Tue, 24 Feb 2015 15:04:52 -0500
8072383: resolve conflicts between open and closed ports
Summary: refactor close to remove references to closed ports
Reviewed-by: kvn, simonis, sgehwolf, dholmes
1 /*
2 * Copyright (c) 1997, 2015, 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 */
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/compileLog.hpp"
33 #include "compiler/disassembler.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/block.hpp"
37 #include "opto/c2compiler.hpp"
38 #include "opto/callGenerator.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/cfgnode.hpp"
41 #include "opto/chaitin.hpp"
42 #include "opto/compile.hpp"
43 #include "opto/connode.hpp"
44 #include "opto/divnode.hpp"
45 #include "opto/escape.hpp"
46 #include "opto/idealGraphPrinter.hpp"
47 #include "opto/loopnode.hpp"
48 #include "opto/machnode.hpp"
49 #include "opto/macro.hpp"
50 #include "opto/matcher.hpp"
51 #include "opto/mathexactnode.hpp"
52 #include "opto/memnode.hpp"
53 #include "opto/mulnode.hpp"
54 #include "opto/node.hpp"
55 #include "opto/opcodes.hpp"
56 #include "opto/output.hpp"
57 #include "opto/parse.hpp"
58 #include "opto/phaseX.hpp"
59 #include "opto/rootnode.hpp"
60 #include "opto/runtime.hpp"
61 #include "opto/stringopts.hpp"
62 #include "opto/type.hpp"
63 #include "opto/vectornode.hpp"
64 #include "runtime/arguments.hpp"
65 #include "runtime/signature.hpp"
66 #include "runtime/stubRoutines.hpp"
67 #include "runtime/timer.hpp"
68 #include "trace/tracing.hpp"
69 #include "utilities/copy.hpp"
70 #if defined AD_MD_HPP
71 # include AD_MD_HPP
72 #elif defined TARGET_ARCH_MODEL_x86_32
73 # include "adfiles/ad_x86_32.hpp"
74 #elif defined TARGET_ARCH_MODEL_x86_64
75 # include "adfiles/ad_x86_64.hpp"
76 #elif defined TARGET_ARCH_MODEL_sparc
77 # include "adfiles/ad_sparc.hpp"
78 #elif defined TARGET_ARCH_MODEL_zero
79 # include "adfiles/ad_zero.hpp"
80 #elif defined TARGET_ARCH_MODEL_ppc_64
81 # include "adfiles/ad_ppc_64.hpp"
82 #endif
85 // -------------------- Compile::mach_constant_base_node -----------------------
86 // Constant table base node singleton.
87 MachConstantBaseNode* Compile::mach_constant_base_node() {
88 if (_mach_constant_base_node == NULL) {
89 _mach_constant_base_node = new (C) MachConstantBaseNode();
90 _mach_constant_base_node->add_req(C->root());
91 }
92 return _mach_constant_base_node;
93 }
96 /// Support for intrinsics.
98 // Return the index at which m must be inserted (or already exists).
99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
100 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
101 #ifdef ASSERT
102 for (int i = 1; i < _intrinsics->length(); i++) {
103 CallGenerator* cg1 = _intrinsics->at(i-1);
104 CallGenerator* cg2 = _intrinsics->at(i);
105 assert(cg1->method() != cg2->method()
106 ? cg1->method() < cg2->method()
107 : cg1->is_virtual() < cg2->is_virtual(),
108 "compiler intrinsics list must stay sorted");
109 }
110 #endif
111 // Binary search sorted list, in decreasing intervals [lo, hi].
112 int lo = 0, hi = _intrinsics->length()-1;
113 while (lo <= hi) {
114 int mid = (uint)(hi + lo) / 2;
115 ciMethod* mid_m = _intrinsics->at(mid)->method();
116 if (m < mid_m) {
117 hi = mid-1;
118 } else if (m > mid_m) {
119 lo = mid+1;
120 } else {
121 // look at minor sort key
122 bool mid_virt = _intrinsics->at(mid)->is_virtual();
123 if (is_virtual < mid_virt) {
124 hi = mid-1;
125 } else if (is_virtual > mid_virt) {
126 lo = mid+1;
127 } else {
128 return mid; // exact match
129 }
130 }
131 }
132 return lo; // inexact match
133 }
135 void Compile::register_intrinsic(CallGenerator* cg) {
136 if (_intrinsics == NULL) {
137 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
138 }
139 // This code is stolen from ciObjectFactory::insert.
140 // Really, GrowableArray should have methods for
141 // insert_at, remove_at, and binary_search.
142 int len = _intrinsics->length();
143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
144 if (index == len) {
145 _intrinsics->append(cg);
146 } else {
147 #ifdef ASSERT
148 CallGenerator* oldcg = _intrinsics->at(index);
149 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
150 #endif
151 _intrinsics->append(_intrinsics->at(len-1));
152 int pos;
153 for (pos = len-2; pos >= index; pos--) {
154 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
155 }
156 _intrinsics->at_put(index, cg);
157 }
158 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
159 }
161 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
162 assert(m->is_loaded(), "don't try this on unloaded methods");
163 if (_intrinsics != NULL) {
164 int index = intrinsic_insertion_index(m, is_virtual);
165 if (index < _intrinsics->length()
166 && _intrinsics->at(index)->method() == m
167 && _intrinsics->at(index)->is_virtual() == is_virtual) {
168 return _intrinsics->at(index);
169 }
170 }
171 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
172 if (m->intrinsic_id() != vmIntrinsics::_none &&
173 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
174 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
175 if (cg != NULL) {
176 // Save it for next time:
177 register_intrinsic(cg);
178 return cg;
179 } else {
180 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
181 }
182 }
183 return NULL;
184 }
186 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
187 // in library_call.cpp.
190 #ifndef PRODUCT
191 // statistics gathering...
193 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
194 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
196 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
197 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
198 int oflags = _intrinsic_hist_flags[id];
199 assert(flags != 0, "what happened?");
200 if (is_virtual) {
201 flags |= _intrinsic_virtual;
202 }
203 bool changed = (flags != oflags);
204 if ((flags & _intrinsic_worked) != 0) {
205 juint count = (_intrinsic_hist_count[id] += 1);
206 if (count == 1) {
207 changed = true; // first time
208 }
209 // increment the overall count also:
210 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
211 }
212 if (changed) {
213 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
214 // Something changed about the intrinsic's virtuality.
215 if ((flags & _intrinsic_virtual) != 0) {
216 // This is the first use of this intrinsic as a virtual call.
217 if (oflags != 0) {
218 // We already saw it as a non-virtual, so note both cases.
219 flags |= _intrinsic_both;
220 }
221 } else if ((oflags & _intrinsic_both) == 0) {
222 // This is the first use of this intrinsic as a non-virtual
223 flags |= _intrinsic_both;
224 }
225 }
226 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
227 }
228 // update the overall flags also:
229 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
230 return changed;
231 }
233 static char* format_flags(int flags, char* buf) {
234 buf[0] = 0;
235 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
236 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
237 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
238 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
239 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
240 if (buf[0] == 0) strcat(buf, ",");
241 assert(buf[0] == ',', "must be");
242 return &buf[1];
243 }
245 void Compile::print_intrinsic_statistics() {
246 char flagsbuf[100];
247 ttyLocker ttyl;
248 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
249 tty->print_cr("Compiler intrinsic usage:");
250 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
251 if (total == 0) total = 1; // avoid div0 in case of no successes
252 #define PRINT_STAT_LINE(name, c, f) \
253 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
254 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
255 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
256 int flags = _intrinsic_hist_flags[id];
257 juint count = _intrinsic_hist_count[id];
258 if ((flags | count) != 0) {
259 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
260 }
261 }
262 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
263 if (xtty != NULL) xtty->tail("statistics");
264 }
266 void Compile::print_statistics() {
267 { ttyLocker ttyl;
268 if (xtty != NULL) xtty->head("statistics type='opto'");
269 Parse::print_statistics();
270 PhaseCCP::print_statistics();
271 PhaseRegAlloc::print_statistics();
272 Scheduling::print_statistics();
273 PhasePeephole::print_statistics();
274 PhaseIdealLoop::print_statistics();
275 if (xtty != NULL) xtty->tail("statistics");
276 }
277 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
278 // put this under its own <statistics> element.
