1 | /* |
2 | * Copyright (c) 1997, 2019, Oracle and/or its affiliates. All rights reserved. |
3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 | * |
5 | * This code is free software; you can redistribute it and/or modify it |
6 | * under the terms of the GNU General Public License version 2 only, as |
7 | * published by the Free Software Foundation. |
8 | * |
9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
12 | * version 2 for more details (a copy is included in the LICENSE file that |
13 | * accompanied this code). |
14 | * |
15 | * You should have received a copy of the GNU General Public License version |
16 | * 2 along with this work; if not, write to the Free Software Foundation, |
17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
18 | * |
19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
20 | * or visit www.oracle.com if you need additional information or have any |
21 | * questions. |
22 | * |
23 | */ |
24 | |
25 | #include "precompiled.hpp" |
26 | #include "asm/macroAssembler.hpp" |
27 | #include "asm/macroAssembler.inline.hpp" |
28 | #include "ci/ciReplay.hpp" |
29 | #include "classfile/systemDictionary.hpp" |
30 | #include "code/exceptionHandlerTable.hpp" |
31 | #include "code/nmethod.hpp" |
32 | #include "compiler/compileBroker.hpp" |
33 | #include "compiler/compileLog.hpp" |
34 | #include "compiler/disassembler.hpp" |
35 | #include "compiler/oopMap.hpp" |
36 | #include "gc/shared/barrierSet.hpp" |
37 | #include "gc/shared/c2/barrierSetC2.hpp" |
38 | #include "memory/resourceArea.hpp" |
39 | #include "opto/addnode.hpp" |
40 | #include "opto/block.hpp" |
41 | #include "opto/c2compiler.hpp" |
42 | #include "opto/callGenerator.hpp" |
43 | #include "opto/callnode.hpp" |
44 | #include "opto/castnode.hpp" |
45 | #include "opto/cfgnode.hpp" |
46 | #include "opto/chaitin.hpp" |
47 | #include "opto/compile.hpp" |
48 | #include "opto/connode.hpp" |
49 | #include "opto/convertnode.hpp" |
50 | #include "opto/divnode.hpp" |
51 | #include "opto/escape.hpp" |
52 | #include "opto/idealGraphPrinter.hpp" |
53 | #include "opto/loopnode.hpp" |
54 | #include "opto/machnode.hpp" |
55 | #include "opto/macro.hpp" |
56 | #include "opto/matcher.hpp" |
57 | #include "opto/mathexactnode.hpp" |
58 | #include "opto/memnode.hpp" |
59 | #include "opto/mulnode.hpp" |
60 | #include "opto/narrowptrnode.hpp" |
61 | #include "opto/node.hpp" |
62 | #include "opto/opcodes.hpp" |
63 | #include "opto/output.hpp" |
64 | #include "opto/parse.hpp" |
65 | #include "opto/phaseX.hpp" |
66 | #include "opto/rootnode.hpp" |
67 | #include "opto/runtime.hpp" |
68 | #include "opto/stringopts.hpp" |
69 | #include "opto/type.hpp" |
70 | #include "opto/vectornode.hpp" |
71 | #include "runtime/arguments.hpp" |
72 | #include "runtime/sharedRuntime.hpp" |
73 | #include "runtime/signature.hpp" |
74 | #include "runtime/stubRoutines.hpp" |
75 | #include "runtime/timer.hpp" |
76 | #include "utilities/align.hpp" |
77 | #include "utilities/copy.hpp" |
78 | #include "utilities/macros.hpp" |
79 | #if INCLUDE_ZGC |
80 | #include "gc/z/c2/zBarrierSetC2.hpp" |
81 | #endif |
82 | |
83 | |
84 | // -------------------- Compile::mach_constant_base_node ----------------------- |
85 | // Constant table base node singleton. |
86 | MachConstantBaseNode* Compile::mach_constant_base_node() { |
87 | if (_mach_constant_base_node == NULL) { |
88 | _mach_constant_base_node = new MachConstantBaseNode(); |
89 | _mach_constant_base_node->add_req(C->root()); |
90 | } |
91 | return _mach_constant_base_node; |
92 | } |
93 | |
94 | |
95 | /// Support for intrinsics. |
96 | |
97 | // Return the index at which m must be inserted (or already exists). |
98 | // The sort order is by the address of the ciMethod, with is_virtual as minor key. |
99 | class IntrinsicDescPair { |
100 | private: |
101 | ciMethod* _m; |
102 | bool _is_virtual; |
103 | public: |
104 | IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} |
105 | static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { |
106 | ciMethod* m= elt->method(); |
107 | ciMethod* key_m = key->_m; |
108 | if (key_m < m) return -1; |
109 | else if (key_m > m) return 1; |
110 | else { |
111 | bool is_virtual = elt->is_virtual(); |
112 | bool key_virtual = key->_is_virtual; |
113 | if (key_virtual < is_virtual) return -1; |
114 | else if (key_virtual > is_virtual) return 1; |
115 | else return 0; |
116 | } |
117 | } |
118 | }; |
119 | int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { |
120 | #ifdef ASSERT |
121 | for (int i = 1; i < _intrinsics->length(); i++) { |
122 | CallGenerator* cg1 = _intrinsics->at(i-1); |
123 | CallGenerator* cg2 = _intrinsics->at(i); |
124 | assert(cg1->method() != cg2->method() |
125 | ? cg1->method() < cg2->method() |
126 | : cg1->is_virtual() < cg2->is_virtual(), |
127 | "compiler intrinsics list must stay sorted" ); |
128 | } |
129 | #endif |
130 | IntrinsicDescPair pair(m, is_virtual); |
131 | return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); |
132 | } |
133 | |
134 | void Compile::register_intrinsic(CallGenerator* cg) { |
135 | if (_intrinsics == NULL) { |
136 | _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); |
137 | } |
138 | int len = _intrinsics->length(); |
139 | bool found = false; |
140 | int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); |
141 | assert(!found, "registering twice" ); |
142 | _intrinsics->insert_before(index, cg); |
143 | assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked" ); |
144 | } |
145 | |
146 | CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { |
147 | assert(m->is_loaded(), "don't try this on unloaded methods" ); |
148 | if (_intrinsics != NULL) { |
149 | bool found = false; |
150 | int index = intrinsic_insertion_index(m, is_virtual, found); |
151 | if (found) { |
152 | return _intrinsics->at(index); |
153 | } |
154 | } |
155 | // Lazily create intrinsics for intrinsic IDs well-known in the runtime. |
156 | if (m->intrinsic_id() != vmIntrinsics::_none && |
157 | m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { |
158 | CallGenerator* cg = make_vm_intrinsic(m, is_virtual); |
159 | if (cg != NULL) { |
160 | // Save it for next time: |
161 | register_intrinsic(cg); |
162 | return cg; |
163 | } else { |
164 | gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); |
165 | } |
166 | } |
167 | return NULL; |
168 | } |
169 | |
170 | // Compile:: register_library_intrinsics and make_vm_intrinsic are defined |
171 | // in library_call.cpp. |
172 | |
173 | |
174 | #ifndef PRODUCT |
175 | // statistics gathering... |
176 | |
177 | juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; |
178 | jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; |
179 | |
180 | bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { |
181 | assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob" ); |
182 | int oflags = _intrinsic_hist_flags[id]; |
183 | assert(flags != 0, "what happened?" ); |
184 | if (is_virtual) { |
185 | flags |= _intrinsic_virtual; |
186 | } |
187 | bool changed = (flags != oflags); |
188 | if ((flags & _intrinsic_worked) != 0) { |
189 | juint count = (_intrinsic_hist_count[id] += 1); |
190 | if (count == 1) { |
191 | changed = true; // first time |
192 | } |
193 | // increment the overall count also: |
194 | _intrinsic_hist_count[vmIntrinsics::_none] += 1; |
195 | } |
196 | if (changed) { |
197 | if (((oflags ^ flags) & _intrinsic_virtual) != 0) { |
198 | // Something changed about the intrinsic's virtuality. |
199 | if ((flags & _intrinsic_virtual) != 0) { |
200 | // This is the first use of this intrinsic as a virtual call. |
201 | if (oflags != 0) { |
202 | // We already saw it as a non-virtual, so note both cases. |
203 | flags |= _intrinsic_both; |
204 | } |
205 | } else if ((oflags & _intrinsic_both) == 0) { |
206 | // This is the first use of this intrinsic as a non-virtual |
207 | flags |= _intrinsic_both; |
208 | } |
209 | } |
210 | _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); |
211 | } |
212 | // update the overall flags also: |
213 | _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; |
214 | return changed; |
215 | } |
216 | |
217 | static char* format_flags(int flags, char* buf) { |
218 | buf[0] = 0; |
219 | if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked" ); |
220 | if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed" ); |
221 | if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled" ); |
222 | if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual" ); |
223 | if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual" ); |
224 | if (buf[0] == 0) strcat(buf, "," ); |
225 | assert(buf[0] == ',', "must be" ); |
226 | return &buf[1]; |
227 | } |
228 | |
229 | void Compile::print_intrinsic_statistics() { |
230 | char flagsbuf[100]; |
231 | ttyLocker ttyl; |
232 | if (xtty != NULL) xtty->head("statistics type='intrinsic'" ); |
233 | tty->print_cr("Compiler intrinsic usage:" ); |
234 | juint total = _intrinsic_hist_count[vmIntrinsics::_none]; |
235 | if (total == 0) total = 1; // avoid div0 in case of no successes |
236 | #define PRINT_STAT_LINE(name, c, f) \ |
237 | tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); |
238 | for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { |
239 | vmIntrinsics::ID id = (vmIntrinsics::ID) index; |
240 | int flags = _intrinsic_hist_flags[id]; |
241 | juint count = _intrinsic_hist_count[id]; |
242 | if ((flags | count) != 0) { |
243 | PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); |
244 | } |
245 | } |
246 | PRINT_STAT_LINE("total" , total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); |
247 | if (xtty != NULL) xtty->tail("statistics" ); |
248 | } |
249 | |
250 | void Compile::print_statistics() { |
251 | { ttyLocker ttyl; |
252 | if (xtty != NULL) xtty->head("statistics type='opto'" ); |
253 | Parse::print_statistics(); |
254 | PhaseCCP::print_statistics(); |
255 | PhaseRegAlloc::print_statistics(); |
256 | Scheduling::print_statistics(); |
257 | PhasePeephole::print_statistics(); |
258 | PhaseIdealLoop::print_statistics(); |
259 | if (xtty != NULL) xtty->tail("statistics" ); |
260 | } |
261 | if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { |
262 | // put this under its own <statistics> element. |
263 | print_intrinsic_statistics(); |
264 | } |
265 | } |
266 | #endif //PRODUCT |
267 | |
268 | // Support for bundling info |
269 | Bundle* Compile::node_bundling(const Node *n) { |
270 | assert(valid_bundle_info(n), "oob" ); |
271 | return &_node_bundling_base[n->_idx]; |
272 | } |
273 | |
274 | bool Compile::valid_bundle_info(const Node *n) { |
275 | return (_node_bundling_limit > n->_idx); |
276 | } |
277 | |
278 | |
279 | void Compile::gvn_replace_by(Node* n, Node* nn) { |
280 | for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { |
281 | Node* use = n->last_out(i); |
282 | bool is_in_table = initial_gvn()->hash_delete(use); |
283 | uint uses_found = 0; |
284 | for (uint j = 0; j < use->len(); j++) { |
285 | if (use->in(j) == n) { |
286 | if (j < use->req()) |
287 | use->set_req(j, nn); |
288 | else |
289 | use->set_prec(j, nn); |
290 | uses_found++; |
291 | } |
292 | } |
293 | if (is_in_table) { |
294 | // reinsert into table |
295 | initial_gvn()->hash_find_insert(use); |
296 | } |
297 | record_for_igvn(use); |
298 | i -= uses_found; // we deleted 1 or more copies of this edge |
299 | } |
300 | } |
301 | |
302 | |
303 | static inline bool not_a_node(const Node* n) { |
304 | if (n == NULL) return true; |
305 | if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. |
306 | if (*(address*)n == badAddress) return true; // kill by Node::destruct |
307 | return false; |
308 | } |
309 | |
310 | // Identify all nodes that are reachable from below, useful. |
311 | // Use breadth-first pass that records state in a Unique_Node_List, |
312 | // recursive traversal is slower. |
313 | void Compile::identify_useful_nodes(Unique_Node_List &useful) { |
314 | int estimated_worklist_size = live_nodes(); |
315 | useful.map( estimated_worklist_size, NULL ); // preallocate space |
316 | |
317 | // Initialize worklist |
318 | if (root() != NULL) { useful.push(root()); } |
319 | // If 'top' is cached, declare it useful to preserve cached node |
320 | if( cached_top_node() ) { useful.push(cached_top_node()); } |
321 | |
322 | // Push all useful nodes onto the list, breadthfirst |
323 | for( uint next = 0; next < useful.size(); ++next ) { |
324 | assert( next < unique(), "Unique useful nodes < total nodes" ); |
325 | Node *n = useful.at(next); |
326 | uint max = n->len(); |
327 | for( uint i = 0; i < max; ++i ) { |
328 | Node *m = n->in(i); |
329 | if (not_a_node(m)) continue; |
330 | useful.push(m); |
331 | } |
332 | } |
333 | } |
334 | |
335 | // Update dead_node_list with any missing dead nodes using useful |
336 | // list. Consider all non-useful nodes to be useless i.e., dead nodes. |
337 | void Compile::update_dead_node_list(Unique_Node_List &useful) { |
338 | uint max_idx = unique(); |
339 | VectorSet& useful_node_set = useful.member_set(); |
340 | |
341 | for (uint node_idx = 0; node_idx < max_idx; node_idx++) { |
342 | // If node with index node_idx is not in useful set, |
343 | // mark it as dead in dead node list. |
344 | if (! useful_node_set.test(node_idx) ) { |
345 | record_dead_node(node_idx); |
346 | } |
347 | } |
348 | } |
349 | |
350 | void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { |
351 | int shift = 0; |
352 | for (int i = 0; i < inlines->length(); i++) { |
353 | CallGenerator* cg = inlines->at(i); |
354 | CallNode* call = cg->call_node(); |
355 | if (shift > 0) { |
356 | inlines->at_put(i-shift, cg); |
357 | } |
358 | if (!useful.member(call)) { |
359 | shift++; |
360 | } |
361 | } |
362 | inlines->trunc_to(inlines->length()-shift); |
363 | } |
364 | |
365 | // Disconnect all useless nodes by disconnecting those at the boundary. |
366 | void Compile::remove_useless_nodes(Unique_Node_List &useful) { |
367 | uint next = 0; |
368 | while (next < useful.size()) { |
369 | Node *n = useful.at(next++); |
370 | if (n->is_SafePoint()) { |
371 | // We're done with a parsing phase. Replaced nodes are not valid |
372 | // beyond that point. |
373 | n->as_SafePoint()->delete_replaced_nodes(); |
374 | } |
375 | // Use raw traversal of out edges since this code removes out edges |
376 | int max = n->outcnt(); |
377 | for (int j = 0; j < max; ++j) { |
378 | Node* child = n->raw_out(j); |
379 | if (! useful.member(child)) { |
380 | assert(!child->is_top() || child != top(), |
381 | "If top is cached in Compile object it is in useful list" ); |
382 | // Only need to remove this out-edge to the useless node |
383 | n->raw_del_out(j); |
384 | --j; |
385 | --max; |
386 | } |
387 | } |
388 | if (n->outcnt() == 1 && n->has_special_unique_user()) { |
389 | record_for_igvn(n->unique_out()); |
390 | } |
391 | } |
392 | // Remove useless macro and predicate opaq nodes |
393 | for (int i = C->macro_count()-1; i >= 0; i--) { |
394 | Node* n = C->macro_node(i); |
395 | if (!useful.member(n)) { |
396 | remove_macro_node(n); |
397 | } |
398 | } |
399 | // Remove useless CastII nodes with range check dependency |
400 | for (int i = range_check_cast_count() - 1; i >= 0; i--) { |
401 | Node* cast = range_check_cast_node(i); |
402 | if (!useful.member(cast)) { |
403 | remove_range_check_cast(cast); |
404 | } |
405 | } |
406 | // Remove useless expensive nodes |
407 | for (int i = C->expensive_count()-1; i >= 0; i--) { |
408 | Node* n = C->expensive_node(i); |
409 | if (!useful.member(n)) { |
410 | remove_expensive_node(n); |
411 | } |
412 | } |
413 | // Remove useless Opaque4 nodes |
414 | for (int i = opaque4_count() - 1; i >= 0; i--) { |
415 | Node* opaq = opaque4_node(i); |
416 | if (!useful.member(opaq)) { |
417 | remove_opaque4_node(opaq); |
418 | } |
419 | } |
420 | BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); |
421 | bs->eliminate_useless_gc_barriers(useful, this); |
422 | // clean up the late inline lists |
423 | remove_useless_late_inlines(&_string_late_inlines, useful); |
424 | remove_useless_late_inlines(&_boxing_late_inlines, useful); |
425 | remove_useless_late_inlines(&_late_inlines, useful); |
426 | debug_only(verify_graph_edges(true/*check for no_dead_code*/);) |
427 | } |
428 | |
429 | //------------------------------frame_size_in_words----------------------------- |
430 | // frame_slots in units of words |
431 | int Compile::frame_size_in_words() const { |
432 | // shift is 0 in LP32 and 1 in LP64 |
433 | const int shift = (LogBytesPerWord - LogBytesPerInt); |
434 | int words = _frame_slots >> shift; |
435 | assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); |
436 | return words; |
437 | } |
438 | |
439 | // To bang the stack of this compiled method we use the stack size |
440 | // that the interpreter would need in case of a deoptimization. This |
441 | // removes the need to bang the stack in the deoptimization blob which |
442 | // in turn simplifies stack overflow handling. |
443 | int Compile::bang_size_in_bytes() const { |
444 | return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); |
445 | } |
446 | |
447 | // ============================================================================ |
448 | //------------------------------CompileWrapper--------------------------------- |
449 | class CompileWrapper : public StackObj { |
450 | Compile *const _compile; |
451 | public: |
452 | CompileWrapper(Compile* compile); |
453 | |
454 | ~CompileWrapper(); |
455 | }; |
456 | |
457 | CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { |
458 | // the Compile* pointer is stored in the current ciEnv: |
459 | ciEnv* env = compile->env(); |
460 | assert(env == ciEnv::current(), "must already be a ciEnv active" ); |
461 | assert(env->compiler_data() == NULL, "compile already active?" ); |
462 | env->set_compiler_data(compile); |
463 | assert(compile == Compile::current(), "sanity" ); |
464 | |
465 | compile->set_type_dict(NULL); |
466 | compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); |
467 | compile->clone_map().set_clone_idx(0); |
468 | compile->set_type_hwm(NULL); |
469 | compile->set_type_last_size(0); |
470 | compile->set_last_tf(NULL, NULL); |
471 | compile->set_indexSet_arena(NULL); |
472 | compile->set_indexSet_free_block_list(NULL); |
473 | compile->init_type_arena(); |
474 | Type::Initialize(compile); |
475 | _compile->set_scratch_buffer_blob(NULL); |
476 | _compile->begin_method(); |
477 | _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); |
478 | } |
479 | CompileWrapper::~CompileWrapper() { |
480 | _compile->end_method(); |
481 | if (_compile->scratch_buffer_blob() != NULL) |
482 | BufferBlob::free(_compile->scratch_buffer_blob()); |
483 | _compile->env()->set_compiler_data(NULL); |
484 | } |
485 | |
486 | |
487 | //----------------------------print_compile_messages--------------------------- |
488 | void Compile::print_compile_messages() { |
489 | #ifndef PRODUCT |
490 | // Check if recompiling |
491 | if (_subsume_loads == false && PrintOpto) { |
492 | // Recompiling without allowing machine instructions to subsume loads |
493 | tty->print_cr("*********************************************************" ); |
494 | tty->print_cr("** Bailout: Recompile without subsuming loads **" ); |
495 | tty->print_cr("*********************************************************" ); |
496 | } |
497 | if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { |
498 | // Recompiling without escape analysis |
499 | tty->print_cr("*********************************************************" ); |
500 | tty->print_cr("** Bailout: Recompile without escape analysis **" ); |
501 | tty->print_cr("*********************************************************" ); |
502 | } |
503 | if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { |
504 | // Recompiling without boxing elimination |
505 | tty->print_cr("*********************************************************" ); |
506 | tty->print_cr("** Bailout: Recompile without boxing elimination **" ); |
507 | tty->print_cr("*********************************************************" ); |
508 | } |
509 | if (C->directive()->BreakAtCompileOption) { |
510 | // Open the debugger when compiling this method. |
511 | tty->print("### Breaking when compiling: " ); |
512 | method()->print_short_name(); |
513 | tty->cr(); |
514 | BREAKPOINT; |
515 | } |
516 | |
517 | if( PrintOpto ) { |
518 | if (is_osr_compilation()) { |
519 | tty->print("[OSR]%3d" , _compile_id); |
520 | } else { |
521 | tty->print("%3d" , _compile_id); |
522 | } |
523 | } |
524 | #endif |
525 | } |
526 | |
527 | |
528 | //-----------------------init_scratch_buffer_blob------------------------------ |
529 | // Construct a temporary BufferBlob and cache it for this compile. |
530 | void Compile::init_scratch_buffer_blob(int const_size) { |
531 | // If there is already a scratch buffer blob allocated and the |
532 | // constant section is big enough, use it. Otherwise free the |
533 | // current and allocate a new one. |
534 | BufferBlob* blob = scratch_buffer_blob(); |
535 | if ((blob != NULL) && (const_size <= _scratch_const_size)) { |
536 | // Use the current blob. |
537 | } else { |
538 | if (blob != NULL) { |
539 | BufferBlob::free(blob); |
540 | } |
541 | |
542 | ResourceMark rm; |
543 | _scratch_const_size = const_size; |
544 | int size = C2Compiler::initial_code_buffer_size(const_size); |
545 | blob = BufferBlob::create("Compile::scratch_buffer" , size); |
546 | // Record the buffer blob for next time. |
547 | set_scratch_buffer_blob(blob); |
548 | // Have we run out of code space? |
549 | if (scratch_buffer_blob() == NULL) { |
550 | // Let CompilerBroker disable further compilations. |
551 | record_failure("Not enough space for scratch buffer in CodeCache" ); |
552 | return; |
553 | } |
554 | } |
555 | |
556 | // Initialize the relocation buffers |
557 | relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; |
558 | set_scratch_locs_memory(locs_buf); |
559 | } |
560 | |
561 | |
562 | //-----------------------scratch_emit_size------------------------------------- |
563 | // Helper function that computes size by emitting code |
564 | uint Compile::scratch_emit_size(const Node* n) { |
565 | // Start scratch_emit_size section. |
566 | set_in_scratch_emit_size(true); |
567 | |
568 | // Emit into a trash buffer and count bytes emitted. |
569 | // This is a pretty expensive way to compute a size, |
570 | // but it works well enough if seldom used. |
571 | // All common fixed-size instructions are given a size |
572 | // method by the AD file. |
573 | // Note that the scratch buffer blob and locs memory are |
574 | // allocated at the beginning of the compile task, and |
575 | // may be shared by several calls to scratch_emit_size. |
576 | // The allocation of the scratch buffer blob is particularly |
577 | // expensive, since it has to grab the code cache lock. |
578 | BufferBlob* blob = this->scratch_buffer_blob(); |
579 | assert(blob != NULL, "Initialize BufferBlob at start" ); |
580 | assert(blob->size() > MAX_inst_size, "sanity" ); |
581 | relocInfo* locs_buf = scratch_locs_memory(); |
582 | address blob_begin = blob->content_begin(); |
583 | address blob_end = (address)locs_buf; |
584 | assert(blob->contains(blob_end), "sanity" ); |
585 | CodeBuffer buf(blob_begin, blob_end - blob_begin); |
