Mercurial > hg > graal-compiler
annotate src/share/vm/opto/compile.cpp @ 2182:9569fdf936ff
Made merge compile.
author | Thomas Wuerthinger <wuerthinger@ssw.jku.at> |
---|---|
date | Wed, 16 Feb 2011 14:21:52 +0100 |
parents | 51bd2d261853 |
children | b92c45f2bc75 |
rev | line source |
---|---|
0 | 1 /* |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1397
diff
changeset
|
2 * Copyright (c) 1997, 2010, Oracle and/or its affiliates. All rights reserved. |
0 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 * | |
5 * This code is free software; you can redistribute it and/or modify it | |
6 * under the terms of the GNU General Public License version 2 only, as | |
7 * published by the Free Software Foundation. | |
8 * | |
9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
12 * version 2 for more details (a copy is included in the LICENSE file that | |
13 * accompanied this code). | |
14 * | |
15 * You should have received a copy of the GNU General Public License version | |
16 * 2 along with this work; if not, write to the Free Software Foundation, | |
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
18 * | |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
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" |
26 #include "asm/assembler.hpp" | |
27 #include "classfile/systemDictionary.hpp" | |
28 #include "code/exceptionHandlerTable.hpp" | |
29 #include "code/nmethod.hpp" | |
30 #include "compiler/compileLog.hpp" | |
31 #include "compiler/oopMap.hpp" | |
32 #include "opto/addnode.hpp" | |
33 #include "opto/block.hpp" | |
34 #include "opto/c2compiler.hpp" | |
35 #include "opto/callGenerator.hpp" | |
36 #include "opto/callnode.hpp" | |
37 #include "opto/cfgnode.hpp" | |
38 #include "opto/chaitin.hpp" | |
39 #include "opto/compile.hpp" | |
40 #include "opto/connode.hpp" | |
41 #include "opto/divnode.hpp" | |
42 #include "opto/escape.hpp" | |
43 #include "opto/idealGraphPrinter.hpp" | |
44 #include "opto/loopnode.hpp" | |
45 #include "opto/machnode.hpp" | |
46 #include "opto/macro.hpp" | |
47 #include "opto/matcher.hpp" | |
48 #include "opto/memnode.hpp" | |
49 #include "opto/mulnode.hpp" | |
50 #include "opto/node.hpp" | |
51 #include "opto/opcodes.hpp" | |
52 #include "opto/output.hpp" | |
53 #include "opto/parse.hpp" | |
54 #include "opto/phaseX.hpp" | |
55 #include "opto/rootnode.hpp" | |
56 #include "opto/runtime.hpp" | |
57 #include "opto/stringopts.hpp" | |
58 #include "opto/type.hpp" | |
59 #include "opto/vectornode.hpp" | |
60 #include "runtime/arguments.hpp" | |
61 #include "runtime/signature.hpp" | |
62 #include "runtime/stubRoutines.hpp" | |
63 #include "runtime/timer.hpp" | |
64 #include "utilities/copy.hpp" | |
65 #ifdef TARGET_ARCH_MODEL_x86_32 | |
66 # include "adfiles/ad_x86_32.hpp" | |
67 #endif | |
68 #ifdef TARGET_ARCH_MODEL_x86_64 | |
69 # include "adfiles/ad_x86_64.hpp" | |
70 #endif | |
71 #ifdef TARGET_ARCH_MODEL_sparc | |
72 # include "adfiles/ad_sparc.hpp" | |
73 #endif | |
74 #ifdef TARGET_ARCH_MODEL_zero | |
75 # include "adfiles/ad_zero.hpp" | |
76 #endif | |
0 | 77 |
2008 | 78 |
79 // -------------------- Compile::mach_constant_base_node ----------------------- | |
80 // Constant table base node singleton. | |
81 MachConstantBaseNode* Compile::mach_constant_base_node() { | |
82 if (_mach_constant_base_node == NULL) { | |
83 _mach_constant_base_node = new (C) MachConstantBaseNode(); | |
84 _mach_constant_base_node->add_req(C->root()); | |
85 } | |
86 return _mach_constant_base_node; | |
87 } | |
88 | |
89 | |
0 | 90 /// Support for intrinsics. |
91 | |
92 // Return the index at which m must be inserted (or already exists). | |
93 // The sort order is by the address of the ciMethod, with is_virtual as minor key. | |
94 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { | |
95 #ifdef ASSERT | |
96 for (int i = 1; i < _intrinsics->length(); i++) { | |
97 CallGenerator* cg1 = _intrinsics->at(i-1); | |
98 CallGenerator* cg2 = _intrinsics->at(i); | |
99 assert(cg1->method() != cg2->method() | |
100 ? cg1->method() < cg2->method() | |
101 : cg1->is_virtual() < cg2->is_virtual(), | |
102 "compiler intrinsics list must stay sorted"); | |
103 } | |
104 #endif | |
105 // Binary search sorted list, in decreasing intervals [lo, hi]. | |
106 int lo = 0, hi = _intrinsics->length()-1; | |
107 while (lo <= hi) { | |
108 int mid = (uint)(hi + lo) / 2; | |
109 ciMethod* mid_m = _intrinsics->at(mid)->method(); | |
110 if (m < mid_m) { | |
111 hi = mid-1; | |
112 } else if (m > mid_m) { | |
113 lo = mid+1; | |
114 } else { | |
115 // look at minor sort key | |
116 bool mid_virt = _intrinsics->at(mid)->is_virtual(); | |
117 if (is_virtual < mid_virt) { | |
118 hi = mid-1; | |
119 } else if (is_virtual > mid_virt) { | |
120 lo = mid+1; | |
121 } else { | |
122 return mid; // exact match | |
123 } | |
124 } | |
125 } | |
126 return lo; // inexact match | |
127 } | |
128 | |
129 void Compile::register_intrinsic(CallGenerator* cg) { | |
130 if (_intrinsics == NULL) { | |
131 _intrinsics = new GrowableArray<CallGenerator*>(60); | |
132 } | |
133 // This code is stolen from ciObjectFactory::insert. | |
134 // Really, GrowableArray should have methods for | |
135 // insert_at, remove_at, and binary_search. | |
136 int len = _intrinsics->length(); | |
137 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); | |
138 if (index == len) { | |
139 _intrinsics->append(cg); | |
140 } else { | |
141 #ifdef ASSERT | |
142 CallGenerator* oldcg = _intrinsics->at(index); | |
143 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); | |
144 #endif | |
145 _intrinsics->append(_intrinsics->at(len-1)); | |
146 int pos; | |
147 for (pos = len-2; pos >= index; pos--) { | |
148 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); | |
149 } | |
150 _intrinsics->at_put(index, cg); | |
151 } | |
152 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); | |
153 } | |
154 | |
155 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { | |
156 assert(m->is_loaded(), "don't try this on unloaded methods"); | |
157 if (_intrinsics != NULL) { | |
158 int index = intrinsic_insertion_index(m, is_virtual); | |
159 if (index < _intrinsics->length() | |
160 && _intrinsics->at(index)->method() == m | |
161 && _intrinsics->at(index)->is_virtual() == is_virtual) { | |
162 return _intrinsics->at(index); | |
163 } | |
164 } | |
165 // 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
|
166 if (m->intrinsic_id() != vmIntrinsics::_none && |
75596850f863
6862576: vmIntrinsics needs cleanup in order to support JSR 292 intrinsics
jrose
parents:
647
diff
changeset
|
167 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { |
0 | 168 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); |
169 if (cg != NULL) { | |
170 // Save it for next time: | |
171 register_intrinsic(cg); | |
172 return cg; | |
173 } else { | |
174 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); | |
175 } | |
176 } | |
177 return NULL; | |
178 } | |
179 | |
180 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined | |
181 // in library_call.cpp. | |
182 | |
183 | |
184 #ifndef PRODUCT | |
185 // statistics gathering... | |
186 | |
187 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; | |
188 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; | |
189 | |
190 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { | |
191 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); | |
192 int oflags = _intrinsic_hist_flags[id]; | |
193 assert(flags != 0, "what happened?"); | |
194 if (is_virtual) { | |
195 flags |= _intrinsic_virtual; | |
196 } | |
197 bool changed = (flags != oflags); | |
198 if ((flags & _intrinsic_worked) != 0) { | |
199 juint count = (_intrinsic_hist_count[id] += 1); | |
200 if (count == 1) { | |
201 changed = true; // first time | |
202 } | |
203 // increment the overall count also: | |
204 _intrinsic_hist_count[vmIntrinsics::_none] += 1; | |
205 } | |
206 if (changed) { | |
207 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { | |
208 // Something changed about the intrinsic's virtuality. | |
209 if ((flags & _intrinsic_virtual) != 0) { | |
210 // This is the first use of this intrinsic as a virtual call. | |
211 if (oflags != 0) { | |
212 // We already saw it as a non-virtual, so note both cases. | |
213 flags |= _intrinsic_both; | |
214 } | |
215 } else if ((oflags & _intrinsic_both) == 0) { | |
216 // This is the first use of this intrinsic as a non-virtual | |
217 flags |= _intrinsic_both; | |
218 } | |
219 } | |
220 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); | |
221 } | |
222 // update the overall flags also: | |
223 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; | |
224 return changed; | |
225 } | |
226 | |
227 static char* format_flags(int flags, char* buf) { | |
228 buf[0] = 0; | |
229 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); | |
230 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); | |
231 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); | |
232 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); | |
233 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); | |
234 if (buf[0] == 0) strcat(buf, ","); | |
235 assert(buf[0] == ',', "must be"); | |
236 return &buf[1]; | |
237 } | |
238 | |
239 void Compile::print_intrinsic_statistics() { | |
240 char flagsbuf[100]; | |
241 ttyLocker ttyl; | |
242 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); | |
243 tty->print_cr("Compiler intrinsic usage:"); | |
244 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; | |
245 if (total == 0) total = 1; // avoid div0 in case of no successes | |
246 #define PRINT_STAT_LINE(name, c, f) \ | |
247 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); | |
248 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { | |
249 vmIntrinsics::ID id = (vmIntrinsics::ID) index; | |
250 int flags = _intrinsic_hist_flags[id]; | |
251 juint count = _intrinsic_hist_count[id]; | |
252 if ((flags | count) != 0) { | |
253 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); | |
254 } | |
255 } | |
256 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); | |
257 if (xtty != NULL) xtty->tail("statistics"); | |
258 } | |
259 | |
260 void Compile::print_statistics() { | |
261 { ttyLocker ttyl; | |
262 if (xtty != NULL) xtty->head("statistics type='opto'"); | |
263 Parse::print_statistics(); | |
264 PhaseCCP::print_statistics(); | |
265 PhaseRegAlloc::print_statistics(); | |
266 Scheduling::print_statistics(); | |
267 PhasePeephole::print_statistics(); | |
268 PhaseIdealLoop::print_statistics(); | |
269 if (xtty != NULL) xtty->tail("statistics"); | |
270 } | |
271 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { | |
272 // put this under its own <statistics> element. | |
273 print_intrinsic_statistics(); | |
274 } | |
275 } | |
276 #endif //PRODUCT | |
277 | |
278 // Support for bundling info | |
279 Bundle* Compile::node_bundling(const Node *n) { | |
280 assert(valid_bundle_info(n), "oob"); | |
281 return &_node_bundling_base[n->_idx]; | |
282 } | |
283 | |
284 bool Compile::valid_bundle_info(const Node *n) { | |
285 return (_node_bundling_limit > n->_idx); | |
286 } | |
287 | |
288 | |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
289 void Compile::gvn_replace_by(Node* n, Node* nn) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
290 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
291 Node* use = n->last_out(i); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
292 bool is_in_table = initial_gvn()->hash_delete(use); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
293 uint uses_found = 0; |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
294 for (uint j = 0; j < use->len(); j++) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
295 if (use->in(j) == n) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
296 if (j < use->req()) |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
297 use->set_req(j, nn); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
298 else |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
299 use->set_prec(j, nn); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
300 uses_found++; |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
301 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
302 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
303 if (is_in_table) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
304 // reinsert into table |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
305 initial_gvn()->hash_find_insert(use); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
306 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
307 record_for_igvn(use); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
308 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
|
309 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
310 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
311 |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
312 |
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 |
0 | 315 // Identify all nodes that are reachable from below, useful. |
316 // Use breadth-first pass that records state in a Unique_Node_List, | |
317 // recursive traversal is slower. | |
318 void Compile::identify_useful_nodes(Unique_Node_List &useful) { | |
319 int estimated_worklist_size = unique(); | |
320 useful.map( estimated_worklist_size, NULL ); // preallocate space | |
321 | |
322 // Initialize worklist | |
323 if (root() != NULL) { useful.push(root()); } | |
324 // If 'top' is cached, declare it useful to preserve cached node | |
325 if( cached_top_node() ) { useful.push(cached_top_node()); } | |
326 | |
327 // Push all useful nodes onto the list, breadthfirst | |
328 for( uint next = 0; next < useful.size(); ++next ) { | |
329 assert( next < unique(), "Unique useful nodes < total nodes"); | |
330 Node *n = useful.at(next); | |
331 uint max = n->len(); | |
332 for( uint i = 0; i < max; ++i ) { | |
333 Node *m = n->in(i); | |
334 if( m == NULL ) continue; | |
335 useful.push(m); | |
336 } | |
337 } | |
338 } | |
339 | |
340 // Disconnect all useless nodes by disconnecting those at the boundary. | |
341 void Compile::remove_useless_nodes(Unique_Node_List &useful) { | |
342 uint next = 0; | |
343 while( next < useful.size() ) { | |
344 Node *n = useful.at(next++); | |
345 // Use raw traversal of out edges since this code removes out edges | |
346 int max = n->outcnt(); | |
347 for (int j = 0; j < max; ++j ) { | |
348 Node* child = n->raw_out(j); | |
349 if( ! useful.member(child) ) { | |
350 assert( !child->is_top() || child != top(), | |
351 "If top is cached in Compile object it is in useful list"); | |
352 // Only need to remove this out-edge to the useless node | |
353 n->raw_del_out(j); | |
354 --j; | |
355 --max; | |
356 } | |
357 } | |
358 if (n->outcnt() == 1 && n->has_special_unique_user()) { | |
359 record_for_igvn( n->unique_out() ); | |
360 } | |
361 } | |
362 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) | |
363 } | |
364 | |
365 //------------------------------frame_size_in_words----------------------------- | |
366 // frame_slots in units of words | |
367 int Compile::frame_size_in_words() const { | |
368 // shift is 0 in LP32 and 1 in LP64 | |
369 const int shift = (LogBytesPerWord - LogBytesPerInt); | |
370 int words = _frame_slots >> shift; | |
371 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); | |
372 return words; | |
373 } | |
374 | |
375 // ============================================================================ | |
376 //------------------------------CompileWrapper--------------------------------- | |
377 class CompileWrapper : public StackObj { | |
378 Compile *const _compile; | |
379 public: | |
380 CompileWrapper(Compile* compile); | |
381 | |
382 ~CompileWrapper(); | |
383 }; | |
384 | |
385 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { | |
386 // the Compile* pointer is stored in the current ciEnv: | |
387 ciEnv* env = compile->env(); | |
388 assert(env == ciEnv::current(), "must already be a ciEnv active"); | |
389 assert(env->compiler_data() == NULL, "compile already active?"); | |
390 env->set_compiler_data(compile); | |
391 assert(compile == Compile::current(), "sanity"); | |
392 | |
393 compile->set_type_dict(NULL); | |
394 compile->set_type_hwm(NULL); | |
395 compile->set_type_last_size(0); | |
396 compile->set_last_tf(NULL, NULL); | |
397 compile->set_indexSet_arena(NULL); | |
398 compile->set_indexSet_free_block_list(NULL); | |
399 compile->init_type_arena(); | |
400 Type::Initialize(compile); | |
401 _compile->set_scratch_buffer_blob(NULL); | |
402 _compile->begin_method(); | |
403 } | |
404 CompileWrapper::~CompileWrapper() { | |
405 _compile->end_method(); | |
406 if (_compile->scratch_buffer_blob() != NULL) | |
407 BufferBlob::free(_compile->scratch_buffer_blob()); | |
408 _compile->env()->set_compiler_data(NULL); | |
409 } | |
410 | |
411 | |
412 //----------------------------print_compile_messages--------------------------- | |
413 void Compile::print_compile_messages() { | |
414 #ifndef PRODUCT | |
415 // Check if recompiling | |
416 if (_subsume_loads == false && PrintOpto) { | |
417 // Recompiling without allowing machine instructions to subsume loads | |
418 tty->print_cr("*********************************************************"); | |
419 tty->print_cr("** Bailout: Recompile without subsuming loads **"); | |
420 tty->print_cr("*********************************************************"); | |
421 } | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
422 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