279 print_intrinsic_statistics();
280 }
281 }
282 #endif //PRODUCT
284 // Support for bundling info
285 Bundle* Compile::node_bundling(const Node *n) {
286 assert(valid_bundle_info(n), "oob");
287 return &_node_bundling_base[n->_idx];
288 }
290 bool Compile::valid_bundle_info(const Node *n) {
291 return (_node_bundling_limit > n->_idx);
292 }
295 void Compile::gvn_replace_by(Node* n, Node* nn) {
296 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
297 Node* use = n->last_out(i);
298 bool is_in_table = initial_gvn()->hash_delete(use);
299 uint uses_found = 0;
300 for (uint j = 0; j < use->len(); j++) {
301 if (use->in(j) == n) {
302 if (j < use->req())
303 use->set_req(j, nn);
304 else
305 use->set_prec(j, nn);
306 uses_found++;
307 }
308 }
309 if (is_in_table) {
310 // reinsert into table
311 initial_gvn()->hash_find_insert(use);
312 }
313 record_for_igvn(use);
314 i -= uses_found; // we deleted 1 or more copies of this edge
315 }
316 }
319 static inline bool not_a_node(const Node* n) {
320 if (n == NULL) return true;
321 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
322 if (*(address*)n == badAddress) return true; // kill by Node::destruct
323 return false;
324 }
326 // Identify all nodes that are reachable from below, useful.
327 // Use breadth-first pass that records state in a Unique_Node_List,
328 // recursive traversal is slower.
329 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
330 int estimated_worklist_size = unique();
331 useful.map( estimated_worklist_size, NULL ); // preallocate space
333 // Initialize worklist
334 if (root() != NULL) { useful.push(root()); }
335 // If 'top' is cached, declare it useful to preserve cached node
336 if( cached_top_node() ) { useful.push(cached_top_node()); }
338 // Push all useful nodes onto the list, breadthfirst
339 for( uint next = 0; next < useful.size(); ++next ) {
340 assert( next < unique(), "Unique useful nodes < total nodes");
341 Node *n = useful.at(next);
342 uint max = n->len();
343 for( uint i = 0; i < max; ++i ) {
344 Node *m = n->in(i);
345 if (not_a_node(m)) continue;
346 useful.push(m);
347 }
348 }
349 }
351 // Update dead_node_list with any missing dead nodes using useful
352 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
353 void Compile::update_dead_node_list(Unique_Node_List &useful) {
354 uint max_idx = unique();
355 VectorSet& useful_node_set = useful.member_set();
357 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
358 // If node with index node_idx is not in useful set,
359 // mark it as dead in dead node list.
360 if (! useful_node_set.test(node_idx) ) {
361 record_dead_node(node_idx);
362 }
363 }
364 }
366 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
367 int shift = 0;
368 for (int i = 0; i < inlines->length(); i++) {
369 CallGenerator* cg = inlines->at(i);
370 CallNode* call = cg->call_node();
371 if (shift > 0) {
372 inlines->at_put(i-shift, cg);
373 }
374 if (!useful.member(call)) {
375 shift++;
376 }
377 }
378 inlines->trunc_to(inlines->length()-shift);
379 }
381 // Disconnect all useless nodes by disconnecting those at the boundary.
382 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
383 uint next = 0;
384 while (next < useful.size()) {
385 Node *n = useful.at(next++);
386 if (n->is_SafePoint()) {
387 // We're done with a parsing phase. Replaced nodes are not valid
388 // beyond that point.
389 n->as_SafePoint()->delete_replaced_nodes();
390 }
391 // Use raw traversal of out edges since this code removes out edges
392 int max = n->outcnt();
393 for (int j = 0; j < max; ++j) {
394 Node* child = n->raw_out(j);
395 if (! useful.member(child)) {
396 assert(!child->is_top() || child != top(),
397 "If top is cached in Compile object it is in useful list");
398 // Only need to remove this out-edge to the useless node
399 n->raw_del_out(j);
400 --j;
401 --max;
402 }
403 }
404 if (n->outcnt() == 1 && n->has_special_unique_user()) {
405 record_for_igvn(n->unique_out());
406 }
407 }
408 // Remove useless macro and predicate opaq nodes
409 for (int i = C->macro_count()-1; i >= 0; i--) {
410 Node* n = C->macro_node(i);
411 if (!useful.member(n)) {
412 remove_macro_node(n);
413 }
414 }
415 // Remove useless expensive node
416 for (int i = C->expensive_count()-1; i >= 0; i--) {
417 Node* n = C->expensive_node(i);
418 if (!useful.member(n)) {
419 remove_expensive_node(n);
420 }
421 }
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 }
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 }
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(_interpreter_frame_size, frame_size_in_bytes());
445 }
447 // ============================================================================
448 //------------------------------CompileWrapper---------------------------------
449 class CompileWrapper : public StackObj {
450 Compile *const _compile;
451 public:
452 CompileWrapper(Compile* compile);
454 ~CompileWrapper();
455 };
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");
465 compile->set_type_dict(NULL);
466 compile->set_type_hwm(NULL);
467 compile->set_type_last_size(0);
468 compile->set_last_tf(NULL, NULL);
469 compile->set_indexSet_arena(NULL);
470 compile->set_indexSet_free_block_list(NULL);
471 compile->init_type_arena();
472 Type::Initialize(compile);
473 _compile->set_scratch_buffer_blob(NULL);
474 _compile->begin_method();
475 }
476 CompileWrapper::~CompileWrapper() {
477 _compile->end_method();
478 if (_compile->scratch_buffer_blob() != NULL)
479 BufferBlob::free(_compile->scratch_buffer_blob());
480 _compile->env()->set_compiler_data(NULL);
481 }
484 //----------------------------print_compile_messages---------------------------
485 void Compile::print_compile_messages() {
486 #ifndef PRODUCT
487 // Check if recompiling
488 if (_subsume_loads == false && PrintOpto) {
489 // Recompiling without allowing machine instructions to subsume loads
490 tty->print_cr("*********************************************************");
491 tty->print_cr("** Bailout: Recompile without subsuming loads **");
492 tty->print_cr("*********************************************************");
493 }
494 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
495 // Recompiling without escape analysis
496 tty->print_cr("*********************************************************");
497 tty->print_cr("** Bailout: Recompile without escape analysis **");
498 tty->print_cr("*********************************************************");
499 }
500 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
501 // Recompiling without boxing elimination
502 tty->print_cr("*********************************************************");
503 tty->print_cr("** Bailout: Recompile without boxing elimination **");
504 tty->print_cr("*********************************************************");
505 }
506 if (env()->break_at_compile()) {
507 // Open the debugger when compiling this method.
508 tty->print("### Breaking when compiling: ");
509 method()->print_short_name();
510 tty->cr();
511 BREAKPOINT;
512 }
514 if( PrintOpto ) {
515 if (is_osr_compilation()) {
516 tty->print("[OSR]%3d", _compile_id);
517 } else {
518 tty->print("%3d", _compile_id);
519 }
520 }
521 #endif
522 }
525 //-----------------------init_scratch_buffer_blob------------------------------
526 // Construct a temporary BufferBlob and cache it for this compile.
527 void Compile::init_scratch_buffer_blob(int const_size) {
528 // If there is already a scratch buffer blob allocated and the
529 // constant section is big enough, use it. Otherwise free the
530 // current and allocate a new one.
531 BufferBlob* blob = scratch_buffer_blob();
532 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
533 // Use the current blob.
534 } else {
535 if (blob != NULL) {
536 BufferBlob::free(blob);
537 }
539 ResourceMark rm;
540 _scratch_const_size = const_size;
541 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
542 blob = BufferBlob::create("Compile::scratch_buffer", size);
543 // Record the buffer blob for next time.
544 set_scratch_buffer_blob(blob);
545 // Have we run out of code space?
546 if (scratch_buffer_blob() == NULL) {
547 // Let CompilerBroker disable further compilations.
548 record_failure("Not enough space for scratch buffer in CodeCache");
549 return;
550 }
551 }
553 // Initialize the relocation buffers
554 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
555 set_scratch_locs_memory(locs_buf);
556 }
559 //-----------------------scratch_emit_size-------------------------------------
560 // Helper function that computes size by emitting code
561 uint Compile::scratch_emit_size(const Node* n) {
562 // Start scratch_emit_size section.
563 set_in_scratch_emit_size(true);
565 // Emit into a trash buffer and count bytes emitted.
566 // This is a pretty expensive way to compute a size,
567 // but it works well enough if seldom used.
568 // All common fixed-size instructions are given a size
569 // method by the AD file.
570 // Note that the scratch buffer blob and locs memory are
571 // allocated at the beginning of the compile task, and
572 // may be shared by several calls to scratch_emit_size.
573 // The allocation of the scratch buffer blob is particularly
574 // expensive, since it has to grab the code cache lock.