586 | buf.initialize_consts_size(_scratch_const_size); |
587 | buf.initialize_stubs_size(MAX_stubs_size); |
588 | assert(locs_buf != NULL, "sanity" ); |
589 | int lsize = MAX_locs_size / 3; |
590 | buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); |
591 | buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); |
592 | buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); |
593 | // Mark as scratch buffer. |
594 | buf.consts()->set_scratch_emit(); |
595 | buf.insts()->set_scratch_emit(); |
596 | buf.stubs()->set_scratch_emit(); |
597 | |
598 | // Do the emission. |
599 | |
600 | Label fakeL; // Fake label for branch instructions. |
601 | Label* saveL = NULL; |
602 | uint save_bnum = 0; |
603 | bool is_branch = n->is_MachBranch(); |
604 | if (is_branch) { |
605 | MacroAssembler masm(&buf); |
606 | masm.bind(fakeL); |
607 | n->as_MachBranch()->save_label(&saveL, &save_bnum); |
608 | n->as_MachBranch()->label_set(&fakeL, 0); |
609 | } |
610 | n->emit(buf, this->regalloc()); |
611 | |
612 | // Emitting into the scratch buffer should not fail |
613 | assert (!failing(), "Must not have pending failure. Reason is: %s" , failure_reason()); |
614 | |
615 | if (is_branch) // Restore label. |
616 | n->as_MachBranch()->label_set(saveL, save_bnum); |
617 | |
618 | // End scratch_emit_size section. |
619 | set_in_scratch_emit_size(false); |
620 | |
621 | return buf.insts_size(); |
622 | } |
623 | |
624 | |
625 | // ============================================================================ |
626 | //------------------------------Compile standard------------------------------- |
627 | debug_only( int Compile::_debug_idx = 100000; ) |
628 | |
629 | // Compile a method. entry_bci is -1 for normal compilations and indicates |
630 | // the continuation bci for on stack replacement. |
631 | |
632 | |
633 | Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, |
634 | bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) |
635 | : Phase(Compiler), |
636 | _compile_id(ci_env->compile_id()), |
637 | _save_argument_registers(false), |
638 | _subsume_loads(subsume_loads), |
639 | _do_escape_analysis(do_escape_analysis), |
640 | _eliminate_boxing(eliminate_boxing), |
641 | _method(target), |
642 | _entry_bci(osr_bci), |
643 | _stub_function(NULL), |
644 | _stub_name(NULL), |
645 | _stub_entry_point(NULL), |
646 | _max_node_limit(MaxNodeLimit), |
647 | _orig_pc_slot(0), |
648 | _orig_pc_slot_offset_in_bytes(0), |
649 | _inlining_progress(false), |
650 | _inlining_incrementally(false), |
651 | _do_cleanup(false), |
652 | _has_reserved_stack_access(target->has_reserved_stack_access()), |
653 | #ifndef PRODUCT |
654 | _trace_opto_output(directive->TraceOptoOutputOption), |
655 | #endif |
656 | _has_method_handle_invokes(false), |
657 | _clinit_barrier_on_entry(false), |
658 | _comp_arena(mtCompiler), |
659 | _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), |
660 | _env(ci_env), |
661 | _directive(directive), |
662 | _log(ci_env->log()), |
663 | _failure_reason(NULL), |
664 | _congraph(NULL), |
665 | #ifndef PRODUCT |
666 | _printer(IdealGraphPrinter::printer()), |
667 | #endif |
668 | _dead_node_list(comp_arena()), |
669 | _dead_node_count(0), |
670 | _node_arena(mtCompiler), |
671 | _old_arena(mtCompiler), |
672 | _mach_constant_base_node(NULL), |
673 | _Compile_types(mtCompiler), |
674 | _initial_gvn(NULL), |
675 | _for_igvn(NULL), |
676 | _warm_calls(NULL), |
677 | _late_inlines(comp_arena(), 2, 0, NULL), |
678 | _string_late_inlines(comp_arena(), 2, 0, NULL), |
679 | _boxing_late_inlines(comp_arena(), 2, 0, NULL), |
680 | _late_inlines_pos(0), |
681 | _number_of_mh_late_inlines(0), |
682 | _print_inlining_stream(NULL), |
683 | _print_inlining_list(NULL), |
684 | _print_inlining_idx(0), |
685 | _print_inlining_output(NULL), |
686 | _replay_inline_data(NULL), |
687 | _java_calls(0), |
688 | _inner_loops(0), |
689 | _interpreter_frame_size(0), |
690 | _node_bundling_limit(0), |
691 | _node_bundling_base(NULL), |
692 | _code_buffer("Compile::Fill_buffer" ), |
693 | _scratch_const_size(-1), |
694 | _in_scratch_emit_size(false) |
695 | #ifndef PRODUCT |
696 | , _in_dump_cnt(0) |
697 | #endif |
698 | { |
699 | C = this; |
700 | #ifndef PRODUCT |
701 | if (_printer != NULL) { |
702 | _printer->set_compile(this); |
703 | } |
704 | #endif |
705 | CompileWrapper cw(this); |
706 | |
707 | if (CITimeVerbose) { |
708 | tty->print(" " ); |
709 | target->holder()->name()->print(); |
710 | tty->print("." ); |
711 | target->print_short_name(); |
712 | tty->print(" " ); |
713 | } |
714 | TraceTime t1("Total compilation time" , &_t_totalCompilation, CITime, CITimeVerbose); |
715 | TraceTime t2(NULL, &_t_methodCompilation, CITime, false); |
716 | |
717 | #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) |
718 | bool print_opto_assembly = directive->PrintOptoAssemblyOption; |
719 | // We can always print a disassembly, either abstract (hex dump) or |
720 | // with the help of a suitable hsdis library. Thus, we should not |
721 | // couple print_assembly and print_opto_assembly controls. |
722 | // But: always print opto and regular assembly on compile command 'print'. |
723 | bool print_assembly = directive->PrintAssemblyOption; |
724 | set_print_assembly(print_opto_assembly || print_assembly); |
725 | #else |
726 | set_print_assembly(false); // must initialize. |
727 | #endif |
728 | |
729 | #ifndef PRODUCT |
730 | set_parsed_irreducible_loop(false); |
731 | |
732 | if (directive->ReplayInlineOption) { |
733 | _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); |
734 | } |
735 | #endif |
736 | set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); |
737 | set_print_intrinsics(directive->PrintIntrinsicsOption); |
738 | set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it |
739 | |
740 | if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { |
741 | // Make sure the method being compiled gets its own MDO, |
742 | // so we can at least track the decompile_count(). |
743 | // Need MDO to record RTM code generation state. |
744 | method()->ensure_method_data(); |
745 | } |
746 | |
747 | Init(::AliasLevel); |
748 | |
749 | |
750 | print_compile_messages(); |
751 | |
752 | _ilt = InlineTree::build_inline_tree_root(); |
753 | |
754 | // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice |
755 | assert(num_alias_types() >= AliasIdxRaw, "" ); |
756 | |
757 | #define MINIMUM_NODE_HASH 1023 |
758 | // Node list that Iterative GVN will start with |
759 | Unique_Node_List for_igvn(comp_arena()); |
760 | set_for_igvn(&for_igvn); |
761 | |
762 | // GVN that will be run immediately on new nodes |
763 | uint estimated_size = method()->code_size()*4+64; |
764 | estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); |
765 | PhaseGVN gvn(node_arena(), estimated_size); |
766 | set_initial_gvn(&gvn); |
767 | |
768 | print_inlining_init(); |
769 | { // Scope for timing the parser |
770 | TracePhase tp("parse" , &timers[_t_parser]); |
771 | |
772 | // Put top into the hash table ASAP. |
773 | initial_gvn()->transform_no_reclaim(top()); |
774 | |
775 | // Set up tf(), start(), and find a CallGenerator. |
776 | CallGenerator* cg = NULL; |
777 | if (is_osr_compilation()) { |
778 | const TypeTuple *domain = StartOSRNode::osr_domain(); |
779 | const TypeTuple *range = TypeTuple::make_range(method()->signature()); |
780 | init_tf(TypeFunc::make(domain, range)); |
781 | StartNode* s = new StartOSRNode(root(), domain); |
782 | initial_gvn()->set_type_bottom(s); |
783 | init_start(s); |
784 | cg = CallGenerator::for_osr(method(), entry_bci()); |
785 | } else { |
786 | // Normal case. |
787 | init_tf(TypeFunc::make(method())); |
788 | StartNode* s = new StartNode(root(), tf()->domain()); |
789 | initial_gvn()->set_type_bottom(s); |
790 | init_start(s); |
791 | if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { |
792 | // With java.lang.ref.reference.get() we must go through the |
793 | // intrinsic - even when get() is the root |
794 | // method of the compile - so that, if necessary, the value in |
795 | // the referent field of the reference object gets recorded by |
796 | // the pre-barrier code. |
797 | cg = find_intrinsic(method(), false); |
798 | } |
799 | if (cg == NULL) { |
800 | float past_uses = method()->interpreter_invocation_count(); |
801 | float expected_uses = past_uses; |
802 | cg = CallGenerator::for_inline(method(), expected_uses); |
803 | } |
804 | } |
805 | if (failing()) return; |
806 | if (cg == NULL) { |
807 | record_method_not_compilable("cannot parse method" ); |
808 | return; |
809 | } |
810 | JVMState* jvms = build_start_state(start(), tf()); |
811 | if ((jvms = cg->generate(jvms)) == NULL) { |
812 | if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { |
813 | record_method_not_compilable("method parse failed" ); |
814 | } |
815 | return; |
816 | } |
817 | GraphKit kit(jvms); |
818 | |
819 | if (!kit.stopped()) { |
820 | // Accept return values, and transfer control we know not where. |
821 | // This is done by a special, unique ReturnNode bound to root. |
822 | return_values(kit.jvms()); |
823 | } |
824 | |
825 | if (kit.has_exceptions()) { |
826 | // Any exceptions that escape from this call must be rethrown |
827 | // to whatever caller is dynamically above us on the stack. |
828 | // This is done by a special, unique RethrowNode bound to root. |
829 | rethrow_exceptions(kit.transfer_exceptions_into_jvms()); |
830 | } |
831 | |
832 | assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off" ); |
833 | |
834 | if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { |
835 | inline_string_calls(true); |
836 | } |
837 | |
838 | if (failing()) return; |
839 | |
840 | print_method(PHASE_BEFORE_REMOVEUSELESS, 3); |
841 | |
842 | // Remove clutter produced by parsing. |
843 | if (!failing()) { |
844 | ResourceMark rm; |
845 | PhaseRemoveUseless pru(initial_gvn(), &for_igvn); |
846 | } |
847 | } |
848 | |
849 | // Note: Large methods are capped off in do_one_bytecode(). |
850 | if (failing()) return; |
851 | |
852 | // After parsing, node notes are no longer automagic. |
853 | // They must be propagated by register_new_node_with_optimizer(), |
854 | // clone(), or the like. |
855 | set_default_node_notes(NULL); |
856 | |
857 | for (;;) { |
858 | int successes = Inline_Warm(); |
859 | if (failing()) return; |
860 | if (successes == 0) break; |
861 | } |
862 | |
863 | // Drain the list. |
864 | Finish_Warm(); |
865 | #ifndef PRODUCT |
866 | if (_printer && _printer->should_print(1)) { |
867 | _printer->print_inlining(); |
868 | } |
869 | #endif |
870 | |
871 | if (failing()) return; |
872 | NOT_PRODUCT( verify_graph_edges(); ) |
873 | |
874 | // Now optimize |
875 | Optimize(); |
876 | if (failing()) return; |
877 | NOT_PRODUCT( verify_graph_edges(); ) |
878 | |
879 | #ifndef PRODUCT |
880 | if (PrintIdeal) { |
881 | ttyLocker ttyl; // keep the following output all in one block |
882 | // This output goes directly to the tty, not the compiler log. |
883 | // To enable tools to match it up with the compilation activity, |
884 | // be sure to tag this tty output with the compile ID. |
885 | if (xtty != NULL) { |
886 | xtty->head("ideal compile_id='%d'%s" , compile_id(), |
887 | is_osr_compilation() ? " compile_kind='osr'" : |
888 | "" ); |
889 | } |
890 | root()->dump(9999); |
891 | if (xtty != NULL) { |
892 | xtty->tail("ideal" ); |
893 | } |
894 | } |
895 | #endif |
896 | |
897 | #ifdef ASSERT |
898 | BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); |
899 | bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); |
900 | #endif |
901 | |
902 | // Dump compilation data to replay it. |
903 | if (directive->DumpReplayOption) { |
904 | env()->dump_replay_data(_compile_id); |
905 | } |
906 | if (directive->DumpInlineOption && (ilt() != NULL)) { |
907 | env()->dump_inline_data(_compile_id); |
908 | } |
909 | |
910 | // Now that we know the size of all the monitors we can add a fixed slot |
911 | // for the original deopt pc. |
912 | |
913 | _orig_pc_slot = fixed_slots(); |
914 | int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); |
915 | set_fixed_slots(next_slot); |
916 | |
917 | // Compute when to use implicit null checks. Used by matching trap based |
918 | // nodes and NullCheck optimization. |
919 | set_allowed_deopt_reasons(); |
920 | |
921 | // Now generate code |
922 | Code_Gen(); |
923 | if (failing()) return; |
924 | |
925 | // Check if we want to skip execution of all compiled code. |
926 | { |
927 | #ifndef PRODUCT |
928 | if (OptoNoExecute) { |
929 | record_method_not_compilable("+OptoNoExecute" ); // Flag as failed |
930 | return; |
931 | } |
932 | #endif |
933 | TracePhase tp("install_code" , &timers[_t_registerMethod]); |
934 | |
935 | if (is_osr_compilation()) { |
936 | _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); |
937 | _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); |
938 | } else { |
939 | _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); |
940 | _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); |
941 | } |
942 | |
943 | env()->register_method(_method, _entry_bci, |
944 | &_code_offsets, |
945 | _orig_pc_slot_offset_in_bytes, |
946 | code_buffer(), |
947 | frame_size_in_words(), _oop_map_set, |
948 | &_handler_table, &_inc_table, |
949 | compiler, |
950 | has_unsafe_access(), |
951 | SharedRuntime::is_wide_vector(max_vector_size()), |
952 | rtm_state() |
953 | ); |
954 | |
955 | if (log() != NULL) // Print code cache state into compiler log |
956 | log()->code_cache_state(); |
957 | } |
958 | } |
959 | |
960 | //------------------------------Compile---------------------------------------- |
961 | // Compile a runtime stub |
962 | Compile::Compile( ciEnv* ci_env, |
963 | TypeFunc_generator generator, |
964 | address stub_function, |
965 | const char *stub_name, |
966 | int is_fancy_jump, |
967 | bool pass_tls, |
968 | bool save_arg_registers, |
969 | bool return_pc, |
970 | DirectiveSet* directive) |
971 | : Phase(Compiler), |
972 | _compile_id(0), |
973 | _save_argument_registers(save_arg_registers), |
974 | _subsume_loads(true), |
975 | _do_escape_analysis(false), |
976 | _eliminate_boxing(false), |
977 | _method(NULL), |
978 | _entry_bci(InvocationEntryBci), |
979 | _stub_function(stub_function), |
980 | _stub_name(stub_name), |
981 | _stub_entry_point(NULL), |
982 | _max_node_limit(MaxNodeLimit), |
983 | _orig_pc_slot(0), |
984 | _orig_pc_slot_offset_in_bytes(0), |
985 | _inlining_progress(false), |
986 | _inlining_incrementally(false), |
987 | _has_reserved_stack_access(false), |
988 | #ifndef PRODUCT |
989 | _trace_opto_output(directive->TraceOptoOutputOption), |
990 | #endif |
991 | _has_method_handle_invokes(false), |
992 | _clinit_barrier_on_entry(false), |
993 | _comp_arena(mtCompiler), |
994 | _env(ci_env), |
995 | _directive(directive), |
996 | _log(ci_env->log()), |
997 | _failure_reason(NULL), |
998 | _congraph(NULL), |
999 | #ifndef PRODUCT |
1000 | _printer(NULL), |
1001 | #endif |
1002 | _dead_node_list(comp_arena()), |
1003 | _dead_node_count(0), |
1004 | _node_arena(mtCompiler), |
1005 | _old_arena(mtCompiler), |
1006 | _mach_constant_base_node(NULL), |
1007 | _Compile_types(mtCompiler), |
1008 | _initial_gvn(NULL), |
1009 | _for_igvn(NULL), |
1010 | _warm_calls(NULL), |
1011 | _number_of_mh_late_inlines(0), |
1012 | _print_inlining_stream(NULL), |
1013 | _print_inlining_list(NULL), |
1014 | _print_inlining_idx(0), |
1015 | _print_inlining_output(NULL), |
1016 | _replay_inline_data(NULL), |
1017 | _java_calls(0), |
1018 | _inner_loops(0), |
1019 | _interpreter_frame_size(0), |
1020 | _node_bundling_limit(0), |
1021 | _node_bundling_base(NULL), |
1022 | _code_buffer("Compile::Fill_buffer" ), |
1023 | #ifndef PRODUCT |
1024 | _in_dump_cnt(0), |
1025 | #endif |
1026 | _allowed_reasons(0) { |
1027 | C = this; |
1028 | |
1029 | TraceTime t1(NULL, &_t_totalCompilation, CITime, false); |
1030 | TraceTime t2(NULL, &_t_stubCompilation, CITime, false); |
1031 | |
1032 | #ifndef PRODUCT |
1033 | set_print_assembly(PrintFrameConverterAssembly); |
1034 | set_parsed_irreducible_loop(false); |
1035 | #else |
1036 | set_print_assembly(false); // Must initialize. |
1037 | #endif |
1038 | set_has_irreducible_loop(false); // no loops |
1039 | |
1040 | CompileWrapper cw(this); |
1041 | Init(/*AliasLevel=*/ 0); |
1042 | init_tf((*generator)()); |
1043 | |
1044 | { |
1045 | // The following is a dummy for the sake of GraphKit::gen_stub |
1046 | Unique_Node_List for_igvn(comp_arena()); |
1047 | set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this |
1048 | PhaseGVN gvn(Thread::current()->resource_area(),255); |
1049 | set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively |
1050 | gvn.transform_no_reclaim(top()); |
1051 | |
1052 | GraphKit kit; |
1053 | kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); |
1054 | } |
1055 | |
1056 | NOT_PRODUCT( verify_graph_edges(); ) |
1057 | Code_Gen(); |
1058 | if (failing()) return; |
1059 | |
1060 | |
1061 | // Entry point will be accessed using compile->stub_entry_point(); |
1062 | if (code_buffer() == NULL) { |
1063 | Matcher::soft_match_failure(); |
1064 | } else { |
1065 | if (PrintAssembly && (WizardMode || Verbose)) |
1066 | tty->print_cr("### Stub::%s" , stub_name); |
1067 | |
1068 | if (!failing()) { |
1069 | assert(_fixed_slots == 0, "no fixed slots used for runtime stubs" ); |
1070 | |
1071 | // Make the NMethod |
1072 | // For now we mark the frame as never safe for profile stackwalking |
1073 | RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, |
1074 | code_buffer(), |
1075 | CodeOffsets::frame_never_safe, |
1076 | // _code_offsets.value(CodeOffsets::Frame_Complete), |
1077 | frame_size_in_words(), |
1078 | _oop_map_set, |
1079 | save_arg_registers); |
1080 | assert(rs != NULL && rs->is_runtime_stub(), "sanity check" ); |
1081 | |
1082 | _stub_entry_point = rs->entry_point(); |
1083 | } |
1084 | } |
1085 | } |
1086 | |
1087 | //------------------------------Init------------------------------------------- |
1088 | // Prepare for a single compilation |
1089 | void Compile::Init(int aliaslevel) { |
1090 | _unique = 0; |
1091 | _regalloc = NULL; |
1092 | |
1093 | _tf = NULL; // filled in later |
1094 | _top = NULL; // cached later |
1095 | _matcher = NULL; // filled in later |
1096 | _cfg = NULL; // filled in later |
1097 | |
1098 | set_24_bit_selection_and_mode(Use24BitFP, false); |
1099 | |
1100 | _node_note_array = NULL; |
1101 | _default_node_notes = NULL; |
1102 | DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() |
1103 | |
1104 | _immutable_memory = NULL; // filled in at first inquiry |
1105 | |
1106 | // Globally visible Nodes |
1107 | // First set TOP to NULL to give safe behavior during creation of RootNode |
1108 | set_cached_top_node(NULL); |
1109 | set_root(new RootNode()); |
1110 | // Now that you have a Root to point to, create the real TOP |
1111 | set_cached_top_node( new ConNode(Type::TOP) ); |
1112 | set_recent_alloc(NULL, NULL); |
1113 | |
1114 | // Create Debug Information Recorder to record scopes, oopmaps, etc. |
1115 | env()->set_oop_recorder(new OopRecorder(env()->arena())); |
1116 | env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); |
1117 | env()->set_dependencies(new Dependencies(env())); |
1118 | |
1119 | _fixed_slots = 0; |
1120 | set_has_split_ifs(false); |
1121 | set_has_loops(has_method() && method()->has_loops()); // first approximation |
1122 | set_has_stringbuilder(false); |
1123 | set_has_boxed_value(false); |
1124 | _trap_can_recompile = false; // no traps emitted yet |
1125 | _major_progress = true; // start out assuming good things will happen |
1126 | set_has_unsafe_access(false); |
1127 | set_max_vector_size(0); |
1128 | set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers |
1129 | Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); |
1130 | set_decompile_count(0); |
1131 | |
1132 | set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); |
1133 | _loop_opts_cnt = LoopOptsCount; |
1134 | set_do_inlining(Inline); |
1135 | set_max_inline_size(MaxInlineSize); |
1136 | set_freq_inline_size(FreqInlineSize); |
1137 | set_do_scheduling(OptoScheduling); |
1138 | set_do_count_invocations(false); |
1139 | set_do_method_data_update(false); |
1140 | |
1141 | set_do_vector_loop(false); |
1142 | |
1143 | if (AllowVectorizeOnDemand) { |
1144 | if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { |
1145 | set_do_vector_loop(true); |
1146 | NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n" , method()->name()->as_quoted_ascii());}) |
1147 | } else if (has_method() && method()->name() != 0 && |
1148 | method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { |
1149 | set_do_vector_loop(true); |
1150 | } |
1151 | } |
1152 | set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally |
1153 | NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n" , method()->name()->as_quoted_ascii());}) |
1154 | |
1155 | set_age_code(has_method() && method()->profile_aging()); |
1156 | set_rtm_state(NoRTM); // No RTM lock eliding by default |
1157 | _max_node_limit = _directive->MaxNodeLimitOption; |
1158 | |
1159 | #if INCLUDE_RTM_OPT |
1160 | if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { |
1161 | int rtm_state = method()->method_data()->rtm_state(); |
1162 | if (method_has_option("NoRTMLockEliding" ) || ((rtm_state & NoRTM) != 0)) { |
1163 | // Don't generate RTM lock eliding code. |
1164 | set_rtm_state(NoRTM); |
1165 | } else if (method_has_option("UseRTMLockEliding" ) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { |
1166 | // Generate RTM lock eliding code without abort ratio calculation code. |
1167 | set_rtm_state(UseRTM); |
1168 | } else if (UseRTMDeopt) { |
1169 | // Generate RTM lock eliding code and include abort ratio calculation |
1170 | // code if UseRTMDeopt is on. |
1171 | set_rtm_state(ProfileRTM); |
1172 | } |
1173 | } |
1174 | #endif |
1175 | if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { |
1176 | set_clinit_barrier_on_entry(true); |
1177 | } |
1178 | if (debug_info()->recording_non_safepoints()) { |
1179 | set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> |
1180 | (comp_arena(), 8, 0, NULL)); |
1181 | set_default_node_notes(Node_Notes::make(this)); |
1182 | } |
1183 | |
1184 | // // -- Initialize types before each compile -- |
1185 | // // Update cached type information |
1186 | // if( _method && _method->constants() ) |
1187 | // Type::update_loaded_types(_method, _method->constants()); |
1188 | |
1189 | // Init alias_type map. |
1190 | if (!_do_escape_analysis && aliaslevel == 3) |
1191 | aliaslevel = 2; // No unique types without escape analysis |
1192 | _AliasLevel = aliaslevel; |
1193 | const int grow_ats = 16; |
1194 | _max_alias_types = grow_ats; |
1195 | _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); |
1196 | AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); |
1197 | Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); |
1198 | { |
1199 | for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; |
1200 | } |
1201 | // Initialize the first few types. |
1202 | _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); |
1203 | _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); |
1204 | _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); |