423 // Recompiling without escape analysis |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
424 tty->print_cr("*********************************************************"); |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
425 tty->print_cr("** Bailout: Recompile without escape analysis **"); |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
426 tty->print_cr("*********************************************************"); |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
427 } |
0 | 428 if (env()->break_at_compile()) { |
605 | 429 // Open the debugger when compiling this method. |
0 | 430 tty->print("### Breaking when compiling: "); |
431 method()->print_short_name(); | |
432 tty->cr(); | |
433 BREAKPOINT; | |
434 } | |
435 | |
436 if( PrintOpto ) { | |
437 if (is_osr_compilation()) { | |
438 tty->print("[OSR]%3d", _compile_id); | |
439 } else { | |
440 tty->print("%3d", _compile_id); | |
441 } | |
442 } | |
443 #endif | |
444 } | |
445 | |
446 | |
2091
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
447 //-----------------------init_scratch_buffer_blob------------------------------ |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
448 // Construct a temporary BufferBlob and cache it for this compile. |
2008 | 449 void Compile::init_scratch_buffer_blob(int const_size) { |
2091
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
450 // 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
|
451 // constant section is big enough, use it. Otherwise free the |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
452 // current and allocate a new one. |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
453 BufferBlob* blob = scratch_buffer_blob(); |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
454 if ((blob != NULL) && (const_size <= _scratch_const_size)) { |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
455 // Use the current blob. |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
456 } else { |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
457 if (blob != NULL) { |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
458 BufferBlob::free(blob); |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
459 } |
0 | 460 |
2091
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
461 ResourceMark rm; |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
462 _scratch_const_size = const_size; |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
463 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
|
464 blob = BufferBlob::create("Compile::scratch_buffer", size); |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
465 // Record the buffer blob for next time. |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
466 set_scratch_buffer_blob(blob); |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
467 // Have we run out of code space? |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
468 if (scratch_buffer_blob() == NULL) { |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
469 // Let CompilerBroker disable further compilations. |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
470 record_failure("Not enough space for scratch buffer in CodeCache"); |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
471 return; |
51bd2d261853
7008325: CodeCache exhausted on sparc starting from hs20b04
kvn
parents:
2008
diff
changeset
|
472 } |
163 | 473 } |
0 | 474 |
475 // Initialize the relocation buffers | |
1748 | 476 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; |
0 | 477 set_scratch_locs_memory(locs_buf); |
478 } | |
479 | |
480 | |
481 //-----------------------scratch_emit_size------------------------------------- | |
482 // Helper function that computes size by emitting code | |
483 uint Compile::scratch_emit_size(const Node* n) { | |
2008 | 484 // Start scratch_emit_size section. |
485 set_in_scratch_emit_size(true); | |
486 | |
0 | 487 // Emit into a trash buffer and count bytes emitted. |
488 // This is a pretty expensive way to compute a size, | |
489 // but it works well enough if seldom used. | |
490 // All common fixed-size instructions are given a size | |
491 // method by the AD file. | |
492 // Note that the scratch buffer blob and locs memory are | |
493 // allocated at the beginning of the compile task, and | |
494 // may be shared by several calls to scratch_emit_size. | |
495 // The allocation of the scratch buffer blob is particularly | |
496 // expensive, since it has to grab the code cache lock. | |
497 BufferBlob* blob = this->scratch_buffer_blob(); | |
498 assert(blob != NULL, "Initialize BufferBlob at start"); | |
499 assert(blob->size() > MAX_inst_size, "sanity"); | |
500 relocInfo* locs_buf = scratch_locs_memory(); | |
1748 | 501 address blob_begin = blob->content_begin(); |
0 | 502 address blob_end = (address)locs_buf; |
1748 | 503 assert(blob->content_contains(blob_end), "sanity"); |
0 | 504 CodeBuffer buf(blob_begin, blob_end - blob_begin); |
2008 | 505 buf.initialize_consts_size(_scratch_const_size); |
0 | 506 buf.initialize_stubs_size(MAX_stubs_size); |
507 assert(locs_buf != NULL, "sanity"); | |
2008 | 508 int lsize = MAX_locs_size / 3; |
509 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); | |
510 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); | |
511 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); | |
512 | |
513 // Do the emission. | |
0 | 514 n->emit(buf, this->regalloc()); |
2008 | 515 |
516 // End scratch_emit_size section. | |
517 set_in_scratch_emit_size(false); | |
518 | |
1748 | 519 return buf.insts_size(); |
0 | 520 } |
521 | |
522 | |
523 // ============================================================================ | |
524 //------------------------------Compile standard------------------------------- | |
525 debug_only( int Compile::_debug_idx = 100000; ) | |
526 | |
527 // Compile a method. entry_bci is -1 for normal compilations and indicates | |
528 // the continuation bci for on stack replacement. | |
529 | |
530 | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
531 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) |
0 | 532 : Phase(Compiler), |
533 _env(ci_env), | |
534 _log(ci_env->log()), | |
535 _compile_id(ci_env->compile_id()), | |
536 _save_argument_registers(false), | |
537 _stub_name(NULL), | |
538 _stub_function(NULL), | |
539 _stub_entry_point(NULL), | |
540 _method(target), | |
541 _entry_bci(osr_bci), | |
542 _initial_gvn(NULL), | |
543 _for_igvn(NULL), | |
544 _warm_calls(NULL), | |
545 _subsume_loads(subsume_loads), | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
546 _do_escape_analysis(do_escape_analysis), |
0 | 547 _failure_reason(NULL), |
548 _code_buffer("Compile::Fill_buffer"), | |
549 _orig_pc_slot(0), | |
550 _orig_pc_slot_offset_in_bytes(0), | |
1265 | 551 _has_method_handle_invokes(false), |
2008 | 552 _mach_constant_base_node(NULL), |
0 | 553 _node_bundling_limit(0), |
554 _node_bundling_base(NULL), | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
555 _java_calls(0), |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
556 _inner_loops(0), |
2008 | 557 _scratch_const_size(-1), |
558 _in_scratch_emit_size(false), | |
0 | 559 #ifndef PRODUCT |
560 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), | |
561 _printer(IdealGraphPrinter::printer()), | |
562 #endif | |
563 _congraph(NULL) { | |
564 C = this; | |
565 | |
566 CompileWrapper cw(this); | |
567 #ifndef PRODUCT | |
568 if (TimeCompiler2) { | |
569 tty->print(" "); | |
570 target->holder()->name()->print(); | |
571 tty->print("."); | |
572 target->print_short_name(); | |
573 tty->print(" "); | |
574 } | |
575 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); | |
576 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); | |
100
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
577 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
578 if (!print_opto_assembly) { |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
579 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
580 if (print_assembly && !Disassembler::can_decode()) { |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
581 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
582 print_opto_assembly = true; |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
583 } |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
584 } |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
585 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
|
586 set_parsed_irreducible_loop(false); |
0 | 587 #endif |
588 | |
589 if (ProfileTraps) { | |
590 // Make sure the method being compiled gets its own MDO, | |
591 // 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
|
592 method()->ensure_method_data(); |
0 | 593 } |
594 | |
595 Init(::AliasLevel); | |
596 | |
597 | |
598 print_compile_messages(); | |
599 | |
600 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) | |
601 _ilt = InlineTree::build_inline_tree_root(); | |
602 else | |
603 _ilt = NULL; | |
604 | |
605 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice | |
606 assert(num_alias_types() >= AliasIdxRaw, ""); | |
607 | |
608 #define MINIMUM_NODE_HASH 1023 | |
609 // Node list that Iterative GVN will start with | |
610 Unique_Node_List for_igvn(comp_arena()); | |
611 set_for_igvn(&for_igvn); | |
612 | |
613 // GVN that will be run immediately on new nodes | |
614 uint estimated_size = method()->code_size()*4+64; | |
615 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); | |
616 PhaseGVN gvn(node_arena(), estimated_size); | |
617 set_initial_gvn(&gvn); | |
618 | |
619 { // Scope for timing the parser | |
620 TracePhase t3("parse", &_t_parser, true); | |
621 | |
622 // Put top into the hash table ASAP. | |
623 initial_gvn()->transform_no_reclaim(top()); | |
624 | |
625 // Set up tf(), start(), and find a CallGenerator. | |
626 CallGenerator* cg; | |
627 if (is_osr_compilation()) { | |
628 const TypeTuple *domain = StartOSRNode::osr_domain(); | |
629 const TypeTuple *range = TypeTuple::make_range(method()->signature()); | |
630 init_tf(TypeFunc::make(domain, range)); | |
631 StartNode* s = new (this, 2) StartOSRNode(root(), domain); | |
632 initial_gvn()->set_type_bottom(s); | |
633 init_start(s); | |
634 cg = CallGenerator::for_osr(method(), entry_bci()); | |
635 } else { | |
636 // Normal case. | |
637 init_tf(TypeFunc::make(method())); | |
638 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); | |
639 initial_gvn()->set_type_bottom(s); | |
640 init_start(s); | |
641 float past_uses = method()->interpreter_invocation_count(); | |
642 float expected_uses = past_uses; | |
643 cg = CallGenerator::for_inline(method(), expected_uses); | |
644 } | |
645 if (failing()) return; | |
646 if (cg == NULL) { | |
647 record_method_not_compilable_all_tiers("cannot parse method"); | |
648 return; | |
649 } | |
650 JVMState* jvms = build_start_state(start(), tf()); | |
651 if ((jvms = cg->generate(jvms)) == NULL) { | |
652 record_method_not_compilable("method parse failed"); | |
653 return; | |
654 } | |
655 GraphKit kit(jvms); | |
656 | |
657 if (!kit.stopped()) { | |
658 // Accept return values, and transfer control we know not where. | |
659 // This is done by a special, unique ReturnNode bound to root. | |
660 return_values(kit.jvms()); | |
661 } | |
662 | |
663 if (kit.has_exceptions()) { | |
664 // Any exceptions that escape from this call must be rethrown | |
665 // to whatever caller is dynamically above us on the stack. | |
666 // This is done by a special, unique RethrowNode bound to root. | |
667 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); | |
668 } | |
669 | |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
670 if (!failing() && has_stringbuilder()) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
671 { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
672 // remove useless nodes to make the usage analysis simpler |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
673 ResourceMark rm; |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
674 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
675 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
676 |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
677 { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
678 ResourceMark rm; |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
679 print_method("Before StringOpts", 3); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
680 PhaseStringOpts pso(initial_gvn(), &for_igvn); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
681 print_method("After StringOpts", 3); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
682 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
683 |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
684 // now inline anything that we skipped the first time around |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
685 while (_late_inlines.length() > 0) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
686 CallGenerator* cg = _late_inlines.pop(); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
687 cg->do_late_inline(); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
688 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
689 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
690 assert(_late_inlines.length() == 0, "should have been processed"); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
691 |
417 | 692 print_method("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
|
693 |
0 | 694 // Remove clutter produced by parsing. |
695 if (!failing()) { | |
696 ResourceMark rm; | |
697 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); | |
698 } | |
699 } | |
700 | |
701 // Note: Large methods are capped off in do_one_bytecode(). | |
702 if (failing()) return; | |
703 | |
704 // After parsing, node notes are no longer automagic. | |
705 // They must be propagated by register_new_node_with_optimizer(), | |
706 // clone(), or the like. | |
707 set_default_node_notes(NULL); | |
708 | |
709 for (;;) { | |
710 int successes = Inline_Warm(); | |
711 if (failing()) return; | |
712 if (successes == 0) break; | |
713 } | |
714 | |
715 // Drain the list. | |
716 Finish_Warm(); | |
717 #ifndef PRODUCT | |
718 if (_printer) { | |
719 _printer->print_inlining(this); | |
720 } | |
721 #endif | |
722 | |
723 if (failing()) return; | |
724 NOT_PRODUCT( verify_graph_edges(); ) | |
725 | |
726 // Now optimize | |
727 Optimize(); | |
728 if (failing()) return; | |
729 NOT_PRODUCT( verify_graph_edges(); ) | |
730 | |
731 #ifndef PRODUCT | |
732 if (PrintIdeal) { | |
733 ttyLocker ttyl; // keep the following output all in one block | |
734 // This output goes directly to the tty, not the compiler log. | |
735 // To enable tools to match it up with the compilation activity, | |
736 // be sure to tag this tty output with the compile ID. | |
737 if (xtty != NULL) { | |
738 xtty->head("ideal compile_id='%d'%s", compile_id(), | |
739 is_osr_compilation() ? " compile_kind='osr'" : | |
740 ""); | |
741 } | |
742 root()->dump(9999); | |
743 if (xtty != NULL) { | |
744 xtty->tail("ideal"); | |
745 } | |
746 } | |
747 #endif | |
748 | |
749 // Now that we know the size of all the monitors we can add a fixed slot | |
750 // for the original deopt pc. | |
751 | |
752 _orig_pc_slot = fixed_slots(); | |
753 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); | |
754 set_fixed_slots(next_slot); | |
755 | |
756 // Now generate code | |
757 Code_Gen(); | |
758 if (failing()) return; | |
759 | |
760 // Check if we want to skip execution of all compiled code. | |
761 { | |
762 #ifndef PRODUCT | |
763 if (OptoNoExecute) { | |
764 record_method_not_compilable("+OptoNoExecute"); // Flag as failed | |
765 return; | |
766 } | |
767 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); | |
768 #endif | |
769 | |
770 if (is_osr_compilation()) { | |
771 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); | |
772 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); | |
773 } else { | |
774 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); | |
775 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); | |
776 } | |
777 | |
778 env()->register_method(_method, _entry_bci, | |
779 &_code_offsets, | |
780 _orig_pc_slot_offset_in_bytes, | |
781 code_buffer(), | |
782 frame_size_in_words(), _oop_map_set, | |
783 &_handler_table, &_inc_table, | |
784 compiler, | |
785 env()->comp_level(), | |
786 true, /*has_debug_info*/ | |
787 has_unsafe_access() | |
788 ); | |
789 } | |
790 } | |
791 | |
792 //------------------------------Compile---------------------------------------- | |
793 // Compile a runtime stub | |
794 Compile::Compile( ciEnv* ci_env, | |
795 TypeFunc_generator generator, | |
796 address stub_function, | |
797 const char *stub_name, | |
798 int is_fancy_jump, | |
799 bool pass_tls, | |
800 bool save_arg_registers, | |
801 bool return_pc ) | |
802 : Phase(Compiler), | |
803 _env(ci_env), | |
804 _log(ci_env->log()), | |
805 _compile_id(-1), | |
806 _save_argument_registers(save_arg_registers), | |
807 _method(NULL), | |
808 _stub_name(stub_name), | |
809 _stub_function(stub_function), | |
810 _stub_entry_point(NULL), | |
811 _entry_bci(InvocationEntryBci), | |
812 _initial_gvn(NULL), | |
813 _for_igvn(NULL), | |
814 _warm_calls(NULL), | |
815 _orig_pc_slot(0), | |
816 _orig_pc_slot_offset_in_bytes(0), | |
817 _subsume_loads(true), | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
818 _do_escape_analysis(false), |
0 | 819 _failure_reason(NULL), |
820 _code_buffer("Compile::Fill_buffer"), | |
1265 | 821 _has_method_handle_invokes(false), |
2008 | 822 _mach_constant_base_node(NULL), |
