endif |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3439 |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3440 int offset = 0; |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3441 for (int i = 0; i < _constants.length(); i++) { |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3442 Constant* con = _constants.adr_at(i); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3443 |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3444 // Align offset for type. |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3445 int typesize = type_to_size_in_bytes(con->type()); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3446 offset = align_size_up(offset, typesize); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3447 con->set_offset(offset); // set constant's offset |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3448 |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3449 if (con->type() == T_VOID) { |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3450 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3451 offset = offset + typesize * n->outcnt(); // expand jump-table |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3452 } else { |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3453 offset = offset + typesize; |
2008 | 3454 } |
3455 } | |
3456 | |
3457 // Align size up to the next section start (which is insts; see | |
3458 // CodeBuffer::align_at_start). | |
3459 assert(_size == -1, "already set?"); | |
4114
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3460 _size = align_size_up(offset, CodeEntryAlignment); |
2008 | 3461 } |
3462 | |
3463 void Compile::ConstantTable::emit(CodeBuffer& cb) { | |
3464 MacroAssembler _masm(&cb); | |
4114
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3465 for (int i = 0; i < _constants.length(); i++) { |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3466 Constant con = _constants.at(i); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3467 address constant_addr; |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3468 switch (con.type()) { |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3469 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3470 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3471 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3472 case T_OBJECT: { |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3473 jobject obj = con.get_jobject(); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3474 int oop_index = _masm.oop_recorder()->find_index(obj); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3475 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3476 break; |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3477 } |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3478 case T_ADDRESS: { |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3479 address addr = (address) con.get_jobject(); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3480 constant_addr = _masm.address_constant(addr); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3481 break; |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3482 } |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3483 // We use T_VOID as marker for jump-table entries (labels) which |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3484 // need an internal word relocation. |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3485 case T_VOID: { |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3486 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3487 // Fill the jump-table with a dummy word. The real value is |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3488 // filled in later in fill_jump_table. |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3489 address dummy = (address) n; |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3490 constant_addr = _masm.address_constant(dummy); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3491 // Expand jump-table |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3492 for (uint i = 1; i < n->outcnt(); i++) { |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3493 address temp_addr = _masm.address_constant(dummy + i); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3494 assert(temp_addr, "consts section too small"); |
2008 | 3495 } |
4114
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3496 break; |
2008 | 3497 } |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3498 case T_METADATA: { |
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3499 Metadata* obj = con.get_metadata(); |
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3500 int metadata_index = _masm.oop_recorder()->find_index(obj); |
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3501 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); |
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3502 break; |
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3503 } |
4114
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3504 default: ShouldNotReachHere(); |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3505 } |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3506 assert(constant_addr, "consts section too small"); |
6268
6c5b7a6becc8
7187454: stack overflow in C2 compiler thread on Solaris x86
kvn
parents:
6179
diff
changeset
|
3507 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); |
2008 | 3508 } |
3509 } | |
3510 | |
3511 int Compile::ConstantTable::find_offset(Constant& con) const { | |
3512 int idx = _constants.find(con); | |
3513 assert(idx != -1, "constant must be in constant table"); | |
3514 int offset = _constants.at(idx).offset(); | |
3515 assert(offset != -1, "constant table not emitted yet?"); | |
3516 return offset; | |
3517 } | |
3518 | |
3519 void Compile::ConstantTable::add(Constant& con) { | |
3520 if (con.can_be_reused()) { | |
3521 int idx = _constants.find(con); | |
3522 if (idx != -1 && _constants.at(idx).can_be_reused()) { | |
4114
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3523 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value |
2008 | 3524 return; |
3525 } | |
3526 } | |
3527 (void) _constants.append(con); | |
3528 } | |
3529 | |
4114
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3530 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3531 Block* b = Compile::current()->cfg()->_bbs[n->_idx]; |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3532 Constant con(type, value, b->_freq); |
2008 | 3533 add(con); |
3534 return con; | |
3535 } | |
3536 | |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3537 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { |
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3538 Constant con(metadata); |
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3539 add(con); |
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3540 return con; |
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3541 } |
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3542 |
4114
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3543 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { |
2008 | 3544 jvalue value; |
3545 BasicType type = oper->type()->basic_type(); | |
3546 switch (type) { | |
3547 case T_LONG: value.j = oper->constantL(); break; | |
3548 case T_FLOAT: value.f = oper->constantF(); break; | |
3549 case T_DOUBLE: value.d = oper->constantD(); break; | |
3550 case T_OBJECT: | |
3551 case T_ADDRESS: value.l = (jobject) oper->constant(); break; | |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3552 case T_METADATA: return add((Metadata*)oper->constant()); break; |
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6619
diff
changeset
|
3553 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); |
2008 | 3554 } |
4114
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3555 return add(n, type, value); |
2008 | 3556 } |
3557 | |
4114
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3558 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { |
2008 | 3559 jvalue value; |
3560 // We can use the node pointer here to identify the right jump-table | |
3561 // as this method is called from Compile::Fill_buffer right before | |