1205 | _num_alias_types = AliasIdxRaw+1; |
1206 | // Zero out the alias type cache. |
1207 | Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); |
1208 | // A NULL adr_type hits in the cache right away. Preload the right answer. |
1209 | probe_alias_cache(NULL)->_index = AliasIdxTop; |
1210 | |
1211 | _intrinsics = NULL; |
1212 | _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); |
1213 | _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); |
1214 | _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); |
1215 | _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); |
1216 | _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); |
1217 | register_library_intrinsics(); |
1218 | } |
1219 | |
1220 | //---------------------------init_start---------------------------------------- |
1221 | // Install the StartNode on this compile object. |
1222 | void Compile::init_start(StartNode* s) { |
1223 | if (failing()) |
1224 | return; // already failing |
1225 | assert(s == start(), "" ); |
1226 | } |
1227 | |
1228 | /** |
1229 | * Return the 'StartNode'. We must not have a pending failure, since the ideal graph |
1230 | * can be in an inconsistent state, i.e., we can get segmentation faults when traversing |
1231 | * the ideal graph. |
1232 | */ |
1233 | StartNode* Compile::start() const { |
1234 | assert (!failing(), "Must not have pending failure. Reason is: %s" , failure_reason()); |
1235 | for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { |
1236 | Node* start = root()->fast_out(i); |
1237 | if (start->is_Start()) { |
1238 | return start->as_Start(); |
1239 | } |
1240 | } |
1241 | fatal("Did not find Start node!" ); |
1242 | return NULL; |
1243 | } |
1244 | |
1245 | //-------------------------------immutable_memory------------------------------------- |
1246 | // Access immutable memory |
1247 | Node* Compile::immutable_memory() { |
1248 | if (_immutable_memory != NULL) { |
1249 | return _immutable_memory; |
1250 | } |
1251 | StartNode* s = start(); |
1252 | for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { |
1253 | Node *p = s->fast_out(i); |
1254 | if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { |
1255 | _immutable_memory = p; |
1256 | return _immutable_memory; |
1257 | } |
1258 | } |
1259 | ShouldNotReachHere(); |
1260 | return NULL; |
1261 | } |
1262 | |
1263 | //----------------------set_cached_top_node------------------------------------ |
1264 | // Install the cached top node, and make sure Node::is_top works correctly. |
1265 | void Compile::set_cached_top_node(Node* tn) { |
1266 | if (tn != NULL) verify_top(tn); |
1267 | Node* old_top = _top; |
1268 | _top = tn; |
1269 | // Calling Node::setup_is_top allows the nodes the chance to adjust |
1270 | // their _out arrays. |
1271 | if (_top != NULL) _top->setup_is_top(); |
1272 | if (old_top != NULL) old_top->setup_is_top(); |
1273 | assert(_top == NULL || top()->is_top(), "" ); |
1274 | } |
1275 | |
1276 | #ifdef ASSERT |
1277 | uint Compile::count_live_nodes_by_graph_walk() { |
1278 | Unique_Node_List useful(comp_arena()); |
1279 | // Get useful node list by walking the graph. |
1280 | identify_useful_nodes(useful); |
1281 | return useful.size(); |
1282 | } |
1283 | |
1284 | void Compile::print_missing_nodes() { |
1285 | |
1286 | // Return if CompileLog is NULL and PrintIdealNodeCount is false. |
1287 | if ((_log == NULL) && (! PrintIdealNodeCount)) { |
1288 | return; |
1289 | } |
1290 | |
1291 | // This is an expensive function. It is executed only when the user |
1292 | // specifies VerifyIdealNodeCount option or otherwise knows the |
1293 | // additional work that needs to be done to identify reachable nodes |
1294 | // by walking the flow graph and find the missing ones using |
1295 | // _dead_node_list. |
1296 | |
1297 | Unique_Node_List useful(comp_arena()); |
1298 | // Get useful node list by walking the graph. |
1299 | identify_useful_nodes(useful); |
1300 | |
1301 | uint l_nodes = C->live_nodes(); |
1302 | uint l_nodes_by_walk = useful.size(); |
1303 | |
1304 | if (l_nodes != l_nodes_by_walk) { |
1305 | if (_log != NULL) { |
1306 | _log->begin_head("mismatched_nodes count='%d'" , abs((int) (l_nodes - l_nodes_by_walk))); |
1307 | _log->stamp(); |
1308 | _log->end_head(); |
1309 | } |
1310 | VectorSet& useful_member_set = useful.member_set(); |
1311 | int last_idx = l_nodes_by_walk; |
1312 | for (int i = 0; i < last_idx; i++) { |
1313 | if (useful_member_set.test(i)) { |
1314 | if (_dead_node_list.test(i)) { |
1315 | if (_log != NULL) { |
1316 | _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'" , i); |
1317 | } |
1318 | if (PrintIdealNodeCount) { |
1319 | // Print the log message to tty |
1320 | tty->print_cr("mismatched_node idx='%d' both live and dead'" , i); |
1321 | useful.at(i)->dump(); |
1322 | } |
1323 | } |
1324 | } |
1325 | else if (! _dead_node_list.test(i)) { |
1326 | if (_log != NULL) { |
1327 | _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'" , i); |
1328 | } |
1329 | if (PrintIdealNodeCount) { |
1330 | // Print the log message to tty |
1331 | tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'" , i); |
1332 | } |
1333 | } |
1334 | } |
1335 | if (_log != NULL) { |
1336 | _log->tail("mismatched_nodes" ); |
1337 | } |
1338 | } |
1339 | } |
1340 | void Compile::record_modified_node(Node* n) { |
1341 | if (_modified_nodes != NULL && !_inlining_incrementally && |
1342 | n->outcnt() != 0 && !n->is_Con()) { |
1343 | _modified_nodes->push(n); |
1344 | } |
1345 | } |
1346 | |
1347 | void Compile::remove_modified_node(Node* n) { |
1348 | if (_modified_nodes != NULL) { |
1349 | _modified_nodes->remove(n); |
1350 | } |
1351 | } |
1352 | #endif |
1353 | |
1354 | #ifndef PRODUCT |
1355 | void Compile::verify_top(Node* tn) const { |
1356 | if (tn != NULL) { |
1357 | assert(tn->is_Con(), "top node must be a constant" ); |
1358 | assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type" ); |
1359 | assert(tn->in(0) != NULL, "must have live top node" ); |
1360 | } |
1361 | } |
1362 | #endif |
1363 | |
1364 | |
1365 | ///-------------------Managing Per-Node Debug & Profile Info------------------- |
1366 | |
1367 | void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { |
1368 | guarantee(arr != NULL, "" ); |
1369 | int num_blocks = arr->length(); |
1370 | if (grow_by < num_blocks) grow_by = num_blocks; |
1371 | int num_notes = grow_by * _node_notes_block_size; |
1372 | Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); |
1373 | Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); |
1374 | while (num_notes > 0) { |
1375 | arr->append(notes); |
1376 | notes += _node_notes_block_size; |
1377 | num_notes -= _node_notes_block_size; |
1378 | } |
1379 | assert(num_notes == 0, "exact multiple, please" ); |
1380 | } |
1381 | |
1382 | bool Compile::copy_node_notes_to(Node* dest, Node* source) { |
1383 | if (source == NULL || dest == NULL) return false; |
1384 | |
1385 | if (dest->is_Con()) |
1386 | return false; // Do not push debug info onto constants. |
1387 | |
1388 | #ifdef ASSERT |
1389 | // Leave a bread crumb trail pointing to the original node: |
1390 | if (dest != NULL && dest != source && dest->debug_orig() == NULL) { |
1391 | dest->set_debug_orig(source); |
1392 | } |
1393 | #endif |
1394 | |
1395 | if (node_note_array() == NULL) |
1396 | return false; // Not collecting any notes now. |
1397 | |
1398 | // This is a copy onto a pre-existing node, which may already have notes. |
1399 | // If both nodes have notes, do not overwrite any pre-existing notes. |
1400 | Node_Notes* source_notes = node_notes_at(source->_idx); |
1401 | if (source_notes == NULL || source_notes->is_clear()) return false; |
1402 | Node_Notes* dest_notes = node_notes_at(dest->_idx); |
1403 | if (dest_notes == NULL || dest_notes->is_clear()) { |
1404 | return set_node_notes_at(dest->_idx, source_notes); |
1405 | } |
1406 | |
1407 | Node_Notes merged_notes = (*source_notes); |
1408 | // The order of operations here ensures that dest notes will win... |
1409 | merged_notes.update_from(dest_notes); |
1410 | return set_node_notes_at(dest->_idx, &merged_notes); |
1411 | } |
1412 | |
1413 | |
1414 | //--------------------------allow_range_check_smearing------------------------- |
1415 | // Gating condition for coalescing similar range checks. |
1416 | // Sometimes we try 'speculatively' replacing a series of a range checks by a |
1417 | // single covering check that is at least as strong as any of them. |
1418 | // If the optimization succeeds, the simplified (strengthened) range check |
1419 | // will always succeed. If it fails, we will deopt, and then give up |
1420 | // on the optimization. |
1421 | bool Compile::allow_range_check_smearing() const { |
1422 | // If this method has already thrown a range-check, |
1423 | // assume it was because we already tried range smearing |
1424 | // and it failed. |
1425 | uint already_trapped = trap_count(Deoptimization::Reason_range_check); |
1426 | return !already_trapped; |
1427 | } |
1428 | |
1429 | |
1430 | //------------------------------flatten_alias_type----------------------------- |
1431 | const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { |
1432 | int offset = tj->offset(); |
1433 | TypePtr::PTR ptr = tj->ptr(); |
1434 | |
1435 | // Known instance (scalarizable allocation) alias only with itself. |
1436 | bool is_known_inst = tj->isa_oopptr() != NULL && |
1437 | tj->is_oopptr()->is_known_instance(); |
1438 | |
1439 | // Process weird unsafe references. |
1440 | if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { |
1441 | assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops" ); |
1442 | assert(!is_known_inst, "scalarizable allocation should not have unsafe references" ); |
1443 | tj = TypeOopPtr::BOTTOM; |
1444 | ptr = tj->ptr(); |
1445 | offset = tj->offset(); |
1446 | } |
1447 | |
1448 | // Array pointers need some flattening |
1449 | const TypeAryPtr *ta = tj->isa_aryptr(); |
1450 | if (ta && ta->is_stable()) { |
1451 | // Erase stability property for alias analysis. |
1452 | tj = ta = ta->cast_to_stable(false); |
1453 | } |
1454 | if( ta && is_known_inst ) { |
1455 | if ( offset != Type::OffsetBot && |
1456 | offset > arrayOopDesc::length_offset_in_bytes() ) { |
1457 | offset = Type::OffsetBot; // Flatten constant access into array body only |
1458 | tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); |
1459 | } |
1460 | } else if( ta && _AliasLevel >= 2 ) { |
1461 | // For arrays indexed by constant indices, we flatten the alias |
1462 | // space to include all of the array body. Only the header, klass |
1463 | // and array length can be accessed un-aliased. |
1464 | if( offset != Type::OffsetBot ) { |
1465 | if( ta->const_oop() ) { // MethodData* or Method* |
1466 | offset = Type::OffsetBot; // Flatten constant access into array body |
1467 | tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); |
1468 | } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { |
1469 | // range is OK as-is. |
1470 | tj = ta = TypeAryPtr::RANGE; |
1471 | } else if( offset == oopDesc::klass_offset_in_bytes() ) { |
1472 | tj = TypeInstPtr::KLASS; // all klass loads look alike |
1473 | ta = TypeAryPtr::RANGE; // generic ignored junk |
1474 | ptr = TypePtr::BotPTR; |
1475 | } else if( offset == oopDesc::mark_offset_in_bytes() ) { |
1476 | tj = TypeInstPtr::MARK; |
1477 | ta = TypeAryPtr::RANGE; // generic ignored junk |
1478 | ptr = TypePtr::BotPTR; |
1479 | } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { |
1480 | ta = tj->isa_aryptr(); |
1481 | } else { // Random constant offset into array body |
1482 | offset = Type::OffsetBot; // Flatten constant access into array body |
1483 | tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); |
1484 | } |
1485 | } |
1486 | // Arrays of fixed size alias with arrays of unknown size. |
1487 | if (ta->size() != TypeInt::POS) { |
1488 | const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); |
1489 | tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); |
1490 | } |
1491 | // Arrays of known objects become arrays of unknown objects. |
1492 | if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { |
1493 | const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); |
1494 | tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); |
1495 | } |
1496 | if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { |
1497 | const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); |
1498 | tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); |
1499 | } |
1500 | // Arrays of bytes and of booleans both use 'bastore' and 'baload' so |
1501 | // cannot be distinguished by bytecode alone. |
1502 | if (ta->elem() == TypeInt::BOOL) { |
1503 | const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); |
1504 | ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); |
1505 | tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); |
1506 | } |
1507 | // During the 2nd round of IterGVN, NotNull castings are removed. |
1508 | // Make sure the Bottom and NotNull variants alias the same. |
1509 | // Also, make sure exact and non-exact variants alias the same. |
1510 | if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { |
1511 | tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); |
1512 | } |
1513 | } |
1514 | |
1515 | // Oop pointers need some flattening |
1516 | const TypeInstPtr *to = tj->isa_instptr(); |
1517 | if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { |
1518 | ciInstanceKlass *k = to->klass()->as_instance_klass(); |
1519 | if( ptr == TypePtr::Constant ) { |
1520 | if (to->klass() != ciEnv::current()->Class_klass() || |
1521 | offset < k->size_helper() * wordSize) { |
1522 | // No constant oop pointers (such as Strings); they alias with |
1523 | // unknown strings. |
1524 | assert(!is_known_inst, "not scalarizable allocation" ); |
1525 | tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); |
1526 | } |
1527 | } else if( is_known_inst ) { |
1528 | tj = to; // Keep NotNull and klass_is_exact for instance type |
1529 | } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { |
1530 | // During the 2nd round of IterGVN, NotNull castings are removed. |
1531 | // Make sure the Bottom and NotNull variants alias the same. |
1532 | // Also, make sure exact and non-exact variants alias the same. |
1533 | tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); |
1534 | } |
1535 | if (to->speculative() != NULL) { |
1536 | tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); |
1537 | } |
1538 | // Canonicalize the holder of this field |
1539 | if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { |
1540 | // First handle header references such as a LoadKlassNode, even if the |
1541 | // object's klass is unloaded at compile time (4965979). |
1542 | if (!is_known_inst) { // Do it only for non-instance types |
1543 | tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); |
1544 | } |
1545 | } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { |
1546 | to = tj->is_instptr(); |
1547 | } else if (offset < 0 || offset >= k->size_helper() * wordSize) { |
1548 | // Static fields are in the space above the normal instance |
1549 | // fields in the java.lang.Class instance. |
1550 | if (to->klass() != ciEnv::current()->Class_klass()) { |
1551 | to = NULL; |
1552 | tj = TypeOopPtr::BOTTOM; |
1553 | offset = tj->offset(); |
1554 | } |
1555 | } else { |
1556 | ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); |
1557 | if (!k->equals(canonical_holder) || tj->offset() != offset) { |
1558 | if( is_known_inst ) { |
1559 | tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); |
1560 | } else { |
1561 | tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); |
1562 | } |
1563 | } |
1564 | } |
1565 | } |
1566 | |
1567 | // Klass pointers to object array klasses need some flattening |
1568 | const TypeKlassPtr *tk = tj->isa_klassptr(); |
1569 | if( tk ) { |
1570 | // If we are referencing a field within a Klass, we need |
1571 | // to assume the worst case of an Object. Both exact and |
1572 | // inexact types must flatten to the same alias class so |
1573 | // use NotNull as the PTR. |
1574 | if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { |
1575 | |
1576 | tj = tk = TypeKlassPtr::make(TypePtr::NotNull, |
1577 | TypeKlassPtr::OBJECT->klass(), |
1578 | offset); |
1579 | } |
1580 | |
1581 | ciKlass* klass = tk->klass(); |
1582 | if( klass->is_obj_array_klass() ) { |
1583 | ciKlass* k = TypeAryPtr::OOPS->klass(); |
1584 | if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs |
1585 | k = TypeInstPtr::BOTTOM->klass(); |
1586 | tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); |
1587 | } |
1588 | |
1589 | // Check for precise loads from the primary supertype array and force them |
1590 | // to the supertype cache alias index. Check for generic array loads from |
1591 | // the primary supertype array and also force them to the supertype cache |
1592 | // alias index. Since the same load can reach both, we need to merge |
1593 | // these 2 disparate memories into the same alias class. Since the |
1594 | // primary supertype array is read-only, there's no chance of confusion |
1595 | // where we bypass an array load and an array store. |
1596 | int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); |
1597 | if (offset == Type::OffsetBot || |
1598 | (offset >= primary_supers_offset && |
1599 | offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || |
1600 | offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { |
1601 | offset = in_bytes(Klass::secondary_super_cache_offset()); |
1602 | tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); |
1603 | } |
1604 | } |
1605 | |
1606 | // Flatten all Raw pointers together. |
1607 | if (tj->base() == Type::RawPtr) |
1608 | tj = TypeRawPtr::BOTTOM; |
1609 | |
1610 | if (tj->base() == Type::AnyPtr) |
1611 | tj = TypePtr::BOTTOM; // An error, which the caller must check for. |
1612 | |
1613 | // Flatten all to bottom for now |
1614 | switch( _AliasLevel ) { |
1615 | case 0: |
1616 | tj = TypePtr::BOTTOM; |
1617 | break; |
1618 | case 1: // Flatten to: oop, static, field or array |
1619 | switch (tj->base()) { |
1620 | //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; |
1621 | case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; |
1622 | case Type::AryPtr: // do not distinguish arrays at all |
1623 | case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; |
1624 | case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; |
1625 | case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it |
1626 | default: ShouldNotReachHere(); |
1627 | } |
1628 | break; |
1629 | case 2: // No collapsing at level 2; keep all splits |
1630 | case 3: // No collapsing at level 3; keep all splits |
1631 | break; |
1632 | default: |
1633 | Unimplemented(); |
1634 | } |
1635 | |
1636 | offset = tj->offset(); |
1637 | assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); |
1638 | |
1639 | assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || |
1640 | (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || |
1641 | (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || |
1642 | (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || |
1643 | (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || |
1644 | (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || |
1645 | (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) || |
1646 | (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)), |
1647 | "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); |
1648 | assert( tj->ptr() != TypePtr::TopPTR && |
1649 | tj->ptr() != TypePtr::AnyNull && |
1650 | tj->ptr() != TypePtr::Null, "No imprecise addresses" ); |
1651 | // assert( tj->ptr() != TypePtr::Constant || |
1652 | // tj->base() == Type::RawPtr || |
1653 | // tj->base() == Type::KlassPtr, "No constant oop addresses" ); |
1654 | |
1655 | return tj; |
1656 | } |
1657 | |
1658 | void Compile::AliasType::Init(int i, const TypePtr* at) { |
1659 | _index = i; |
1660 | _adr_type = at; |
1661 | _field = NULL; |
1662 | _element = NULL; |
1663 | _is_rewritable = true; // default |
1664 | const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; |
1665 | if (atoop != NULL && atoop->is_known_instance()) { |
1666 | const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); |
1667 | _general_index = Compile::current()->get_alias_index(gt); |
1668 | } else { |
1669 | _general_index = 0; |
1670 | } |
1671 | } |
1672 | |
1673 | BasicType Compile::AliasType::basic_type() const { |
1674 | if (element() != NULL) { |
1675 | const Type* element = adr_type()->is_aryptr()->elem(); |
1676 | return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); |
1677 | } if (field() != NULL) { |
1678 | return field()->layout_type(); |
1679 | } else { |
1680 | return T_ILLEGAL; // unknown |
1681 | } |
1682 | } |
1683 | |
1684 | //---------------------------------print_on------------------------------------ |
1685 | #ifndef PRODUCT |
1686 | void Compile::AliasType::print_on(outputStream* st) { |
1687 | if (index() < 10) |
1688 | st->print("@ <%d> " , index()); |
1689 | else st->print("@ <%d>" , index()); |
1690 | st->print(is_rewritable() ? " " : " RO" ); |
1691 | int offset = adr_type()->offset(); |
1692 | if (offset == Type::OffsetBot) |
1693 | st->print(" +any" ); |
1694 | else st->print(" +%-3d" , offset); |
1695 | st->print(" in " ); |
1696 | adr_type()->dump_on(st); |
1697 | const TypeOopPtr* tjp = adr_type()->isa_oopptr(); |
1698 | if (field() != NULL && tjp) { |
1699 | if (tjp->klass() != field()->holder() || |
1700 | tjp->offset() != field()->offset_in_bytes()) { |
1701 | st->print(" != " ); |
1702 | field()->print(); |
1703 | st->print(" ***" ); |
1704 | } |
1705 | } |
1706 | } |
1707 | |
1708 | void print_alias_types() { |
1709 | Compile* C = Compile::current(); |
1710 | tty->print_cr("--- Alias types, AliasIdxBot .. %d" , C->num_alias_types()-1); |
1711 | for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { |
1712 | C->alias_type(idx)->print_on(tty); |
1713 | tty->cr(); |
1714 | } |
1715 | } |
1716 | #endif |
1717 | |
1718 | |
1719 | //----------------------------probe_alias_cache-------------------------------- |
1720 | Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { |
1721 | intptr_t key = (intptr_t) adr_type; |
1722 | key ^= key >> logAliasCacheSize; |
1723 | return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; |
1724 | } |
1725 | |
1726 | |
1727 | //-----------------------------grow_alias_types-------------------------------- |
1728 | void Compile::grow_alias_types() { |
1729 | const int old_ats = _max_alias_types; // how many before? |
1730 | const int new_ats = old_ats; // how many more? |
1731 | const int grow_ats = old_ats+new_ats; // how many now? |
1732 | _max_alias_types = grow_ats; |
1733 | _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); |
1734 | AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); |
1735 | Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); |
1736 | for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; |
1737 | } |
1738 | |
1739 | |
1740 | //--------------------------------find_alias_type------------------------------ |
1741 | Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { |
1742 | if (_AliasLevel == 0) |
1743 | return alias_type(AliasIdxBot); |
1744 | |