0 | 823 _node_bundling_limit(0), |
824 _node_bundling_base(NULL), | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
825 _java_calls(0), |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
826 _inner_loops(0), |
0 | 827 #ifndef PRODUCT |
828 _trace_opto_output(TraceOptoOutput), | |
829 _printer(NULL), | |
830 #endif | |
831 _congraph(NULL) { | |
832 C = this; | |
833 | |
834 #ifndef PRODUCT | |
835 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); | |
836 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); | |
837 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
|
838 set_parsed_irreducible_loop(false); |
0 | 839 #endif |
840 CompileWrapper cw(this); | |
841 Init(/*AliasLevel=*/ 0); | |
842 init_tf((*generator)()); | |
843 | |
844 { | |
845 // The following is a dummy for the sake of GraphKit::gen_stub | |
846 Unique_Node_List for_igvn(comp_arena()); | |
847 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this | |
848 PhaseGVN gvn(Thread::current()->resource_area(),255); | |
849 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively | |
850 gvn.transform_no_reclaim(top()); | |
851 | |
852 GraphKit kit; | |
853 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); | |
854 } | |
855 | |
856 NOT_PRODUCT( verify_graph_edges(); ) | |
857 Code_Gen(); | |
858 if (failing()) return; | |
859 | |
860 | |
861 // Entry point will be accessed using compile->stub_entry_point(); | |
862 if (code_buffer() == NULL) { | |
863 Matcher::soft_match_failure(); | |
864 } else { | |
865 if (PrintAssembly && (WizardMode || Verbose)) | |
866 tty->print_cr("### Stub::%s", stub_name); | |
867 | |
868 if (!failing()) { | |
869 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); | |
870 | |
871 // Make the NMethod | |
872 // For now we mark the frame as never safe for profile stackwalking | |
873 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, | |
874 code_buffer(), | |
875 CodeOffsets::frame_never_safe, | |
876 // _code_offsets.value(CodeOffsets::Frame_Complete), | |
877 frame_size_in_words(), | |
878 _oop_map_set, | |
879 save_arg_registers); | |
880 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); | |
881 | |
882 _stub_entry_point = rs->entry_point(); | |
883 } | |
884 } | |
885 } | |
886 | |
887 #ifndef PRODUCT | |
888 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { | |
889 if(PrintOpto && Verbose) { | |
890 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); | |
891 } | |
892 } | |
893 #endif | |
894 | |
895 void Compile::print_codes() { | |
896 } | |
897 | |
898 //------------------------------Init------------------------------------------- | |
899 // Prepare for a single compilation | |
900 void Compile::Init(int aliaslevel) { | |
901 _unique = 0; | |
902 _regalloc = NULL; | |
903 | |
904 _tf = NULL; // filled in later | |
905 _top = NULL; // cached later | |
906 _matcher = NULL; // filled in later | |
907 _cfg = NULL; // filled in later | |
908 | |
909 set_24_bit_selection_and_mode(Use24BitFP, false); | |
910 | |
911 _node_note_array = NULL; | |
912 _default_node_notes = NULL; | |
913 | |
914 _immutable_memory = NULL; // filled in at first inquiry | |
915 | |
916 // Globally visible Nodes | |
917 // First set TOP to NULL to give safe behavior during creation of RootNode | |
918 set_cached_top_node(NULL); | |
919 set_root(new (this, 3) RootNode()); | |
920 // Now that you have a Root to point to, create the real TOP | |
921 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); | |
922 set_recent_alloc(NULL, NULL); | |
923 | |
924 // Create Debug Information Recorder to record scopes, oopmaps, etc. | |
925 env()->set_oop_recorder(new OopRecorder(comp_arena())); | |
926 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); | |
927 env()->set_dependencies(new Dependencies(env())); | |
928 | |
929 _fixed_slots = 0; | |
930 set_has_split_ifs(false); | |
931 set_has_loops(has_method() && method()->has_loops()); // first approximation | |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
932 set_has_stringbuilder(false); |
0 | 933 _trap_can_recompile = false; // no traps emitted yet |
934 _major_progress = true; // start out assuming good things will happen | |
935 set_has_unsafe_access(false); | |
936 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); | |
937 set_decompile_count(0); | |
938 | |
418 | 939 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); |
1783 | 940 set_num_loop_opts(LoopOptsCount); |
941 set_do_inlining(Inline); | |
942 set_max_inline_size(MaxInlineSize); | |
943 set_freq_inline_size(FreqInlineSize); | |
944 set_do_scheduling(OptoScheduling); | |
945 set_do_count_invocations(false); | |
946 set_do_method_data_update(false); | |
0 | 947 |
948 if (debug_info()->recording_non_safepoints()) { | |
949 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> | |
950 (comp_arena(), 8, 0, NULL)); | |
951 set_default_node_notes(Node_Notes::make(this)); | |
952 } | |
953 | |
954 // // -- Initialize types before each compile -- | |
955 // // Update cached type information | |
956 // if( _method && _method->constants() ) | |
957 // Type::update_loaded_types(_method, _method->constants()); | |
958 | |
959 // Init alias_type map. | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
960 if (!_do_escape_analysis && aliaslevel == 3) |
0 | 961 aliaslevel = 2; // No unique types without escape analysis |
962 _AliasLevel = aliaslevel; | |
963 const int grow_ats = 16; | |
964 _max_alias_types = grow_ats; | |
965 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); | |
966 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); | |
967 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); | |
968 { | |
969 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; | |
970 } | |
971 // Initialize the first few types. | |
972 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); | |
973 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); | |
974 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); | |
975 _num_alias_types = AliasIdxRaw+1; | |
976 // Zero out the alias type cache. | |
977 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); | |
978 // A NULL adr_type hits in the cache right away. Preload the right answer. | |
979 probe_alias_cache(NULL)->_index = AliasIdxTop; | |
980 | |
981 _intrinsics = NULL; | |
1685 | 982 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); |
983 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); | |
0 | 984 register_library_intrinsics(); |
985 } | |
986 | |
987 //---------------------------init_start---------------------------------------- | |
988 // Install the StartNode on this compile object. | |
989 void Compile::init_start(StartNode* s) { | |
990 if (failing()) | |
991 return; // already failing | |
992 assert(s == start(), ""); | |
993 } | |
994 | |
995 StartNode* Compile::start() const { | |
996 assert(!failing(), ""); | |
997 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { | |
998 Node* start = root()->fast_out(i); | |
999 if( start->is_Start() ) | |
1000 return start->as_Start(); | |
1001 } | |
1002 ShouldNotReachHere(); | |
1003 return NULL; | |
1004 } | |
1005 | |
1006 //-------------------------------immutable_memory------------------------------------- | |
1007 // Access immutable memory | |
1008 Node* Compile::immutable_memory() { | |
1009 if (_immutable_memory != NULL) { | |
1010 return _immutable_memory; | |
1011 } | |
1012 StartNode* s = start(); | |
1013 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { | |
1014 Node *p = s->fast_out(i); | |
1015 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { | |
1016 _immutable_memory = p; | |
1017 return _immutable_memory; | |
1018 } | |
1019 } | |
1020 ShouldNotReachHere(); | |
1021 return NULL; | |
1022 } | |
1023 | |
1024 //----------------------set_cached_top_node------------------------------------ | |
1025 // Install the cached top node, and make sure Node::is_top works correctly. | |
1026 void Compile::set_cached_top_node(Node* tn) { | |
1027 if (tn != NULL) verify_top(tn); | |
1028 Node* old_top = _top; | |
1029 _top = tn; | |
1030 // Calling Node::setup_is_top allows the nodes the chance to adjust | |
1031 // their _out arrays. | |
1032 if (_top != NULL) _top->setup_is_top(); | |
1033 if (old_top != NULL) old_top->setup_is_top(); | |
1034 assert(_top == NULL || top()->is_top(), ""); | |
1035 } | |
1036 | |
1037 #ifndef PRODUCT | |
1038 void Compile::verify_top(Node* tn) const { | |
1039 if (tn != NULL) { | |
1040 assert(tn->is_Con(), "top node must be a constant"); | |
1041 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); | |
1042 assert(tn->in(0) != NULL, "must have live top node"); | |
1043 } | |
1044 } | |
1045 #endif | |
1046 | |
1047 | |
1048 ///-------------------Managing Per-Node Debug & Profile Info------------------- | |
1049 | |
1050 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { | |
1051 guarantee(arr != NULL, ""); | |
1052 int num_blocks = arr->length(); | |
1053 if (grow_by < num_blocks) grow_by = num_blocks; | |
1054 int num_notes = grow_by * _node_notes_block_size; | |
1055 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); | |
1056 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); | |
1057 while (num_notes > 0) { | |
1058 arr->append(notes); | |
1059 notes += _node_notes_block_size; | |
1060 num_notes -= _node_notes_block_size; | |
1061 } | |
1062 assert(num_notes == 0, "exact multiple, please"); | |
1063 } | |
1064 | |
1065 bool Compile::copy_node_notes_to(Node* dest, Node* source) { | |
1066 if (source == NULL || dest == NULL) return false; | |
1067 | |
1068 if (dest->is_Con()) | |
1069 return false; // Do not push debug info onto constants. | |
1070 | |
1071 #ifdef ASSERT | |
1072 // Leave a bread crumb trail pointing to the original node: | |
1073 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { | |
1074 dest->set_debug_orig(source); | |
1075 } | |
1076 #endif | |
1077 | |
1078 if (node_note_array() == NULL) | |
1079 return false; // Not collecting any notes now. | |
1080 | |
1081 // This is a copy onto a pre-existing node, which may already have notes. | |
1082 // If both nodes have notes, do not overwrite any pre-existing notes. | |
1083 Node_Notes* source_notes = node_notes_at(source->_idx); | |
1084 if (source_notes == NULL || source_notes->is_clear()) return false; | |
1085 Node_Notes* dest_notes = node_notes_at(dest->_idx); | |
1086 if (dest_notes == NULL || dest_notes->is_clear()) { | |
1087 return set_node_notes_at(dest->_idx, source_notes); | |
1088 } | |
1089 | |
1090 Node_Notes merged_notes = (*source_notes); | |
1091 // The order of operations here ensures that dest notes will win... | |
1092 merged_notes.update_from(dest_notes); | |
1093 return set_node_notes_at(dest->_idx, &merged_notes); | |
1094 } | |
1095 | |
1096 | |
1097 //--------------------------allow_range_check_smearing------------------------- | |
1098 // Gating condition for coalescing similar range checks. | |
1099 // Sometimes we try 'speculatively' replacing a series of a range checks by a | |
1100 // single covering check that is at least as strong as any of them. | |
1101 // If the optimization succeeds, the simplified (strengthened) range check | |
1102 // will always succeed. If it fails, we will deopt, and then give up | |
1103 // on the optimization. | |
1104 bool Compile::allow_range_check_smearing() const { | |
1105 // If this method has already thrown a range-check, | |
1106 // assume it was because we already tried range smearing | |
1107 // and it failed. | |
1108 uint already_trapped = trap_count(Deoptimization::Reason_range_check); | |
1109 return !already_trapped; | |
1110 } | |
1111 | |
1112 | |
1113 //------------------------------flatten_alias_type----------------------------- | |
1114 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { | |
1115 int offset = tj->offset(); | |
1116 TypePtr::PTR ptr = tj->ptr(); | |
1117 | |
247 | 1118 // Known instance (scalarizable allocation) alias only with itself. |
1119 bool is_known_inst = tj->isa_oopptr() != NULL && | |
1120 tj->is_oopptr()->is_known_instance(); | |
1121 | |
0 | 1122 // Process weird unsafe references. |
1123 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { | |
1124 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); | |
247 | 1125 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); |
0 | 1126 tj = TypeOopPtr::BOTTOM; |
1127 ptr = tj->ptr(); | |
1128 offset = tj->offset(); | |
1129 } | |
1130 | |
1131 // Array pointers need some flattening | |
1132 const TypeAryPtr *ta = tj->isa_aryptr(); | |
247 | 1133 if( ta && is_known_inst ) { |
1134 if ( offset != Type::OffsetBot && | |
1135 offset > arrayOopDesc::length_offset_in_bytes() ) { | |
1136 offset = Type::OffsetBot; // Flatten constant access into array body only | |
1137 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); | |
1138 } | |
1139 } else if( ta && _AliasLevel >= 2 ) { | |
0 | 1140 // For arrays indexed by constant indices, we flatten the alias |
1141 // space to include all of the array body. Only the header, klass | |
1142 // and array length can be accessed un-aliased. | |
1143 if( offset != Type::OffsetBot ) { | |
1144 if( ta->const_oop() ) { // methodDataOop or methodOop | |
1145 offset = Type::OffsetBot; // Flatten constant access into array body | |
247 | 1146 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); |
0 | 1147 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { |
1148 // range is OK as-is. | |
1149 tj = ta = TypeAryPtr::RANGE; | |
1150 } else if( offset == oopDesc::klass_offset_in_bytes() ) { | |
1151 tj = TypeInstPtr::KLASS; // all klass loads look alike | |
1152 ta = TypeAryPtr::RANGE; // generic ignored junk | |
1153 ptr = TypePtr::BotPTR; | |
1154 } else if( offset == oopDesc::mark_offset_in_bytes() ) { | |
1155 tj = TypeInstPtr::MARK; | |
1156 ta = TypeAryPtr::RANGE; // generic ignored junk | |
1157 ptr = TypePtr::BotPTR; | |
1158 } else { // Random constant offset into array body | |
1159 offset = Type::OffsetBot; // Flatten constant access into array body | |
247 | 1160 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); |
0 | 1161 } |
1162 } | |
1163 // Arrays of fixed size alias with arrays of unknown size. | |
1164 if (ta->size() != TypeInt::POS) { | |
1165 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); | |
247 | 1166 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); |
0 | 1167 } |
1168 // 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
|
1169 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
|
1170 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); |
247 | 1171 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
|
1172 } |
0 | 1173 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { |
1174 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); | |
247 | 1175 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); |
0 | 1176 } |
1177 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so | |
1178 // cannot be distinguished by bytecode alone. | |
1179 if (ta->elem() == TypeInt::BOOL) { | |
1180 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); | |
1181 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); | |
247 | 1182 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); |
0 | 1183 } |
1184 // During the 2nd round of IterGVN, NotNull castings are removed. | |
1185 // Make sure the Bottom and NotNull variants alias the same. | |
1186 // Also, make sure exact and non-exact variants alias the same. | |
1187 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { | |
1188 if (ta->const_oop()) { | |
1189 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); | |
1190 } else { | |
1191 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); | |
1192 } | |
1193 } | |
1194 } | |
1195 | |
1196 // Oop pointers need some flattening | |
1197 const TypeInstPtr *to = tj->isa_instptr(); | |
1198 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { | |
1199 if( ptr == TypePtr::Constant ) { | |
1200 // No constant oop pointers (such as Strings); they alias with | |
1201 // unknown strings. | |
247 | 1202 assert(!is_known_inst, "not scalarizable allocation"); |
0 | 1203 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); |
247 | 1204 } else if( is_known_inst ) { |
163 | 1205 tj = to; // Keep NotNull and klass_is_exact for instance type |
0 | 1206 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { |
1207 // During the 2nd round of IterGVN, NotNull castings are removed. | |
1208 // Make sure the Bottom and NotNull variants alias the same. | |
1209 // Also, make sure exact and non-exact variants alias the same. | |
247 | 1210 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); |
0 | 1211 } |
1212 // Canonicalize the holder of this field | |
1213 ciInstanceKlass *k = to->klass()->as_instance_klass(); | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1214 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { |
0 | 1215 // First handle header references such as a LoadKlassNode, even if the |
1216 // object's klass is unloaded at compile time (4965979). | |
247 | 1217 if (!is_known_inst) { // Do it only for non-instance types |
1218 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); | |
1219 } | |
0 | 1220 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { |
1221 to = NULL; | |
1222 tj = TypeOopPtr::BOTTOM; | |
1223 offset = tj->offset(); | |
1224 } else { | |
1225 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); | |
1226 if (!k->equals(canonical_holder) || tj->offset() != offset) { | |
247 | 1227 if( is_known_inst ) { |
1228 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); | |