575 BufferBlob* blob = this->scratch_buffer_blob();
576 assert(blob != NULL, "Initialize BufferBlob at start");
577 assert(blob->size() > MAX_inst_size, "sanity");
578 relocInfo* locs_buf = scratch_locs_memory();
579 address blob_begin = blob->content_begin();
580 address blob_end = (address)locs_buf;
581 assert(blob->content_contains(blob_end), "sanity");
582 CodeBuffer buf(blob_begin, blob_end - blob_begin);
583 buf.initialize_consts_size(_scratch_const_size);
584 buf.initialize_stubs_size(MAX_stubs_size);
585 assert(locs_buf != NULL, "sanity");
586 int lsize = MAX_locs_size / 3;
587 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
588 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
589 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
591 // Do the emission.
593 Label fakeL; // Fake label for branch instructions.
594 Label* saveL = NULL;
595 uint save_bnum = 0;
596 bool is_branch = n->is_MachBranch();
597 if (is_branch) {
598 MacroAssembler masm(&buf);
599 masm.bind(fakeL);
600 n->as_MachBranch()->save_label(&saveL, &save_bnum);
601 n->as_MachBranch()->label_set(&fakeL, 0);
602 }
603 n->emit(buf, this->regalloc());
604 if (is_branch) // Restore label.
605 n->as_MachBranch()->label_set(saveL, save_bnum);
607 // End scratch_emit_size section.
608 set_in_scratch_emit_size(false);
610 return buf.insts_size();
611 }
614 // ============================================================================
615 //------------------------------Compile standard-------------------------------
616 debug_only( int Compile::_debug_idx = 100000; )
618 // Compile a method. entry_bci is -1 for normal compilations and indicates
619 // the continuation bci for on stack replacement.
622 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
623 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
624 : Phase(Compiler),
625 _env(ci_env),
626 _log(ci_env->log()),
627 _compile_id(ci_env->compile_id()),
628 _save_argument_registers(false),
629 _stub_name(NULL),
630 _stub_function(NULL),
631 _stub_entry_point(NULL),
632 _method(target),
633 _entry_bci(osr_bci),
634 _initial_gvn(NULL),
635 _for_igvn(NULL),
636 _warm_calls(NULL),
637 _subsume_loads(subsume_loads),
638 _do_escape_analysis(do_escape_analysis),
639 _eliminate_boxing(eliminate_boxing),
640 _failure_reason(NULL),
641 _code_buffer("Compile::Fill_buffer"),
642 _orig_pc_slot(0),
643 _orig_pc_slot_offset_in_bytes(0),
644 _has_method_handle_invokes(false),
645 _mach_constant_base_node(NULL),
646 _node_bundling_limit(0),
647 _node_bundling_base(NULL),
648 _java_calls(0),
649 _inner_loops(0),
650 _scratch_const_size(-1),
651 _in_scratch_emit_size(false),
652 _dead_node_list(comp_arena()),
653 _dead_node_count(0),
654 #ifndef PRODUCT
655 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
656 _in_dump_cnt(0),
657 _printer(IdealGraphPrinter::printer()),
658 #endif
659 _congraph(NULL),
660 _comp_arena(mtCompiler),
661 _node_arena(mtCompiler),
662 _old_arena(mtCompiler),
663 _Compile_types(mtCompiler),
664 _replay_inline_data(NULL),
665 _late_inlines(comp_arena(), 2, 0, NULL),
666 _string_late_inlines(comp_arena(), 2, 0, NULL),
667 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
668 _late_inlines_pos(0),
669 _number_of_mh_late_inlines(0),
670 _inlining_progress(false),
671 _inlining_incrementally(false),
672 _print_inlining_list(NULL),
673 _print_inlining_idx(0),
674 _interpreter_frame_size(0),
675 _max_node_limit(MaxNodeLimit) {
676 C = this;
678 CompileWrapper cw(this);
679 #ifndef PRODUCT
680 if (TimeCompiler2) {
681 tty->print(" ");
682 target->holder()->name()->print();
683 tty->print(".");
684 target->print_short_name();
685 tty->print(" ");
686 }
687 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
688 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
689 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
690 if (!print_opto_assembly) {
691 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
692 if (print_assembly && !Disassembler::can_decode()) {
693 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
694 print_opto_assembly = true;
695 }
696 }
697 set_print_assembly(print_opto_assembly);
698 set_parsed_irreducible_loop(false);
700 if (method()->has_option("ReplayInline")) {
701 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
702 }
703 #endif
704 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
705 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
706 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
708 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
709 // Make sure the method being compiled gets its own MDO,
710 // so we can at least track the decompile_count().
711 // Need MDO to record RTM code generation state.
712 method()->ensure_method_data();
713 }
715 Init(::AliasLevel);
718 print_compile_messages();
720 _ilt = InlineTree::build_inline_tree_root();
722 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
723 assert(num_alias_types() >= AliasIdxRaw, "");
725 #define MINIMUM_NODE_HASH 1023
726 // Node list that Iterative GVN will start with
727 Unique_Node_List for_igvn(comp_arena());
728 set_for_igvn(&for_igvn);
730 // GVN that will be run immediately on new nodes
731 uint estimated_size = method()->code_size()*4+64;
732 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
733 PhaseGVN gvn(node_arena(), estimated_size);
734 set_initial_gvn(&gvn);
736 if (print_inlining() || print_intrinsics()) {
737 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
738 }
739 { // Scope for timing the parser
740 TracePhase t3("parse", &_t_parser, true);
742 // Put top into the hash table ASAP.
743 initial_gvn()->transform_no_reclaim(top());
745 // Set up tf(), start(), and find a CallGenerator.
746 CallGenerator* cg = NULL;
747 if (is_osr_compilation()) {
748 const TypeTuple *domain = StartOSRNode::osr_domain();
749 const TypeTuple *range = TypeTuple::make_range(method()->signature());
750 init_tf(TypeFunc::make(domain, range));
751 StartNode* s = new (this) StartOSRNode(root(), domain);
752 initial_gvn()->set_type_bottom(s);
753 init_start(s);
754 cg = CallGenerator::for_osr(method(), entry_bci());
755 } else {
756 // Normal case.
757 init_tf(TypeFunc::make(method()));
758 StartNode* s = new (this) StartNode(root(), tf()->domain());
759 initial_gvn()->set_type_bottom(s);
760 init_start(s);
761 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
762 // With java.lang.ref.reference.get() we must go through the
763 // intrinsic when G1 is enabled - even when get() is the root
764 // method of the compile - so that, if necessary, the value in
765 // the referent field of the reference object gets recorded by
766 // the pre-barrier code.
767 // Specifically, if G1 is enabled, the value in the referent
768 // field is recorded by the G1 SATB pre barrier. This will
769 // result in the referent being marked live and the reference
770 // object removed from the list of discovered references during
771 // reference processing.
772 cg = find_intrinsic(method(), false);
773 }
774 if (cg == NULL) {
775 float past_uses = method()->interpreter_invocation_count();
776 float expected_uses = past_uses;
777 cg = CallGenerator::for_inline(method(), expected_uses);
778 }
779 }
780 if (failing()) return;
781 if (cg == NULL) {
782 record_method_not_compilable_all_tiers("cannot parse method");
783 return;
784 }
785 JVMState* jvms = build_start_state(start(), tf());
786 if ((jvms = cg->generate(jvms)) == NULL) {
787 record_method_not_compilable("method parse failed");
788 return;
789 }
790 GraphKit kit(jvms);
792 if (!kit.stopped()) {
793 // Accept return values, and transfer control we know not where.
794 // This is done by a special, unique ReturnNode bound to root.
795 return_values(kit.jvms());
796 }
798 if (kit.has_exceptions()) {
799 // Any exceptions that escape from this call must be rethrown
800 // to whatever caller is dynamically above us on the stack.
801 // This is done by a special, unique RethrowNode bound to root.
802 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
803 }
805 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
807 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
808 inline_string_calls(true);
809 }
811 if (failing()) return;
813 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
815 // Remove clutter produced by parsing.
816 if (!failing()) {
817 ResourceMark rm;
818 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
819 }
820 }
822 // Note: Large methods are capped off in do_one_bytecode().
823 if (failing()) return;
825 // After parsing, node notes are no longer automagic.
826 // They must be propagated by register_new_node_with_optimizer(),
827 // clone(), or the like.
828 set_default_node_notes(NULL);
830 for (;;) {
831 int successes = Inline_Warm();
832 if (failing()) return;
833 if (successes == 0) break;
834 }
836 // Drain the list.