1745 | AliasCacheEntry* ace = probe_alias_cache(adr_type); |
1746 | if (ace->_adr_type == adr_type) { |
1747 | return alias_type(ace->_index); |
1748 | } |
1749 | |
1750 | // Handle special cases. |
1751 | if (adr_type == NULL) return alias_type(AliasIdxTop); |
1752 | if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); |
1753 | |
1754 | // Do it the slow way. |
1755 | const TypePtr* flat = flatten_alias_type(adr_type); |
1756 | |
1757 | #ifdef ASSERT |
1758 | { |
1759 | ResourceMark rm; |
1760 | assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s" , |
1761 | Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); |
1762 | assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s" , |
1763 | Type::str(adr_type)); |
1764 | if (flat->isa_oopptr() && !flat->isa_klassptr()) { |
1765 | const TypeOopPtr* foop = flat->is_oopptr(); |
1766 | // Scalarizable allocations have exact klass always. |
1767 | bool exact = !foop->klass_is_exact() || foop->is_known_instance(); |
1768 | const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); |
1769 | assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s" , |
1770 | Type::str(foop), Type::str(xoop)); |
1771 | } |
1772 | } |
1773 | #endif |
1774 | |
1775 | int idx = AliasIdxTop; |
1776 | for (int i = 0; i < num_alias_types(); i++) { |
1777 | if (alias_type(i)->adr_type() == flat) { |
1778 | idx = i; |
1779 | break; |
1780 | } |
1781 | } |
1782 | |
1783 | if (idx == AliasIdxTop) { |
1784 | if (no_create) return NULL; |
1785 | // Grow the array if necessary. |
1786 | if (_num_alias_types == _max_alias_types) grow_alias_types(); |
1787 | // Add a new alias type. |
1788 | idx = _num_alias_types++; |
1789 | _alias_types[idx]->Init(idx, flat); |
1790 | if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); |
1791 | if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); |
1792 | if (flat->isa_instptr()) { |
1793 | if (flat->offset() == java_lang_Class::klass_offset_in_bytes() |
1794 | && flat->is_instptr()->klass() == env()->Class_klass()) |
1795 | alias_type(idx)->set_rewritable(false); |
1796 | } |
1797 | if (flat->isa_aryptr()) { |
1798 | #ifdef ASSERT |
1799 | const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); |
1800 | // (T_BYTE has the weakest alignment and size restrictions...) |
1801 | assert(flat->offset() < header_size_min, "array body reference must be OffsetBot" ); |
1802 | #endif |
1803 | if (flat->offset() == TypePtr::OffsetBot) { |
1804 | alias_type(idx)->set_element(flat->is_aryptr()->elem()); |
1805 | } |
1806 | } |
1807 | if (flat->isa_klassptr()) { |
1808 | if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) |
1809 | alias_type(idx)->set_rewritable(false); |
1810 | if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) |
1811 | alias_type(idx)->set_rewritable(false); |
1812 | if (flat->offset() == in_bytes(Klass::access_flags_offset())) |
1813 | alias_type(idx)->set_rewritable(false); |
1814 | if (flat->offset() == in_bytes(Klass::java_mirror_offset())) |
1815 | alias_type(idx)->set_rewritable(false); |
1816 | } |
1817 | // %%% (We would like to finalize JavaThread::threadObj_offset(), |
1818 | // but the base pointer type is not distinctive enough to identify |
1819 | // references into JavaThread.) |
1820 | |
1821 | // Check for final fields. |
1822 | const TypeInstPtr* tinst = flat->isa_instptr(); |
1823 | if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { |
1824 | ciField* field; |
1825 | if (tinst->const_oop() != NULL && |
1826 | tinst->klass() == ciEnv::current()->Class_klass() && |
1827 | tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { |
1828 | // static field |
1829 | ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); |
1830 | field = k->get_field_by_offset(tinst->offset(), true); |
1831 | } else { |
1832 | ciInstanceKlass *k = tinst->klass()->as_instance_klass(); |
1833 | field = k->get_field_by_offset(tinst->offset(), false); |
1834 | } |
1835 | assert(field == NULL || |
1836 | original_field == NULL || |
1837 | (field->holder() == original_field->holder() && |
1838 | field->offset() == original_field->offset() && |
1839 | field->is_static() == original_field->is_static()), "wrong field?" ); |
1840 | // Set field() and is_rewritable() attributes. |
1841 | if (field != NULL) alias_type(idx)->set_field(field); |
1842 | } |
1843 | } |
1844 | |
1845 | // Fill the cache for next time. |
1846 | ace->_adr_type = adr_type; |
1847 | ace->_index = idx; |
1848 | assert(alias_type(adr_type) == alias_type(idx), "type must be installed" ); |
1849 | |
1850 | // Might as well try to fill the cache for the flattened version, too. |
1851 | AliasCacheEntry* face = probe_alias_cache(flat); |
1852 | if (face->_adr_type == NULL) { |
1853 | face->_adr_type = flat; |
1854 | face->_index = idx; |
1855 | assert(alias_type(flat) == alias_type(idx), "flat type must work too" ); |
1856 | } |
1857 | |
1858 | return alias_type(idx); |
1859 | } |
1860 | |
1861 | |
1862 | Compile::AliasType* Compile::alias_type(ciField* field) { |
1863 | const TypeOopPtr* t; |
1864 | if (field->is_static()) |
1865 | t = TypeInstPtr::make(field->holder()->java_mirror()); |
1866 | else |
1867 | t = TypeOopPtr::make_from_klass_raw(field->holder()); |
1868 | AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); |
1869 | assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct" ); |
1870 | return atp; |
1871 | } |
1872 | |
1873 | |
1874 | //------------------------------have_alias_type-------------------------------- |
1875 | bool Compile::have_alias_type(const TypePtr* adr_type) { |
1876 | AliasCacheEntry* ace = probe_alias_cache(adr_type); |
1877 | if (ace->_adr_type == adr_type) { |
1878 | return true; |
1879 | } |
1880 | |
1881 | // Handle special cases. |
1882 | if (adr_type == NULL) return true; |
1883 | if (adr_type == TypePtr::BOTTOM) return true; |
1884 | |
1885 | return find_alias_type(adr_type, true, NULL) != NULL; |
1886 | } |
1887 | |
1888 | //-----------------------------must_alias-------------------------------------- |
1889 | // True if all values of the given address type are in the given alias category. |
1890 | bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { |
1891 | if (alias_idx == AliasIdxBot) return true; // the universal category |
1892 | if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP |
1893 | if (alias_idx == AliasIdxTop) return false; // the empty category |
1894 | if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins |
1895 | |
1896 | // the only remaining possible overlap is identity |
1897 | int adr_idx = get_alias_index(adr_type); |
1898 | assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "" ); |
1899 | assert(adr_idx == alias_idx || |
1900 | (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM |
1901 | && adr_type != TypeOopPtr::BOTTOM), |
1902 | "should not be testing for overlap with an unsafe pointer" ); |
1903 | return adr_idx == alias_idx; |
1904 | } |
1905 | |
1906 | //------------------------------can_alias-------------------------------------- |
1907 | // True if any values of the given address type are in the given alias category. |
1908 | bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { |
1909 | if (alias_idx == AliasIdxTop) return false; // the empty category |
1910 | if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP |
1911 | if (alias_idx == AliasIdxBot) return true; // the universal category |
1912 | if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins |
1913 | |
1914 | // the only remaining possible overlap is identity |
1915 | int adr_idx = get_alias_index(adr_type); |
1916 | assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "" ); |
1917 | return adr_idx == alias_idx; |
1918 | } |
1919 | |
1920 | |
1921 | |
1922 | //---------------------------pop_warm_call------------------------------------- |
1923 | WarmCallInfo* Compile::pop_warm_call() { |
1924 | WarmCallInfo* wci = _warm_calls; |
1925 | if (wci != NULL) _warm_calls = wci->remove_from(wci); |
1926 | return wci; |
1927 | } |
1928 | |
1929 | //----------------------------Inline_Warm-------------------------------------- |
1930 | int Compile::Inline_Warm() { |
1931 | // If there is room, try to inline some more warm call sites. |
1932 | // %%% Do a graph index compaction pass when we think we're out of space? |
1933 | if (!InlineWarmCalls) return 0; |
1934 | |
1935 | int calls_made_hot = 0; |
1936 | int room_to_grow = NodeCountInliningCutoff - unique(); |
1937 | int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); |
1938 | int amount_grown = 0; |
1939 | WarmCallInfo* call; |
1940 | while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { |
1941 | int est_size = (int)call->size(); |
1942 | if (est_size > (room_to_grow - amount_grown)) { |
1943 | // This one won't fit anyway. Get rid of it. |
1944 | call->make_cold(); |
1945 | continue; |
1946 | } |
1947 | call->make_hot(); |
1948 | calls_made_hot++; |
1949 | amount_grown += est_size; |
1950 | amount_to_grow -= est_size; |
1951 | } |
1952 | |
1953 | if (calls_made_hot > 0) set_major_progress(); |
1954 | return calls_made_hot; |
1955 | } |
1956 | |
1957 | |
1958 | //----------------------------Finish_Warm-------------------------------------- |
1959 | void Compile::Finish_Warm() { |
1960 | if (!InlineWarmCalls) return; |
1961 | if (failing()) return; |
1962 | if (warm_calls() == NULL) return; |
1963 | |
1964 | // Clean up loose ends, if we are out of space for inlining. |
1965 | WarmCallInfo* call; |
1966 | while ((call = pop_warm_call()) != NULL) { |
1967 | call->make_cold(); |
1968 | } |
1969 | } |
1970 | |
1971 | //---------------------cleanup_loop_predicates----------------------- |
1972 | // Remove the opaque nodes that protect the predicates so that all unused |
1973 | // checks and uncommon_traps will be eliminated from the ideal graph |
1974 | void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { |
1975 | if (predicate_count()==0) return; |
1976 | for (int i = predicate_count(); i > 0; i--) { |
1977 | Node * n = predicate_opaque1_node(i-1); |
1978 | assert(n->Opcode() == Op_Opaque1, "must be" ); |
1979 | igvn.replace_node(n, n->in(1)); |
1980 | } |
1981 | assert(predicate_count()==0, "should be clean!" ); |
1982 | } |
1983 | |
1984 | void Compile::add_range_check_cast(Node* n) { |
1985 | assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency" ); |
1986 | assert(!_range_check_casts->contains(n), "duplicate entry in range check casts" ); |
1987 | _range_check_casts->append(n); |
1988 | } |
1989 | |
1990 | // Remove all range check dependent CastIINodes. |
1991 | void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { |
1992 | for (int i = range_check_cast_count(); i > 0; i--) { |
1993 | Node* cast = range_check_cast_node(i-1); |
1994 | assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency" ); |
1995 | igvn.replace_node(cast, cast->in(1)); |
1996 | } |
1997 | assert(range_check_cast_count() == 0, "should be empty" ); |
1998 | } |
1999 | |
2000 | void Compile::add_opaque4_node(Node* n) { |
2001 | assert(n->Opcode() == Op_Opaque4, "Opaque4 only" ); |
2002 | assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list" ); |
2003 | _opaque4_nodes->append(n); |
2004 | } |
2005 | |
2006 | // Remove all Opaque4 nodes. |
2007 | void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { |
2008 | for (int i = opaque4_count(); i > 0; i--) { |
2009 | Node* opaq = opaque4_node(i-1); |
2010 | assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only" ); |
2011 | igvn.replace_node(opaq, opaq->in(2)); |
2012 | } |
2013 | assert(opaque4_count() == 0, "should be empty" ); |
2014 | } |
2015 | |
2016 | // StringOpts and late inlining of string methods |
2017 | void Compile::inline_string_calls(bool parse_time) { |
2018 | { |
2019 | // remove useless nodes to make the usage analysis simpler |
2020 | ResourceMark rm; |
2021 | PhaseRemoveUseless pru(initial_gvn(), for_igvn()); |
2022 | } |
2023 | |
2024 | { |
2025 | ResourceMark rm; |
2026 | print_method(PHASE_BEFORE_STRINGOPTS, 3); |
2027 | PhaseStringOpts pso(initial_gvn(), for_igvn()); |
2028 | print_method(PHASE_AFTER_STRINGOPTS, 3); |
2029 | } |
2030 | |
2031 | // now inline anything that we skipped the first time around |
2032 | if (!parse_time) { |
2033 | _late_inlines_pos = _late_inlines.length(); |
2034 | } |
2035 | |
2036 | while (_string_late_inlines.length() > 0) { |
2037 | CallGenerator* cg = _string_late_inlines.pop(); |
2038 | cg->do_late_inline(); |
2039 | if (failing()) return; |
2040 | } |
2041 | _string_late_inlines.trunc_to(0); |
2042 | } |
2043 | |
2044 | // Late inlining of boxing methods |
2045 | void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { |
2046 | if (_boxing_late_inlines.length() > 0) { |
2047 | assert(has_boxed_value(), "inconsistent" ); |
2048 | |
2049 | PhaseGVN* gvn = initial_gvn(); |
2050 | set_inlining_incrementally(true); |
2051 | |
2052 | assert( igvn._worklist.size() == 0, "should be done with igvn" ); |
2053 | for_igvn()->clear(); |
2054 | gvn->replace_with(&igvn); |
2055 | |
2056 | _late_inlines_pos = _late_inlines.length(); |
2057 | |
2058 | while (_boxing_late_inlines.length() > 0) { |
2059 | CallGenerator* cg = _boxing_late_inlines.pop(); |
2060 | cg->do_late_inline(); |
2061 | if (failing()) return; |
2062 | } |
2063 | _boxing_late_inlines.trunc_to(0); |
2064 | |
2065 | inline_incrementally_cleanup(igvn); |
2066 | |
2067 | set_inlining_incrementally(false); |
2068 | } |
2069 | } |
2070 | |
2071 | bool Compile::inline_incrementally_one() { |
2072 | assert(IncrementalInline, "incremental inlining should be on" ); |
2073 | |
2074 | TracePhase tp("incrementalInline_inline" , &timers[_t_incrInline_inline]); |
2075 | set_inlining_progress(false); |
2076 | set_do_cleanup(false); |
2077 | int i = 0; |
2078 | for (; i <_late_inlines.length() && !inlining_progress(); i++) { |
2079 | CallGenerator* cg = _late_inlines.at(i); |
2080 | _late_inlines_pos = i+1; |
2081 | cg->do_late_inline(); |
2082 | if (failing()) return false; |
2083 | } |
2084 | int j = 0; |
2085 | for (; i < _late_inlines.length(); i++, j++) { |
2086 | _late_inlines.at_put(j, _late_inlines.at(i)); |
2087 | } |
2088 | _late_inlines.trunc_to(j); |
2089 | assert(inlining_progress() || _late_inlines.length() == 0, "" ); |
2090 | |
2091 | bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); |
2092 | |
2093 | set_inlining_progress(false); |
2094 | set_do_cleanup(false); |
2095 | return (_late_inlines.length() > 0) && !needs_cleanup; |
2096 | } |
2097 | |
2098 | void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { |
2099 | { |
2100 | TracePhase tp("incrementalInline_pru" , &timers[_t_incrInline_pru]); |
2101 | ResourceMark rm; |
2102 | PhaseRemoveUseless pru(initial_gvn(), for_igvn()); |
2103 | } |
2104 | { |
2105 | TracePhase tp("incrementalInline_igvn" , &timers[_t_incrInline_igvn]); |
2106 | igvn = PhaseIterGVN(initial_gvn()); |
2107 | igvn.optimize(); |
2108 | } |
2109 | } |
2110 | |
2111 | // Perform incremental inlining until bound on number of live nodes is reached |
2112 | void Compile::inline_incrementally(PhaseIterGVN& igvn) { |
2113 | TracePhase tp("incrementalInline" , &timers[_t_incrInline]); |
2114 | |
2115 | set_inlining_incrementally(true); |
2116 | uint low_live_nodes = 0; |
2117 | |
2118 | while (_late_inlines.length() > 0) { |
2119 | if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { |
2120 | if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { |
2121 | TracePhase tp("incrementalInline_ideal" , &timers[_t_incrInline_ideal]); |
2122 | // PhaseIdealLoop is expensive so we only try it once we are |
2123 | // out of live nodes and we only try it again if the previous |
2124 | // helped got the number of nodes down significantly |
2125 | PhaseIdealLoop::optimize(igvn, LoopOptsNone); |
2126 | if (failing()) return; |
2127 | low_live_nodes = live_nodes(); |
2128 | _major_progress = true; |
2129 | } |
2130 | |
2131 | if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { |
2132 | break; // finish |
2133 | } |
2134 | } |
2135 | |
2136 | for_igvn()->clear(); |
2137 | initial_gvn()->replace_with(&igvn); |
2138 | |
2139 | while (inline_incrementally_one()) { |
2140 | assert(!failing(), "inconsistent" ); |
2141 | } |
2142 | |
2143 | if (failing()) return; |
2144 | |
2145 | inline_incrementally_cleanup(igvn); |
2146 | |
2147 | if (failing()) return; |
2148 | } |
2149 | assert( igvn._worklist.size() == 0, "should be done with igvn" ); |
2150 | |
2151 | if (_string_late_inlines.length() > 0) { |
2152 | assert(has_stringbuilder(), "inconsistent" ); |
2153 | for_igvn()->clear(); |
2154 | initial_gvn()->replace_with(&igvn); |
2155 | |
2156 | inline_string_calls(false); |
2157 | |
2158 | if (failing()) return; |
2159 | |
2160 | inline_incrementally_cleanup(igvn); |
2161 | } |
2162 | |
2163 | set_inlining_incrementally(false); |
2164 | } |
2165 | |
2166 | |
2167 | bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { |
2168 | if(_loop_opts_cnt > 0) { |
2169 | debug_only( int cnt = 0; ); |
2170 | while(major_progress() && (_loop_opts_cnt > 0)) { |
2171 | TracePhase tp("idealLoop" , &timers[_t_idealLoop]); |
2172 | assert( cnt++ < 40, "infinite cycle in loop optimization" ); |
2173 | PhaseIdealLoop::optimize(igvn, mode); |
2174 | _loop_opts_cnt--; |
2175 | if (failing()) return false; |
2176 | if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); |
2177 | } |
2178 | } |
2179 | return true; |
2180 | } |
2181 | |
2182 | // Remove edges from "root" to each SafePoint at a backward branch. |
2183 | // They were inserted during parsing (see add_safepoint()) to make |
2184 | // infinite loops without calls or exceptions visible to root, i.e., |
2185 | // useful. |
2186 | void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { |
2187 | Node *r = root(); |
2188 | if (r != NULL) { |
2189 | for (uint i = r->req(); i < r->len(); ++i) { |
2190 | Node *n = r->in(i); |
2191 | if (n != NULL && n->is_SafePoint()) { |
2192 | r->rm_prec(i); |
2193 | if (n->outcnt() == 0) { |
2194 | igvn.remove_dead_node(n); |
2195 | } |
2196 | --i; |
2197 | } |
2198 | } |
2199 | } |
2200 | } |
2201 | |
2202 | //------------------------------Optimize--------------------------------------- |
2203 | // Given a graph, optimize it. |
2204 | void Compile::Optimize() { |
2205 | TracePhase tp("optimizer" , &timers[_t_optimizer]); |
2206 | |
2207 | #ifndef PRODUCT |
2208 | if (_directive->BreakAtCompileOption) { |
2209 | BREAKPOINT; |
2210 | } |
2211 | |
2212 | #endif |
2213 | |
2214 | BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); |
2215 | #ifdef ASSERT |
2216 | bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); |
2217 | #endif |
2218 | |
2219 | ResourceMark rm; |
2220 | |
2221 | print_inlining_reinit(); |
2222 | |
2223 | NOT_PRODUCT( verify_graph_edges(); ) |
2224 | |
2225 | print_method(PHASE_AFTER_PARSING); |
2226 | |
2227 | { |
2228 | // Iterative Global Value Numbering, including ideal transforms |
2229 | // Initialize IterGVN with types and values from parse-time GVN |
2230 | PhaseIterGVN igvn(initial_gvn()); |
2231 | #ifdef ASSERT |
2232 | _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); |
2233 | #endif |
2234 | { |
2235 | TracePhase tp("iterGVN" , &timers[_t_iterGVN]); |
2236 | igvn.optimize(); |
2237 | } |
2238 | |
2239 | if (failing()) return; |
2240 | |
2241 | print_method(PHASE_ITER_GVN1, 2); |
2242 | |
2243 | inline_incrementally(igvn); |
2244 | |
2245 | print_method(PHASE_INCREMENTAL_INLINE, 2); |
2246 | |
2247 | if (failing()) return; |
2248 | |
2249 | if (eliminate_boxing()) { |
2250 | // Inline valueOf() methods now. |
2251 | inline_boxing_calls(igvn); |
2252 | |
2253 | if (AlwaysIncrementalInline) { |
2254 | inline_incrementally(igvn); |
2255 | } |
2256 | |
2257 | print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); |
2258 | |
2259 | if (failing()) return; |
2260 | } |
2261 | |
2262 | // Now that all inlining is over, cut edge from root to loop |
2263 | // safepoints |
2264 | remove_root_to_sfpts_edges(igvn); |
2265 | |
2266 | // Remove the speculative part of types and clean up the graph from |
2267 | // the extra CastPP nodes whose only purpose is to carry them. Do |
2268 | // that early so that optimizations are not disrupted by the extra |
2269 | // CastPP nodes. |
2270 | remove_speculative_types(igvn); |
2271 | |
2272 | // No more new expensive nodes will be added to the list from here |
2273 | // so keep only the actual candidates for optimizations. |
2274 | cleanup_expensive_nodes(igvn); |
2275 | |
2276 | if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { |
2277 | Compile::TracePhase tp("" , &timers[_t_renumberLive]); |
2278 | initial_gvn()->replace_with(&igvn); |
2279 | for_igvn()->clear(); |
2280 | Unique_Node_List new_worklist(C->comp_arena()); |
2281 | { |
2282 | ResourceMark rm; |
2283 | PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); |
2284 | } |
2285 | set_for_igvn(&new_worklist); |
2286 | igvn = PhaseIterGVN(initial_gvn()); |
2287 | igvn.optimize(); |
2288 | } |
2289 | |
2290 | // Perform escape analysis |
2291 | if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { |
2292 | if (has_loops()) { |
2293 | // Cleanup graph (remove dead nodes). |
2294 | TracePhase tp("idealLoop" , &timers[_t_idealLoop]); |
2295 | PhaseIdealLoop::optimize(igvn, LoopOptsNone); |
2296 | if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); |
2297 | if (failing()) return; |
2298 | } |
2299 | ConnectionGraph::do_analysis(this, &igvn); |
2300 | |
2301 | if (failing()) return; |
2302 | |
2303 | // Optimize out fields loads from scalar replaceable allocations. |
2304 | igvn.optimize(); |
2305 | print_method(PHASE_ITER_GVN_AFTER_EA, 2); |
2306 | |
2307 | if (failing()) return; |
2308 | |
2309 | if (congraph() != NULL && macro_count() > 0) { |
2310 | TracePhase tp("macroEliminate" , &timers[_t_macroEliminate]); |
2311 | PhaseMacroExpand mexp(igvn); |
2312 | mexp.eliminate_macro_nodes(); |
2313 | igvn.set_delay_transform(false); |
2314 | |
2315 | igvn.optimize(); |
2316 | print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); |
2317 | |
2318 | if (failing()) return; |
2319 | } |
2320 | } |
2321 | |
2322 | // Loop transforms on the ideal graph. Range Check Elimination, |
2323 | // peeling, unrolling, etc. |
2324 | |
2325 | // Set loop opts counter |
2326 | if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { |
2327 | { |
2328 | TracePhase tp("idealLoop" , &timers[_t_idealLoop]); |
2329 | PhaseIdealLoop::optimize(igvn, LoopOptsDefault); |
2330 | _loop_opts_cnt--; |
2331 | if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); |
2332 | if (failing()) return; |
2333 | } |
2334 | // Loop opts pass if partial peeling occurred in previous pass |
2335 | if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { |
2336 | TracePhase tp("idealLoop" , &timers[_t_idealLoop]); |
2337 | PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); |
2338 | _loop_opts_cnt--; |
2339 | if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); |
2340 | if (failing()) return; |
2341 | } |
2342 | // Loop opts pass for loop-unrolling before CCP |
2343 | if(major_progress() && (_loop_opts_cnt > 0)) { |
2344 | TracePhase tp("idealLoop" , &timers[_t_idealLoop]); |
2345 | PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); |
2346 | _loop_opts_cnt--; |
2347 | if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); |