3562 // the MachNodes are emitted and the jump-table is filled (means the | |
3563 // MachNode pointers do not change anymore). | |
3564 value.l = (jobject) n; | |
4114
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3565 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. |
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3566 add(con); |
2008 | 3567 return con; |
3568 } | |
3569 | |
3570 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { | |
3571 // If called from Compile::scratch_emit_size do nothing. | |
3572 if (Compile::current()->in_scratch_emit_size()) return; | |
3573 | |
3574 assert(labels.is_nonempty(), "must be"); | |
6268
6c5b7a6becc8
7187454: stack overflow in C2 compiler thread on Solaris x86
kvn
parents:
6179
diff
changeset
|
3575 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); |
2008 | 3576 |
3577 // Since MachConstantNode::constant_offset() also contains | |
3578 // table_base_offset() we need to subtract the table_base_offset() | |
3579 // to get the plain offset into the constant table. | |
3580 int offset = n->constant_offset() - table_base_offset(); | |
3581 | |
3582 MacroAssembler _masm(&cb); | |
3583 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); | |
3584 | |
4114
6729bbc1fcd6
7003454: order constants in constant table by number of references in code
twisti
parents:
4064
diff
changeset
|
3585 for (uint i = 0; i < n->outcnt(); i++) { |
2008 | 3586 address* constant_addr = &jump_table_base[i]; |
6268
6c5b7a6becc8
7187454: stack overflow in C2 compiler thread on Solaris x86
kvn
parents:
6179
diff
changeset
|
3587 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i))); |
2008 | 3588 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); |
3589 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); | |
3590 } | |
3591 } | |
7421
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
3592 |
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
3593 void Compile::dump_inlining() { |
7605 | 3594 if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) { |
7473 | 3595 // Print inlining message for candidates that we couldn't inline |
3596 // for lack of space or non constant receiver | |
3597 for (int i = 0; i < _late_inlines.length(); i++) { | |
3598 CallGenerator* cg = _late_inlines.at(i); | |
3599 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); | |
3600 } | |
3601 Unique_Node_List useful; | |
3602 useful.push(root()); | |
3603 for (uint next = 0; next < useful.size(); ++next) { | |
3604 Node* n = useful.at(next); | |
3605 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { | |
3606 CallNode* call = n->as_Call(); | |
3607 CallGenerator* cg = call->generator(); | |
3608 cg->print_inlining_late("receiver not constant"); | |
3609 } | |
3610 uint max = n->len(); | |
3611 for ( uint i = 0; i < max; ++i ) { | |
3612 Node *m = n->in(i); | |
3613 if ( m == NULL ) continue; | |
3614 useful.push(m); | |
3615 } | |
3616 } | |
7421
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
3617 for (int i = 0; i < _print_inlining_list->length(); i++) { |
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
3618 tty->print(_print_inlining_list->at(i).ss()->as_string()); |
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
3619 } |
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
3620 } |
ad5dd04754ee
8005031: Some cleanup in c2 to prepare for incremental inlining support
roland
parents:
7199
diff
changeset
|
3621 } |
8048
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3622 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3623 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3624 if (n1->Opcode() < n2->Opcode()) return -1; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3625 else if (n1->Opcode() > n2->Opcode()) return 1; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3626 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3627 assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req())); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3628 for (uint i = 1; i < n1->req(); i++) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3629 if (n1->in(i) < n2->in(i)) return -1; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3630 else if (n1->in(i) > n2->in(i)) return 1; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3631 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3632 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3633 return 0; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3634 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3635 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3636 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3637 Node* n1 = *n1p; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3638 Node* n2 = *n2p; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3639 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3640 return cmp_expensive_nodes(n1, n2); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3641 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3642 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3643 void Compile::sort_expensive_nodes() { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3644 if (!expensive_nodes_sorted()) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3645 _expensive_nodes->sort(cmp_expensive_nodes); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3646 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3647 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3648 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3649 bool Compile::expensive_nodes_sorted() const { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3650 for (int i = 1; i < _expensive_nodes->length(); i++) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3651 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3652 return false; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3653 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3654 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3655 return true; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3656 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3657 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3658 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3659 if (_expensive_nodes->length() == 0) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3660 return false; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3661 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3662 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3663 assert(OptimizeExpensiveOps, "optimization off?"); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3664 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3665 // Take this opportunity to remove dead nodes from the list |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3666 int j = 0; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3667 for (int i = 0; i < _expensive_nodes->length(); i++) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3668 Node* n = _expensive_nodes->at(i); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3669 if (!n->is_unreachable(igvn)) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3670 assert(n->is_expensive(), "should be expensive"); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3671 _expensive_nodes->at_put(j, n); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3672 j++; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3673 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3674 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3675 _expensive_nodes->trunc_to(j); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3676 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3677 // Then sort the list so that similar nodes are next to each other |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3678 // and check for at least two nodes of identical kind with same data |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3679 // inputs. |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3680 sort_expensive_nodes(); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3681 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3682 for (int i = 0; i < _expensive_nodes->length()-1; i++) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3683 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3684 return true; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3685 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3686 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3687 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3688 return false; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3689 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3690 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3691 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3692 if (_expensive_nodes->length() == 0) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3693 return; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3694 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3695 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3696 assert(OptimizeExpensiveOps, "optimization off?"); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3697 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3698 // Sort to bring similar nodes next to each other and clear the |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3699 // control input of nodes for which there's only a single copy. |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3700 sort_expensive_nodes(); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3701 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3702 int j = 0; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3703 int identical = 0; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3704 int i = 0; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3705 for (; i < _expensive_nodes->length()-1; i++) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3706 assert(j <= i, "can't write beyond current index"); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3707 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3708 identical++; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3709 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3710 continue; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3711 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3712 if (identical > 0) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3713 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3714 identical = 0; |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3715 } else { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3716 Node* n = _expensive_nodes->at(i); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3717 igvn.hash_delete(n); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3718 n->set_req(0, NULL); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3719 igvn.hash_insert(n); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3720 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3721 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3722 if (identical > 0) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3723 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3724 } else if (_expensive_nodes->length() >= 1) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3725 Node* n = _expensive_nodes->at(i); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3726 igvn.hash_delete(n); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3727 n->set_req(0, NULL); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3728 igvn.hash_insert(n); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3729 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3730 _expensive_nodes->trunc_to(j); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3731 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3732 |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3733 void Compile::add_expensive_node(Node * n) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3734 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3735 assert(n->is_expensive(), "expensive nodes with non-null control here only"); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3736 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3737 if (OptimizeExpensiveOps) { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3738 _expensive_nodes->append(n); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3739 } else { |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3740 // Clear control input and let IGVN optimize expensive nodes if |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3741 // OptimizeExpensiveOps is off. |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3742 n->set_req(0, NULL); |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3743 } |
8b3da8d14c93
7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob
roland
parents:
7605
diff
changeset
|
3744 } |
8691
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3745 |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3746 // Auxiliary method to support randomized stressing/fuzzing. |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3747 // |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3748 // This method can be called the arbitrary number of times, with current count |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3749 // as the argument. The logic allows selecting a single candidate from the |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3750 // running list of candidates as follows: |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3751 // int count = 0; |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3752 // Cand* selected = null; |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3753 // while(cand = cand->next()) { |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3754 // if (randomized_select(++count)) { |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3755 // selected = cand; |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3756 // } |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3757 // } |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3758 // |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3759 // Including count equalizes the chances any candidate is "selected". |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3760 // This is useful when we don't have the complete list of candidates to choose |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3761 // from uniformly. In this case, we need to adjust the randomicity of the |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3762 // selection, or else we will end up biasing the selection towards the latter |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3763 // candidates. |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3764 // |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3765 // Quick back-envelope calculation shows that for the list of n candidates |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3766 // the equal probability for the candidate to persist as "best" can be |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3767 // achieved by replacing it with "next" k-th candidate with the probability |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3768 // of 1/k. It can be easily shown that by the end of the run, the |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3769 // probability for any candidate is converged to 1/n, thus giving the |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3770 // uniform distribution among all the candidates. |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3771 // |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3772 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3773 #define RANDOMIZED_DOMAIN_POW 29 |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3774 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3775 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3776 bool Compile::randomized_select(int count) { |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3777 assert(count > 0, "only positive"); |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3778 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); |
571076d3c79d
8009120: Fuzz instruction scheduling in HotSpot compilers
shade
parents:
8048
diff
changeset
|
3779 } |