837 Finish_Warm();
838 #ifndef PRODUCT
839 if (_printer) {
840 _printer->print_inlining(this);
841 }
842 #endif
844 if (failing()) return;
845 NOT_PRODUCT( verify_graph_edges(); )
847 // Now optimize
848 Optimize();
849 if (failing()) return;
850 NOT_PRODUCT( verify_graph_edges(); )
852 #ifndef PRODUCT
853 if (PrintIdeal) {
854 ttyLocker ttyl; // keep the following output all in one block
855 // This output goes directly to the tty, not the compiler log.
856 // To enable tools to match it up with the compilation activity,
857 // be sure to tag this tty output with the compile ID.
858 if (xtty != NULL) {
859 xtty->head("ideal compile_id='%d'%s", compile_id(),
860 is_osr_compilation() ? " compile_kind='osr'" :
861 "");
862 }
863 root()->dump(9999);
864 if (xtty != NULL) {
865 xtty->tail("ideal");
866 }
867 }
868 #endif
870 NOT_PRODUCT( verify_barriers(); )
872 // Dump compilation data to replay it.
873 if (method()->has_option("DumpReplay")) {
874 env()->dump_replay_data(_compile_id);
875 }
876 if (method()->has_option("DumpInline") && (ilt() != NULL)) {
877 env()->dump_inline_data(_compile_id);
878 }
880 // Now that we know the size of all the monitors we can add a fixed slot
881 // for the original deopt pc.
883 _orig_pc_slot = fixed_slots();
884 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
885 set_fixed_slots(next_slot);
887 // Compute when to use implicit null checks. Used by matching trap based
888 // nodes and NullCheck optimization.
889 set_allowed_deopt_reasons();
891 // Now generate code
892 Code_Gen();
893 if (failing()) return;
895 // Check if we want to skip execution of all compiled code.
896 {
897 #ifndef PRODUCT
898 if (OptoNoExecute) {
899 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
900 return;
901 }
902 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
903 #endif
905 if (is_osr_compilation()) {
906 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
907 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
908 } else {
909 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
910 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
911 }
913 env()->register_method(_method, _entry_bci,
914 &_code_offsets,
915 _orig_pc_slot_offset_in_bytes,
916 code_buffer(),
917 frame_size_in_words(), _oop_map_set,
918 &_handler_table, &_inc_table,
919 compiler,
920 env()->comp_level(),
921 has_unsafe_access(),
922 SharedRuntime::is_wide_vector(max_vector_size()),
923 rtm_state()
924 );
926 if (log() != NULL) // Print code cache state into compiler log
927 log()->code_cache_state();
928 }
929 }
931 //------------------------------Compile----------------------------------------
932 // Compile a runtime stub
933 Compile::Compile( ciEnv* ci_env,
934 TypeFunc_generator generator,
935 address stub_function,
936 const char *stub_name,
937 int is_fancy_jump,
938 bool pass_tls,
939 bool save_arg_registers,
940 bool return_pc )
941 : Phase(Compiler),
942 _env(ci_env),
943 _log(ci_env->log()),
944 _compile_id(0),
945 _save_argument_registers(save_arg_registers),
946 _method(NULL),
947 _stub_name(stub_name),
948 _stub_function(stub_function),
949 _stub_entry_point(NULL),
950 _entry_bci(InvocationEntryBci),
951 _initial_gvn(NULL),
952 _for_igvn(NULL),
953 _warm_calls(NULL),
954 _orig_pc_slot(0),
955 _orig_pc_slot_offset_in_bytes(0),
956 _subsume_loads(true),
957 _do_escape_analysis(false),
958 _eliminate_boxing(false),
959 _failure_reason(NULL),
960 _code_buffer("Compile::Fill_buffer"),
961 _has_method_handle_invokes(false),
962 _mach_constant_base_node(NULL),
963 _node_bundling_limit(0),
964 _node_bundling_base(NULL),
965 _java_calls(0),
966 _inner_loops(0),
967 #ifndef PRODUCT
968 _trace_opto_output(TraceOptoOutput),
969 _in_dump_cnt(0),
970 _printer(NULL),
971 #endif
972 _comp_arena(mtCompiler),
973 _node_arena(mtCompiler),
974 _old_arena(mtCompiler),
975 _Compile_types(mtCompiler),
976 _dead_node_list(comp_arena()),
977 _dead_node_count(0),
978 _congraph(NULL),
979 _replay_inline_data(NULL),
980 _number_of_mh_late_inlines(0),
981 _inlining_progress(false),
982 _inlining_incrementally(false),
983 _print_inlining_list(NULL),
984 _print_inlining_idx(0),
985 _allowed_reasons(0),
986 _interpreter_frame_size(0),
987 _max_node_limit(MaxNodeLimit) {
988 C = this;
990 #ifndef PRODUCT
991 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
992 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
993 set_print_assembly(PrintFrameConverterAssembly);
994 set_parsed_irreducible_loop(false);
995 #endif
996 set_has_irreducible_loop(false); // no loops
998 CompileWrapper cw(this);
999 Init(/*AliasLevel=*/ 0);
1000 init_tf((*generator)());
1002 {
1003 // The following is a dummy for the sake of GraphKit::gen_stub
1004 Unique_Node_List for_igvn(comp_arena());
1005 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1006 PhaseGVN gvn(Thread::current()->resource_area(),255);
1007 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1008 gvn.transform_no_reclaim(top());
1010 GraphKit kit;
1011 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1012 }
1014 NOT_PRODUCT( verify_graph_edges(); )
1015 Code_Gen();
1016 if (failing()) return;
1019 // Entry point will be accessed using compile->stub_entry_point();
1020 if (code_buffer() == NULL) {
1021 Matcher::soft_match_failure();
1022 } else {
1023 if (PrintAssembly && (WizardMode || Verbose))
1024 tty->print_cr("### Stub::%s", stub_name);
1026 if (!failing()) {
1027 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1029 // Make the NMethod
1030 // For now we mark the frame as never safe for profile stackwalking
1031 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1032 code_buffer(),
1033 CodeOffsets::frame_never_safe,
1034 // _code_offsets.value(CodeOffsets::Frame_Complete),
1035 frame_size_in_words(),
1036 _oop_map_set,
1037 save_arg_registers);
1038 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1040 _stub_entry_point = rs->entry_point();
1041 }
1042 }
1043 }
1045 //------------------------------Init-------------------------------------------
1046 // Prepare for a single compilation
1047 void Compile::Init(int aliaslevel) {
1048 _unique = 0;
1049 _regalloc = NULL;
1051 _tf = NULL; // filled in later
1052 _top = NULL; // cached later
1053 _matcher = NULL; // filled in later
1054 _cfg = NULL; // filled in later
1056 set_24_bit_selection_and_mode(Use24BitFP, false);
1058 _node_note_array = NULL;
1059 _default_node_notes = NULL;
1061 _immutable_memory = NULL; // filled in at first inquiry
1063 // Globally visible Nodes
1064 // First set TOP to NULL to give safe behavior during creation of RootNode
1065 set_cached_top_node(NULL);
1066 set_root(new (this) RootNode());
1067 // Now that you have a Root to point to, create the real TOP
1068 set_cached_top_node( new (this) ConNode(Type::TOP) );
1069 set_recent_alloc(NULL, NULL);
1071 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1072 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1073 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1074 env()->set_dependencies(new Dependencies(env()));
1076 _fixed_slots = 0;
1077 set_has_split_ifs(false);
1078 set_has_loops(has_method() && method()->has_loops()); // first approximation
1079 set_has_stringbuilder(false);
1080 set_has_boxed_value(false);
1081 _trap_can_recompile = false; // no traps emitted yet
1082 _major_progress = true; // start out assuming good things will happen
1083 set_has_unsafe_access(false);
1084 set_max_vector_size(0);
1085 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1086 set_decompile_count(0);
1088 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1089 set_num_loop_opts(LoopOptsCount);
1090 set_do_inlining(Inline);
1091 set_max_inline_size(MaxInlineSize);
1092 set_freq_inline_size(FreqInlineSize);
1093 set_do_scheduling(OptoScheduling);
1094 set_do_count_invocations(false);
1095 set_do_method_data_update(false);
1096 set_rtm_state(NoRTM); // No RTM lock eliding by default
1097 method_has_option_value("MaxNodeLimit", _max_node_limit);
1098 #if INCLUDE_RTM_OPT
1099 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1100 int rtm_state = method()->method_data()->rtm_state();
1101 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1102 // Don't generate RTM lock eliding code.
1103 set_rtm_state(NoRTM);
1104 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1105 // Generate RTM lock eliding code without abort ratio calculation code.