2348 | } |
2349 | if (!failing()) { |
2350 | // Verify that last round of loop opts produced a valid graph |
2351 | TracePhase tp("idealLoopVerify" , &timers[_t_idealLoopVerify]); |
2352 | PhaseIdealLoop::verify(igvn); |
2353 | } |
2354 | } |
2355 | if (failing()) return; |
2356 | |
2357 | // Conditional Constant Propagation; |
2358 | PhaseCCP ccp( &igvn ); |
2359 | assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)" ); |
2360 | { |
2361 | TracePhase tp("ccp" , &timers[_t_ccp]); |
2362 | ccp.do_transform(); |
2363 | } |
2364 | print_method(PHASE_CPP1, 2); |
2365 | |
2366 | assert( true, "Break here to ccp.dump_old2new_map()" ); |
2367 | |
2368 | // Iterative Global Value Numbering, including ideal transforms |
2369 | { |
2370 | TracePhase tp("iterGVN2" , &timers[_t_iterGVN2]); |
2371 | igvn = ccp; |
2372 | igvn.optimize(); |
2373 | } |
2374 | print_method(PHASE_ITER_GVN2, 2); |
2375 | |
2376 | if (failing()) return; |
2377 | |
2378 | // Loop transforms on the ideal graph. Range Check Elimination, |
2379 | // peeling, unrolling, etc. |
2380 | if (!optimize_loops(igvn, LoopOptsDefault)) { |
2381 | return; |
2382 | } |
2383 | |
2384 | if (failing()) return; |
2385 | |
2386 | // Ensure that major progress is now clear |
2387 | C->clear_major_progress(); |
2388 | |
2389 | { |
2390 | // Verify that all previous optimizations produced a valid graph |
2391 | // at least to this point, even if no loop optimizations were done. |
2392 | TracePhase tp("idealLoopVerify" , &timers[_t_idealLoopVerify]); |
2393 | PhaseIdealLoop::verify(igvn); |
2394 | } |
2395 | |
2396 | if (range_check_cast_count() > 0) { |
2397 | // No more loop optimizations. Remove all range check dependent CastIINodes. |
2398 | C->remove_range_check_casts(igvn); |
2399 | igvn.optimize(); |
2400 | } |
2401 | |
2402 | #ifdef ASSERT |
2403 | bs->verify_gc_barriers(this, BarrierSetC2::BeforeLateInsertion); |
2404 | #endif |
2405 | |
2406 | bs->barrier_insertion_phase(C, igvn); |
2407 | if (failing()) return; |
2408 | |
2409 | #ifdef ASSERT |
2410 | bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); |
2411 | #endif |
2412 | |
2413 | { |
2414 | TracePhase tp("macroExpand" , &timers[_t_macroExpand]); |
2415 | PhaseMacroExpand mex(igvn); |
2416 | if (mex.expand_macro_nodes()) { |
2417 | assert(failing(), "must bail out w/ explicit message" ); |
2418 | return; |
2419 | } |
2420 | print_method(PHASE_MACRO_EXPANSION, 2); |
2421 | } |
2422 | |
2423 | { |
2424 | TracePhase tp("barrierExpand" , &timers[_t_barrierExpand]); |
2425 | if (bs->expand_barriers(this, igvn)) { |
2426 | assert(failing(), "must bail out w/ explicit message" ); |
2427 | return; |
2428 | } |
2429 | print_method(PHASE_BARRIER_EXPANSION, 2); |
2430 | } |
2431 | |
2432 | if (opaque4_count() > 0) { |
2433 | C->remove_opaque4_nodes(igvn); |
2434 | igvn.optimize(); |
2435 | } |
2436 | |
2437 | DEBUG_ONLY( _modified_nodes = NULL; ) |
2438 | } // (End scope of igvn; run destructor if necessary for asserts.) |
2439 | |
2440 | process_print_inlining(); |
2441 | // A method with only infinite loops has no edges entering loops from root |
2442 | { |
2443 | TracePhase tp("graphReshape" , &timers[_t_graphReshaping]); |
2444 | if (final_graph_reshaping()) { |
2445 | assert(failing(), "must bail out w/ explicit message" ); |
2446 | return; |
2447 | } |
2448 | } |
2449 | |
2450 | print_method(PHASE_OPTIMIZE_FINISHED, 2); |
2451 | } |
2452 | |
2453 | |
2454 | //------------------------------Code_Gen--------------------------------------- |
2455 | // Given a graph, generate code for it |
2456 | void Compile::Code_Gen() { |
2457 | if (failing()) { |
2458 | return; |
2459 | } |
2460 | |
2461 | // Perform instruction selection. You might think we could reclaim Matcher |
2462 | // memory PDQ, but actually the Matcher is used in generating spill code. |
2463 | // Internals of the Matcher (including some VectorSets) must remain live |
2464 | // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage |
2465 | // set a bit in reclaimed memory. |
2466 | |
2467 | // In debug mode can dump m._nodes.dump() for mapping of ideal to machine |
2468 | // nodes. Mapping is only valid at the root of each matched subtree. |
2469 | NOT_PRODUCT( verify_graph_edges(); ) |
2470 | |
2471 | Matcher matcher; |
2472 | _matcher = &matcher; |
2473 | { |
2474 | TracePhase tp("matcher" , &timers[_t_matcher]); |
2475 | matcher.match(); |
2476 | } |
2477 | // In debug mode can dump m._nodes.dump() for mapping of ideal to machine |
2478 | // nodes. Mapping is only valid at the root of each matched subtree. |
2479 | NOT_PRODUCT( verify_graph_edges(); ) |
2480 | |
2481 | // If you have too many nodes, or if matching has failed, bail out |
2482 | check_node_count(0, "out of nodes matching instructions" ); |
2483 | if (failing()) { |
2484 | return; |
2485 | } |
2486 | |
2487 | print_method(PHASE_MATCHING, 2); |
2488 | |
2489 | // Build a proper-looking CFG |
2490 | PhaseCFG cfg(node_arena(), root(), matcher); |
2491 | _cfg = &cfg; |
2492 | { |
2493 | TracePhase tp("scheduler" , &timers[_t_scheduler]); |
2494 | bool success = cfg.do_global_code_motion(); |
2495 | if (!success) { |
2496 | return; |
2497 | } |
2498 | |
2499 | print_method(PHASE_GLOBAL_CODE_MOTION, 2); |
2500 | NOT_PRODUCT( verify_graph_edges(); ) |
2501 | debug_only( cfg.verify(); ) |
2502 | } |
2503 | |
2504 | PhaseChaitin regalloc(unique(), cfg, matcher, false); |
2505 | _regalloc = ®alloc; |
2506 | { |
2507 | TracePhase tp("regalloc" , &timers[_t_registerAllocation]); |
2508 | // Perform register allocation. After Chaitin, use-def chains are |
2509 | // no longer accurate (at spill code) and so must be ignored. |
2510 | // Node->LRG->reg mappings are still accurate. |
2511 | _regalloc->Register_Allocate(); |
2512 | |
2513 | // Bail out if the allocator builds too many nodes |
2514 | if (failing()) { |
2515 | return; |
2516 | } |
2517 | } |
2518 | |
2519 | // Prior to register allocation we kept empty basic blocks in case the |
2520 | // the allocator needed a place to spill. After register allocation we |
2521 | // are not adding any new instructions. If any basic block is empty, we |
2522 | // can now safely remove it. |
2523 | { |
2524 | TracePhase tp("blockOrdering" , &timers[_t_blockOrdering]); |
2525 | cfg.remove_empty_blocks(); |
2526 | if (do_freq_based_layout()) { |
2527 | PhaseBlockLayout layout(cfg); |
2528 | } else { |
2529 | cfg.set_loop_alignment(); |
2530 | } |
2531 | cfg.fixup_flow(); |
2532 | } |
2533 | |
2534 | // Apply peephole optimizations |
2535 | if( OptoPeephole ) { |
2536 | TracePhase tp("peephole" , &timers[_t_peephole]); |
2537 | PhasePeephole peep( _regalloc, cfg); |
2538 | peep.do_transform(); |
2539 | } |
2540 | |
2541 | // Do late expand if CPU requires this. |
2542 | if (Matcher::require_postalloc_expand) { |
2543 | TracePhase tp("postalloc_expand" , &timers[_t_postalloc_expand]); |
2544 | cfg.postalloc_expand(_regalloc); |
2545 | } |
2546 | |
2547 | // Convert Nodes to instruction bits in a buffer |
2548 | { |
2549 | TraceTime tp("output" , &timers[_t_output], CITime); |
2550 | Output(); |
2551 | } |
2552 | |
2553 | print_method(PHASE_FINAL_CODE); |
2554 | |
2555 | // He's dead, Jim. |
2556 | _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); |
2557 | _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); |
2558 | } |
2559 | |
2560 | |
2561 | //------------------------------dump_asm--------------------------------------- |
2562 | // Dump formatted assembly |
2563 | #if defined(SUPPORT_OPTO_ASSEMBLY) |
2564 | void Compile::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) { |
2565 | |
2566 | int pc_digits = 3; // #chars required for pc |
2567 | int sb_chars = 3; // #chars for "start bundle" indicator |
2568 | int tab_size = 8; |
2569 | if (pcs != NULL) { |
2570 | int max_pc = 0; |
2571 | for (uint i = 0; i < pc_limit; i++) { |
2572 | max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc; |
2573 | } |
2574 | pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc |
2575 | } |
2576 | int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size; |
2577 | |
2578 | bool cut_short = false; |
2579 | st->print_cr("#" ); |
2580 | st->print("# " ); _tf->dump_on(st); st->cr(); |
2581 | st->print_cr("#" ); |
2582 | |
2583 | // For all blocks |
2584 | int pc = 0x0; // Program counter |
2585 | char starts_bundle = ' '; |
2586 | _regalloc->dump_frame(); |
2587 | |
2588 | Node *n = NULL; |
2589 | for (uint i = 0; i < _cfg->number_of_blocks(); i++) { |
2590 | if (VMThread::should_terminate()) { |
2591 | cut_short = true; |
2592 | break; |
2593 | } |
2594 | Block* block = _cfg->get_block(i); |
2595 | if (block->is_connector() && !Verbose) { |
2596 | continue; |
2597 | } |
2598 | n = block->head(); |
2599 | if ((pcs != NULL) && (n->_idx < pc_limit)) { |
2600 | pc = pcs[n->_idx]; |
2601 | st->print("%*.*x" , pc_digits, pc_digits, pc); |
2602 | } |
2603 | st->fill_to(prefix_len); |
2604 | block->dump_head(_cfg, st); |
2605 | if (block->is_connector()) { |
2606 | st->fill_to(prefix_len); |
2607 | st->print_cr("# Empty connector block" ); |
2608 | } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { |
2609 | st->fill_to(prefix_len); |
2610 | st->print_cr("# Block is sole successor of call" ); |
2611 | } |
2612 | |
2613 | // For all instructions |
2614 | Node *delay = NULL; |
2615 | for (uint j = 0; j < block->number_of_nodes(); j++) { |
2616 | if (VMThread::should_terminate()) { |
2617 | cut_short = true; |
2618 | break; |
2619 | } |
2620 | n = block->get_node(j); |
2621 | if (valid_bundle_info(n)) { |
2622 | Bundle* bundle = node_bundling(n); |
2623 | if (bundle->used_in_unconditional_delay()) { |
2624 | delay = n; |
2625 | continue; |
2626 | } |
2627 | if (bundle->starts_bundle()) { |
2628 | starts_bundle = '+'; |
2629 | } |
2630 | } |
2631 | |
2632 | if (WizardMode) { |
2633 | n->dump(); |
2634 | } |
2635 | |
2636 | if( !n->is_Region() && // Dont print in the Assembly |
2637 | !n->is_Phi() && // a few noisely useless nodes |
2638 | !n->is_Proj() && |
2639 | !n->is_MachTemp() && |
2640 | !n->is_SafePointScalarObject() && |
2641 | !n->is_Catch() && // Would be nice to print exception table targets |
2642 | !n->is_MergeMem() && // Not very interesting |
2643 | !n->is_top() && // Debug info table constants |
2644 | !(n->is_Con() && !n->is_Mach())// Debug info table constants |
2645 | ) { |
2646 | if ((pcs != NULL) && (n->_idx < pc_limit)) { |
2647 | pc = pcs[n->_idx]; |
2648 | st->print("%*.*x" , pc_digits, pc_digits, pc); |
2649 | } else { |
2650 | st->fill_to(pc_digits); |
2651 | } |
2652 | st->print(" %c " , starts_bundle); |
2653 | starts_bundle = ' '; |
2654 | st->fill_to(prefix_len); |
2655 | n->format(_regalloc, st); |
2656 | st->cr(); |
2657 | } |
2658 | |
2659 | // If we have an instruction with a delay slot, and have seen a delay, |
2660 | // then back up and print it |
2661 | if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { |
2662 | // Coverity finding - Explicit null dereferenced. |
2663 | guarantee(delay != NULL, "no unconditional delay instruction" ); |
2664 | if (WizardMode) delay->dump(); |
2665 | |
2666 | if (node_bundling(delay)->starts_bundle()) |
2667 | starts_bundle = '+'; |
2668 | if ((pcs != NULL) && (n->_idx < pc_limit)) { |
2669 | pc = pcs[n->_idx]; |
2670 | st->print("%*.*x" , pc_digits, pc_digits, pc); |
2671 | } else { |
2672 | st->fill_to(pc_digits); |
2673 | } |
2674 | st->print(" %c " , starts_bundle); |
2675 | starts_bundle = ' '; |
2676 | st->fill_to(prefix_len); |
2677 | delay->format(_regalloc, st); |
2678 | st->cr(); |
2679 | delay = NULL; |
2680 | } |
2681 | |
2682 | // Dump the exception table as well |
2683 | if( n->is_Catch() && (Verbose || WizardMode) ) { |
2684 | // Print the exception table for this offset |
2685 | _handler_table.print_subtable_for(pc); |
2686 | } |
2687 | st->bol(); // Make sure we start on a new line |
2688 | } |
2689 | st->cr(); // one empty line between blocks |
2690 | assert(cut_short || delay == NULL, "no unconditional delay branch" ); |
2691 | } // End of per-block dump |
2692 | |
2693 | if (cut_short) st->print_cr("*** disassembly is cut short ***" ); |
2694 | } |
2695 | #endif |
2696 | |
2697 | //------------------------------Final_Reshape_Counts--------------------------- |
2698 | // This class defines counters to help identify when a method |
2699 | // may/must be executed using hardware with only 24-bit precision. |
2700 | struct Final_Reshape_Counts : public StackObj { |
2701 | int _call_count; // count non-inlined 'common' calls |
2702 | int _float_count; // count float ops requiring 24-bit precision |
2703 | int _double_count; // count double ops requiring more precision |
2704 | int _java_call_count; // count non-inlined 'java' calls |
2705 | int _inner_loop_count; // count loops which need alignment |
2706 | VectorSet _visited; // Visitation flags |
2707 | Node_List _tests; // Set of IfNodes & PCTableNodes |
2708 | |
2709 | Final_Reshape_Counts() : |
2710 | _call_count(0), _float_count(0), _double_count(0), |
2711 | _java_call_count(0), _inner_loop_count(0), |
2712 | _visited( Thread::current()->resource_area() ) { } |
2713 | |
2714 | void inc_call_count () { _call_count ++; } |
2715 | void inc_float_count () { _float_count ++; } |
2716 | void inc_double_count() { _double_count++; } |
2717 | void inc_java_call_count() { _java_call_count++; } |
2718 | void inc_inner_loop_count() { _inner_loop_count++; } |
2719 | |
2720 | int get_call_count () const { return _call_count ; } |
2721 | int get_float_count () const { return _float_count ; } |
2722 | int get_double_count() const { return _double_count; } |
2723 | int get_java_call_count() const { return _java_call_count; } |
2724 | int get_inner_loop_count() const { return _inner_loop_count; } |
2725 | }; |
2726 | |
2727 | #ifdef ASSERT |
2728 | static bool oop_offset_is_sane(const TypeInstPtr* tp) { |
2729 | ciInstanceKlass *k = tp->klass()->as_instance_klass(); |
2730 | // Make sure the offset goes inside the instance layout. |
2731 | return k->contains_field_offset(tp->offset()); |
2732 | // Note that OffsetBot and OffsetTop are very negative. |
2733 | } |
2734 | #endif |
2735 | |
2736 | // Eliminate trivially redundant StoreCMs and accumulate their |
2737 | // precedence edges. |
2738 | void Compile::eliminate_redundant_card_marks(Node* n) { |
2739 | assert(n->Opcode() == Op_StoreCM, "expected StoreCM" ); |
2740 | if (n->in(MemNode::Address)->outcnt() > 1) { |
2741 | // There are multiple users of the same address so it might be |
2742 | // possible to eliminate some of the StoreCMs |
2743 | Node* mem = n->in(MemNode::Memory); |
2744 | Node* adr = n->in(MemNode::Address); |
2745 | Node* val = n->in(MemNode::ValueIn); |
2746 | Node* prev = n; |
2747 | bool done = false; |
2748 | // Walk the chain of StoreCMs eliminating ones that match. As |
2749 | // long as it's a chain of single users then the optimization is |
2750 | // safe. Eliminating partially redundant StoreCMs would require |
2751 | // cloning copies down the other paths. |
2752 | while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { |
2753 | if (adr == mem->in(MemNode::Address) && |
2754 | val == mem->in(MemNode::ValueIn)) { |
2755 | // redundant StoreCM |
2756 | if (mem->req() > MemNode::OopStore) { |
2757 | // Hasn't been processed by this code yet. |
2758 | n->add_prec(mem->in(MemNode::OopStore)); |
2759 | } else { |
2760 | // Already converted to precedence edge |
2761 | for (uint i = mem->req(); i < mem->len(); i++) { |
2762 | // Accumulate any precedence edges |
2763 | if (mem->in(i) != NULL) { |
2764 | n->add_prec(mem->in(i)); |
2765 | } |
2766 | } |
2767 | // Everything above this point has been processed. |
2768 | done = true; |
2769 | } |
2770 | // Eliminate the previous StoreCM |
2771 | prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); |
2772 | assert(mem->outcnt() == 0, "should be dead" ); |
2773 | mem->disconnect_inputs(NULL, this); |
2774 | } else { |
2775 | prev = mem; |
2776 | } |
2777 | mem = prev->in(MemNode::Memory); |
2778 | } |
2779 | } |
2780 | } |
2781 | |
2782 | //------------------------------final_graph_reshaping_impl---------------------- |
2783 | // Implement items 1-5 from final_graph_reshaping below. |
2784 | void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { |
2785 | |
2786 | if ( n->outcnt() == 0 ) return; // dead node |
2787 | uint nop = n->Opcode(); |
2788 | |
2789 | // Check for 2-input instruction with "last use" on right input. |
2790 | // Swap to left input. Implements item (2). |
2791 | if( n->req() == 3 && // two-input instruction |
2792 | n->in(1)->outcnt() > 1 && // left use is NOT a last use |
2793 | (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop |
2794 | n->in(2)->outcnt() == 1 &&// right use IS a last use |
2795 | !n->in(2)->is_Con() ) { // right use is not a constant |
2796 | // Check for commutative opcode |
2797 | switch( nop ) { |
2798 | case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: |
2799 | case Op_MaxI: case Op_MinI: |
2800 | case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: |
2801 | case Op_AndL: case Op_XorL: case Op_OrL: |
2802 | case Op_AndI: case Op_XorI: case Op_OrI: { |
2803 | // Move "last use" input to left by swapping inputs |
2804 | n->swap_edges(1, 2); |
2805 | break; |
2806 | } |
2807 | default: |
2808 | break; |
2809 | } |
2810 | } |
2811 | |
2812 | #ifdef ASSERT |
2813 | if( n->is_Mem() ) { |
2814 | int alias_idx = get_alias_index(n->as_Mem()->adr_type()); |
2815 | assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || |
2816 | // oop will be recorded in oop map if load crosses safepoint |
2817 | n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || |
2818 | LoadNode::is_immutable_value(n->in(MemNode::Address))), |
2819 | "raw memory operations should have control edge" ); |
2820 | } |
2821 | if (n->is_MemBar()) { |
2822 | MemBarNode* mb = n->as_MemBar(); |
2823 | if (mb->trailing_store() || mb->trailing_load_store()) { |
2824 | assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair" ); |
2825 | Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); |
2826 | assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || |
2827 | (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op" ); |
2828 | } else if (mb->leading()) { |
2829 | assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair" ); |
2830 | } |
2831 | } |
2832 | #endif |
2833 | // Count FPU ops and common calls, implements item (3) |
2834 | bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop); |
2835 | if (!gc_handled) { |
2836 | final_graph_reshaping_main_switch(n, frc, nop); |
2837 | } |
2838 | |
2839 | // Collect CFG split points |
2840 | if (n->is_MultiBranch() && !n->is_RangeCheck()) { |
2841 | frc._tests.push(n); |
2842 | } |
2843 | } |
2844 | |
2845 | void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) { |
2846 | switch( nop ) { |
2847 | // Count all float operations that may use FPU |
2848 | case Op_AddF: |
2849 | case Op_SubF: |
2850 | case Op_MulF: |
2851 | case Op_DivF: |
2852 | case Op_NegF: |
2853 | case Op_ModF: |
2854 | case Op_ConvI2F: |
2855 | case Op_ConF: |
2856 | case Op_CmpF: |
2857 | case Op_CmpF3: |
2858 | // case Op_ConvL2F: // longs are split into 32-bit halves |
2859 | frc.inc_float_count(); |
2860 | break; |
2861 | |
2862 | case Op_ConvF2D: |
2863 | case Op_ConvD2F: |
2864 | frc.inc_float_count(); |
2865 | frc.inc_double_count(); |
2866 | break; |
2867 | |
2868 | // Count all double operations that may use FPU |
2869 | case Op_AddD: |
2870 | case Op_SubD: |
2871 | case Op_MulD: |
2872 | case Op_DivD: |
2873 | case Op_NegD: |
2874 | case Op_ModD: |
2875 | case Op_ConvI2D: |
2876 | case Op_ConvD2I: |
2877 | // case Op_ConvL2D: // handled by leaf call |
2878 | // case Op_ConvD2L: // handled by leaf call |
2879 | case Op_ConD: |
2880 | case Op_CmpD: |
2881 | case Op_CmpD3: |
2882 | frc.inc_double_count(); |
2883 | break; |
2884 | case Op_Opaque1: // Remove Opaque Nodes before matching |
2885 | case Op_Opaque2: // Remove Opaque Nodes before matching |
2886 | case Op_Opaque3: |
2887 | n->subsume_by(n->in(1), this); |
2888 | break; |
2889 | case Op_CallStaticJava: |
2890 | case Op_CallJava: |
2891 | case Op_CallDynamicJava: |
2892 | frc.inc_java_call_count(); // Count java call site; |
2893 | case Op_CallRuntime: |
2894 | case Op_CallLeaf: |
2895 | case Op_CallLeafNoFP: { |
2896 | assert (n->is_Call(), "" ); |
2897 | CallNode *call = n->as_Call(); |
2898 | // Count call sites where the FP mode bit would have to be flipped. |
2899 | // Do not count uncommon runtime calls: |
2900 | // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, |
2901 | // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... |
2902 | if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { |
2903 | frc.inc_call_count(); // Count the call site |
2904 | } else { // See if uncommon argument is shared |
2905 | Node *n = call->in(TypeFunc::Parms); |
2906 | int nop = n->Opcode(); |
2907 | // Clone shared simple arguments to uncommon calls, item (1). |
2908 | if (n->outcnt() > 1 && |
2909 | !n->is_Proj() && |
2910 | nop != Op_CreateEx && |
2911 | nop != Op_CheckCastPP && |
2912 | nop != Op_DecodeN && |
2913 | nop != Op_DecodeNKlass && |
2914 | !n->is_Mem() && |
2915 | !n->is_Phi()) { |
2916 | Node *x = n->clone(); |
2917 | call->set_req(TypeFunc::Parms, x); |
2918 | } |
2919 | } |
2920 | break; |
2921 | } |
2922 | |
2923 | case Op_StoreD: |
2924 | case Op_LoadD: |
2925 | case Op_LoadD_unaligned: |
2926 | frc.inc_double_count(); |
2927 | goto handle_mem; |
2928 | case Op_StoreF: |
2929 | case Op_LoadF: |
2930 | frc.inc_float_count(); |
2931 | goto handle_mem; |
2932 | |
2933 | case Op_StoreCM: |
2934 | { |
2935 | // Convert OopStore dependence into precedence edge |
2936 | Node* prec = n->in(MemNode::OopStore); |
2937 | n->del_req(MemNode::OopStore); |
2938 | n->add_prec(prec); |
2939 | eliminate_redundant_card_marks(n); |
2940 | } |
2941 | |
2942 | // fall through |
2943 | |
2944 | case Op_StoreB: |
2945 | case Op_StoreC: |
2946 | case Op_StorePConditional: |
2947 | case Op_StoreI: |
2948 | case Op_StoreL: |
2949 | case Op_StoreIConditional: |
2950 | case Op_StoreLConditional: |
2951 | case Op_CompareAndSwapB: |
2952 | case Op_CompareAndSwapS: |
2953 | case Op_CompareAndSwapI: |
2954 | case Op_CompareAndSwapL: |
2955 | case Op_CompareAndSwapP: |
2956 | case Op_CompareAndSwapN: |
2957 | case Op_WeakCompareAndSwapB: |
2958 | case Op_WeakCompareAndSwapS: |
2959 | case Op_WeakCompareAndSwapI: |
2960 | case Op_WeakCompareAndSwapL: |
2961 | case Op_WeakCompareAndSwapP: |
2962 | case Op_WeakCompareAndSwapN: |
2963 | case Op_CompareAndExchangeB: |
2964 | case Op_CompareAndExchangeS: |
2965 | case Op_CompareAndExchangeI: |
2966 | case Op_CompareAndExchangeL: |
2967 | case Op_CompareAndExchangeP: |
2968 | case Op_CompareAndExchangeN: |
2969 | case Op_GetAndAddS: |
2970 | case Op_GetAndAddB: |
2971 | case Op_GetAndAddI: |
2972 | case Op_GetAndAddL: |
2973 | case Op_GetAndSetS: |
2974 | case Op_GetAndSetB: |
2975 | case Op_GetAndSetI: |
2976 | case Op_GetAndSetL: |
2977 | case Op_GetAndSetP: |
2978 | case Op_GetAndSetN: |
2979 | case Op_StoreP: |
2980 | case Op_StoreN: |
2981 | case Op_StoreNKlass: |
2982 | case Op_LoadB: |
2983 | case Op_LoadUB: |
2984 | case Op_LoadUS: |