1229 } else { | |
1230 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); | |
1231 } | |
0 | 1232 } |
1233 } | |
1234 } | |
1235 | |
1236 // Klass pointers to object array klasses need some flattening | |
1237 const TypeKlassPtr *tk = tj->isa_klassptr(); | |
1238 if( tk ) { | |
1239 // If we are referencing a field within a Klass, we need | |
1240 // to assume the worst case of an Object. Both exact and | |
1241 // inexact types must flatten to the same alias class. | |
1242 // Since the flattened result for a klass is defined to be | |
1243 // precisely java.lang.Object, use a constant ptr. | |
1244 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { | |
1245 | |
1246 tj = tk = TypeKlassPtr::make(TypePtr::Constant, | |
1247 TypeKlassPtr::OBJECT->klass(), | |
1248 offset); | |
1249 } | |
1250 | |
1251 ciKlass* klass = tk->klass(); | |
1252 if( klass->is_obj_array_klass() ) { | |
1253 ciKlass* k = TypeAryPtr::OOPS->klass(); | |
1254 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs | |
1255 k = TypeInstPtr::BOTTOM->klass(); | |
1256 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); | |
1257 } | |
1258 | |
1259 // Check for precise loads from the primary supertype array and force them | |
1260 // to the supertype cache alias index. Check for generic array loads from | |
1261 // the primary supertype array and also force them to the supertype cache | |
1262 // alias index. Since the same load can reach both, we need to merge | |
1263 // these 2 disparate memories into the same alias class. Since the | |
1264 // primary supertype array is read-only, there's no chance of confusion | |
1265 // where we bypass an array load and an array store. | |
1266 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); | |
1267 if( offset == Type::OffsetBot || | |
1268 off2 < Klass::primary_super_limit()*wordSize ) { | |
1269 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); | |
1270 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); | |
1271 } | |
1272 } | |
1273 | |
1274 // Flatten all Raw pointers together. | |
1275 if (tj->base() == Type::RawPtr) | |
1276 tj = TypeRawPtr::BOTTOM; | |
1277 | |
1278 if (tj->base() == Type::AnyPtr) | |
1279 tj = TypePtr::BOTTOM; // An error, which the caller must check for. | |
1280 | |
1281 // Flatten all to bottom for now | |
1282 switch( _AliasLevel ) { | |
1283 case 0: | |
1284 tj = TypePtr::BOTTOM; | |
1285 break; | |
1286 case 1: // Flatten to: oop, static, field or array | |
1287 switch (tj->base()) { | |
1288 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; | |
1289 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; | |
1290 case Type::AryPtr: // do not distinguish arrays at all | |
1291 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; | |
1292 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; | |
1293 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it | |
1294 default: ShouldNotReachHere(); | |
1295 } | |
1296 break; | |
605 | 1297 case 2: // No collapsing at level 2; keep all splits |
1298 case 3: // No collapsing at level 3; keep all splits | |
0 | 1299 break; |
1300 default: | |
1301 Unimplemented(); | |
1302 } | |
1303 | |
1304 offset = tj->offset(); | |
1305 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); | |
1306 | |
1307 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || | |
1308 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || | |
1309 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || | |
1310 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || | |
1311 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || | |
1312 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || | |
1313 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , | |
1314 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); | |
1315 assert( tj->ptr() != TypePtr::TopPTR && | |
1316 tj->ptr() != TypePtr::AnyNull && | |
1317 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); | |
1318 // assert( tj->ptr() != TypePtr::Constant || | |
1319 // tj->base() == Type::RawPtr || | |
1320 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); | |
1321 | |
1322 return tj; | |
1323 } | |
1324 | |
1325 void Compile::AliasType::Init(int i, const TypePtr* at) { | |
1326 _index = i; | |
1327 _adr_type = at; | |
1328 _field = NULL; | |
1329 _is_rewritable = true; // default | |
1330 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; | |
223 | 1331 if (atoop != NULL && atoop->is_known_instance()) { |
1332 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); | |
0 | 1333 _general_index = Compile::current()->get_alias_index(gt); |
1334 } else { | |
1335 _general_index = 0; | |
1336 } | |
1337 } | |
1338 | |
1339 //---------------------------------print_on------------------------------------ | |
1340 #ifndef PRODUCT | |
1341 void Compile::AliasType::print_on(outputStream* st) { | |
1342 if (index() < 10) | |
1343 st->print("@ <%d> ", index()); | |
1344 else st->print("@ <%d>", index()); | |
1345 st->print(is_rewritable() ? " " : " RO"); | |
1346 int offset = adr_type()->offset(); | |
1347 if (offset == Type::OffsetBot) | |
1348 st->print(" +any"); | |
1349 else st->print(" +%-3d", offset); | |
1350 st->print(" in "); | |
1351 adr_type()->dump_on(st); | |
1352 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); | |
1353 if (field() != NULL && tjp) { | |
1354 if (tjp->klass() != field()->holder() || | |
1355 tjp->offset() != field()->offset_in_bytes()) { | |
1356 st->print(" != "); | |
1357 field()->print(); | |
1358 st->print(" ***"); | |
1359 } | |
1360 } | |
1361 } | |
1362 | |
1363 void print_alias_types() { | |
1364 Compile* C = Compile::current(); | |
1365 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); | |
1366 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { | |
1367 C->alias_type(idx)->print_on(tty); | |
1368 tty->cr(); | |
1369 } | |
1370 } | |
1371 #endif | |
1372 | |
1373 | |
1374 //----------------------------probe_alias_cache-------------------------------- | |
1375 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { | |
1376 intptr_t key = (intptr_t) adr_type; | |
1377 key ^= key >> logAliasCacheSize; | |
1378 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; | |
1379 } | |
1380 | |
1381 | |
1382 //-----------------------------grow_alias_types-------------------------------- | |
1383 void Compile::grow_alias_types() { | |
1384 const int old_ats = _max_alias_types; // how many before? | |
1385 const int new_ats = old_ats; // how many more? | |
1386 const int grow_ats = old_ats+new_ats; // how many now? | |
1387 _max_alias_types = grow_ats; | |
1388 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); | |
1389 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); | |
1390 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); | |
1391 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; | |
1392 } | |
1393 | |
1394 | |
1395 //--------------------------------find_alias_type------------------------------ | |
1396 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { | |
1397 if (_AliasLevel == 0) | |
1398 return alias_type(AliasIdxBot); | |
1399 | |
1400 AliasCacheEntry* ace = probe_alias_cache(adr_type); | |
1401 if (ace->_adr_type == adr_type) { | |
1402 return alias_type(ace->_index); | |
1403 } | |
1404 | |
1405 // Handle special cases. | |
1406 if (adr_type == NULL) return alias_type(AliasIdxTop); | |
1407 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); | |
1408 | |
1409 // Do it the slow way. | |
1410 const TypePtr* flat = flatten_alias_type(adr_type); | |
1411 | |
1412 #ifdef ASSERT | |
1413 assert(flat == flatten_alias_type(flat), "idempotent"); | |
1414 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); | |
1415 if (flat->isa_oopptr() && !flat->isa_klassptr()) { | |
1416 const TypeOopPtr* foop = flat->is_oopptr(); | |
247 | 1417 // Scalarizable allocations have exact klass always. |
1418 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); | |
1419 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); | |
0 | 1420 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); |
1421 } | |
1422 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); | |
1423 #endif | |
1424 | |
1425 int idx = AliasIdxTop; | |
1426 for (int i = 0; i < num_alias_types(); i++) { | |
1427 if (alias_type(i)->adr_type() == flat) { | |
1428 idx = i; | |
1429 break; | |
1430 } | |
1431 } | |
1432 | |
1433 if (idx == AliasIdxTop) { | |
1434 if (no_create) return NULL; | |
1435 // Grow the array if necessary. | |
1436 if (_num_alias_types == _max_alias_types) grow_alias_types(); | |
1437 // Add a new alias type. | |
1438 idx = _num_alias_types++; | |
1439 _alias_types[idx]->Init(idx, flat); | |
1440 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); | |
1441 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); | |
1442 if (flat->isa_instptr()) { | |
1443 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() | |
1444 && flat->is_instptr()->klass() == env()->Class_klass()) | |
1445 alias_type(idx)->set_rewritable(false); | |
1446 } | |
1447 if (flat->isa_klassptr()) { | |
1448 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) | |
1449 alias_type(idx)->set_rewritable(false); | |
1450 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) | |
1451 alias_type(idx)->set_rewritable(false); | |
1452 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) | |
1453 alias_type(idx)->set_rewritable(false); | |
1454 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) | |
1455 alias_type(idx)->set_rewritable(false); | |
1456 } | |
1457 // %%% (We would like to finalize JavaThread::threadObj_offset(), | |
1458 // but the base pointer type is not distinctive enough to identify | |
1459 // references into JavaThread.) | |
1460 | |
1461 // Check for final instance fields. | |
1462 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
|
1463 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { |
0 | 1464 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); |
1465 ciField* field = k->get_field_by_offset(tinst->offset(), false); | |
1466 // Set field() and is_rewritable() attributes. | |
1467 if (field != NULL) alias_type(idx)->set_field(field); | |
1468 } | |
1469 const TypeKlassPtr* tklass = flat->isa_klassptr(); | |
1470 // Check for final static fields. | |
1471 if (tklass && tklass->klass()->is_instance_klass()) { | |
1472 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); | |
1473 ciField* field = k->get_field_by_offset(tklass->offset(), true); | |
1474 // Set field() and is_rewritable() attributes. | |
1475 if (field != NULL) alias_type(idx)->set_field(field); | |
1476 } | |
1477 } | |
1478 | |
1479 // Fill the cache for next time. | |
1480 ace->_adr_type = adr_type; | |
1481 ace->_index = idx; | |
1482 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); | |
1483 | |
1484 // Might as well try to fill the cache for the flattened version, too. | |
1485 AliasCacheEntry* face = probe_alias_cache(flat); | |
1486 if (face->_adr_type == NULL) { | |
1487 face->_adr_type = flat; | |
1488 face->_index = idx; | |
1489 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); | |
1490 } | |
1491 | |
1492 return alias_type(idx); | |
1493 } | |
1494 | |
1495 | |
1496 Compile::AliasType* Compile::alias_type(ciField* field) { | |
1497 const TypeOopPtr* t; | |
1498 if (field->is_static()) | |
1499 t = TypeKlassPtr::make(field->holder()); | |
1500 else | |
1501 t = TypeOopPtr::make_from_klass_raw(field->holder()); | |
1502 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); | |
1503 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); | |
1504 return atp; | |
1505 } | |
1506 | |
1507 | |
1508 //------------------------------have_alias_type-------------------------------- | |
1509 bool Compile::have_alias_type(const TypePtr* adr_type) { | |
1510 AliasCacheEntry* ace = probe_alias_cache(adr_type); | |
1511 if (ace->_adr_type == adr_type) { | |
1512 return true; | |
1513 } | |
1514 | |
1515 // Handle special cases. | |
1516 if (adr_type == NULL) return true; | |
1517 if (adr_type == TypePtr::BOTTOM) return true; | |
1518 | |
1519 return find_alias_type(adr_type, true) != NULL; | |
1520 } | |
1521 | |
1522 //-----------------------------must_alias-------------------------------------- | |
1523 // True if all values of the given address type are in the given alias category. | |
1524 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { | |
1525 if (alias_idx == AliasIdxBot) return true; // the universal category | |
1526 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP | |
1527 if (alias_idx == AliasIdxTop) return false; // the empty category | |
1528 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins | |
1529 | |
1530 // the only remaining possible overlap is identity | |
1531 int adr_idx = get_alias_index(adr_type); | |
1532 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); | |
1533 assert(adr_idx == alias_idx || | |
1534 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM | |
1535 && adr_type != TypeOopPtr::BOTTOM), | |
1536 "should not be testing for overlap with an unsafe pointer"); | |
1537 return adr_idx == alias_idx; | |
1538 } | |
1539 | |
1540 //------------------------------can_alias-------------------------------------- | |
1541 // True if any values of the given address type are in the given alias category. | |
1542 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { | |
1543 if (alias_idx == AliasIdxTop) return false; // the empty category | |
1544 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP | |
1545 if (alias_idx == AliasIdxBot) return true; // the universal category | |
1546 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins | |
1547 | |
1548 // the only remaining possible overlap is identity | |
1549 int adr_idx = get_alias_index(adr_type); | |
1550 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); | |
1551 return adr_idx == alias_idx; | |
1552 } | |
1553 | |
1554 | |
1555 | |
1556 //---------------------------pop_warm_call------------------------------------- | |
1557 WarmCallInfo* Compile::pop_warm_call() { | |
1558 WarmCallInfo* wci = _warm_calls; | |
1559 if (wci != NULL) _warm_calls = wci->remove_from(wci); | |
1560 return wci; | |
1561 } | |
1562 | |
1563 //----------------------------Inline_Warm-------------------------------------- | |
1564 int Compile::Inline_Warm() { | |
1565 // If there is room, try to inline some more warm call sites. | |
1566 // %%% Do a graph index compaction pass when we think we're out of space? | |
1567 if (!InlineWarmCalls) return 0; | |
1568 | |
1569 int calls_made_hot = 0; | |
1570 int room_to_grow = NodeCountInliningCutoff - unique(); | |
1571 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); | |
1572 int amount_grown = 0; | |
1573 WarmCallInfo* call; | |
1574 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { | |
1575 int est_size = (int)call->size(); | |
1576 if (est_size > (room_to_grow - amount_grown)) { | |
1577 // This one won't fit anyway. Get rid of it. | |
1578 call->make_cold(); | |
1579 continue; | |
1580 } | |
1581 call->make_hot(); | |
1582 calls_made_hot++; | |
1583 amount_grown += est_size; | |
1584 amount_to_grow -= est_size; | |
1585 } | |
1586 | |
1587 if (calls_made_hot > 0) set_major_progress(); | |
1588 return calls_made_hot; | |
1589 } | |
1590 | |
1591 | |
1592 //----------------------------Finish_Warm-------------------------------------- | |
1593 void Compile::Finish_Warm() { | |
1594 if (!InlineWarmCalls) return; | |
1595 if (failing()) return; | |
1596 if (warm_calls() == NULL) return; | |
1597 | |
1598 // Clean up loose ends, if we are out of space for inlining. | |
1599 WarmCallInfo* call; | |
1600 while ((call = pop_warm_call()) != NULL) { | |
1601 call->make_cold(); | |
1602 } | |
1603 } | |
1604 | |
1172 | 1605 //---------------------cleanup_loop_predicates----------------------- |
1606 // Remove the opaque nodes that protect the predicates so that all unused | |
1607 // checks and uncommon_traps will be eliminated from the ideal graph | |
1608 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { | |
1609 if (predicate_count()==0) return; | |
1610 for (int i = predicate_count(); i > 0; i--) { | |
1611 Node * n = predicate_opaque1_node(i-1); | |
1612 assert(n->Opcode() == Op_Opaque1, "must be"); | |
1613 igvn.replace_node(n, n->in(1)); | |
1614 } | |
1615 assert(predicate_count()==0, "should be clean!"); | |
1616 igvn.optimize(); | |
1617 } | |
0 | 1618 |
1619 //------------------------------Optimize--------------------------------------- | |
1620 // Given a graph, optimize it. | |
1621 void Compile::Optimize() { | |
1622 TracePhase t1("optimizer", &_t_optimizer, true); | |
1623 | |
1624 #ifndef PRODUCT | |
1625 if (env()->break_at_compile()) { | |
1626 BREAKPOINT; | |
1627 } | |
1628 | |
1629 #endif | |
1630 | |
1631 ResourceMark rm; | |
1632 int loop_opts_cnt; | |
1633 | |
1634 NOT_PRODUCT( verify_graph_edges(); ) | |
1635 | |
222 | 1636 print_method("After Parsing"); |
0 | 1637 |
1638 { | |
1639 // Iterative Global Value Numbering, including ideal transforms | |
1640 // Initialize IterGVN with types and values from parse-time GVN | |
1641 PhaseIterGVN igvn(initial_gvn()); | |
1642 { | |
1643 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) | |
1644 igvn.optimize(); | |
1645 } | |
1646 | |
1647 print_method("Iter GVN 1", 2); | |
1648 | |
1649 if (failing()) return; | |
1650 | |
1634
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
1651 // Perform escape analysis |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
1652 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