1106 set_rtm_state(UseRTM);
1107 } else if (UseRTMDeopt) {
1108 // Generate RTM lock eliding code and include abort ratio calculation
1109 // code if UseRTMDeopt is on.
1110 set_rtm_state(ProfileRTM);
1111 }
1112 }
1113 #endif
1114 if (debug_info()->recording_non_safepoints()) {
1115 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1116 (comp_arena(), 8, 0, NULL));
1117 set_default_node_notes(Node_Notes::make(this));
1118 }
1120 // // -- Initialize types before each compile --
1121 // // Update cached type information
1122 // if( _method && _method->constants() )
1123 // Type::update_loaded_types(_method, _method->constants());
1125 // Init alias_type map.
1126 if (!_do_escape_analysis && aliaslevel == 3)
1127 aliaslevel = 2; // No unique types without escape analysis
1128 _AliasLevel = aliaslevel;
1129 const int grow_ats = 16;
1130 _max_alias_types = grow_ats;
1131 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1132 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1133 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1134 {
1135 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1136 }
1137 // Initialize the first few types.
1138 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1139 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1140 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1141 _num_alias_types = AliasIdxRaw+1;
1142 // Zero out the alias type cache.
1143 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1144 // A NULL adr_type hits in the cache right away. Preload the right answer.
1145 probe_alias_cache(NULL)->_index = AliasIdxTop;
1147 _intrinsics = NULL;
1148 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1149 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1150 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1151 register_library_intrinsics();
1152 }
1154 //---------------------------init_start----------------------------------------
1155 // Install the StartNode on this compile object.
1156 void Compile::init_start(StartNode* s) {
1157 if (failing())
1158 return; // already failing
1159 assert(s == start(), "");
1160 }
1162 StartNode* Compile::start() const {
1163 assert(!failing(), "");
1164 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1165 Node* start = root()->fast_out(i);
1166 if( start->is_Start() )
1167 return start->as_Start();
1168 }
1169 fatal("Did not find Start node!");
1170 return NULL;
1171 }
1173 //-------------------------------immutable_memory-------------------------------------
1174 // Access immutable memory
1175 Node* Compile::immutable_memory() {
1176 if (_immutable_memory != NULL) {
1177 return _immutable_memory;
1178 }
1179 StartNode* s = start();
1180 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1181 Node *p = s->fast_out(i);
1182 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1183 _immutable_memory = p;
1184 return _immutable_memory;
1185 }
1186 }
1187 ShouldNotReachHere();
1188 return NULL;
1189 }
1191 //----------------------set_cached_top_node------------------------------------
1192 // Install the cached top node, and make sure Node::is_top works correctly.
1193 void Compile::set_cached_top_node(Node* tn) {
1194 if (tn != NULL) verify_top(tn);
1195 Node* old_top = _top;
1196 _top = tn;
1197 // Calling Node::setup_is_top allows the nodes the chance to adjust
1198 // their _out arrays.
1199 if (_top != NULL) _top->setup_is_top();
1200 if (old_top != NULL) old_top->setup_is_top();
1201 assert(_top == NULL || top()->is_top(), "");
1202 }
1204 #ifdef ASSERT
1205 uint Compile::count_live_nodes_by_graph_walk() {
1206 Unique_Node_List useful(comp_arena());
1207 // Get useful node list by walking the graph.
1208 identify_useful_nodes(useful);
1209 return useful.size();
1210 }
1212 void Compile::print_missing_nodes() {
1214 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1215 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1216 return;
1217 }
1219 // This is an expensive function. It is executed only when the user
1220 // specifies VerifyIdealNodeCount option or otherwise knows the
1221 // additional work that needs to be done to identify reachable nodes
1222 // by walking the flow graph and find the missing ones using
1223 // _dead_node_list.
1225 Unique_Node_List useful(comp_arena());
1226 // Get useful node list by walking the graph.
1227 identify_useful_nodes(useful);
1229 uint l_nodes = C->live_nodes();
1230 uint l_nodes_by_walk = useful.size();
1232 if (l_nodes != l_nodes_by_walk) {
1233 if (_log != NULL) {
1234 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1235 _log->stamp();
1236 _log->end_head();
1237 }
1238 VectorSet& useful_member_set = useful.member_set();
1239 int last_idx = l_nodes_by_walk;
1240 for (int i = 0; i < last_idx; i++) {
1241 if (useful_member_set.test(i)) {
1242 if (_dead_node_list.test(i)) {
1243 if (_log != NULL) {
1244 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1245 }
1246 if (PrintIdealNodeCount) {
1247 // Print the log message to tty
1248 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1249 useful.at(i)->dump();
1250 }
1251 }
1252 }
1253 else if (! _dead_node_list.test(i)) {
1254 if (_log != NULL) {
1255 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1256 }
1257 if (PrintIdealNodeCount) {
1258 // Print the log message to tty
1259 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1260 }
1261 }
1262 }
1263 if (_log != NULL) {
1264 _log->tail("mismatched_nodes");
1265 }
1266 }
1267 }
1268 #endif
1270 #ifndef PRODUCT
1271 void Compile::verify_top(Node* tn) const {
1272 if (tn != NULL) {
1273 assert(tn->is_Con(), "top node must be a constant");
1274 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1275 assert(tn->in(0) != NULL, "must have live top node");
1276 }
1277 }
1278 #endif
1281 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1283 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1284 guarantee(arr != NULL, "");
1285 int num_blocks = arr->length();
1286 if (grow_by < num_blocks) grow_by = num_blocks;
1287 int num_notes = grow_by * _node_notes_block_size;
1288 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1289 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1290 while (num_notes > 0) {
1291 arr->append(notes);
1292 notes += _node_notes_block_size;
1293 num_notes -= _node_notes_block_size;
1294 }
1295 assert(num_notes == 0, "exact multiple, please");
1296 }
1298 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1299 if (source == NULL || dest == NULL) return false;
1301 if (dest->is_Con())
1302 return false; // Do not push debug info onto constants.
1304 #ifdef ASSERT
1305 // Leave a bread crumb trail pointing to the original node:
1306 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1307 dest->set_debug_orig(source);
1308 }
1309 #endif
1311 if (node_note_array() == NULL)
1312 return false; // Not collecting any notes now.
1314 // This is a copy onto a pre-existing node, which may already have notes.
1315 // If both nodes have notes, do not overwrite any pre-existing notes.
1316 Node_Notes* source_notes = node_notes_at(source->_idx);
1317 if (source_notes == NULL || source_notes->is_clear()) return false;
1318 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1319 if (dest_notes == NULL || dest_notes->is_clear()) {
1320 return set_node_notes_at(dest->_idx, source_notes);
1321 }
1323 Node_Notes merged_notes = (*source_notes);
1324 // The order of operations here ensures that dest notes will win...
1325 merged_notes.update_from(dest_notes);
1326 return set_node_notes_at(dest->_idx, &merged_notes);
1327 }
1330 //--------------------------allow_range_check_smearing-------------------------
1331 // Gating condition for coalescing similar range checks.
1332 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1333 // single covering check that is at least as strong as any of them.
1334 // If the optimization succeeds, the simplified (strengthened) range check
1335 // will always succeed. If it fails, we will deopt, and then give up
1336 // on the optimization.
1337 bool Compile::allow_range_check_smearing() const {
1338 // If this method has already thrown a range-check,
1339 // assume it was because we already tried range smearing
1340 // and it failed.
1341 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1342 return !already_trapped;
1343 }
1346 //------------------------------flatten_alias_type-----------------------------
1347 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1348 int offset = tj->offset();
1349 TypePtr::PTR ptr = tj->ptr();
1351 // Known instance (scalarizable allocation) alias only with itself.
1352 bool is_known_inst = tj->isa_oopptr() != NULL &&
1353 tj->is_oopptr()->is_known_instance();
1355 // Process weird unsafe references.
1356 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1357 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1358 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1359 tj = TypeOopPtr::BOTTOM;
1360 ptr = tj->ptr();
1361 offset = tj->offset();
1362 }
1364 // Array pointers need some flattening
1365 const TypeAryPtr *ta = tj->isa_aryptr();
1366 if (ta && ta->is_stable()) {
1367 // Erase stability property for alias analysis.
1368 tj = ta = ta->cast_to_stable(false);
1369 }
1370 if( ta && is_known_inst ) {
1371 if ( offset != Type::OffsetBot &&
1372 offset > arrayOopDesc::length_offset_in_bytes() ) {
1373 offset = Type::OffsetBot; // Flatten constant access into array body only
1374 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1375 }
1376 } else if( ta && _AliasLevel >= 2 ) {
1377 // For arrays indexed by constant indices, we flatten the alias
1378 // space to include all of the array body. Only the header, klass
1379 // and array length can be accessed un-aliased.