2985 | case Op_LoadI: |
2986 | case Op_LoadKlass: |
2987 | case Op_LoadNKlass: |
2988 | case Op_LoadL: |
2989 | case Op_LoadL_unaligned: |
2990 | case Op_LoadPLocked: |
2991 | case Op_LoadP: |
2992 | case Op_LoadN: |
2993 | case Op_LoadRange: |
2994 | case Op_LoadS: { |
2995 | handle_mem: |
2996 | #ifdef ASSERT |
2997 | if( VerifyOptoOopOffsets ) { |
2998 | MemNode* mem = n->as_Mem(); |
2999 | // Check to see if address types have grounded out somehow. |
3000 | const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); |
3001 | assert( !tp || oop_offset_is_sane(tp), "" ); |
3002 | } |
3003 | #endif |
3004 | break; |
3005 | } |
3006 | |
3007 | case Op_AddP: { // Assert sane base pointers |
3008 | Node *addp = n->in(AddPNode::Address); |
3009 | assert( !addp->is_AddP() || |
3010 | addp->in(AddPNode::Base)->is_top() || // Top OK for allocation |
3011 | addp->in(AddPNode::Base) == n->in(AddPNode::Base), |
3012 | "Base pointers must match (addp %u)" , addp->_idx ); |
3013 | #ifdef _LP64 |
3014 | if ((UseCompressedOops || UseCompressedClassPointers) && |
3015 | addp->Opcode() == Op_ConP && |
3016 | addp == n->in(AddPNode::Base) && |
3017 | n->in(AddPNode::Offset)->is_Con()) { |
3018 | // If the transformation of ConP to ConN+DecodeN is beneficial depends |
3019 | // on the platform and on the compressed oops mode. |
3020 | // Use addressing with narrow klass to load with offset on x86. |
3021 | // Some platforms can use the constant pool to load ConP. |
3022 | // Do this transformation here since IGVN will convert ConN back to ConP. |
3023 | const Type* t = addp->bottom_type(); |
3024 | bool is_oop = t->isa_oopptr() != NULL; |
3025 | bool is_klass = t->isa_klassptr() != NULL; |
3026 | |
3027 | if ((is_oop && Matcher::const_oop_prefer_decode() ) || |
3028 | (is_klass && Matcher::const_klass_prefer_decode())) { |
3029 | Node* nn = NULL; |
3030 | |
3031 | int op = is_oop ? Op_ConN : Op_ConNKlass; |
3032 | |
3033 | // Look for existing ConN node of the same exact type. |
3034 | Node* r = root(); |
3035 | uint cnt = r->outcnt(); |
3036 | for (uint i = 0; i < cnt; i++) { |
3037 | Node* m = r->raw_out(i); |
3038 | if (m!= NULL && m->Opcode() == op && |
3039 | m->bottom_type()->make_ptr() == t) { |
3040 | nn = m; |
3041 | break; |
3042 | } |
3043 | } |
3044 | if (nn != NULL) { |
3045 | // Decode a narrow oop to match address |
3046 | // [R12 + narrow_oop_reg<<3 + offset] |
3047 | if (is_oop) { |
3048 | nn = new DecodeNNode(nn, t); |
3049 | } else { |
3050 | nn = new DecodeNKlassNode(nn, t); |
3051 | } |
3052 | // Check for succeeding AddP which uses the same Base. |
3053 | // Otherwise we will run into the assertion above when visiting that guy. |
3054 | for (uint i = 0; i < n->outcnt(); ++i) { |
3055 | Node *out_i = n->raw_out(i); |
3056 | if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { |
3057 | out_i->set_req(AddPNode::Base, nn); |
3058 | #ifdef ASSERT |
3059 | for (uint j = 0; j < out_i->outcnt(); ++j) { |
3060 | Node *out_j = out_i->raw_out(j); |
3061 | assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, |
3062 | "more than 2 AddP nodes in a chain (out_j %u)" , out_j->_idx); |
3063 | } |
3064 | #endif |
3065 | } |
3066 | } |
3067 | n->set_req(AddPNode::Base, nn); |
3068 | n->set_req(AddPNode::Address, nn); |
3069 | if (addp->outcnt() == 0) { |
3070 | addp->disconnect_inputs(NULL, this); |
3071 | } |
3072 | } |
3073 | } |
3074 | } |
3075 | #endif |
3076 | // platform dependent reshaping of the address expression |
3077 | reshape_address(n->as_AddP()); |
3078 | break; |
3079 | } |
3080 | |
3081 | case Op_CastPP: { |
3082 | // Remove CastPP nodes to gain more freedom during scheduling but |
3083 | // keep the dependency they encode as control or precedence edges |
3084 | // (if control is set already) on memory operations. Some CastPP |
3085 | // nodes don't have a control (don't carry a dependency): skip |
3086 | // those. |
3087 | if (n->in(0) != NULL) { |
3088 | ResourceMark rm; |
3089 | Unique_Node_List wq; |
3090 | wq.push(n); |
3091 | for (uint next = 0; next < wq.size(); ++next) { |
3092 | Node *m = wq.at(next); |
3093 | for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { |
3094 | Node* use = m->fast_out(i); |
3095 | if (use->is_Mem() || use->is_EncodeNarrowPtr()) { |
3096 | use->ensure_control_or_add_prec(n->in(0)); |
3097 | } else { |
3098 | switch(use->Opcode()) { |
3099 | case Op_AddP: |
3100 | case Op_DecodeN: |
3101 | case Op_DecodeNKlass: |
3102 | case Op_CheckCastPP: |
3103 | case Op_CastPP: |
3104 | wq.push(use); |
3105 | break; |
3106 | } |
3107 | } |
3108 | } |
3109 | } |
3110 | } |
3111 | const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); |
3112 | if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { |
3113 | Node* in1 = n->in(1); |
3114 | const Type* t = n->bottom_type(); |
3115 | Node* new_in1 = in1->clone(); |
3116 | new_in1->as_DecodeN()->set_type(t); |
3117 | |
3118 | if (!Matcher::narrow_oop_use_complex_address()) { |
3119 | // |
3120 | // x86, ARM and friends can handle 2 adds in addressing mode |
3121 | // and Matcher can fold a DecodeN node into address by using |
3122 | // a narrow oop directly and do implicit NULL check in address: |
3123 | // |
3124 | // [R12 + narrow_oop_reg<<3 + offset] |
3125 | // NullCheck narrow_oop_reg |
3126 | // |
3127 | // On other platforms (Sparc) we have to keep new DecodeN node and |
3128 | // use it to do implicit NULL check in address: |
3129 | // |
3130 | // decode_not_null narrow_oop_reg, base_reg |
3131 | // [base_reg + offset] |
3132 | // NullCheck base_reg |
3133 | // |
3134 | // Pin the new DecodeN node to non-null path on these platform (Sparc) |
3135 | // to keep the information to which NULL check the new DecodeN node |
3136 | // corresponds to use it as value in implicit_null_check(). |
3137 | // |
3138 | new_in1->set_req(0, n->in(0)); |
3139 | } |
3140 | |
3141 | n->subsume_by(new_in1, this); |
3142 | if (in1->outcnt() == 0) { |
3143 | in1->disconnect_inputs(NULL, this); |
3144 | } |
3145 | } else { |
3146 | n->subsume_by(n->in(1), this); |
3147 | if (n->outcnt() == 0) { |
3148 | n->disconnect_inputs(NULL, this); |
3149 | } |
3150 | } |
3151 | break; |
3152 | } |
3153 | #ifdef _LP64 |
3154 | case Op_CmpP: |
3155 | // Do this transformation here to preserve CmpPNode::sub() and |
3156 | // other TypePtr related Ideal optimizations (for example, ptr nullness). |
3157 | if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { |
3158 | Node* in1 = n->in(1); |
3159 | Node* in2 = n->in(2); |
3160 | if (!in1->is_DecodeNarrowPtr()) { |
3161 | in2 = in1; |
3162 | in1 = n->in(2); |
3163 | } |
3164 | assert(in1->is_DecodeNarrowPtr(), "sanity" ); |
3165 | |
3166 | Node* new_in2 = NULL; |
3167 | if (in2->is_DecodeNarrowPtr()) { |
3168 | assert(in2->Opcode() == in1->Opcode(), "must be same node type" ); |
3169 | new_in2 = in2->in(1); |
3170 | } else if (in2->Opcode() == Op_ConP) { |
3171 | const Type* t = in2->bottom_type(); |
3172 | if (t == TypePtr::NULL_PTR) { |
3173 | assert(in1->is_DecodeN(), "compare klass to null?" ); |
3174 | // Don't convert CmpP null check into CmpN if compressed |
3175 | // oops implicit null check is not generated. |
3176 | // This will allow to generate normal oop implicit null check. |
3177 | if (Matcher::gen_narrow_oop_implicit_null_checks()) |
3178 | new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); |
3179 | // |
3180 | // This transformation together with CastPP transformation above |
3181 | // will generated code for implicit NULL checks for compressed oops. |
3182 | // |
3183 | // The original code after Optimize() |
3184 | // |
3185 | // LoadN memory, narrow_oop_reg |
3186 | // decode narrow_oop_reg, base_reg |
3187 | // CmpP base_reg, NULL |
3188 | // CastPP base_reg // NotNull |
3189 | // Load [base_reg + offset], val_reg |
3190 | // |
3191 | // after these transformations will be |
3192 | // |
3193 | // LoadN memory, narrow_oop_reg |
3194 | // CmpN narrow_oop_reg, NULL |
3195 | // decode_not_null narrow_oop_reg, base_reg |
3196 | // Load [base_reg + offset], val_reg |
3197 | // |
3198 | // and the uncommon path (== NULL) will use narrow_oop_reg directly |
3199 | // since narrow oops can be used in debug info now (see the code in |
3200 | // final_graph_reshaping_walk()). |
3201 | // |
3202 | // At the end the code will be matched to |
3203 | // on x86: |
3204 | // |
3205 | // Load_narrow_oop memory, narrow_oop_reg |
3206 | // Load [R12 + narrow_oop_reg<<3 + offset], val_reg |
3207 | // NullCheck narrow_oop_reg |
3208 | // |
3209 | // and on sparc: |
3210 | // |
3211 | // Load_narrow_oop memory, narrow_oop_reg |
3212 | // decode_not_null narrow_oop_reg, base_reg |
3213 | // Load [base_reg + offset], val_reg |
3214 | // NullCheck base_reg |
3215 | // |
3216 | } else if (t->isa_oopptr()) { |
3217 | new_in2 = ConNode::make(t->make_narrowoop()); |
3218 | } else if (t->isa_klassptr()) { |
3219 | new_in2 = ConNode::make(t->make_narrowklass()); |
3220 | } |
3221 | } |
3222 | if (new_in2 != NULL) { |
3223 | Node* cmpN = new CmpNNode(in1->in(1), new_in2); |
3224 | n->subsume_by(cmpN, this); |
3225 | if (in1->outcnt() == 0) { |
3226 | in1->disconnect_inputs(NULL, this); |
3227 | } |
3228 | if (in2->outcnt() == 0) { |
3229 | in2->disconnect_inputs(NULL, this); |
3230 | } |
3231 | } |
3232 | } |
3233 | break; |
3234 | |
3235 | case Op_DecodeN: |
3236 | case Op_DecodeNKlass: |
3237 | assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out" ); |
3238 | // DecodeN could be pinned when it can't be fold into |
3239 | // an address expression, see the code for Op_CastPP above. |
3240 | assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control" ); |
3241 | break; |
3242 | |
3243 | case Op_EncodeP: |
3244 | case Op_EncodePKlass: { |
3245 | Node* in1 = n->in(1); |
3246 | if (in1->is_DecodeNarrowPtr()) { |
3247 | n->subsume_by(in1->in(1), this); |
3248 | } else if (in1->Opcode() == Op_ConP) { |
3249 | const Type* t = in1->bottom_type(); |
3250 | if (t == TypePtr::NULL_PTR) { |
3251 | assert(t->isa_oopptr(), "null klass?" ); |
3252 | n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); |
3253 | } else if (t->isa_oopptr()) { |
3254 | n->subsume_by(ConNode::make(t->make_narrowoop()), this); |
3255 | } else if (t->isa_klassptr()) { |
3256 | n->subsume_by(ConNode::make(t->make_narrowklass()), this); |
3257 | } |
3258 | } |
3259 | if (in1->outcnt() == 0) { |
3260 | in1->disconnect_inputs(NULL, this); |
3261 | } |
3262 | break; |
3263 | } |
3264 | |
3265 | case Op_Proj: { |
3266 | if (OptimizeStringConcat) { |
3267 | ProjNode* p = n->as_Proj(); |
3268 | if (p->_is_io_use) { |
3269 | // Separate projections were used for the exception path which |
3270 | // are normally removed by a late inline. If it wasn't inlined |
3271 | // then they will hang around and should just be replaced with |
3272 | // the original one. |
3273 | Node* proj = NULL; |
3274 | // Replace with just one |
3275 | for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { |
3276 | Node *use = i.get(); |
3277 | if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { |
3278 | proj = use; |
3279 | break; |
3280 | } |
3281 | } |
3282 | assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop" ); |
3283 | if (proj != NULL) { |
3284 | p->subsume_by(proj, this); |
3285 | } |
3286 | } |
3287 | } |
3288 | break; |
3289 | } |
3290 | |
3291 | case Op_Phi: |
3292 | if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { |
3293 | // The EncodeP optimization may create Phi with the same edges |
3294 | // for all paths. It is not handled well by Register Allocator. |
3295 | Node* unique_in = n->in(1); |
3296 | assert(unique_in != NULL, "" ); |
3297 | uint cnt = n->req(); |
3298 | for (uint i = 2; i < cnt; i++) { |
3299 | Node* m = n->in(i); |
3300 | assert(m != NULL, "" ); |
3301 | if (unique_in != m) |
3302 | unique_in = NULL; |
3303 | } |
3304 | if (unique_in != NULL) { |
3305 | n->subsume_by(unique_in, this); |
3306 | } |
3307 | } |
3308 | break; |
3309 | |
3310 | #endif |
3311 | |
3312 | #ifdef ASSERT |
3313 | case Op_CastII: |
3314 | // Verify that all range check dependent CastII nodes were removed. |
3315 | if (n->isa_CastII()->has_range_check()) { |
3316 | n->dump(3); |
3317 | assert(false, "Range check dependent CastII node was not removed" ); |
3318 | } |
3319 | break; |
3320 | #endif |
3321 | |
3322 | case Op_ModI: |
3323 | if (UseDivMod) { |
3324 | // Check if a%b and a/b both exist |
3325 | Node* d = n->find_similar(Op_DivI); |
3326 | if (d) { |
3327 | // Replace them with a fused divmod if supported |
3328 | if (Matcher::has_match_rule(Op_DivModI)) { |
3329 | DivModINode* divmod = DivModINode::make(n); |
3330 | d->subsume_by(divmod->div_proj(), this); |
3331 | n->subsume_by(divmod->mod_proj(), this); |
3332 | } else { |
3333 | // replace a%b with a-((a/b)*b) |
3334 | Node* mult = new MulINode(d, d->in(2)); |
3335 | Node* sub = new SubINode(d->in(1), mult); |
3336 | n->subsume_by(sub, this); |
3337 | } |
3338 | } |
3339 | } |
3340 | break; |
3341 | |
3342 | case Op_ModL: |
3343 | if (UseDivMod) { |
3344 | // Check if a%b and a/b both exist |
3345 | Node* d = n->find_similar(Op_DivL); |
3346 | if (d) { |
3347 | // Replace them with a fused divmod if supported |
3348 | if (Matcher::has_match_rule(Op_DivModL)) { |
3349 | DivModLNode* divmod = DivModLNode::make(n); |
3350 | d->subsume_by(divmod->div_proj(), this); |
3351 | n->subsume_by(divmod->mod_proj(), this); |
3352 | } else { |
3353 | // replace a%b with a-((a/b)*b) |
3354 | Node* mult = new MulLNode(d, d->in(2)); |
3355 | Node* sub = new SubLNode(d->in(1), mult); |
3356 | n->subsume_by(sub, this); |
3357 | } |
3358 | } |
3359 | } |
3360 | break; |
3361 | |
3362 | case Op_LoadVector: |
3363 | case Op_StoreVector: |
3364 | break; |
3365 | |
3366 | case Op_AddReductionVI: |
3367 | case Op_AddReductionVL: |
3368 | case Op_AddReductionVF: |
3369 | case Op_AddReductionVD: |
3370 | case Op_MulReductionVI: |
3371 | case Op_MulReductionVL: |
3372 | case Op_MulReductionVF: |
3373 | case Op_MulReductionVD: |
3374 | case Op_MinReductionV: |
3375 | case Op_MaxReductionV: |
3376 | break; |
3377 | |
3378 | case Op_PackB: |
3379 | case Op_PackS: |
3380 | case Op_PackI: |
3381 | case Op_PackF: |
3382 | case Op_PackL: |
3383 | case Op_PackD: |
3384 | if (n->req()-1 > 2) { |
3385 | // Replace many operand PackNodes with a binary tree for matching |
3386 | PackNode* p = (PackNode*) n; |
3387 | Node* btp = p->binary_tree_pack(1, n->req()); |
3388 | n->subsume_by(btp, this); |
3389 | } |
3390 | break; |
3391 | case Op_Loop: |
3392 | case Op_CountedLoop: |
3393 | case Op_OuterStripMinedLoop: |
3394 | if (n->as_Loop()->is_inner_loop()) { |
3395 | frc.inc_inner_loop_count(); |
3396 | } |
3397 | n->as_Loop()->verify_strip_mined(0); |
3398 | break; |
3399 | case Op_LShiftI: |
3400 | case Op_RShiftI: |
3401 | case Op_URShiftI: |
3402 | case Op_LShiftL: |
3403 | case Op_RShiftL: |
3404 | case Op_URShiftL: |
3405 | if (Matcher::need_masked_shift_count) { |
3406 | // The cpu's shift instructions don't restrict the count to the |
3407 | // lower 5/6 bits. We need to do the masking ourselves. |
3408 | Node* in2 = n->in(2); |
3409 | juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); |
3410 | const TypeInt* t = in2->find_int_type(); |
3411 | if (t != NULL && t->is_con()) { |
3412 | juint shift = t->get_con(); |
3413 | if (shift > mask) { // Unsigned cmp |
3414 | n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); |
3415 | } |
3416 | } else { |
3417 | if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { |
3418 | Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); |
3419 | n->set_req(2, shift); |
3420 | } |
3421 | } |
3422 | if (in2->outcnt() == 0) { // Remove dead node |
3423 | in2->disconnect_inputs(NULL, this); |
3424 | } |
3425 | } |
3426 | break; |
3427 | case Op_MemBarStoreStore: |
3428 | case Op_MemBarRelease: |
3429 | // Break the link with AllocateNode: it is no longer useful and |
3430 | // confuses register allocation. |
3431 | if (n->req() > MemBarNode::Precedent) { |
3432 | n->set_req(MemBarNode::Precedent, top()); |
3433 | } |
3434 | break; |
3435 | case Op_MemBarAcquire: { |
3436 | if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { |
3437 | // At parse time, the trailing MemBarAcquire for a volatile load |
3438 | // is created with an edge to the load. After optimizations, |
3439 | // that input may be a chain of Phis. If those phis have no |
3440 | // other use, then the MemBarAcquire keeps them alive and |
3441 | // register allocation can be confused. |
3442 | ResourceMark rm; |
3443 | Unique_Node_List wq; |
3444 | wq.push(n->in(MemBarNode::Precedent)); |
3445 | n->set_req(MemBarNode::Precedent, top()); |
3446 | while (wq.size() > 0) { |
3447 | Node* m = wq.pop(); |
3448 | if (m->outcnt() == 0) { |
3449 | for (uint j = 0; j < m->req(); j++) { |
3450 | Node* in = m->in(j); |
3451 | if (in != NULL) { |
3452 | wq.push(in); |
3453 | } |
3454 | } |
3455 | m->disconnect_inputs(NULL, this); |
3456 | } |
3457 | } |
3458 | } |
3459 | break; |
3460 | } |
3461 | case Op_RangeCheck: { |
3462 | RangeCheckNode* rc = n->as_RangeCheck(); |
3463 | Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); |
3464 | n->subsume_by(iff, this); |
3465 | frc._tests.push(iff); |
3466 | break; |
3467 | } |
3468 | case Op_ConvI2L: { |
3469 | if (!Matcher::convi2l_type_required) { |
3470 | // Code generation on some platforms doesn't need accurate |
3471 | // ConvI2L types. Widening the type can help remove redundant |
3472 | // address computations. |
3473 | n->as_Type()->set_type(TypeLong::INT); |
3474 | ResourceMark rm; |
3475 | Node_List wq; |
3476 | wq.push(n); |
3477 | for (uint next = 0; next < wq.size(); next++) { |
3478 | Node *m = wq.at(next); |
3479 | |
3480 | for(;;) { |
3481 | // Loop over all nodes with identical inputs edges as m |
3482 | Node* k = m->find_similar(m->Opcode()); |
3483 | if (k == NULL) { |
3484 | break; |
3485 | } |
3486 | // Push their uses so we get a chance to remove node made |
3487 | // redundant |
3488 | for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { |
3489 | Node* u = k->fast_out(i); |
3490 | assert(!wq.contains(u), "shouldn't process one node several times" ); |
3491 | if (u->Opcode() == Op_LShiftL || |
3492 | u->Opcode() == Op_AddL || |
3493 | u->Opcode() == Op_SubL || |
3494 | u->Opcode() == Op_AddP) { |
3495 | wq.push(u); |
3496 | } |
3497 | } |
3498 | // Replace all nodes with identical edges as m with m |
3499 | k->subsume_by(m, this); |
3500 | } |
3501 | } |
3502 | } |
3503 | break; |
3504 | } |
3505 | case Op_CmpUL: { |
3506 | if (!Matcher::has_match_rule(Op_CmpUL)) { |
3507 | // No support for unsigned long comparisons |
3508 | ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); |
3509 | Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); |
3510 | Node* orl = new OrLNode(n->in(1), sign_bit_mask); |
3511 | ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); |
3512 | Node* andl = new AndLNode(orl, remove_sign_mask); |
3513 | Node* cmp = new CmpLNode(andl, n->in(2)); |
3514 | n->subsume_by(cmp, this); |
3515 | } |
3516 | break; |
3517 | } |
3518 | default: |
3519 | assert(!n->is_Call(), "" ); |
3520 | assert(!n->is_Mem(), "" ); |
3521 | assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN" ); |
3522 | break; |
3523 | } |
3524 | } |
3525 | |
3526 | //------------------------------final_graph_reshaping_walk--------------------- |
3527 | // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), |
3528 | // requires that the walk visits a node's inputs before visiting the node. |
3529 | void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { |
3530 | ResourceArea *area = Thread::current()->resource_area(); |
3531 | Unique_Node_List sfpt(area); |
3532 | |
3533 | frc._visited.set(root->_idx); // first, mark node as visited |
3534 | uint cnt = root->req(); |
3535 | Node *n = root; |
3536 | uint i = 0; |
3537 | while (true) { |
3538 | if (i < cnt) { |
3539 | // Place all non-visited non-null inputs onto stack |
3540 | Node* m = n->in(i); |
3541 | ++i; |
3542 | if (m != NULL && !frc._visited.test_set(m->_idx)) { |
3543 | if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { |
3544 | // compute worst case interpreter size in case of a deoptimization |
3545 | update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); |
3546 | |
3547 | sfpt.push(m); |
3548 | } |
3549 | cnt = m->req(); |
3550 | nstack.push(n, i); // put on stack parent and next input's index |
3551 | n = m; |
3552 | i = 0; |
3553 | } |
3554 | } else { |
3555 | // Now do post-visit work |
3556 | final_graph_reshaping_impl( n, frc ); |
3557 | if (nstack.is_empty()) |
3558 | break; // finished |
3559 | n = nstack.node(); // Get node from stack |
3560 | cnt = n->req(); |
3561 | i = nstack.index(); |
3562 | nstack.pop(); // Shift to the next node on stack |
3563 | } |
3564 | } |
3565 | |
3566 | // Skip next transformation if compressed oops are not used. |
3567 | if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || |
3568 | (!UseCompressedOops && !UseCompressedClassPointers)) |
3569 | return; |
3570 | |
3571 | // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. |
3572 | // It could be done for an uncommon traps or any safepoints/calls |
3573 | // if the DecodeN/DecodeNKlass node is referenced only in a debug info. |
3574 | while (sfpt.size() > 0) { |
3575 | n = sfpt.pop(); |
3576 | JVMState *jvms = n->as_SafePoint()->jvms(); |
3577 | assert(jvms != NULL, "sanity" ); |
3578 | int start = jvms->debug_start(); |
3579 | int end = n->req(); |
3580 | bool is_uncommon = (n->is_CallStaticJava() && |
3581 | n->as_CallStaticJava()->uncommon_trap_request() != 0); |
3582 | for (int j = start; j < end; j++) { |
3583 | Node* in = n->in(j); |
3584 | if (in->is_DecodeNarrowPtr()) { |
3585 | bool safe_to_skip = true; |
3586 | if (!is_uncommon ) { |
3587 | // Is it safe to skip? |
3588 | for (uint i = 0; i < in->outcnt(); i++) { |
3589 | Node* u = in->raw_out(i); |
3590 | if (!u->is_SafePoint() || |
3591 | (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { |
3592 | safe_to_skip = false; |
3593 | } |
3594 | } |
3595 | } |
3596 | if (safe_to_skip) { |
3597 | n->set_req(j, in->in(1)); |
3598 | } |
3599 | if (in->outcnt() == 0) { |
3600 | in->disconnect_inputs(NULL, this); |
3601 | } |
3602 | } |
3603 | } |
3604 | } |
3605 | } |
3606 | |
3607 | //------------------------------final_graph_reshaping-------------------------- |
3608 | // Final Graph Reshaping. |
3609 | // |
3610 | // (1) Clone simple inputs to uncommon calls, so they can be scheduled late |
3611 | // and not commoned up and forced early. Must come after regular |
3612 | // optimizations to avoid GVN undoing the cloning. Clone constant |
3613 | // inputs to Loop Phis; these will be split by the allocator anyways. |
3614 | // Remove Opaque nodes. |
3615 | // (2) Move last-uses by commutative operations to the left input to encourage |
3616 | // Intel update-in-place two-address operations and better register usage |
3617 | // on RISCs. Must come after regular optimizations to avoid GVN Ideal |
3618 | // calls canonicalizing them back. |
3619 | // (3) Count the number of double-precision FP ops, single-precision FP ops |
3620 | // and call sites. On Intel, we can get correct rounding either by |
3621 | // forcing singles to memory (requires extra stores and loads after each |