1653 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true); |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
1654 ConnectionGraph::do_analysis(this, &igvn); |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
1655 |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
1656 if (failing()) return; |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
1657 |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
1658 igvn.optimize(); |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
1659 print_method("Iter GVN 3", 2); |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
1660 |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
1661 if (failing()) return; |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
1662 |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
1663 } |
60a14ad85270
6966411: escape.cpp:450 assert(base->Opcode() == Op_ConP
kvn
parents:
1609
diff
changeset
|
1664 |
0 | 1665 // Loop transforms on the ideal graph. Range Check Elimination, |
1666 // peeling, unrolling, etc. | |
1667 | |
1668 // Set loop opts counter | |
1669 loop_opts_cnt = num_loop_opts(); | |
1670 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { | |
1671 { | |
1672 TracePhase t2("idealLoop", &_t_idealLoop, true); | |
1172 | 1673 PhaseIdealLoop ideal_loop( igvn, true, UseLoopPredicate); |
0 | 1674 loop_opts_cnt--; |
1675 if (major_progress()) print_method("PhaseIdealLoop 1", 2); | |
1676 if (failing()) return; | |
1677 } | |
1678 // Loop opts pass if partial peeling occurred in previous pass | |
1679 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { | |
1680 TracePhase t3("idealLoop", &_t_idealLoop, true); | |
1172 | 1681 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate); |
0 | 1682 loop_opts_cnt--; |
1683 if (major_progress()) print_method("PhaseIdealLoop 2", 2); | |
1684 if (failing()) return; | |
1685 } | |
1686 // Loop opts pass for loop-unrolling before CCP | |
1687 if(major_progress() && (loop_opts_cnt > 0)) { | |
1688 TracePhase t4("idealLoop", &_t_idealLoop, true); | |
1172 | 1689 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate); |
0 | 1690 loop_opts_cnt--; |
1691 if (major_progress()) print_method("PhaseIdealLoop 3", 2); | |
1692 } | |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
1693 if (!failing()) { |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
1694 // 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
|
1695 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
1696 PhaseIdealLoop::verify(igvn); |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
1697 } |
0 | 1698 } |
1699 if (failing()) return; | |
1700 | |
1701 // Conditional Constant Propagation; | |
1702 PhaseCCP ccp( &igvn ); | |
1703 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); | |
1704 { | |
1705 TracePhase t2("ccp", &_t_ccp, true); | |
1706 ccp.do_transform(); | |
1707 } | |
1708 print_method("PhaseCPP 1", 2); | |
1709 | |
1710 assert( true, "Break here to ccp.dump_old2new_map()"); | |
1711 | |
1712 // Iterative Global Value Numbering, including ideal transforms | |
1713 { | |
1714 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) | |
1715 igvn = ccp; | |
1716 igvn.optimize(); | |
1717 } | |
1718 | |
1719 print_method("Iter GVN 2", 2); | |
1720 | |
1721 if (failing()) return; | |
1722 | |
1723 // Loop transforms on the ideal graph. Range Check Elimination, | |
1724 // peeling, unrolling, etc. | |
1725 if(loop_opts_cnt > 0) { | |
1726 debug_only( int cnt = 0; ); | |
1172 | 1727 bool loop_predication = UseLoopPredicate; |
0 | 1728 while(major_progress() && (loop_opts_cnt > 0)) { |
1729 TracePhase t2("idealLoop", &_t_idealLoop, true); | |
1730 assert( cnt++ < 40, "infinite cycle in loop optimization" ); | |
1172 | 1731 PhaseIdealLoop ideal_loop( igvn, true, loop_predication); |
0 | 1732 loop_opts_cnt--; |
1733 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); | |
1734 if (failing()) return; | |
1172 | 1735 // Perform loop predication optimization during first iteration after CCP. |
1736 // After that switch it off and cleanup unused loop predicates. | |
1737 if (loop_predication) { | |
1738 loop_predication = false; | |
1739 cleanup_loop_predicates(igvn); | |
1740 if (failing()) return; | |
1741 } | |
0 | 1742 } |
1743 } | |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
1744 |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
1745 { |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
1746 // Verify that all previous optimizations produced a valid graph |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
1747 // 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
|
1748 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
1749 PhaseIdealLoop::verify(igvn); |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
1750 } |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
859
diff
changeset
|
1751 |
0 | 1752 { |
1753 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) | |
1754 PhaseMacroExpand mex(igvn); | |
1755 if (mex.expand_macro_nodes()) { | |
1756 assert(failing(), "must bail out w/ explicit message"); | |
1757 return; | |
1758 } | |
1759 } | |
1760 | |
1761 } // (End scope of igvn; run destructor if necessary for asserts.) | |
1762 | |
1763 // A method with only infinite loops has no edges entering loops from root | |
1764 { | |
1765 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) | |
1766 if (final_graph_reshaping()) { | |
1767 assert(failing(), "must bail out w/ explicit message"); | |
1768 return; | |
1769 } | |
1770 } | |
1771 | |
1772 print_method("Optimize finished", 2); | |
1773 } | |
1774 | |
1775 | |
1776 //------------------------------Code_Gen--------------------------------------- | |
1777 // Given a graph, generate code for it | |
1778 void Compile::Code_Gen() { | |
1779 if (failing()) return; | |
1780 | |
1781 // Perform instruction selection. You might think we could reclaim Matcher | |
1782 // memory PDQ, but actually the Matcher is used in generating spill code. | |
1783 // Internals of the Matcher (including some VectorSets) must remain live | |
1784 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage | |
1785 // set a bit in reclaimed memory. | |
1786 | |
1787 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine | |
1788 // nodes. Mapping is only valid at the root of each matched subtree. | |
1789 NOT_PRODUCT( verify_graph_edges(); ) | |
1790 | |
1791 Node_List proj_list; | |
1792 Matcher m(proj_list); | |
1793 _matcher = &m; | |
1794 { | |
1795 TracePhase t2("matcher", &_t_matcher, true); | |
1796 m.match(); | |
1797 } | |
1798 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine | |
1799 // nodes. Mapping is only valid at the root of each matched subtree. | |
1800 NOT_PRODUCT( verify_graph_edges(); ) | |
1801 | |
1802 // If you have too many nodes, or if matching has failed, bail out | |
1803 check_node_count(0, "out of nodes matching instructions"); | |
1804 if (failing()) return; | |
1805 | |
1806 // Build a proper-looking CFG | |
1807 PhaseCFG cfg(node_arena(), root(), m); | |
1808 _cfg = &cfg; | |
1809 { | |
1810 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) | |
1811 cfg.Dominators(); | |
1812 if (failing()) return; | |
1813 | |
1814 NOT_PRODUCT( verify_graph_edges(); ) | |
1815 | |
1816 cfg.Estimate_Block_Frequency(); | |
1817 cfg.GlobalCodeMotion(m,unique(),proj_list); | |
1818 | |
1819 print_method("Global code motion", 2); | |
1820 | |
1821 if (failing()) return; | |
1822 NOT_PRODUCT( verify_graph_edges(); ) | |
1823 | |
1824 debug_only( cfg.verify(); ) | |
1825 } | |
1826 NOT_PRODUCT( verify_graph_edges(); ) | |
1827 | |
1828 PhaseChaitin regalloc(unique(),cfg,m); | |
1829 _regalloc = ®alloc; | |
1830 { | |
1831 TracePhase t2("regalloc", &_t_registerAllocation, true); | |
1832 // Perform any platform dependent preallocation actions. This is used, | |
1833 // for example, to avoid taking an implicit null pointer exception | |
1834 // using the frame pointer on win95. | |
1835 _regalloc->pd_preallocate_hook(); | |
1836 | |
1837 // Perform register allocation. After Chaitin, use-def chains are | |
1838 // no longer accurate (at spill code) and so must be ignored. | |
1839 // Node->LRG->reg mappings are still accurate. | |
1840 _regalloc->Register_Allocate(); | |
1841 | |
1842 // Bail out if the allocator builds too many nodes | |
1843 if (failing()) return; | |
1844 } | |
1845 | |
1846 // Prior to register allocation we kept empty basic blocks in case the | |
1847 // the allocator needed a place to spill. After register allocation we | |
1848 // are not adding any new instructions. If any basic block is empty, we | |
1849 // can now safely remove it. | |
1850 { | |
418 | 1851 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) |
1852 cfg.remove_empty(); | |
1853 if (do_freq_based_layout()) { | |
1854 PhaseBlockLayout layout(cfg); | |
1855 } else { | |
1856 cfg.set_loop_alignment(); | |
1857 } | |
1858 cfg.fixup_flow(); | |
0 | 1859 } |
1860 | |
1861 // Perform any platform dependent postallocation verifications. | |
1862 debug_only( _regalloc->pd_postallocate_verify_hook(); ) | |
1863 | |
1864 // Apply peephole optimizations | |
1865 if( OptoPeephole ) { | |
1866 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) | |
1867 PhasePeephole peep( _regalloc, cfg); | |
1868 peep.do_transform(); | |
1869 } | |
1870 | |
1871 // Convert Nodes to instruction bits in a buffer | |
1872 { | |
1873 // %%%% workspace merge brought two timers together for one job | |
1874 TracePhase t2a("output", &_t_output, true); | |
1875 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) | |
1876 Output(); | |
1877 } | |
1878 | |
222 | 1879 print_method("Final Code"); |
0 | 1880 |
1881 // He's dead, Jim. | |
1882 _cfg = (PhaseCFG*)0xdeadbeef; | |
1883 _regalloc = (PhaseChaitin*)0xdeadbeef; | |
1884 } | |
1885 | |
1886 | |
1887 //------------------------------dump_asm--------------------------------------- | |
1888 // Dump formatted assembly | |
1889 #ifndef PRODUCT | |
1890 void Compile::dump_asm(int *pcs, uint pc_limit) { | |
1891 bool cut_short = false; | |
1892 tty->print_cr("#"); | |
1893 tty->print("# "); _tf->dump(); tty->cr(); | |
1894 tty->print_cr("#"); | |
1895 | |
1896 // For all blocks | |
1897 int pc = 0x0; // Program counter | |
1898 char starts_bundle = ' '; | |
1899 _regalloc->dump_frame(); | |
1900 | |
1901 Node *n = NULL; | |
1902 for( uint i=0; i<_cfg->_num_blocks; i++ ) { | |
1903 if (VMThread::should_terminate()) { cut_short = true; break; } | |
1904 Block *b = _cfg->_blocks[i]; | |
1905 if (b->is_connector() && !Verbose) continue; | |
1906 n = b->_nodes[0]; | |
1907 if (pcs && n->_idx < pc_limit) | |
1908 tty->print("%3.3x ", pcs[n->_idx]); | |
1909 else | |
1910 tty->print(" "); | |
1911 b->dump_head( &_cfg->_bbs ); | |
1912 if (b->is_connector()) { | |
1913 tty->print_cr(" # Empty connector block"); | |
1914 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { | |
1915 tty->print_cr(" # Block is sole successor of call"); | |
1916 } | |
1917 | |
1918 // For all instructions | |
1919 Node *delay = NULL; | |
1920 for( uint j = 0; j<b->_nodes.size(); j++ ) { | |
1921 if (VMThread::should_terminate()) { cut_short = true; break; } | |
1922 n = b->_nodes[j]; | |
1923 if (valid_bundle_info(n)) { | |
1924 Bundle *bundle = node_bundling(n); | |
1925 if (bundle->used_in_unconditional_delay()) { | |
1926 delay = n; | |
1927 continue; | |
1928 } | |
1929 if (bundle->starts_bundle()) | |
1930 starts_bundle = '+'; | |
1931 } | |
1932 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1933 if (WizardMode) n->dump(); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1934 |
0 | 1935 if( !n->is_Region() && // Dont print in the Assembly |
1936 !n->is_Phi() && // a few noisely useless nodes | |
1937 !n->is_Proj() && | |
1938 !n->is_MachTemp() && | |
1100
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
1939 !n->is_SafePointScalarObject() && |
0 | 1940 !n->is_Catch() && // Would be nice to print exception table targets |
1941 !n->is_MergeMem() && // Not very interesting | |
1942 !n->is_top() && // Debug info table constants | |
1943 !(n->is_Con() && !n->is_Mach())// Debug info table constants | |
1944 ) { | |
1945 if (pcs && n->_idx < pc_limit) | |
1946 tty->print("%3.3x", pcs[n->_idx]); | |
1947 else | |
1948 tty->print(" "); | |
1949 tty->print(" %c ", starts_bundle); | |
1950 starts_bundle = ' '; | |
1951 tty->print("\t"); | |
1952 n->format(_regalloc, tty); | |
1953 tty->cr(); | |
1954 } | |
1955 | |
1956 // If we have an instruction with a delay slot, and have seen a delay, | |
1957 // then back up and print it | |
1958 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { | |
1959 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
|
1960 if (WizardMode) delay->dump(); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1961 |
0 | 1962 if (node_bundling(delay)->starts_bundle()) |
1963 starts_bundle = '+'; | |
1964 if (pcs && n->_idx < pc_limit) | |
1965 tty->print("%3.3x", pcs[n->_idx]); | |
1966 else | |
1967 tty->print(" "); | |
1968 tty->print(" %c ", starts_bundle); | |
1969 starts_bundle = ' '; | |
1970 tty->print("\t"); | |
1971 delay->format(_regalloc, tty); | |
1972 tty->print_cr(""); | |
1973 delay = NULL; | |
1974 } | |
1975 | |
1976 // Dump the exception table as well | |
1977 if( n->is_Catch() && (Verbose || WizardMode) ) { | |
1978 // Print the exception table for this offset | |
1979 _handler_table.print_subtable_for(pc); | |
1980 } | |
1981 } | |
1982 | |
1983 if (pcs && n->_idx < pc_limit) | |
1984 tty->print_cr("%3.3x", pcs[n->_idx]); | |
1985 else | |
1986 tty->print_cr(""); | |
1987 | |
1988 assert(cut_short || delay == NULL, "no unconditional delay branch"); | |
1989 | |
1990 } // End of per-block dump | |
1991 tty->print_cr(""); | |
1992 | |
1993 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); | |
1994 } | |
1995 #endif | |
1996 | |
1997 //------------------------------Final_Reshape_Counts--------------------------- | |
1998 // This class defines counters to help identify when a method | |
1999 // may/must be executed using hardware with only 24-bit precision. | |
2000 struct Final_Reshape_Counts : public StackObj { | |
2001 int _call_count; // count non-inlined 'common' calls | |
2002 int _float_count; // count float ops requiring 24-bit precision | |
2003 int _double_count; // count double ops requiring more precision | |
2004 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
|
2005 int _inner_loop_count; // count loops which need alignment |
0 | 2006 VectorSet _visited; // Visitation flags |
2007 Node_List _tests; // Set of IfNodes & PCTableNodes | |
2008 | |
2009 Final_Reshape_Counts() : | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2010 _call_count(0), _float_count(0), _double_count(0), |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2011 _java_call_count(0), _inner_loop_count(0), |
0 | 2012 _visited( Thread::current()->resource_area() ) { } |
2013 | |
2014 void inc_call_count () { _call_count ++; } | |
2015 void inc_float_count () { _float_count ++; } | |
2016 void inc_double_count() { _double_count++; } | |
2017 void inc_java_call_count() { _java_call_count++; } | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2018 void inc_inner_loop_count() { _inner_loop_count++; } |
0 | 2019 |
2020 int get_call_count () const { return _call_count ; } | |
2021 int get_float_count () const { return _float_count ; } | |
2022 int get_double_count() const { return _double_count; } | |
2023 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
|
2024 int get_inner_loop_count() const { return _inner_loop_count; } |
0 | 2025 }; |
2026 | |
2027 static bool oop_offset_is_sane(const TypeInstPtr* tp) { | |
2028 ciInstanceKlass *k = tp->klass()->as_instance_klass(); | |
2029 // 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
|
2030 return k->contains_field_offset(tp->offset()); |
0 | 2031 // Note that OffsetBot and OffsetTop are very negative. |
2032 } | |
2033 | |
2034 //------------------------------final_graph_reshaping_impl---------------------- | |
2035 // Implement items 1-5 from final_graph_reshaping below. | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2036 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) { |
0 | 2037 |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2038 if ( n->outcnt() == 0 ) return; // dead node |
0 | 2039 uint nop = n->Opcode(); |
2040 | |
2041 // Check for 2-input instruction with "last use" on right input. | |
2042 // Swap to left input. Implements item (2). | |
2043 if( n->req() == 3 && // two-input instruction | |
2044 n->in(1)->outcnt() > 1 && // left use is NOT a last use | |
2045 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop | |
2046 n->in(2)->outcnt() == 1 &&// right use IS a last use | |
2047 !n->in(2)->is_Con() ) { // right use is not a constant | |
2048 // Check for commutative opcode | |
2049 switch( nop ) { | |
2050 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: | |
2051 case Op_MaxI: case Op_MinI: | |
2052 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: | |
2053 case Op_AndL: case Op_XorL: case Op_OrL: | |
2054 case Op_AndI: case Op_XorI: case Op_OrI: { | |
2055 // Move "last use" input to left by swapping inputs | |
2056 n->swap_edges(1, 2); | |
2057 break; | |
2058 } | |
2059 default: | |
2060 break; | |
2061 } | |
2062 } | |
2063 | |
1609 | 2064 #ifdef ASSERT |
2065 if( n->is_Mem() ) { | |
2066 Compile* C = Compile::current(); | |
2067 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type()); | |
2068 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || | |
2069 // oop will be recorded in oop map if load crosses safepoint | |
2070 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || | |
2071 LoadNode::is_immutable_value(n->in(MemNode::Address))), | |
2072 "raw memory operations should have control edge"); | |
2073 } | |
2074 #endif | |
0 | 2075 // Count FPU ops and common calls, implements item (3) |
2076 switch( nop ) { | |
2077 // Count all float operations that may use FPU | |
2078 case Op_AddF: | |
2079 case Op_SubF: | |
2080 case Op_MulF: | |
2081 case Op_DivF: | |
2082 case Op_NegF: | |
2083 case Op_ModF: | |
2084 case Op_ConvI2F: | |
2085 case Op_ConF: | |
2086 case Op_CmpF: | |
2087 case Op_CmpF3: | |
2088 // 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
|
2089 frc.inc_float_count(); |
0 | 2090 break; |
2091 | |
2092 case Op_ConvF2D: | |
2093 case Op_ConvD2F: | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2094 frc.inc_float_count(); |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2095 frc.inc_double_count(); |
0 | 2096 break; |
2097 | |
2098 // Count all double operations that may use FPU | |
2099 case Op_AddD: | |
2100 case Op_SubD: | |
2101 case Op_MulD: | |
2102 case Op_DivD: | |
2103 case Op_NegD: | |
2104 case Op_ModD: | |
2105 case Op_ConvI2D: | |
2106 case Op_ConvD2I: | |
2107 // case Op_ConvL2D: // handled by leaf call | |
2108 // case Op_ConvD2L: // handled by leaf call | |
2109 case Op_ConD: | |
2110 case Op_CmpD: | |
2111 case Op_CmpD3: | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2112 frc.inc_double_count(); |
0 | 2113 break; |
2114 case Op_Opaque1: // Remove Opaque Nodes before matching | |
2115 case Op_Opaque2: // Remove Opaque Nodes before matching | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2116 n->subsume_by(n->in(1)); |
0 | 2117 break; |
2118 case Op_CallStaticJava: | |
2119 case Op_CallJava: | |
2120 case Op_CallDynamicJava: | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2121 frc.inc_java_call_count(); // Count java call site; |
0 | 2122 case Op_CallRuntime: |
2123 case Op_CallLeaf: | |
2124 case Op_CallLeafNoFP: { | |
2125 assert( n->is_Call(), "" ); | |
2126 CallNode *call = n->as_Call(); | |
2127 // Count call sites where the FP mode bit would have to be flipped. | |
2128 // Do not count uncommon runtime calls: | |
2129 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, | |
2130 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... | |
2131 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2132 frc.inc_call_count(); // Count the call site |
0 | 2133 } else { // See if uncommon argument is shared |
2134 Node *n = call->in(TypeFunc::Parms); | |
2135 int nop = n->Opcode(); | |
2136 // Clone shared simple arguments to uncommon calls, item (1). | |
2137 if( n->outcnt() > 1 && | |
2138 !n->is_Proj() && | |
2139 nop != Op_CreateEx && | |
2140 nop != Op_CheckCastPP && | |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2141 nop != Op_DecodeN && |
0 | 2142 !n->is_Mem() ) { |
2143 Node *x = n->clone(); | |
2144 call->set_req( TypeFunc::Parms, x ); | |
2145 } | |
2146 } | |
2147 break; | |
2148 } | |
2149 | |
2150 case Op_StoreD: | |
2151 case Op_LoadD: | |
2152 case Op_LoadD_unaligned: | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2153 frc.inc_double_count(); |
0 | 2154 goto handle_mem; |
2155 case Op_StoreF: | |
2156 case Op_LoadF: | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2157 frc.inc_float_count(); |
0 | 2158 goto handle_mem; |
2159 | |
2160 case Op_StoreB: | |
2161 case Op_StoreC: | |
2162 case Op_StoreCM: | |
2163 case Op_StorePConditional: | |
2164 case Op_StoreI: | |
2165 case Op_StoreL: | |
420
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
418
diff
changeset
|
2166 case Op_StoreIConditional: |
0 | 2167 case Op_StoreLConditional: |
2168 case Op_CompareAndSwapI: | |
2169 case Op_CompareAndSwapL: | |
2170 case Op_CompareAndSwapP: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
2171 case Op_CompareAndSwapN: |
0 | 2172 case Op_StoreP: |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
2173 case Op_StoreN: |
0 | 2174 case Op_LoadB: |
624 | 2175 case Op_LoadUB: |
558
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
492
diff
changeset
|
2176 case Op_LoadUS: |
0 | 2177 case Op_LoadI: |
624 | 2178 case Op_LoadUI2L: |
0 | 2179 case Op_LoadKlass: |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2180 case Op_LoadNKlass: |
0 | 2181 case Op_LoadL: |
2182 case Op_LoadL_unaligned: | |
2183 case Op_LoadPLocked: | |
2184 case Op_LoadLLocked: | |
2185 case Op_LoadP: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
2186 case Op_LoadN: |
0 | 2187 case Op_LoadRange: |
2188 case Op_LoadS: { | |
2189 handle_mem: | |
2190 #ifdef ASSERT | |
2191 if( VerifyOptoOopOffsets ) { | |
2192 assert( n->is_Mem(), "" ); | |
2193 MemNode *mem = (MemNode*)n; | |
2194 // Check to see if address types have grounded out somehow. | |
2195 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); | |
2196 assert( !tp || oop_offset_is_sane(tp), "" ); | |
2197 } | |
2198 #endif | |
2199 break; | |
2200 } | |
2201 | |
2202 case Op_AddP: { // Assert sane base pointers | |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2203 Node *addp = n->in(AddPNode::Address); |
0 | 2204 assert( !addp->is_AddP() || |
2205 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation | |
2206 addp->in(AddPNode::Base) == n->in(AddPNode::Base), | |
2207 "Base pointers must match" ); | |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2208 #ifdef _LP64 |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2209 if (UseCompressedOops && |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2210 addp->Opcode() == Op_ConP && |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2211 addp == n->in(AddPNode::Base) && |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2212 n->in(AddPNode::Offset)->is_Con()) { |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2213 // 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
|
2214 // 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
|
2215 // instructions (4) then load 64-bits constant (7). |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2216 // 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
|
2217 const Type* t = addp->bottom_type(); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2218 if (t->isa_oopptr()) { |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2219 Node* nn = NULL; |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2220 |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2221 // Look for existing ConN node of the same exact type. |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2222 Compile* C = Compile::current(); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2223 Node* r = C->root(); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2224 uint cnt = r->outcnt(); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2225 for (uint i = 0; i < cnt; i++) { |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2226 Node* m = r->raw_out(i); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2227 if (m!= NULL && m->Opcode() == Op_ConN && |
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
182
diff
changeset
|
2228 m->bottom_type()->make_ptr() == t) { |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2229 nn = m; |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2230 break; |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2231 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2232 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2233 if (nn != NULL) { |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2234 // Decode a narrow oop to match address |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2235 // [R12 + narrow_oop_reg<<3 + offset] |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2236 nn = new (C, 2) DecodeNNode(nn, t); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2237 n->set_req(AddPNode::Base, nn); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2238 n->set_req(AddPNode::Address, nn); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2239 if (addp->outcnt() == 0) { |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2240 addp->disconnect_inputs(NULL); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2241 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2242 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2243 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2244 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2245 #endif |
0 | 2246 break; |
2247 } | |
2248 | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2249 #ifdef _LP64 |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2250 case Op_CastPP: |
1575
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2251 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
|
2252 Compile* C = Compile::current(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2253 Node* in1 = n->in(1); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2254 const Type* t = n->bottom_type(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2255 Node* new_in1 = in1->clone(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2256 new_in1->as_DecodeN()->set_type(t); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2257 |
1575
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2258 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
|
2259 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2260 // 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
|
2261 // 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
|
2262 // 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
|
2263 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2264 // [R12 + narrow_oop_reg<<3 + offset] |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2265 // NullCheck narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2266 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2267 // 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
|
2268 // 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
|
2269 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2270 // 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
|
2271 // [base_reg + offset] |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2272 // NullCheck base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2273 // |
605 | 2274 // 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
|
2275 // 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
|
2276 // 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
|
2277 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2278 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
|
2279 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2280 |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2281 n->subsume_by(new_in1); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2282 if (in1->outcnt() == 0) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2283 in1->disconnect_inputs(NULL); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2284 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2285 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2286 break; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2287 |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2288 case Op_CmpP: |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2289 // 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
|
2290 // other TypePtr related Ideal optimizations (for example, ptr nullness). |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2291 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2292 Node* in1 = n->in(1); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2293 Node* in2 = n->in(2); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2294 if (!in1->is_DecodeN()) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2295 in2 = in1; |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2296 in1 = n->in(2); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2297 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2298 assert(in1->is_DecodeN(), "sanity"); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2299 |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2300 Compile* C = Compile::current(); |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2301 Node* new_in2 = NULL; |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2302 if (in2->is_DecodeN()) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2303 new_in2 = in2->in(1); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2304 } else if (in2->Opcode() == Op_ConP) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2305 const Type* t = in2->bottom_type(); |
1575
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2306 if (t == TypePtr::NULL_PTR) { |
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2307 // 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
|
2308 // oops implicit null check is not generated. |
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2309 // 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
|
2310 if (Matcher::gen_narrow_oop_implicit_null_checks()) |
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2311 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2312 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2313 // This transformation together with CastPP transformation above |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2314 // 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
|
2315 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2316 // The original code after Optimize() |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2317 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2318 // LoadN memory, narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2319 // decode narrow_oop_reg, base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2320 // CmpP base_reg, NULL |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2321 // CastPP base_reg // NotNull |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2322 // Load [base_reg + offset], val_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2323 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2324 // after these transformations will be |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2325 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2326 // LoadN memory, narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2327 // CmpN narrow_oop_reg, NULL |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2328 // 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
|
2329 // Load [base_reg + offset], val_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2330 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2331 // 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
|
2332 // 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
|
2333 // final_graph_reshaping_walk()). |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2334 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2335 // 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
|
2336 // on x86: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2337 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2338 // Load_narrow_oop memory, narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2339 // 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
|
2340 // NullCheck narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2341 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2342 // and on sparc: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2343 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2344 // Load_narrow_oop memory, narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2345 // 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
|
2346 // Load [base_reg + offset], val_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2347 // NullCheck base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2348 // |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2349 } else if (t->isa_oopptr()) { |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2350 new_in2 = ConNode::make(C, t->make_narrowoop()); |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2351 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2352 } |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2353 if (new_in2 != NULL) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2354 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2355 n->subsume_by( cmpN ); |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2356 if (in1->outcnt() == 0) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2357 in1->disconnect_inputs(NULL); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2358 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2359 if (in2->outcnt() == 0) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2360 in2->disconnect_inputs(NULL); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2361 } |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2362 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2363 } |
293
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
253
diff
changeset
|
2364 break; |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2365 |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2366 case Op_DecodeN: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2367 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); |
1575
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2368 // 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
|
2369 // an address expression, see the code for Op_CastPP above. |
1575
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2370 assert(n->in(0) == NULL || !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