1380 if( offset != Type::OffsetBot ) {
1381 if( ta->const_oop() ) { // MethodData* or Method*
1382 offset = Type::OffsetBot; // Flatten constant access into array body
1383 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1384 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1385 // range is OK as-is.
1386 tj = ta = TypeAryPtr::RANGE;
1387 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1388 tj = TypeInstPtr::KLASS; // all klass loads look alike
1389 ta = TypeAryPtr::RANGE; // generic ignored junk
1390 ptr = TypePtr::BotPTR;
1391 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1392 tj = TypeInstPtr::MARK;
1393 ta = TypeAryPtr::RANGE; // generic ignored junk
1394 ptr = TypePtr::BotPTR;
1395 } else { // Random constant offset into array body
1396 offset = Type::OffsetBot; // Flatten constant access into array body
1397 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1398 }
1399 }
1400 // Arrays of fixed size alias with arrays of unknown size.
1401 if (ta->size() != TypeInt::POS) {
1402 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1403 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1404 }
1405 // Arrays of known objects become arrays of unknown objects.
1406 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1407 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1408 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1409 }
1410 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1411 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1412 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1413 }
1414 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1415 // cannot be distinguished by bytecode alone.
1416 if (ta->elem() == TypeInt::BOOL) {
1417 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1418 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1419 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1420 }
1421 // During the 2nd round of IterGVN, NotNull castings are removed.
1422 // Make sure the Bottom and NotNull variants alias the same.
1423 // Also, make sure exact and non-exact variants alias the same.
1424 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1425 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1426 }
1427 }
1429 // Oop pointers need some flattening
1430 const TypeInstPtr *to = tj->isa_instptr();
1431 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1432 ciInstanceKlass *k = to->klass()->as_instance_klass();
1433 if( ptr == TypePtr::Constant ) {
1434 if (to->klass() != ciEnv::current()->Class_klass() ||
1435 offset < k->size_helper() * wordSize) {
1436 // No constant oop pointers (such as Strings); they alias with
1437 // unknown strings.
1438 assert(!is_known_inst, "not scalarizable allocation");
1439 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1440 }
1441 } else if( is_known_inst ) {
1442 tj = to; // Keep NotNull and klass_is_exact for instance type
1443 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1444 // During the 2nd round of IterGVN, NotNull castings are removed.
1445 // Make sure the Bottom and NotNull variants alias the same.
1446 // Also, make sure exact and non-exact variants alias the same.
1447 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1448 }
1449 if (to->speculative() != NULL) {
1450 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1451 }
1452 // Canonicalize the holder of this field
1453 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1454 // First handle header references such as a LoadKlassNode, even if the
1455 // object's klass is unloaded at compile time (4965979).
1456 if (!is_known_inst) { // Do it only for non-instance types
1457 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1458 }
1459 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1460 // Static fields are in the space above the normal instance
1461 // fields in the java.lang.Class instance.
1462 if (to->klass() != ciEnv::current()->Class_klass()) {
1463 to = NULL;
1464 tj = TypeOopPtr::BOTTOM;
1465 offset = tj->offset();
1466 }
1467 } else {
1468 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1469 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1470 if( is_known_inst ) {
1471 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1472 } else {
1473 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1474 }
1475 }
1476 }
1477 }
1479 // Klass pointers to object array klasses need some flattening
1480 const TypeKlassPtr *tk = tj->isa_klassptr();
1481 if( tk ) {
1482 // If we are referencing a field within a Klass, we need
1483 // to assume the worst case of an Object. Both exact and
1484 // inexact types must flatten to the same alias class so
1485 // use NotNull as the PTR.
1486 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1488 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1489 TypeKlassPtr::OBJECT->klass(),
1490 offset);
1491 }
1493 ciKlass* klass = tk->klass();
1494 if( klass->is_obj_array_klass() ) {
1495 ciKlass* k = TypeAryPtr::OOPS->klass();
1496 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1497 k = TypeInstPtr::BOTTOM->klass();
1498 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1499 }
1501 // Check for precise loads from the primary supertype array and force them
1502 // to the supertype cache alias index. Check for generic array loads from
1503 // the primary supertype array and also force them to the supertype cache
1504 // alias index. Since the same load can reach both, we need to merge
1505 // these 2 disparate memories into the same alias class. Since the
1506 // primary supertype array is read-only, there's no chance of confusion
1507 // where we bypass an array load and an array store.
1508 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1509 if (offset == Type::OffsetBot ||
1510 (offset >= primary_supers_offset &&
1511 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1512 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1513 offset = in_bytes(Klass::secondary_super_cache_offset());
1514 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1515 }
1516 }
1518 // Flatten all Raw pointers together.
1519 if (tj->base() == Type::RawPtr)
1520 tj = TypeRawPtr::BOTTOM;
1522 if (tj->base() == Type::AnyPtr)
1523 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1525 // Flatten all to bottom for now
1526 switch( _AliasLevel ) {
1527 case 0:
1528 tj = TypePtr::BOTTOM;
1529 break;
1530 case 1: // Flatten to: oop, static, field or array
1531 switch (tj->base()) {
1532 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1533 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1534 case Type::AryPtr: // do not distinguish arrays at all
1535 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1536 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1537 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1538 default: ShouldNotReachHere();
1539 }
1540 break;
1541 case 2: // No collapsing at level 2; keep all splits
1542 case 3: // No collapsing at level 3; keep all splits
1543 break;
1544 default:
1545 Unimplemented();
1546 }
1548 offset = tj->offset();
1549 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1551 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1552 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1553 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1554 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1555 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1556 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1557 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1558 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1559 assert( tj->ptr() != TypePtr::TopPTR &&
1560 tj->ptr() != TypePtr::AnyNull &&
1561 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1562 // assert( tj->ptr() != TypePtr::Constant ||
1563 // tj->base() == Type::RawPtr ||
1564 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1566 return tj;
1567 }
1569 void Compile::AliasType::Init(int i, const TypePtr* at) {
1570 _index = i;
1571 _adr_type = at;
1572 _field = NULL;
1573 _element = NULL;
1574 _is_rewritable = true; // default
1575 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1576 if (atoop != NULL && atoop->is_known_instance()) {
1577 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1578 _general_index = Compile::current()->get_alias_index(gt);
1579 } else {
1580 _general_index = 0;
1581 }
1582 }
1584 //---------------------------------print_on------------------------------------
1585 #ifndef PRODUCT
1586 void Compile::AliasType::print_on(outputStream* st) {
1587 if (index() < 10)
1588 st->print("@ <%d> ", index());
1589 else st->print("@ <%d>", index());
1590 st->print(is_rewritable() ? " " : " RO");
1591 int offset = adr_type()->offset();
1592 if (offset == Type::OffsetBot)
1593 st->print(" +any");
1594 else st->print(" +%-3d", offset);
1595 st->print(" in ");
1596 adr_type()->dump_on(st);
1597 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1598 if (field() != NULL && tjp) {
1599 if (tjp->klass() != field()->holder() ||
1600 tjp->offset() != field()->offset_in_bytes()) {
1601 st->print(" != ");
1602 field()->print();
1603 st->print(" ***");
1604 }
1605 }
1606 }
1608 void print_alias_types() {
1609 Compile* C = Compile::current();
1610 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1611 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1612 C->alias_type(idx)->print_on(tty);
1613 tty->cr();
1614 }
1615 }
1616 #endif
1619 //----------------------------probe_alias_cache--------------------------------
1620 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1621 intptr_t key = (intptr_t) adr_type;
1622 key ^= key >> logAliasCacheSize;
1623 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1624 }
1627 //-----------------------------grow_alias_types--------------------------------
1628 void Compile::grow_alias_types() {
1629 const int old_ats = _max_alias_types; // how many before?
1630 const int new_ats = old_ats; // how many more?
1631 const int grow_ats = old_ats+new_ats; // how many now?
1632 _max_alias_types = grow_ats;
1633 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1634 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1635 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1636 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1637 }
1640 //--------------------------------find_alias_type------------------------------
1641 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1642 if (_AliasLevel == 0)
1643 return alias_type(AliasIdxBot);
1645 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1646 if (ace->_adr_type == adr_type) {
1647 return alias_type(ace->_index);
1648 }
1650 // Handle special cases.
1651 if (adr_type == NULL) return alias_type(AliasIdxTop);
1652 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1654 // Do it the slow way.