3622 | // FP bytecode) or we can set a rounding mode bit (requires setting and |
3623 | // clearing the mode bit around call sites). The mode bit is only used |
3624 | // if the relative frequency of single FP ops to calls is low enough. |
3625 | // This is a key transform for SPEC mpeg_audio. |
3626 | // (4) Detect infinite loops; blobs of code reachable from above but not |
3627 | // below. Several of the Code_Gen algorithms fail on such code shapes, |
3628 | // so we simply bail out. Happens a lot in ZKM.jar, but also happens |
3629 | // from time to time in other codes (such as -Xcomp finalizer loops, etc). |
3630 | // Detection is by looking for IfNodes where only 1 projection is |
3631 | // reachable from below or CatchNodes missing some targets. |
3632 | // (5) Assert for insane oop offsets in debug mode. |
3633 | |
3634 | bool Compile::final_graph_reshaping() { |
3635 | // an infinite loop may have been eliminated by the optimizer, |
3636 | // in which case the graph will be empty. |
3637 | if (root()->req() == 1) { |
3638 | record_method_not_compilable("trivial infinite loop" ); |
3639 | return true; |
3640 | } |
3641 | |
3642 | // Expensive nodes have their control input set to prevent the GVN |
3643 | // from freely commoning them. There's no GVN beyond this point so |
3644 | // no need to keep the control input. We want the expensive nodes to |
3645 | // be freely moved to the least frequent code path by gcm. |
3646 | assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?" ); |
3647 | for (int i = 0; i < expensive_count(); i++) { |
3648 | _expensive_nodes->at(i)->set_req(0, NULL); |
3649 | } |
3650 | |
3651 | Final_Reshape_Counts frc; |
3652 | |
3653 | // Visit everybody reachable! |
3654 | // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc |
3655 | Node_Stack nstack(live_nodes() >> 1); |
3656 | final_graph_reshaping_walk(nstack, root(), frc); |
3657 | |
3658 | // Check for unreachable (from below) code (i.e., infinite loops). |
3659 | for( uint i = 0; i < frc._tests.size(); i++ ) { |
3660 | MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); |
3661 | // Get number of CFG targets. |
3662 | // Note that PCTables include exception targets after calls. |
3663 | uint required_outcnt = n->required_outcnt(); |
3664 | if (n->outcnt() != required_outcnt) { |
3665 | // Check for a few special cases. Rethrow Nodes never take the |
3666 | // 'fall-thru' path, so expected kids is 1 less. |
3667 | if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { |
3668 | if (n->in(0)->in(0)->is_Call()) { |
3669 | CallNode *call = n->in(0)->in(0)->as_Call(); |
3670 | if (call->entry_point() == OptoRuntime::rethrow_stub()) { |
3671 | required_outcnt--; // Rethrow always has 1 less kid |
3672 | } else if (call->req() > TypeFunc::Parms && |
3673 | call->is_CallDynamicJava()) { |
3674 | // Check for null receiver. In such case, the optimizer has |
3675 | // detected that the virtual call will always result in a null |
3676 | // pointer exception. The fall-through projection of this CatchNode |
3677 | // will not be populated. |
3678 | Node *arg0 = call->in(TypeFunc::Parms); |
3679 | if (arg0->is_Type() && |
3680 | arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { |
3681 | required_outcnt--; |
3682 | } |
3683 | } else if (call->entry_point() == OptoRuntime::new_array_Java() && |
3684 | call->req() > TypeFunc::Parms+1 && |
3685 | call->is_CallStaticJava()) { |
3686 | // Check for negative array length. In such case, the optimizer has |
3687 | // detected that the allocation attempt will always result in an |
3688 | // exception. There is no fall-through projection of this CatchNode . |
3689 | Node *arg1 = call->in(TypeFunc::Parms+1); |
3690 | if (arg1->is_Type() && |
3691 | arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { |
3692 | required_outcnt--; |
3693 | } |
3694 | } |
3695 | } |
3696 | } |
3697 | // Recheck with a better notion of 'required_outcnt' |
3698 | if (n->outcnt() != required_outcnt) { |
3699 | record_method_not_compilable("malformed control flow" ); |
3700 | return true; // Not all targets reachable! |
3701 | } |
3702 | } |
3703 | // Check that I actually visited all kids. Unreached kids |
3704 | // must be infinite loops. |
3705 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) |
3706 | if (!frc._visited.test(n->fast_out(j)->_idx)) { |
3707 | record_method_not_compilable("infinite loop" ); |
3708 | return true; // Found unvisited kid; must be unreach |
3709 | } |
3710 | |
3711 | // Here so verification code in final_graph_reshaping_walk() |
3712 | // always see an OuterStripMinedLoopEnd |
3713 | if (n->is_OuterStripMinedLoopEnd()) { |
3714 | IfNode* init_iff = n->as_If(); |
3715 | Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); |
3716 | n->subsume_by(iff, this); |
3717 | } |
3718 | } |
3719 | |
3720 | // If original bytecodes contained a mixture of floats and doubles |
3721 | // check if the optimizer has made it homogenous, item (3). |
3722 | if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && |
3723 | frc.get_float_count() > 32 && |
3724 | frc.get_double_count() == 0 && |
3725 | (10 * frc.get_call_count() < frc.get_float_count()) ) { |
3726 | set_24_bit_selection_and_mode( false, true ); |
3727 | } |
3728 | |
3729 | set_java_calls(frc.get_java_call_count()); |
3730 | set_inner_loops(frc.get_inner_loop_count()); |
3731 | |
3732 | // No infinite loops, no reason to bail out. |
3733 | return false; |
3734 | } |
3735 | |
3736 | //-----------------------------too_many_traps---------------------------------- |
3737 | // Report if there are too many traps at the current method and bci. |
3738 | // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. |
3739 | bool Compile::too_many_traps(ciMethod* method, |
3740 | int bci, |
3741 | Deoptimization::DeoptReason reason) { |
3742 | ciMethodData* md = method->method_data(); |
3743 | if (md->is_empty()) { |
3744 | // Assume the trap has not occurred, or that it occurred only |
3745 | // because of a transient condition during start-up in the interpreter. |
3746 | return false; |
3747 | } |
3748 | ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; |
3749 | if (md->has_trap_at(bci, m, reason) != 0) { |
3750 | // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. |
3751 | // Also, if there are multiple reasons, or if there is no per-BCI record, |
3752 | // assume the worst. |
3753 | if (log()) |
3754 | log()->elem("observe trap='%s' count='%d'" , |
3755 | Deoptimization::trap_reason_name(reason), |
3756 | md->trap_count(reason)); |
3757 | return true; |
3758 | } else { |
3759 | // Ignore method/bci and see if there have been too many globally. |
3760 | return too_many_traps(reason, md); |
3761 | } |
3762 | } |
3763 | |
3764 | // Less-accurate variant which does not require a method and bci. |
3765 | bool Compile::too_many_traps(Deoptimization::DeoptReason reason, |
3766 | ciMethodData* logmd) { |
3767 | if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { |
3768 | // Too many traps globally. |
3769 | // Note that we use cumulative trap_count, not just md->trap_count. |
3770 | if (log()) { |
3771 | int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); |
3772 | log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'" , |
3773 | Deoptimization::trap_reason_name(reason), |
3774 | mcount, trap_count(reason)); |
3775 | } |
3776 | return true; |
3777 | } else { |
3778 | // The coast is clear. |
3779 | return false; |
3780 | } |
3781 | } |
3782 | |
3783 | //--------------------------too_many_recompiles-------------------------------- |
3784 | // Report if there are too many recompiles at the current method and bci. |
3785 | // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. |
3786 | // Is not eager to return true, since this will cause the compiler to use |
3787 | // Action_none for a trap point, to avoid too many recompilations. |
3788 | bool Compile::too_many_recompiles(ciMethod* method, |
3789 | int bci, |
3790 | Deoptimization::DeoptReason reason) { |
3791 | ciMethodData* md = method->method_data(); |
3792 | if (md->is_empty()) { |
3793 | // Assume the trap has not occurred, or that it occurred only |
3794 | // because of a transient condition during start-up in the interpreter. |
3795 | return false; |
3796 | } |
3797 | // Pick a cutoff point well within PerBytecodeRecompilationCutoff. |
3798 | uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; |
3799 | uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero |
3800 | Deoptimization::DeoptReason per_bc_reason |
3801 | = Deoptimization::reason_recorded_per_bytecode_if_any(reason); |
3802 | ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; |
3803 | if ((per_bc_reason == Deoptimization::Reason_none |
3804 | || md->has_trap_at(bci, m, reason) != 0) |
3805 | // The trap frequency measure we care about is the recompile count: |
3806 | && md->trap_recompiled_at(bci, m) |
3807 | && md->overflow_recompile_count() >= bc_cutoff) { |
3808 | // Do not emit a trap here if it has already caused recompilations. |
3809 | // Also, if there are multiple reasons, or if there is no per-BCI record, |
3810 | // assume the worst. |
3811 | if (log()) |
3812 | log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'" , |
3813 | Deoptimization::trap_reason_name(reason), |
3814 | md->trap_count(reason), |
3815 | md->overflow_recompile_count()); |
3816 | return true; |
3817 | } else if (trap_count(reason) != 0 |
3818 | && decompile_count() >= m_cutoff) { |
3819 | // Too many recompiles globally, and we have seen this sort of trap. |
3820 | // Use cumulative decompile_count, not just md->decompile_count. |
3821 | if (log()) |
3822 | log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'" , |
3823 | Deoptimization::trap_reason_name(reason), |
3824 | md->trap_count(reason), trap_count(reason), |
3825 | md->decompile_count(), decompile_count()); |
3826 | return true; |
3827 | } else { |
3828 | // The coast is clear. |
3829 | return false; |
3830 | } |
3831 | } |
3832 | |
3833 | // Compute when not to trap. Used by matching trap based nodes and |
3834 | // NullCheck optimization. |
3835 | void Compile::set_allowed_deopt_reasons() { |
3836 | _allowed_reasons = 0; |
3837 | if (is_method_compilation()) { |
3838 | for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { |
3839 | assert(rs < BitsPerInt, "recode bit map" ); |
3840 | if (!too_many_traps((Deoptimization::DeoptReason) rs)) { |
3841 | _allowed_reasons |= nth_bit(rs); |
3842 | } |
3843 | } |
3844 | } |
3845 | } |
3846 | |
3847 | bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { |
3848 | return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); |
3849 | } |
3850 | |
3851 | bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { |
3852 | return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); |
3853 | } |
3854 | |
3855 | bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { |
3856 | if (holder->is_initialized()) { |
3857 | return false; |
3858 | } |
3859 | if (holder->is_being_initialized()) { |
3860 | if (accessing_method->holder() == holder) { |
3861 | // Access inside a class. The barrier can be elided when access happens in <clinit>, |
3862 | // <init>, or a static method. In all those cases, there was an initialization |
3863 | // barrier on the holder klass passed. |
3864 | if (accessing_method->is_static_initializer() || |
3865 | accessing_method->is_object_initializer() || |
3866 | accessing_method->is_static()) { |
3867 | return false; |
3868 | } |
3869 | } else if (accessing_method->holder()->is_subclass_of(holder)) { |
3870 | // Access from a subclass. The barrier can be elided only when access happens in <clinit>. |
3871 | // In case of <init> or a static method, the barrier is on the subclass is not enough: |
3872 | // child class can become fully initialized while its parent class is still being initialized. |
3873 | if (accessing_method->is_static_initializer()) { |
3874 | return false; |
3875 | } |
3876 | } |
3877 | ciMethod* root = method(); // the root method of compilation |
3878 | if (root != accessing_method) { |
3879 | return needs_clinit_barrier(holder, root); // check access in the context of compilation root |
3880 | } |
3881 | } |
3882 | return true; |
3883 | } |
3884 | |
3885 | #ifndef PRODUCT |
3886 | //------------------------------verify_graph_edges--------------------------- |
3887 | // Walk the Graph and verify that there is a one-to-one correspondence |
3888 | // between Use-Def edges and Def-Use edges in the graph. |
3889 | void Compile::verify_graph_edges(bool no_dead_code) { |
3890 | if (VerifyGraphEdges) { |
3891 | ResourceArea *area = Thread::current()->resource_area(); |
3892 | Unique_Node_List visited(area); |
3893 | // Call recursive graph walk to check edges |
3894 | _root->verify_edges(visited); |
3895 | if (no_dead_code) { |
3896 | // Now make sure that no visited node is used by an unvisited node. |
3897 | bool dead_nodes = false; |
3898 | Unique_Node_List checked(area); |
3899 | while (visited.size() > 0) { |
3900 | Node* n = visited.pop(); |
3901 | checked.push(n); |
3902 | for (uint i = 0; i < n->outcnt(); i++) { |
3903 | Node* use = n->raw_out(i); |
3904 | if (checked.member(use)) continue; // already checked |
3905 | if (visited.member(use)) continue; // already in the graph |
3906 | if (use->is_Con()) continue; // a dead ConNode is OK |
3907 | // At this point, we have found a dead node which is DU-reachable. |
3908 | if (!dead_nodes) { |
3909 | tty->print_cr("*** Dead nodes reachable via DU edges:" ); |
3910 | dead_nodes = true; |
3911 | } |
3912 | use->dump(2); |
3913 | tty->print_cr("---" ); |
3914 | checked.push(use); // No repeats; pretend it is now checked. |
3915 | } |
3916 | } |
3917 | assert(!dead_nodes, "using nodes must be reachable from root" ); |
3918 | } |
3919 | } |
3920 | } |
3921 | #endif |
3922 | |
3923 | // The Compile object keeps track of failure reasons separately from the ciEnv. |
3924 | // This is required because there is not quite a 1-1 relation between the |
3925 | // ciEnv and its compilation task and the Compile object. Note that one |
3926 | // ciEnv might use two Compile objects, if C2Compiler::compile_method decides |
3927 | // to backtrack and retry without subsuming loads. Other than this backtracking |
3928 | // behavior, the Compile's failure reason is quietly copied up to the ciEnv |
3929 | // by the logic in C2Compiler. |
3930 | void Compile::record_failure(const char* reason) { |
3931 | if (log() != NULL) { |
3932 | log()->elem("failure reason='%s' phase='compile'" , reason); |
3933 | } |
3934 | if (_failure_reason == NULL) { |
3935 | // Record the first failure reason. |
3936 | _failure_reason = reason; |
3937 | } |
3938 | |
3939 | if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { |
3940 | C->print_method(PHASE_FAILURE); |
3941 | } |
3942 | _root = NULL; // flush the graph, too |
3943 | } |
3944 | |
3945 | Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) |
3946 | : TraceTime(name, accumulator, CITime, CITimeVerbose), |
3947 | _phase_name(name), _dolog(CITimeVerbose) |
3948 | { |
3949 | if (_dolog) { |
3950 | C = Compile::current(); |
3951 | _log = C->log(); |
3952 | } else { |
3953 | C = NULL; |
3954 | _log = NULL; |
3955 | } |
3956 | if (_log != NULL) { |
3957 | _log->begin_head("phase name='%s' nodes='%d' live='%d'" , _phase_name, C->unique(), C->live_nodes()); |
3958 | _log->stamp(); |
3959 | _log->end_head(); |
3960 | } |
3961 | } |
3962 | |
3963 | Compile::TracePhase::~TracePhase() { |
3964 | |
3965 | C = Compile::current(); |
3966 | if (_dolog) { |
3967 | _log = C->log(); |
3968 | } else { |
3969 | _log = NULL; |
3970 | } |
3971 | |
3972 | #ifdef ASSERT |
3973 | if (PrintIdealNodeCount) { |
3974 | tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'" , |
3975 | _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); |
3976 | } |
3977 | |
3978 | if (VerifyIdealNodeCount) { |
3979 | Compile::current()->print_missing_nodes(); |
3980 | } |
3981 | #endif |
3982 | |
3983 | if (_log != NULL) { |
3984 | _log->done("phase name='%s' nodes='%d' live='%d'" , _phase_name, C->unique(), C->live_nodes()); |
3985 | } |
3986 | } |
3987 | |
3988 | //============================================================================= |
3989 | // Two Constant's are equal when the type and the value are equal. |
3990 | bool Compile::Constant::operator==(const Constant& other) { |
3991 | if (type() != other.type() ) return false; |
3992 | if (can_be_reused() != other.can_be_reused()) return false; |
3993 | // For floating point values we compare the bit pattern. |
3994 | switch (type()) { |
3995 | case T_INT: |
3996 | case T_FLOAT: return (_v._value.i == other._v._value.i); |
3997 | case T_LONG: |
3998 | case T_DOUBLE: return (_v._value.j == other._v._value.j); |
3999 | case T_OBJECT: |
4000 | case T_ADDRESS: return (_v._value.l == other._v._value.l); |
4001 | case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries |
4002 | case T_METADATA: return (_v._metadata == other._v._metadata); |
4003 | default: ShouldNotReachHere(); return false; |
4004 | } |
4005 | } |
4006 | |
4007 | static int type_to_size_in_bytes(BasicType t) { |
4008 | switch (t) { |
4009 | case T_INT: return sizeof(jint ); |
4010 | case T_LONG: return sizeof(jlong ); |
4011 | case T_FLOAT: return sizeof(jfloat ); |
4012 | case T_DOUBLE: return sizeof(jdouble); |
4013 | case T_METADATA: return sizeof(Metadata*); |
4014 | // We use T_VOID as marker for jump-table entries (labels) which |
4015 | // need an internal word relocation. |
4016 | case T_VOID: |
4017 | case T_ADDRESS: |
4018 | case T_OBJECT: return sizeof(jobject); |
4019 | default: |
4020 | ShouldNotReachHere(); |
4021 | return -1; |
4022 | } |
4023 | } |
4024 | |
4025 | int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { |
4026 | // sort descending |
4027 | if (a->freq() > b->freq()) return -1; |
4028 | if (a->freq() < b->freq()) return 1; |
4029 | return 0; |
4030 | } |
4031 | |
4032 | void Compile::ConstantTable::calculate_offsets_and_size() { |
4033 | // First, sort the array by frequencies. |
4034 | _constants.sort(qsort_comparator); |
4035 | |
4036 | #ifdef ASSERT |
4037 | // Make sure all jump-table entries were sorted to the end of the |
4038 | // array (they have a negative frequency). |
4039 | bool found_void = false; |
4040 | for (int i = 0; i < _constants.length(); i++) { |
4041 | Constant con = _constants.at(i); |
4042 | if (con.type() == T_VOID) |
4043 | found_void = true; // jump-tables |
4044 | else |
4045 | assert(!found_void, "wrong sorting" ); |
4046 | } |
4047 | #endif |
4048 | |
4049 | int offset = 0; |
4050 | for (int i = 0; i < _constants.length(); i++) { |
4051 | Constant* con = _constants.adr_at(i); |
4052 | |
4053 | // Align offset for type. |
4054 | int typesize = type_to_size_in_bytes(con->type()); |
4055 | offset = align_up(offset, typesize); |
4056 | con->set_offset(offset); // set constant's offset |
4057 | |
4058 | if (con->type() == T_VOID) { |
4059 | MachConstantNode* n = (MachConstantNode*) con->get_jobject(); |
4060 | offset = offset + typesize * n->outcnt(); // expand jump-table |
4061 | } else { |
4062 | offset = offset + typesize; |
4063 | } |
4064 | } |
4065 | |
4066 | // Align size up to the next section start (which is insts; see |
4067 | // CodeBuffer::align_at_start). |
4068 | assert(_size == -1, "already set?" ); |
4069 | _size = align_up(offset, (int)CodeEntryAlignment); |
4070 | } |
4071 | |
4072 | void Compile::ConstantTable::emit(CodeBuffer& cb) { |
4073 | MacroAssembler _masm(&cb); |
4074 | for (int i = 0; i < _constants.length(); i++) { |
4075 | Constant con = _constants.at(i); |
4076 | address constant_addr = NULL; |
4077 | switch (con.type()) { |
4078 | case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; |
4079 | case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; |
4080 | case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; |
4081 | case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; |
4082 | case T_OBJECT: { |
4083 | jobject obj = con.get_jobject(); |
4084 | int oop_index = _masm.oop_recorder()->find_index(obj); |
4085 | constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); |
4086 | break; |
4087 | } |
4088 | case T_ADDRESS: { |
4089 | address addr = (address) con.get_jobject(); |
4090 | constant_addr = _masm.address_constant(addr); |
4091 | break; |
4092 | } |
4093 | // We use T_VOID as marker for jump-table entries (labels) which |
4094 | // need an internal word relocation. |
4095 | case T_VOID: { |
4096 | MachConstantNode* n = (MachConstantNode*) con.get_jobject(); |
4097 | // Fill the jump-table with a dummy word. The real value is |
4098 | // filled in later in fill_jump_table. |
4099 | address dummy = (address) n; |
4100 | constant_addr = _masm.address_constant(dummy); |
4101 | // Expand jump-table |
4102 | for (uint i = 1; i < n->outcnt(); i++) { |
4103 | address temp_addr = _masm.address_constant(dummy + i); |
4104 | assert(temp_addr, "consts section too small" ); |
4105 | } |
4106 | break; |
4107 | } |
4108 | case T_METADATA: { |
4109 | Metadata* obj = con.get_metadata(); |
4110 | int metadata_index = _masm.oop_recorder()->find_index(obj); |
4111 | constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); |
4112 | break; |
4113 | } |
4114 | default: ShouldNotReachHere(); |
4115 | } |
4116 | assert(constant_addr, "consts section too small" ); |
4117 | assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), |
4118 | "must be: %d == %d" , (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); |
4119 | } |
4120 | } |
4121 | |
4122 | int Compile::ConstantTable::find_offset(Constant& con) const { |
4123 | int idx = _constants.find(con); |
4124 | guarantee(idx != -1, "constant must be in constant table" ); |
4125 | int offset = _constants.at(idx).offset(); |
4126 | guarantee(offset != -1, "constant table not emitted yet?" ); |
4127 | return offset; |
4128 | } |
4129 | |
4130 | void Compile::ConstantTable::add(Constant& con) { |
4131 | if (con.can_be_reused()) { |
4132 | int idx = _constants.find(con); |
4133 | if (idx != -1 && _constants.at(idx).can_be_reused()) { |
4134 | _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value |
4135 | return; |
4136 | } |
4137 | } |
4138 | (void) _constants.append(con); |
4139 | } |
4140 | |
4141 | Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { |
4142 | Block* b = Compile::current()->cfg()->get_block_for_node(n); |
4143 | Constant con(type, value, b->_freq); |
4144 | add(con); |
4145 | return con; |
4146 | } |
4147 | |
4148 | Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { |
4149 | Constant con(metadata); |
4150 | add(con); |
4151 | return con; |
4152 | } |
4153 | |
4154 | Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { |
4155 | jvalue value; |
4156 | BasicType type = oper->type()->basic_type(); |
4157 | switch (type) { |
4158 | case T_LONG: value.j = oper->constantL(); break; |
4159 | case T_FLOAT: value.f = oper->constantF(); break; |
4160 | case T_DOUBLE: value.d = oper->constantD(); break; |
4161 | case T_OBJECT: |
4162 | case T_ADDRESS: value.l = (jobject) oper->constant(); break; |
4163 | case T_METADATA: return add((Metadata*)oper->constant()); break; |
4164 | default: guarantee(false, "unhandled type: %s" , type2name(type)); |