|
2371 break; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2372 |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2373 case Op_EncodeP: { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2374 Node* in1 = n->in(1); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2375 if (in1->is_DecodeN()) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2376 n->subsume_by(in1->in(1)); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2377 } else if (in1->Opcode() == Op_ConP) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2378 Compile* C = Compile::current(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2379 const Type* t = in1->bottom_type(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2380 if (t == TypePtr::NULL_PTR) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2381 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2382 } else if (t->isa_oopptr()) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2383 n->subsume_by(ConNode::make(C, t->make_narrowoop())); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2384 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2385 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2386 if (in1->outcnt() == 0) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2387 in1->disconnect_inputs(NULL); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2388 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2389 break; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2390 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2391 |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2392 case Op_Proj: { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2393 if (OptimizeStringConcat) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2394 ProjNode* p = n->as_Proj(); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2395 if (p->_is_io_use) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2396 // Separate projections were used for the exception path which |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2397 // 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
|
2398 // then they will hang around and should just be replaced with |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2399 // the original one. |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2400 Node* proj = NULL; |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2401 // Replace with just one |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2402 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2403 Node *use = i.get(); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2404 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
|
2405 proj = use; |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2406 break; |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2407 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2408 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2409 assert(p != NULL, "must be found"); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2410 p->subsume_by(proj); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2411 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2412 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2413 break; |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2414 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
929
diff
changeset
|
2415 |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2416 case Op_Phi: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2417 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2418 // 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
|
2419 // 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
|
2420 Node* unique_in = n->in(1); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2421 assert(unique_in != NULL, ""); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2422 uint cnt = n->req(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2423 for (uint i = 2; i < cnt; i++) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2424 Node* m = n->in(i); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2425 assert(m != NULL, ""); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2426 if (unique_in != m) |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2427 unique_in = NULL; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2428 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2429 if (unique_in != NULL) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2430 n->subsume_by(unique_in); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2431 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2432 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2433 break; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2434 |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2435 #endif |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2436 |
0 | 2437 case Op_ModI: |
2438 if (UseDivMod) { | |
2439 // Check if a%b and a/b both exist | |
2440 Node* d = n->find_similar(Op_DivI); | |
2441 if (d) { | |
2442 // Replace them with a fused divmod if supported | |
2443 Compile* C = Compile::current(); | |
2444 if (Matcher::has_match_rule(Op_DivModI)) { | |
2445 DivModINode* divmod = DivModINode::make(C, n); | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2446 d->subsume_by(divmod->div_proj()); |
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2447 n->subsume_by(divmod->mod_proj()); |
0 | 2448 } else { |
2449 // replace a%b with a-((a/b)*b) | |
2450 Node* mult = new (C, 3) MulINode(d, d->in(2)); | |
2451 Node* sub = new (C, 3) SubINode(d->in(1), mult); | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2452 n->subsume_by( sub ); |
0 | 2453 } |
2454 } | |
2455 } | |
2456 break; | |
2457 | |
2458 case Op_ModL: | |
2459 if (UseDivMod) { | |
2460 // Check if a%b and a/b both exist | |
2461 Node* d = n->find_similar(Op_DivL); | |
2462 if (d) { | |
2463 // Replace them with a fused divmod if supported | |
2464 Compile* C = Compile::current(); | |
2465 if (Matcher::has_match_rule(Op_DivModL)) { | |
2466 DivModLNode* divmod = DivModLNode::make(C, n); | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2467 d->subsume_by(divmod->div_proj()); |
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2468 n->subsume_by(divmod->mod_proj()); |
0 | 2469 } else { |
2470 // replace a%b with a-((a/b)*b) | |
2471 Node* mult = new (C, 3) MulLNode(d, d->in(2)); | |
2472 Node* sub = new (C, 3) SubLNode(d->in(1), mult); | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2473 n->subsume_by( sub ); |
0 | 2474 } |
2475 } | |
2476 } | |
2477 break; | |
2478 | |
2479 case Op_Load16B: | |
2480 case Op_Load8B: | |
2481 case Op_Load4B: | |
2482 case Op_Load8S: | |
2483 case Op_Load4S: | |
2484 case Op_Load2S: | |
2485 case Op_Load8C: | |
2486 case Op_Load4C: | |
2487 case Op_Load2C: | |
2488 case Op_Load4I: | |
2489 case Op_Load2I: | |
2490 case Op_Load2L: | |
2491 case Op_Load4F: | |
2492 case Op_Load2F: | |
2493 case Op_Load2D: | |
2494 case Op_Store16B: | |
2495 case Op_Store8B: | |
2496 case Op_Store4B: | |
2497 case Op_Store8C: | |
2498 case Op_Store4C: | |
2499 case Op_Store2C: | |
2500 case Op_Store4I: | |
2501 case Op_Store2I: | |
2502 case Op_Store2L: | |
2503 case Op_Store4F: | |
2504 case Op_Store2F: | |
2505 case Op_Store2D: | |
2506 break; | |
2507 | |
2508 case Op_PackB: | |
2509 case Op_PackS: | |
2510 case Op_PackC: | |
2511 case Op_PackI: | |
2512 case Op_PackF: | |
2513 case Op_PackL: | |
2514 case Op_PackD: | |
2515 if (n->req()-1 > 2) { | |
2516 // Replace many operand PackNodes with a binary tree for matching | |
2517 PackNode* p = (PackNode*) n; | |
2518 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2519 n->subsume_by(btp); |
0 | 2520 } |
2521 break; | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2522 case Op_Loop: |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2523 case Op_CountedLoop: |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2524 if (n->as_Loop()->is_inner_loop()) { |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2525 frc.inc_inner_loop_count(); |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2526 } |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2527 break; |
0 | 2528 default: |
2529 assert( !n->is_Call(), "" ); | |
2530 assert( !n->is_Mem(), "" ); | |
2531 break; | |
2532 } | |
127 | 2533 |
2534 // Collect CFG split points | |
2535 if (n->is_MultiBranch()) | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2536 frc._tests.push(n); |
0 | 2537 } |
2538 | |
2539 //------------------------------final_graph_reshaping_walk--------------------- | |
2540 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), | |
2541 // requires that the walk visits a node's inputs before visiting the node. | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2542 static void 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
|
2543 ResourceArea *area = Thread::current()->resource_area(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2544 Unique_Node_List sfpt(area); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2545 |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2546 frc._visited.set(root->_idx); // first, mark node as visited |
0 | 2547 uint cnt = root->req(); |
2548 Node *n = root; | |
2549 uint i = 0; | |
2550 while (true) { | |
2551 if (i < cnt) { | |
2552 // Place all non-visited non-null inputs onto stack | |
2553 Node* m = n->in(i); | |
2554 ++i; | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2555 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
|
2556 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2557 sfpt.push(m); |
0 | 2558 cnt = m->req(); |
2559 nstack.push(n, i); // put on stack parent and next input's index | |
2560 n = m; | |
2561 i = 0; | |
2562 } | |
2563 } else { | |
2564 // Now do post-visit work | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2565 final_graph_reshaping_impl( n, frc ); |
0 | 2566 if (nstack.is_empty()) |
2567 break; // finished | |
2568 n = nstack.node(); // Get node from stack | |
2569 cnt = n->req(); | |
2570 i = nstack.index(); | |
2571 nstack.pop(); // Shift to the next node on stack | |
2572 } | |
2573 } | |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2574 |
1575
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2575 // Skip next transformation if compressed oops are not used. |
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2576 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks()) |
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2577 return; |
3657cb01ffc5
6954029: Improve implicit null check generation with compressed oops
kvn
parents:
1397
diff
changeset
|
2578 |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2579 // Go over safepoints nodes to skip DecodeN nodes for debug edges. |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2580 // It could be done for an uncommon traps or any safepoints/calls |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2581 // if the DecodeN node is referenced only in a debug info. |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2582 while (sfpt.size() > 0) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2583 n = sfpt.pop(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2584 JVMState *jvms = n->as_SafePoint()->jvms(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2585 assert(jvms != NULL, "sanity"); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2586 int start = jvms->debug_start(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2587 int end = n->req(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2588 bool is_uncommon = (n->is_CallStaticJava() && |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2589 n->as_CallStaticJava()->uncommon_trap_request() != 0); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2590 for (int j = start; j < end; j++) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2591 Node* in = n->in(j); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2592 if (in->is_DecodeN()) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2593 bool safe_to_skip = true; |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2594 if (!is_uncommon ) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2595 // Is it safe to skip? |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2596 for (uint i = 0; i < in->outcnt(); i++) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2597 Node* u = in->raw_out(i); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2598 if (!u->is_SafePoint() || |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2599 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
|
2600 safe_to_skip = false; |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2601 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2602 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2603 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2604 if (safe_to_skip) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2605 n->set_req(j, in->in(1)); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2606 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2607 if (in->outcnt() == 0) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2608 in->disconnect_inputs(NULL); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2609 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2610 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2611 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2612 } |
0 | 2613 } |
2614 | |
2615 //------------------------------final_graph_reshaping-------------------------- | |
2616 // Final Graph Reshaping. | |
2617 // | |
2618 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late | |
2619 // and not commoned up and forced early. Must come after regular | |
2620 // optimizations to avoid GVN undoing the cloning. Clone constant | |
2621 // inputs to Loop Phis; these will be split by the allocator anyways. | |
2622 // Remove Opaque nodes. | |
2623 // (2) Move last-uses by commutative operations to the left input to encourage | |
2624 // Intel update-in-place two-address operations and better register usage | |
2625 // on RISCs. Must come after regular optimizations to avoid GVN Ideal | |
2626 // calls canonicalizing them back. | |
2627 // (3) Count the number of double-precision FP ops, single-precision FP ops | |
2628 // and call sites. On Intel, we can get correct rounding either by | |
2629 // forcing singles to memory (requires extra stores and loads after each | |
2630 // FP bytecode) or we can set a rounding mode bit (requires setting and | |
2631 // clearing the mode bit around call sites). The mode bit is only used | |
2632 // if the relative frequency of single FP ops to calls is low enough. | |
2633 // This is a key transform for SPEC mpeg_audio. | |
2634 // (4) Detect infinite loops; blobs of code reachable from above but not | |
2635 // below. Several of the Code_Gen algorithms fail on such code shapes, | |
2636 // so we simply bail out. Happens a lot in ZKM.jar, but also happens | |
2637 // from time to time in other codes (such as -Xcomp finalizer loops, etc). | |
2638 // Detection is by looking for IfNodes where only 1 projection is | |
2639 // reachable from below or CatchNodes missing some targets. | |
2640 // (5) Assert for insane oop offsets in debug mode. | |
2641 | |
2642 bool Compile::final_graph_reshaping() { | |
2643 // an infinite loop may have been eliminated by the optimizer, | |
2644 // in which case the graph will be empty. | |
2645 if (root()->req() == 1) { | |
2646 record_method_not_compilable("trivial infinite loop"); | |
2647 return true; | |
2648 } | |
2649 | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2650 Final_Reshape_Counts frc; |
0 | 2651 |
2652 // Visit everybody reachable! | |
2653 // Allocate stack of size C->unique()/2 to avoid frequent realloc | |
2654 Node_Stack nstack(unique() >> 1); | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2655 final_graph_reshaping_walk(nstack, root(), frc); |
0 | 2656 |
2657 // 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
|
2658 for( uint i = 0; i < frc._tests.size(); i++ ) { |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2659 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); |
127 | 2660 // Get number of CFG targets. |
0 | 2661 // Note that PCTables include exception targets after calls. |
127 | 2662 uint required_outcnt = n->required_outcnt(); |
2663 if (n->outcnt() != required_outcnt) { | |
0 | 2664 // Check for a few special cases. Rethrow Nodes never take the |
2665 // 'fall-thru' path, so expected kids is 1 less. | |
2666 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { | |
2667 if (n->in(0)->in(0)->is_Call()) { | |
2668 CallNode *call = n->in(0)->in(0)->as_Call(); | |
2669 if (call->entry_point() == OptoRuntime::rethrow_stub()) { | |
127 | 2670 required_outcnt--; // Rethrow always has 1 less kid |
0 | 2671 } else if (call->req() > TypeFunc::Parms && |
2672 call->is_CallDynamicJava()) { | |
2673 // Check for null receiver. In such case, the optimizer has | |
2674 // detected that the virtual call will always result in a null | |
2675 // pointer exception. The fall-through projection of this CatchNode | |
2676 // will not be populated. | |
2677 Node *arg0 = call->in(TypeFunc::Parms); | |
2678 if (arg0->is_Type() && | |
2679 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { | |
127 | 2680 required_outcnt--; |
0 | 2681 } |
2682 } else if (call->entry_point() == OptoRuntime::new_array_Java() && | |
2683 call->req() > TypeFunc::Parms+1 && | |
2684 call->is_CallStaticJava()) { | |