1655 const TypePtr* flat = flatten_alias_type(adr_type);
1657 #ifdef ASSERT
1658 assert(flat == flatten_alias_type(flat), "idempotent");
1659 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1660 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1661 const TypeOopPtr* foop = flat->is_oopptr();
1662 // Scalarizable allocations have exact klass always.
1663 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1664 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1665 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1666 }
1667 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1668 #endif
1670 int idx = AliasIdxTop;
1671 for (int i = 0; i < num_alias_types(); i++) {
1672 if (alias_type(i)->adr_type() == flat) {
1673 idx = i;
1674 break;
1675 }
1676 }
1678 if (idx == AliasIdxTop) {
1679 if (no_create) return NULL;
1680 // Grow the array if necessary.
1681 if (_num_alias_types == _max_alias_types) grow_alias_types();
1682 // Add a new alias type.
1683 idx = _num_alias_types++;
1684 _alias_types[idx]->Init(idx, flat);
1685 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1686 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1687 if (flat->isa_instptr()) {
1688 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1689 && flat->is_instptr()->klass() == env()->Class_klass())
1690 alias_type(idx)->set_rewritable(false);
1691 }
1692 if (flat->isa_aryptr()) {
1693 #ifdef ASSERT
1694 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1695 // (T_BYTE has the weakest alignment and size restrictions...)
1696 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1697 #endif
1698 if (flat->offset() == TypePtr::OffsetBot) {
1699 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1700 }
1701 }
1702 if (flat->isa_klassptr()) {
1703 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1704 alias_type(idx)->set_rewritable(false);
1705 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1706 alias_type(idx)->set_rewritable(false);
1707 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1708 alias_type(idx)->set_rewritable(false);
1709 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1710 alias_type(idx)->set_rewritable(false);
1711 }
1712 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1713 // but the base pointer type is not distinctive enough to identify
1714 // references into JavaThread.)
1716 // Check for final fields.
1717 const TypeInstPtr* tinst = flat->isa_instptr();
1718 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1719 ciField* field;
1720 if (tinst->const_oop() != NULL &&
1721 tinst->klass() == ciEnv::current()->Class_klass() &&
1722 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1723 // static field
1724 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1725 field = k->get_field_by_offset(tinst->offset(), true);
1726 } else {
1727 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1728 field = k->get_field_by_offset(tinst->offset(), false);
1729 }
1730 assert(field == NULL ||
1731 original_field == NULL ||
1732 (field->holder() == original_field->holder() &&
1733 field->offset() == original_field->offset() &&
1734 field->is_static() == original_field->is_static()), "wrong field?");
1735 // Set field() and is_rewritable() attributes.
1736 if (field != NULL) alias_type(idx)->set_field(field);
1737 }
1738 }
1740 // Fill the cache for next time.
1741 ace->_adr_type = adr_type;
1742 ace->_index = idx;
1743 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1745 // Might as well try to fill the cache for the flattened version, too.
1746 AliasCacheEntry* face = probe_alias_cache(flat);
1747 if (face->_adr_type == NULL) {
1748 face->_adr_type = flat;
1749 face->_index = idx;
1750 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1751 }
1753 return alias_type(idx);
1754 }
1757 Compile::AliasType* Compile::alias_type(ciField* field) {
1758 const TypeOopPtr* t;
1759 if (field->is_static())
1760 t = TypeInstPtr::make(field->holder()->java_mirror());
1761 else
1762 t = TypeOopPtr::make_from_klass_raw(field->holder());
1763 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1764 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1765 return atp;
1766 }
1769 //------------------------------have_alias_type--------------------------------
1770 bool Compile::have_alias_type(const TypePtr* adr_type) {
1771 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1772 if (ace->_adr_type == adr_type) {
1773 return true;
1774 }
1776 // Handle special cases.
1777 if (adr_type == NULL) return true;
1778 if (adr_type == TypePtr::BOTTOM) return true;
1780 return find_alias_type(adr_type, true, NULL) != NULL;
1781 }
1783 //-----------------------------must_alias--------------------------------------
1784 // True if all values of the given address type are in the given alias category.
1785 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1786 if (alias_idx == AliasIdxBot) return true; // the universal category
1787 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1788 if (alias_idx == AliasIdxTop) return false; // the empty category
1789 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1791 // the only remaining possible overlap is identity
1792 int adr_idx = get_alias_index(adr_type);
1793 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1794 assert(adr_idx == alias_idx ||
1795 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1796 && adr_type != TypeOopPtr::BOTTOM),
1797 "should not be testing for overlap with an unsafe pointer");
1798 return adr_idx == alias_idx;
1799 }
1801 //------------------------------can_alias--------------------------------------
1802 // True if any values of the given address type are in the given alias category.
1803 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1804 if (alias_idx == AliasIdxTop) return false; // the empty category
1805 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1806 if (alias_idx == AliasIdxBot) return true; // the universal category
1807 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1809 // the only remaining possible overlap is identity
1810 int adr_idx = get_alias_index(adr_type);
1811 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1812 return adr_idx == alias_idx;
1813 }
1817 //---------------------------pop_warm_call-------------------------------------
1818 WarmCallInfo* Compile::pop_warm_call() {
1819 WarmCallInfo* wci = _warm_calls;
1820 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1821 return wci;
1822 }
1824 //----------------------------Inline_Warm--------------------------------------
1825 int Compile::Inline_Warm() {
1826 // If there is room, try to inline some more warm call sites.
1827 // %%% Do a graph index compaction pass when we think we're out of space?
1828 if (!InlineWarmCalls) return 0;
1830 int calls_made_hot = 0;
1831 int room_to_grow = NodeCountInliningCutoff - unique();
1832 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1833 int amount_grown = 0;
1834 WarmCallInfo* call;
1835 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1836 int est_size = (int)call->size();
1837 if (est_size > (room_to_grow - amount_grown)) {
1838 // This one won't fit anyway. Get rid of it.
1839 call->make_cold();
1840 continue;
1841 }
1842 call->make_hot();
1843 calls_made_hot++;
1844 amount_grown += est_size;
1845 amount_to_grow -= est_size;
1846 }
1848 if (calls_made_hot > 0) set_major_progress();
1849 return calls_made_hot;
1850 }
1853 //----------------------------Finish_Warm--------------------------------------
1854 void Compile::Finish_Warm() {
1855 if (!InlineWarmCalls) return;
1856 if (failing()) return;
1857 if (warm_calls() == NULL) return;
1859 // Clean up loose ends, if we are out of space for inlining.