4165 | } |
4166 | return add(n, type, value); |
4167 | } |
4168 | |
4169 | Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { |
4170 | jvalue value; |
4171 | // We can use the node pointer here to identify the right jump-table |
4172 | // as this method is called from Compile::Fill_buffer right before |
4173 | // the MachNodes are emitted and the jump-table is filled (means the |
4174 | // MachNode pointers do not change anymore). |
4175 | value.l = (jobject) n; |
4176 | Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. |
4177 | add(con); |
4178 | return con; |
4179 | } |
4180 | |
4181 | void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { |
4182 | // If called from Compile::scratch_emit_size do nothing. |
4183 | if (Compile::current()->in_scratch_emit_size()) return; |
4184 | |
4185 | assert(labels.is_nonempty(), "must be" ); |
4186 | assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d" , labels.length(), n->outcnt()); |
4187 | |
4188 | // Since MachConstantNode::constant_offset() also contains |
4189 | // table_base_offset() we need to subtract the table_base_offset() |
4190 | // to get the plain offset into the constant table. |
4191 | int offset = n->constant_offset() - table_base_offset(); |
4192 | |
4193 | MacroAssembler _masm(&cb); |
4194 | address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); |
4195 | |
4196 | for (uint i = 0; i < n->outcnt(); i++) { |
4197 | address* constant_addr = &jump_table_base[i]; |
4198 | assert(*constant_addr == (((address) n) + i), "all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i)); |
4199 | *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); |
4200 | cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); |
4201 | } |
4202 | } |
4203 | |
4204 | //----------------------------static_subtype_check----------------------------- |
4205 | // Shortcut important common cases when superklass is exact: |
4206 | // (0) superklass is java.lang.Object (can occur in reflective code) |
4207 | // (1) subklass is already limited to a subtype of superklass => always ok |
4208 | // (2) subklass does not overlap with superklass => always fail |
4209 | // (3) superklass has NO subtypes and we can check with a simple compare. |
4210 | int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { |
4211 | if (StressReflectiveCode) { |
4212 | return SSC_full_test; // Let caller generate the general case. |
4213 | } |
4214 | |
4215 | if (superk == env()->Object_klass()) { |
4216 | return SSC_always_true; // (0) this test cannot fail |
4217 | } |
4218 | |
4219 | ciType* superelem = superk; |
4220 | if (superelem->is_array_klass()) |
4221 | superelem = superelem->as_array_klass()->base_element_type(); |
4222 | |
4223 | if (!subk->is_interface()) { // cannot trust static interface types yet |
4224 | if (subk->is_subtype_of(superk)) { |
4225 | return SSC_always_true; // (1) false path dead; no dynamic test needed |
4226 | } |
4227 | if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && |
4228 | !superk->is_subtype_of(subk)) { |
4229 | return SSC_always_false; |
4230 | } |
4231 | } |
4232 | |
4233 | // If casting to an instance klass, it must have no subtypes |
4234 | if (superk->is_interface()) { |
4235 | // Cannot trust interfaces yet. |
4236 | // %%% S.B. superk->nof_implementors() == 1 |
4237 | } else if (superelem->is_instance_klass()) { |
4238 | ciInstanceKlass* ik = superelem->as_instance_klass(); |
4239 | if (!ik->has_subklass() && !ik->is_interface()) { |
4240 | if (!ik->is_final()) { |
4241 | // Add a dependency if there is a chance of a later subclass. |
4242 | dependencies()->assert_leaf_type(ik); |
4243 | } |
4244 | return SSC_easy_test; // (3) caller can do a simple ptr comparison |
4245 | } |
4246 | } else { |
4247 | // A primitive array type has no subtypes. |
4248 | return SSC_easy_test; // (3) caller can do a simple ptr comparison |
4249 | } |
4250 | |
4251 | return SSC_full_test; |
4252 | } |
4253 | |
4254 | Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { |
4255 | #ifdef _LP64 |
4256 | // The scaled index operand to AddP must be a clean 64-bit value. |
4257 | // Java allows a 32-bit int to be incremented to a negative |
4258 | // value, which appears in a 64-bit register as a large |
4259 | // positive number. Using that large positive number as an |
4260 | // operand in pointer arithmetic has bad consequences. |
4261 | // On the other hand, 32-bit overflow is rare, and the possibility |
4262 | // can often be excluded, if we annotate the ConvI2L node with |
4263 | // a type assertion that its value is known to be a small positive |
4264 | // number. (The prior range check has ensured this.) |
4265 | // This assertion is used by ConvI2LNode::Ideal. |
4266 | int index_max = max_jint - 1; // array size is max_jint, index is one less |
4267 | if (sizetype != NULL) index_max = sizetype->_hi - 1; |
4268 | const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); |
4269 | idx = constrained_convI2L(phase, idx, iidxtype, ctrl); |
4270 | #endif |
4271 | return idx; |
4272 | } |
4273 | |
4274 | // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) |
4275 | Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { |
4276 | if (ctrl != NULL) { |
4277 | // Express control dependency by a CastII node with a narrow type. |
4278 | value = new CastIINode(value, itype, false, true /* range check dependency */); |
4279 | // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L |
4280 | // node from floating above the range check during loop optimizations. Otherwise, the |
4281 | // ConvI2L node may be eliminated independently of the range check, causing the data path |
4282 | // to become TOP while the control path is still there (although it's unreachable). |
4283 | value->set_req(0, ctrl); |
4284 | // Save CastII node to remove it after loop optimizations. |
4285 | phase->C->add_range_check_cast(value); |
4286 | value = phase->transform(value); |
4287 | } |
4288 | const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); |
4289 | return phase->transform(new ConvI2LNode(value, ltype)); |
4290 | } |
4291 | |
4292 | // The message about the current inlining is accumulated in |
4293 | // _print_inlining_stream and transfered into the _print_inlining_list |
4294 | // once we know whether inlining succeeds or not. For regular |
4295 | // inlining, messages are appended to the buffer pointed by |
4296 | // _print_inlining_idx in the _print_inlining_list. For late inlining, |
4297 | // a new buffer is added after _print_inlining_idx in the list. This |
4298 | // way we can update the inlining message for late inlining call site |
4299 | // when the inlining is attempted again. |
4300 | void Compile::print_inlining_init() { |
4301 | if (print_inlining() || print_intrinsics()) { |
4302 | _print_inlining_stream = new stringStream(); |
4303 | _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); |
4304 | } |
4305 | } |
4306 | |
4307 | void Compile::print_inlining_reinit() { |
4308 | if (print_inlining() || print_intrinsics()) { |
4309 | // Re allocate buffer when we change ResourceMark |
4310 | _print_inlining_stream = new stringStream(); |
4311 | } |
4312 | } |
4313 | |
4314 | void Compile::print_inlining_reset() { |
4315 | _print_inlining_stream->reset(); |
4316 | } |
4317 | |
4318 | void Compile::print_inlining_commit() { |
4319 | assert(print_inlining() || print_intrinsics(), "PrintInlining off?" ); |
4320 | // Transfer the message from _print_inlining_stream to the current |
4321 | // _print_inlining_list buffer and clear _print_inlining_stream. |
4322 | _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); |
4323 | print_inlining_reset(); |
4324 | } |
4325 | |
4326 | void Compile::print_inlining_push() { |
4327 | // Add new buffer to the _print_inlining_list at current position |
4328 | _print_inlining_idx++; |
4329 | _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); |
4330 | } |
4331 | |
4332 | Compile::PrintInliningBuffer& Compile::print_inlining_current() { |
4333 | return _print_inlining_list->at(_print_inlining_idx); |
4334 | } |
4335 | |
4336 | void Compile::print_inlining_update(CallGenerator* cg) { |
4337 | if (print_inlining() || print_intrinsics()) { |
4338 | if (!cg->is_late_inline()) { |
4339 | if (print_inlining_current().cg() != NULL) { |
4340 | print_inlining_push(); |
4341 | } |
4342 | print_inlining_commit(); |
4343 | } else { |
4344 | if (print_inlining_current().cg() != cg && |
4345 | (print_inlining_current().cg() != NULL || |
4346 | print_inlining_current().ss()->size() != 0)) { |
4347 | print_inlining_push(); |
4348 | } |
4349 | print_inlining_commit(); |
4350 | print_inlining_current().set_cg(cg); |
4351 | } |
4352 | } |
4353 | } |
4354 | |
4355 | void Compile::print_inlining_move_to(CallGenerator* cg) { |
4356 | // We resume inlining at a late inlining call site. Locate the |
4357 | // corresponding inlining buffer so that we can update it. |
4358 | if (print_inlining()) { |
4359 | for (int i = 0; i < _print_inlining_list->length(); i++) { |
4360 | if (_print_inlining_list->adr_at(i)->cg() == cg) { |
4361 | _print_inlining_idx = i; |
4362 | return; |
4363 | } |
4364 | } |
4365 | ShouldNotReachHere(); |
4366 | } |
4367 | } |
4368 | |
4369 | void Compile::print_inlining_update_delayed(CallGenerator* cg) { |
4370 | if (print_inlining()) { |
4371 | assert(_print_inlining_stream->size() > 0, "missing inlining msg" ); |
4372 | assert(print_inlining_current().cg() == cg, "wrong entry" ); |
4373 | // replace message with new message |
4374 | _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); |
4375 | print_inlining_commit(); |
4376 | print_inlining_current().set_cg(cg); |
4377 | } |
4378 | } |
4379 | |
4380 | void Compile::print_inlining_assert_ready() { |
4381 | assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data" ); |
4382 | } |
4383 | |
4384 | void Compile::process_print_inlining() { |
4385 | bool do_print_inlining = print_inlining() || print_intrinsics(); |
4386 | if (do_print_inlining || log() != NULL) { |
4387 | // Print inlining message for candidates that we couldn't inline |
4388 | // for lack of space |
4389 | for (int i = 0; i < _late_inlines.length(); i++) { |
4390 | CallGenerator* cg = _late_inlines.at(i); |
4391 | if (!cg->is_mh_late_inline()) { |
4392 | const char* msg = "live nodes > LiveNodeCountInliningCutoff" ; |
4393 | if (do_print_inlining) { |
4394 | cg->print_inlining_late(msg); |
4395 | } |
4396 | log_late_inline_failure(cg, msg); |
4397 | } |
4398 | } |
4399 | } |
4400 | if (do_print_inlining) { |
4401 | ResourceMark rm; |
4402 | stringStream ss; |
4403 | for (int i = 0; i < _print_inlining_list->length(); i++) { |
4404 | ss.print("%s" , _print_inlining_list->adr_at(i)->ss()->as_string()); |
4405 | } |
4406 | size_t end = ss.size(); |
4407 | _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); |
4408 | strncpy(_print_inlining_output, ss.base(), end+1); |
4409 | _print_inlining_output[end] = 0; |
4410 | } |
4411 | } |
4412 | |
4413 | void Compile::dump_print_inlining() { |
4414 | if (_print_inlining_output != NULL) { |
4415 | tty->print_raw(_print_inlining_output); |
4416 | } |
4417 | } |
4418 | |
4419 | void Compile::log_late_inline(CallGenerator* cg) { |
4420 | if (log() != NULL) { |
4421 | log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'" , log()->identify(cg->method()), |
4422 | cg->unique_id()); |
4423 | JVMState* p = cg->call_node()->jvms(); |
4424 | while (p != NULL) { |
4425 | log()->elem("jvms bci='%d' method='%d'" , p->bci(), log()->identify(p->method())); |
4426 | p = p->caller(); |
4427 | } |
4428 | log()->tail("late_inline" ); |
4429 | } |
4430 | } |
4431 | |
4432 | void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { |
4433 | log_late_inline(cg); |
4434 | if (log() != NULL) { |
4435 | log()->inline_fail(msg); |
4436 | } |
4437 | } |
4438 | |
4439 | void Compile::log_inline_id(CallGenerator* cg) { |
4440 | if (log() != NULL) { |
4441 | // The LogCompilation tool needs a unique way to identify late |
4442 | // inline call sites. This id must be unique for this call site in |
4443 | // this compilation. Try to have it unique across compilations as |
4444 | // well because it can be convenient when grepping through the log |
4445 | // file. |
4446 | // Distinguish OSR compilations from others in case CICountOSR is |
4447 | // on. |
4448 | jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); |
4449 | cg->set_unique_id(id); |
4450 | log()->elem("inline_id id='" JLONG_FORMAT "'" , id); |
4451 | } |
4452 | } |
4453 | |
4454 | void Compile::log_inline_failure(const char* msg) { |
4455 | if (C->log() != NULL) { |
4456 | C->log()->inline_fail(msg); |
4457 | } |
4458 | } |
4459 | |
4460 | |
4461 | // Dump inlining replay data to the stream. |
4462 | // Don't change thread state and acquire any locks. |
4463 | void Compile::dump_inline_data(outputStream* out) { |
4464 | InlineTree* inl_tree = ilt(); |
4465 | if (inl_tree != NULL) { |
4466 | out->print(" inline %d" , inl_tree->count()); |
4467 | inl_tree->dump_replay_data(out); |
4468 | } |
4469 | } |
4470 | |
4471 | int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { |
4472 | if (n1->Opcode() < n2->Opcode()) return -1; |
4473 | else if (n1->Opcode() > n2->Opcode()) return 1; |
4474 | |
4475 | assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d" , NodeClassNames[n1->Opcode()], n1->req(), n2->req()); |
4476 | for (uint i = 1; i < n1->req(); i++) { |
4477 | if (n1->in(i) < n2->in(i)) return -1; |
4478 | else if (n1->in(i) > n2->in(i)) return 1; |
4479 | } |
4480 | |
4481 | return 0; |
4482 | } |
4483 | |
4484 | int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { |
4485 | Node* n1 = *n1p; |
4486 | Node* n2 = *n2p; |
4487 | |
4488 | return cmp_expensive_nodes(n1, n2); |
4489 | } |
4490 | |
4491 | void Compile::sort_expensive_nodes() { |
4492 | if (!expensive_nodes_sorted()) { |
4493 | _expensive_nodes->sort(cmp_expensive_nodes); |
4494 | } |
4495 | } |
4496 | |
4497 | bool Compile::expensive_nodes_sorted() const { |
4498 | for (int i = 1; i < _expensive_nodes->length(); i++) { |
4499 | if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { |
4500 | return false; |
4501 | } |
4502 | } |
4503 | return true; |
4504 | } |
4505 | |
4506 | bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { |
4507 | if (_expensive_nodes->length() == 0) { |
4508 | return false; |
4509 | } |
4510 | |
4511 | assert(OptimizeExpensiveOps, "optimization off?" ); |
4512 | |
4513 | // Take this opportunity to remove dead nodes from the list |
4514 | int j = 0; |
4515 | for (int i = 0; i < _expensive_nodes->length(); i++) { |
4516 | Node* n = _expensive_nodes->at(i); |
4517 | if (!n->is_unreachable(igvn)) { |
4518 | assert(n->is_expensive(), "should be expensive" ); |
4519 | _expensive_nodes->at_put(j, n); |
4520 | j++; |
4521 | } |
4522 | } |
4523 | _expensive_nodes->trunc_to(j); |
4524 | |
4525 | // Then sort the list so that similar nodes are next to each other |
4526 | // and check for at least two nodes of identical kind with same data |
4527 | // inputs. |
4528 | sort_expensive_nodes(); |
4529 | |
4530 | for (int i = 0; i < _expensive_nodes->length()-1; i++) { |
4531 | if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { |
4532 | return true; |
4533 | } |
4534 | } |
4535 | |
4536 | return false; |
4537 | } |
4538 | |
4539 | void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { |
4540 | if (_expensive_nodes->length() == 0) { |
4541 | return; |
4542 | } |
4543 | |
4544 | assert(OptimizeExpensiveOps, "optimization off?" ); |
4545 | |
4546 | // Sort to bring similar nodes next to each other and clear the |
4547 | // control input of nodes for which there's only a single copy. |
4548 | sort_expensive_nodes(); |
4549 | |
4550 | int j = 0; |
4551 | int identical = 0; |
4552 | int i = 0; |
4553 | bool modified = false; |
4554 | for (; i < _expensive_nodes->length()-1; i++) { |
4555 | assert(j <= i, "can't write beyond current index" ); |
4556 | if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { |
4557 | identical++; |
4558 | _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); |
4559 | continue; |
4560 | } |
4561 | if (identical > 0) { |
4562 | _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); |
4563 | identical = 0; |
4564 | } else { |
4565 | Node* n = _expensive_nodes->at(i); |
4566 | igvn.replace_input_of(n, 0, NULL); |
4567 | igvn.hash_insert(n); |
4568 | modified = true; |
4569 | } |
4570 | } |
4571 | if (identical > 0) { |
4572 | _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); |
4573 | } else if (_expensive_nodes->length() >= 1) { |
4574 | Node* n = _expensive_nodes->at(i); |
4575 | igvn.replace_input_of(n, 0, NULL); |
4576 | igvn.hash_insert(n); |
4577 | modified = true; |
4578 | } |
4579 | _expensive_nodes->trunc_to(j); |
4580 | if (modified) { |
4581 | igvn.optimize(); |
4582 | } |
4583 | } |
4584 | |
4585 | void Compile::add_expensive_node(Node * n) { |
4586 | assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list" ); |
4587 | assert(n->is_expensive(), "expensive nodes with non-null control here only" ); |
4588 | assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here" ); |
4589 | if (OptimizeExpensiveOps) { |
4590 | _expensive_nodes->append(n); |
4591 | } else { |
4592 | // Clear control input and let IGVN optimize expensive nodes if |
4593 | // OptimizeExpensiveOps is off. |
4594 | n->set_req(0, NULL); |
4595 | } |
4596 | } |
4597 | |
4598 | /** |
4599 | * Remove the speculative part of types and clean up the graph |
4600 | */ |
4601 | void Compile::remove_speculative_types(PhaseIterGVN &igvn) { |
4602 | if (UseTypeSpeculation) { |
4603 | Unique_Node_List worklist; |
4604 | worklist.push(root()); |
4605 | int modified = 0; |
4606 | // Go over all type nodes that carry a speculative type, drop the |
4607 | // speculative part of the type and enqueue the node for an igvn |
4608 | // which may optimize it out. |
4609 | for (uint next = 0; next < worklist.size(); ++next) { |
4610 | Node *n = worklist.at(next); |
4611 | if (n->is_Type()) { |
4612 | TypeNode* tn = n->as_Type(); |
4613 | const Type* t = tn->type(); |
4614 | const Type* t_no_spec = t->remove_speculative(); |
4615 | if (t_no_spec != t) { |
4616 | bool in_hash = igvn.hash_delete(n); |
4617 | assert(in_hash, "node should be in igvn hash table" ); |
4618 | tn->set_type(t_no_spec); |
4619 | igvn.hash_insert(n); |
4620 | igvn._worklist.push(n); // give it a chance to go away |
4621 | modified++; |
4622 | } |
4623 | } |
4624 | uint max = n->len(); |
4625 | for( uint i = 0; i < max; ++i ) { |
4626 | Node *m = n->in(i); |
4627 | if (not_a_node(m)) continue; |
4628 | worklist.push(m); |
4629 | } |
4630 | } |
4631 | // Drop the speculative part of all types in the igvn's type table |
4632 | igvn.remove_speculative_types(); |
4633 | if (modified > 0) { |
4634 | igvn.optimize(); |
4635 | } |
4636 | #ifdef ASSERT |
4637 | // Verify that after the IGVN is over no speculative type has resurfaced |
4638 | worklist.clear(); |
4639 | worklist.push(root()); |
4640 | for (uint next = 0; next < worklist.size(); ++next) { |
4641 | Node *n = worklist.at(next); |
4642 | const Type* t = igvn.type_or_null(n); |
4643 | assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types" ); |
4644 | if (n->is_Type()) { |
4645 | t = n->as_Type()->type(); |
4646 | assert(t == t->remove_speculative(), "no more speculative types" ); |
4647 | } |
4648 | uint max = n->len(); |
4649 | for( uint i = 0; i < max; ++i ) { |
4650 | Node *m = n->in(i); |
4651 | if (not_a_node(m)) continue; |
4652 | worklist.push(m); |
4653 | } |
4654 | } |
4655 | igvn.check_no_speculative_types(); |
4656 | #endif |
4657 | } |
4658 | } |
4659 | |
4660 | // Auxiliary method to support randomized stressing/fuzzing. |
4661 | // |
4662 | // This method can be called the arbitrary number of times, with current count |
4663 | // as the argument. The logic allows selecting a single candidate from the |
4664 | // running list of candidates as follows: |
4665 | // int count = 0; |
4666 | // Cand* selected = null; |
4667 | // while(cand = cand->next()) { |
4668 | // if (randomized_select(++count)) { |
4669 | // selected = cand; |
4670 | // } |
4671 | // } |
4672 | // |
4673 | // Including count equalizes the chances any candidate is "selected". |
4674 | // This is useful when we don't have the complete list of candidates to choose |
4675 | // from uniformly. In this case, we need to adjust the randomicity of the |
4676 | // selection, or else we will end up biasing the selection towards the latter |
4677 | // candidates. |
4678 | // |
4679 | // Quick back-envelope calculation shows that for the list of n candidates |
4680 | // the equal probability for the candidate to persist as "best" can be |
4681 | // achieved by replacing it with "next" k-th candidate with the probability |
4682 | // of 1/k. It can be easily shown that by the end of the run, the |
4683 | // probability for any candidate is converged to 1/n, thus giving the |
4684 | // uniform distribution among all the candidates. |
4685 | // |
4686 | // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. |
4687 | #define RANDOMIZED_DOMAIN_POW 29 |
4688 | #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) |
4689 | #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) |
4690 | bool Compile::randomized_select(int count) { |
4691 | assert(count > 0, "only positive" ); |
4692 | return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); |
4693 | } |
4694 | |
4695 | CloneMap& Compile::clone_map() { return _clone_map; } |
4696 | void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } |
4697 | |
4698 | void NodeCloneInfo::dump() const { |
4699 | tty->print(" {%d:%d} " , idx(), gen()); |
4700 | } |
4701 | |
4702 | void CloneMap::clone(Node* old, Node* nnn, int gen) { |
4703 | uint64_t val = value(old->_idx); |
4704 | NodeCloneInfo cio(val); |
4705 | assert(val != 0, "old node should be in the map" ); |
4706 | NodeCloneInfo cin(cio.idx(), gen + cio.gen()); |
4707 | insert(nnn->_idx, cin.get()); |
4708 | #ifndef PRODUCT |
4709 | if (is_debug()) { |
4710 | tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap" , nnn->_idx, cin.idx(), cin.gen()); |
4711 | } |
4712 | #endif |
4713 | } |
4714 | |
4715 | void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { |
4716 | NodeCloneInfo cio(value(old->_idx)); |
4717 | if (cio.get() == 0) { |
4718 | cio.set(old->_idx, 0); |
4719 | insert(old->_idx, cio.get()); |
4720 | #ifndef PRODUCT |
4721 | if (is_debug()) { |
4722 | tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap" , old->_idx, cio.idx(), cio.gen()); |
4723 | } |
4724 | #endif |
4725 | } |
4726 | clone(old, nnn, gen); |
4727 | } |
4728 | |
4729 | int CloneMap::max_gen() const { |
4730 | int g = 0; |
4731 | DictI di(_dict); |
4732 | for(; di.test(); ++di) { |
4733 | int t = gen(di._key); |
4734 | if (g < t) { |
4735 | g = t; |
4736 | #ifndef PRODUCT |
4737 | if (is_debug()) { |
4738 | tty->print_cr("CloneMap::max_gen() update max=%d from %d" , g, _2_node_idx_t(di._key)); |
4739 | } |
4740 | #endif |
4741 | } |
4742 | } |
4743 | return g; |
4744 | } |
4745 | |
4746 | void CloneMap::dump(node_idx_t key) const { |
4747 | uint64_t val = value(key); |
4748 | if (val != 0) { |
4749 | NodeCloneInfo ni(val); |
4750 | ni.dump(); |
4751 | } |
4752 | } |
4753 | |