2685 // Check for negative array length. In such case, the optimizer has | |
2686 // detected that the allocation attempt will always result in an | |
2687 // exception. There is no fall-through projection of this CatchNode . | |
2688 Node *arg1 = call->in(TypeFunc::Parms+1); | |
2689 if (arg1->is_Type() && | |
2690 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { | |
127 | 2691 required_outcnt--; |
0 | 2692 } |
2693 } | |
2694 } | |
2695 } | |
127 | 2696 // Recheck with a better notion of 'required_outcnt' |
2697 if (n->outcnt() != required_outcnt) { | |
0 | 2698 record_method_not_compilable("malformed control flow"); |
2699 return true; // Not all targets reachable! | |
2700 } | |
2701 } | |
2702 // Check that I actually visited all kids. Unreached kids | |
2703 // must be infinite loops. | |
2704 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
|
2705 if (!frc._visited.test(n->fast_out(j)->_idx)) { |
0 | 2706 record_method_not_compilable("infinite loop"); |
2707 return true; // Found unvisited kid; must be unreach | |
2708 } | |
2709 } | |
2710 | |
2711 // If original bytecodes contained a mixture of floats and doubles | |
2712 // 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
|
2713 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2714 frc.get_float_count() > 32 && |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2715 frc.get_double_count() == 0 && |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2716 (10 * frc.get_call_count() < frc.get_float_count()) ) { |
0 | 2717 set_24_bit_selection_and_mode( false, true ); |
2718 } | |
2719 | |
859
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2720 set_java_calls(frc.get_java_call_count()); |
ea3f9723b5cf
6860599: nodes limit could be reached during Output phase
kvn
parents:
856
diff
changeset
|
2721 set_inner_loops(frc.get_inner_loop_count()); |
0 | 2722 |
2723 // No infinite loops, no reason to bail out. | |
2724 return false; | |
2725 } | |
2726 | |
2727 //-----------------------------too_many_traps---------------------------------- | |
2728 // Report if there are too many traps at the current method and bci. | |
2729 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. | |
2730 bool Compile::too_many_traps(ciMethod* method, | |
2731 int bci, | |
2732 Deoptimization::DeoptReason reason) { | |
2733 ciMethodData* md = method->method_data(); | |
2734 if (md->is_empty()) { | |
2735 // Assume the trap has not occurred, or that it occurred only | |
2736 // because of a transient condition during start-up in the interpreter. | |
2737 return false; | |
2738 } | |
2739 if (md->has_trap_at(bci, reason) != 0) { | |
2740 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. | |
2741 // Also, if there are multiple reasons, or if there is no per-BCI record, | |
2742 // assume the worst. | |
2743 if (log()) | |
2744 log()->elem("observe trap='%s' count='%d'", | |
2745 Deoptimization::trap_reason_name(reason), | |
2746 md->trap_count(reason)); | |
2747 return true; | |
2748 } else { | |
2749 // Ignore method/bci and see if there have been too many globally. | |
2750 return too_many_traps(reason, md); | |
2751 } | |
2752 } | |
2753 | |
2754 // Less-accurate variant which does not require a method and bci. | |
2755 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, | |
2756 ciMethodData* logmd) { | |
2757 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { | |
2758 // Too many traps globally. | |
2759 // Note that we use cumulative trap_count, not just md->trap_count. | |
2760 if (log()) { | |
2761 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); | |
2762 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", | |
2763 Deoptimization::trap_reason_name(reason), | |
2764 mcount, trap_count(reason)); | |
2765 } | |
2766 return true; | |
2767 } else { | |
2768 // The coast is clear. | |
2769 return false; | |
2770 } | |
2771 } | |
2772 | |
2773 //--------------------------too_many_recompiles-------------------------------- | |
2774 // Report if there are too many recompiles at the current method and bci. | |
2775 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. | |
2776 // Is not eager to return true, since this will cause the compiler to use | |
2777 // Action_none for a trap point, to avoid too many recompilations. | |
2778 bool Compile::too_many_recompiles(ciMethod* method, | |
2779 int bci, | |
2780 Deoptimization::DeoptReason reason) { | |
2781 ciMethodData* md = method->method_data(); | |
2782 if (md->is_empty()) { | |
2783 // Assume the trap has not occurred, or that it occurred only | |
2784 // because of a transient condition during start-up in the interpreter. | |
2785 return false; | |
2786 } | |
2787 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. | |
2788 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; | |
2789 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero | |
2790 Deoptimization::DeoptReason per_bc_reason | |
2791 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); | |
2792 if ((per_bc_reason == Deoptimization::Reason_none | |
2793 || md->has_trap_at(bci, reason) != 0) | |
2794 // The trap frequency measure we care about is the recompile count: | |
2795 && md->trap_recompiled_at(bci) | |
2796 && md->overflow_recompile_count() >= bc_cutoff) { | |
2797 // Do not emit a trap here if it has already caused recompilations. | |
2798 // Also, if there are multiple reasons, or if there is no per-BCI record, | |
2799 // assume the worst. | |
2800 if (log()) | |
2801 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", | |
2802 Deoptimization::trap_reason_name(reason), | |
2803 md->trap_count(reason), | |
2804 md->overflow_recompile_count()); | |
2805 return true; | |
2806 } else if (trap_count(reason) != 0 | |
2807 && decompile_count() >= m_cutoff) { | |
2808 // Too many recompiles globally, and we have seen this sort of trap. | |
2809 // Use cumulative decompile_count, not just md->decompile_count. | |
2810 if (log()) | |
2811 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", | |
2812 Deoptimization::trap_reason_name(reason), | |
2813 md->trap_count(reason), trap_count(reason), | |
2814 md->decompile_count(), decompile_count()); | |
2815 return true; | |
2816 } else { | |
2817 // The coast is clear. | |
2818 return false; | |
2819 } | |
2820 } | |
2821 | |
2822 | |
2823 #ifndef PRODUCT | |
2824 //------------------------------verify_graph_edges--------------------------- | |
2825 // Walk the Graph and verify that there is a one-to-one correspondence | |
2826 // between Use-Def edges and Def-Use edges in the graph. | |
2827 void Compile::verify_graph_edges(bool no_dead_code) { | |
2828 if (VerifyGraphEdges) { | |
2829 ResourceArea *area = Thread::current()->resource_area(); | |
2830 Unique_Node_List visited(area); | |
2831 // Call recursive graph walk to check edges | |
2832 _root->verify_edges(visited); | |
2833 if (no_dead_code) { | |
2834 // Now make sure that no visited node is used by an unvisited node. | |
2835 bool dead_nodes = 0; | |
2836 Unique_Node_List checked(area); | |
2837 while (visited.size() > 0) { | |
2838 Node* n = visited.pop(); | |
2839 checked.push(n); | |
2840 for (uint i = 0; i < n->outcnt(); i++) { | |
2841 Node* use = n->raw_out(i); | |
2842 if (checked.member(use)) continue; // already checked | |
2843 if (visited.member(use)) continue; // already in the graph | |
2844 if (use->is_Con()) continue; // a dead ConNode is OK | |
2845 // At this point, we have found a dead node which is DU-reachable. | |
2846 if (dead_nodes++ == 0) | |
2847 tty->print_cr("*** Dead nodes reachable via DU edges:"); | |
2848 use->dump(2); | |
2849 tty->print_cr("---"); | |
2850 checked.push(use); // No repeats; pretend it is now checked. | |
2851 } | |
2852 } | |
2853 assert(dead_nodes == 0, "using nodes must be reachable from root"); | |
2854 } | |
2855 } | |
2856 } | |
2857 #endif | |
2858 | |
2859 // The Compile object keeps track of failure reasons separately from the ciEnv. | |
2860 // This is required because there is not quite a 1-1 relation between the | |
2861 // ciEnv and its compilation task and the Compile object. Note that one | |
2862 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides | |
2863 // to backtrack and retry without subsuming loads. Other than this backtracking | |
2864 // behavior, the Compile's failure reason is quietly copied up to the ciEnv | |
2865 // by the logic in C2Compiler. | |
2866 void Compile::record_failure(const char* reason) { | |
2867 if (log() != NULL) { | |
2868 log()->elem("failure reason='%s' phase='compile'", reason); | |
2869 } | |
2870 if (_failure_reason == NULL) { | |
2871 // Record the first failure reason. | |
2872 _failure_reason = reason; | |
2873 } | |
222 | 2874 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { |
2875 C->print_method(_failure_reason); | |
2876 } | |
0 | 2877 _root = NULL; // flush the graph, too |
2878 } | |
2879 | |
2880 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) | |
2881 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) | |
2882 { | |
2883 if (dolog) { | |
2884 C = Compile::current(); | |
2885 _log = C->log(); | |
2886 } else { | |
2887 C = NULL; | |
2888 _log = NULL; | |
2889 } | |
2890 if (_log != NULL) { | |
2891 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); | |
2892 _log->stamp(); | |
2893 _log->end_head(); | |
2894 } | |
2895 } | |
2896 | |
2897 Compile::TracePhase::~TracePhase() { | |
2898 if (_log != NULL) { | |
2899 _log->done("phase nodes='%d'", C->unique()); | |
2900 } | |
2901 } | |
2008 | 2902 |
2903 //============================================================================= | |
2904 // Two Constant's are equal when the type and the value are equal. | |
2905 bool Compile::Constant::operator==(const Constant& other) { | |
2906 if (type() != other.type() ) return false; | |
2907 if (can_be_reused() != other.can_be_reused()) return false; | |
2908 // For floating point values we compare the bit pattern. | |
2909 switch (type()) { | |
2910 case T_FLOAT: return (_value.i == other._value.i); | |
2911 case T_LONG: | |
2912 case T_DOUBLE: return (_value.j == other._value.j); | |
2913 case T_OBJECT: | |
2914 case T_ADDRESS: return (_value.l == other._value.l); | |
2915 case T_VOID: return (_value.l == other._value.l); // jump-table entries | |
2916 default: ShouldNotReachHere(); | |
2917 } | |
2918 return false; | |
2919 } | |
2920 | |
2921 // Emit constants grouped in the following order: | |
2922 static BasicType type_order[] = { | |
2923 T_FLOAT, // 32-bit | |
2924 T_OBJECT, // 32 or 64-bit | |
2925 T_ADDRESS, // 32 or 64-bit | |
2926 T_DOUBLE, // 64-bit | |
2927 T_LONG, // 64-bit | |
2928 T_VOID, // 32 or 64-bit (jump-tables are at the end of the constant table for code emission reasons) | |
2929 T_ILLEGAL | |
2930 }; | |
2931 | |
2932 static int type_to_size_in_bytes(BasicType t) { | |
2933 switch (t) { | |
2934 case T_LONG: return sizeof(jlong ); | |
2935 case T_FLOAT: return sizeof(jfloat ); | |
2936 case T_DOUBLE: return sizeof(jdouble); | |
2937 // We use T_VOID as marker for jump-table entries (labels) which | |
2938 // need an interal word relocation. | |
2939 case T_VOID: | |
2940 case T_ADDRESS: | |
2941 case T_OBJECT: return sizeof(jobject); | |
2942 } | |
2943 | |
2944 ShouldNotReachHere(); | |
2945 return -1; | |
2946 } | |
2947 | |
2948 void Compile::ConstantTable::calculate_offsets_and_size() { | |
2949 int size = 0; | |
2950 for (int t = 0; type_order[t] != T_ILLEGAL; t++) { | |
2951 BasicType type = type_order[t]; | |
2952 | |
2953 for (int i = 0; i < _constants.length(); i++) { | |
2954 Constant con = _constants.at(i); | |
2955 if (con.type() != type) continue; // Skip other types. | |
2956 | |
2957 // Align size for type. | |
2958 int typesize = type_to_size_in_bytes(con.type()); | |
2959 size = align_size_up(size, typesize); | |
2960 | |
2961 // Set offset. | |
2962 con.set_offset(size); | |
2963 _constants.at_put(i, con); | |
2964 | |
2965 // Add type size. | |
2966 size = size + typesize; | |
2967 } | |
2968 } | |
2969 | |
2970 // Align size up to the next section start (which is insts; see | |
2971 // CodeBuffer::align_at_start). | |
2972 assert(_size == -1, "already set?"); | |
2973 _size = align_size_up(size, CodeEntryAlignment); | |
2974 | |
2975 if (Matcher::constant_table_absolute_addressing) { | |
2976 set_table_base_offset(0); // No table base offset required | |
2977 } else { | |
2978 if (UseRDPCForConstantTableBase) { | |
2979 // table base offset is set in MachConstantBaseNode::emit | |
2980 } else { | |
2981 // When RDPC is not used, the table base is set into the middle of | |
2982 // the constant table. | |
2983 int half_size = _size / 2; | |
2984 assert(half_size * 2 == _size, "sanity"); | |
2985 set_table_base_offset(-half_size); | |
2986 } | |
2987 } | |
2988 } | |
2989 | |
2990 void Compile::ConstantTable::emit(CodeBuffer& cb) { | |
2991 MacroAssembler _masm(&cb); | |
2992 for (int t = 0; type_order[t] != T_ILLEGAL; t++) { | |
2993 BasicType type = type_order[t]; | |
2994 | |
2995 for (int i = 0; i < _constants.length(); i++) { | |
2996 Constant con = _constants.at(i); | |
2997 if (con.type() != type) continue; // Skip other types. | |
2998 | |
2999 address constant_addr; | |
3000 switch (con.type()) { | |
3001 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; | |
3002 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; | |
3003 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; | |
3004 case T_OBJECT: { | |
3005 jobject obj = con.get_jobject(); | |
3006 int oop_index = _masm.oop_recorder()->find_index(obj); | |
3007 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); | |
3008 break; | |
3009 } | |
3010 case T_ADDRESS: { | |
3011 address addr = (address) con.get_jobject(); | |
3012 constant_addr = _masm.address_constant(addr); | |
3013 break; | |
3014 } | |
3015 // We use T_VOID as marker for jump-table entries (labels) which | |
3016 // need an interal word relocation. | |
3017 case T_VOID: { | |
3018 // Write a dummy word. The real value is filled in later | |
3019 // in fill_jump_table_in_constant_table. | |
3020 address addr = (address) con.get_jobject(); | |
3021 constant_addr = _masm.address_constant(addr); | |
3022 break; | |
3023 } | |
3024 default: ShouldNotReachHere(); | |
3025 } | |
3026 assert(constant_addr != NULL, "consts section too small"); | |
3027 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); | |
3028 } | |
3029 } | |
3030 } | |
3031 | |
3032 int Compile::ConstantTable::find_offset(Constant& con) const { | |
3033 int idx = _constants.find(con); | |
3034 assert(idx != -1, "constant must be in constant table"); | |
3035 int offset = _constants.at(idx).offset(); | |
3036 assert(offset != -1, "constant table not emitted yet?"); | |
3037 return offset; | |
3038 } | |
3039 | |
3040 void Compile::ConstantTable::add(Constant& con) { | |
3041 if (con.can_be_reused()) { | |
3042 int idx = _constants.find(con); | |
3043 if (idx != -1 && _constants.at(idx).can_be_reused()) { | |
3044 return; | |
3045 } | |
3046 } | |
3047 (void) _constants.append(con); | |
3048 } | |
3049 | |
3050 Compile::Constant Compile::ConstantTable::add(BasicType type, jvalue value) { | |
3051 Constant con(type, value); | |
3052 add(con); | |
3053 return con; | |
3054 } | |
3055 | |
3056 Compile::Constant Compile::ConstantTable::add(MachOper* oper) { | |
3057 jvalue value; | |
3058 BasicType type = oper->type()->basic_type(); | |
3059 switch (type) { | |
3060 case T_LONG: value.j = oper->constantL(); break; | |
3061 case T_FLOAT: value.f = oper->constantF(); break; | |
3062 case T_DOUBLE: value.d = oper->constantD(); break; | |
3063 case T_OBJECT: | |
3064 case T_ADDRESS: value.l = (jobject) oper->constant(); break; | |
3065 default: ShouldNotReachHere(); | |
3066 } | |
3067 return add(type, value); | |
3068 } | |
3069 | |
3070 Compile::Constant Compile::ConstantTable::allocate_jump_table(MachConstantNode* n) { | |
3071 jvalue value; | |
3072 // We can use the node pointer here to identify the right jump-table | |
3073 // as this method is called from Compile::Fill_buffer right before | |
3074 // the MachNodes are emitted and the jump-table is filled (means the | |
3075 // MachNode pointers do not change anymore). | |
3076 value.l = (jobject) n; | |
3077 Constant con(T_VOID, value, false); // Labels of a jump-table cannot be reused. | |
3078 for (uint i = 0; i < n->outcnt(); i++) { | |
3079 add(con); | |
3080 } | |
3081 return con; | |
3082 } | |
3083 | |
3084 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { | |
3085 // If called from Compile::scratch_emit_size do nothing. | |
3086 if (Compile::current()->in_scratch_emit_size()) return; | |
3087 | |
3088 assert(labels.is_nonempty(), "must be"); | |
3089 assert((uint) labels.length() == n->outcnt(), err_msg("must be equal: %d == %d", labels.length(), n->outcnt())); | |
3090 | |
3091 // Since MachConstantNode::constant_offset() also contains | |
3092 // table_base_offset() we need to subtract the table_base_offset() | |
3093 // to get the plain offset into the constant table. | |
3094 int offset = n->constant_offset() - table_base_offset(); | |
3095 | |
3096 MacroAssembler _masm(&cb); | |
3097 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); | |
3098 | |
3099 for (int i = 0; i < labels.length(); i++) { | |
3100 address* constant_addr = &jump_table_base[i]; | |
3101 assert(*constant_addr == (address) n, "all jump-table entries must contain node pointer"); | |
3102 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); | |
3103 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); | |
3104 } | |
3105 } |