1860 WarmCallInfo* call;
1861 while ((call = pop_warm_call()) != NULL) {
1862 call->make_cold();
1863 }
1864 }
1866 //---------------------cleanup_loop_predicates-----------------------
1867 // Remove the opaque nodes that protect the predicates so that all unused
1868 // checks and uncommon_traps will be eliminated from the ideal graph
1869 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1870 if (predicate_count()==0) return;
1871 for (int i = predicate_count(); i > 0; i--) {
1872 Node * n = predicate_opaque1_node(i-1);
1873 assert(n->Opcode() == Op_Opaque1, "must be");
1874 igvn.replace_node(n, n->in(1));
1875 }
1876 assert(predicate_count()==0, "should be clean!");
1877 }
1879 // StringOpts and late inlining of string methods
1880 void Compile::inline_string_calls(bool parse_time) {
1881 {
1882 // remove useless nodes to make the usage analysis simpler
1883 ResourceMark rm;
1884 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1885 }
1887 {
1888 ResourceMark rm;
1889 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1890 PhaseStringOpts pso(initial_gvn(), for_igvn());
1891 print_method(PHASE_AFTER_STRINGOPTS, 3);
1892 }
1894 // now inline anything that we skipped the first time around
1895 if (!parse_time) {
1896 _late_inlines_pos = _late_inlines.length();
1897 }
1899 while (_string_late_inlines.length() > 0) {
1900 CallGenerator* cg = _string_late_inlines.pop();
1901 cg->do_late_inline();
1902 if (failing()) return;
1903 }
1904 _string_late_inlines.trunc_to(0);
1905 }
1907 // Late inlining of boxing methods
1908 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1909 if (_boxing_late_inlines.length() > 0) {
1910 assert(has_boxed_value(), "inconsistent");
1912 PhaseGVN* gvn = initial_gvn();
1913 set_inlining_incrementally(true);
1915 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1916 for_igvn()->clear();
1917 gvn->replace_with(&igvn);
1919 _late_inlines_pos = _late_inlines.length();
1921 while (_boxing_late_inlines.length() > 0) {
1922 CallGenerator* cg = _boxing_late_inlines.pop();
1923 cg->do_late_inline();
1924 if (failing()) return;
1925 }
1926 _boxing_late_inlines.trunc_to(0);
1928 {
1929 ResourceMark rm;
1930 PhaseRemoveUseless pru(gvn, for_igvn());
1931 }
1933 igvn = PhaseIterGVN(gvn);
1934 igvn.optimize();
1936 set_inlining_progress(false);
1937 set_inlining_incrementally(false);
1938 }
1939 }
1941 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1942 assert(IncrementalInline, "incremental inlining should be on");
1943 PhaseGVN* gvn = initial_gvn();
1945 set_inlining_progress(false);
1946 for_igvn()->clear();
1947 gvn->replace_with(&igvn);
1949 int i = 0;
1951 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1952 CallGenerator* cg = _late_inlines.at(i);
1953 _late_inlines_pos = i+1;
1954 cg->do_late_inline();
1955 if (failing()) return;
1956 }
1957 int j = 0;
1958 for (; i < _late_inlines.length(); i++, j++) {
1959 _late_inlines.at_put(j, _late_inlines.at(i));
1960 }
1961 _late_inlines.trunc_to(j);
1963 {
1964 ResourceMark rm;
1965 PhaseRemoveUseless pru(gvn, for_igvn());
1966 }
1968 igvn = PhaseIterGVN(gvn);
1969 }
1971 // Perform incremental inlining until bound on number of live nodes is reached
1972 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1973 PhaseGVN* gvn = initial_gvn();
1975 set_inlining_incrementally(true);
1976 set_inlining_progress(true);
1977 uint low_live_nodes = 0;
1979 while(inlining_progress() && _late_inlines.length() > 0) {
1981 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1982 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1983 // PhaseIdealLoop is expensive so we only try it once we are
1984 // out of live nodes and we only try it again if the previous
1985 // helped got the number of nodes down significantly
1986 PhaseIdealLoop ideal_loop( igvn, false, true );
1987 if (failing()) return;
1988 low_live_nodes = live_nodes();
1989 _major_progress = true;
1990 }
1992 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1993 break;
1994 }
1995 }
1997 inline_incrementally_one(igvn);
1999 if (failing()) return;
2001 igvn.optimize();
2003 if (failing()) return;
2004 }
2006 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2008 if (_string_late_inlines.length() > 0) {
2009 assert(has_stringbuilder(), "inconsistent");
2010 for_igvn()->clear();
2011 initial_gvn()->replace_with(&igvn);
2013 inline_string_calls(false);
2015 if (failing()) return;
2017 {
2018 ResourceMark rm;
2019 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2020 }
2022 igvn = PhaseIterGVN(gvn);
2024 igvn.optimize();
2025 }
2027 set_inlining_incrementally(false);
2028 }
2031 //------------------------------Optimize---------------------------------------
2032 // Given a graph, optimize it.
2033 void Compile::Optimize() {
2034 TracePhase t1("optimizer", &_t_optimizer, true);
2036 #ifndef PRODUCT
2037 if (env()->break_at_compile()) {
2038 BREAKPOINT;
2039 }
2041 #endif
2043 ResourceMark rm;
2044 int loop_opts_cnt;
2046 NOT_PRODUCT( verify_graph_edges(); )
2048 print_method(PHASE_AFTER_PARSING);
2050 {
2051 // Iterative Global Value Numbering, including ideal transforms
2052 // Initialize IterGVN with types and values from parse-time GVN
2053 PhaseIterGVN igvn(initial_gvn());
2054 {
2055 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2056 igvn.optimize();
2057 }
2059 print_method(PHASE_ITER_GVN1, 2);
2061 if (failing()) return;
2063 {
2064 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2065 inline_incrementally(igvn);
2066 }
2068 print_method(PHASE_INCREMENTAL_INLINE, 2);
2070 if (failing()) return;
2072 if (eliminate_boxing()) {
2073 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2074 // Inline valueOf() methods now.
2075 inline_boxing_calls(igvn);
2077 if (AlwaysIncrementalInline) {
2078 inline_incrementally(igvn);
2079 }
2081 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2083 if (failing()) return;
2084 }
2086 // Remove the speculative part of types and clean up the graph from
2087 // the extra CastPP nodes whose only purpose is to carry them. Do
2088 // that early so that optimizations are not disrupted by the extra
2089 // CastPP nodes.
2090 remove_speculative_types(igvn);
2092 // No more new expensive nodes will be added to the list from here
2093 // so keep only the actual candidates for optimizations.
2094 cleanup_expensive_nodes(igvn);
2096 // Perform escape analysis
2097 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2098 if (has_loops()) {
2099 // Cleanup graph (remove dead nodes).
2100 TracePhase t2("idealLoop", &_t_idealLoop, true);
2101 PhaseIdealLoop ideal_loop( igvn, false, true );
2102 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2103 if (failing()) return;
2104 }
2105 ConnectionGraph::do_analysis(this, &igvn);
2107 if (failing()) return;
2109 // Optimize out fields loads from scalar replaceable allocations.
2110 igvn.optimize();
2111 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2113 if (failing()) return;
2115 if (congraph() != NULL && macro_count() > 0) {
2116 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2117 PhaseMacroExpand mexp(igvn);
2118 mexp.eliminate_macro_nodes();
2119 igvn.set_delay_transform(false);
2121 igvn.optimize();
2122 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2124 if (failing()) return;
2125 }
2126 }
2128 // Loop transforms on the ideal graph. Range Check Elimination,
2129 // peeling, unrolling, etc.
2131 // Set loop opts counter
2132 loop_opts_cnt = num_loop_opts();
2133 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2134 {
2135 TracePhase t2("idealLoop", &_t_idealLoop, true);
2136 PhaseIdealLoop ideal_loop( igvn, true );
2137 loop_opts_cnt--;
2138 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2139 if (failing()) return;
2140 }
2141 // Loop opts pass if partial peeling occurred in previous pass
2142 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2143 TracePhase t3("idealLoop", &_t_idealLoop, true);
2144 PhaseIdealLoop ideal_loop( igvn, false );
2145 loop_opts_cnt--;
2146 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2147 if (failing()) return;
2148 }
2149 // Loop opts pass for loop-unrolling before CCP
2150 if(major_progress() && (loop_opts_cnt > 0)) {
2151 TracePhase t4("idealLoop", &_t_idealLoop, true);
2152 PhaseIdealLoop ideal_loop( igvn, false );
2153 loop_opts_cnt--;
2154 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2155 }
2156 if (!failing()) {
2157 // Verify that last round of loop opts produced a valid graph
2158 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2159 PhaseIdealLoop::verify(igvn);
2160 }
2161 }
2162 if (failing()) return;
2164 // Conditional Constant Propagation;
2165 PhaseCCP ccp( &igvn );
2166 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2167 {
2168 TracePhase t2("ccp", &_t_ccp, true);
2169 ccp.do_transform();
2170 }
2171 print_method(PHASE_CPP1, 2);
2173 assert( true, "Break here to ccp.dump_old2new_map()");
2175 // Iterative Global Value Numbering, including ideal transforms
2176 {
2177 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2178 igvn = ccp;
2179 igvn.optimize();
2180 }
2182 print_method(PHASE_ITER_GVN2, 2);
2184 if (failing()) return;
2186 // Loop transforms on the ideal graph. Range Check Elimination,
2187 // peeling, unrolling, etc.
2188 if(loop_opts_cnt > 0) {
2189 debug_only( int cnt = 0; );
2190 while(major_progress() && (loop_opts_cnt > 0)) {
2191 TracePhase t2("idealLoop", &_t_idealLoop, true);
2192 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2193 PhaseIdealLoop ideal_loop( igvn, true);
2194 loop_opts_cnt--;
2195 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2196 if (failing()) return;
2197 }
2198 }
2200 {
2201 // Verify that all previous optimizations produced a valid graph
2202 // at least to this point, even if no loop optimizations were done.
2203 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2204 PhaseIdealLoop::verify(igvn);
2205 }
2207 {
2208 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2209 PhaseMacroExpand mex(igvn);
2210 if (mex.expand_macro_nodes()) {
2211 assert(failing(), "must bail out w/ explicit message");
2212 return;
2213 }
2214 }
2216 } // (End scope of igvn; run destructor if necessary for asserts.)
2218 dump_inlining();
2219 // A method with only infinite loops has no edges entering loops from root
2220 {
2221 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2222 if (final_graph_reshaping()) {
2223 assert(failing(), "must bail out w/ explicit message");
2224 return;
2225 }
2226 }
2228 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2229 }
2232 //------------------------------Code_Gen---------------------------------------
2233 // Given a graph, generate code for it
2234 void Compile::Code_Gen() {
2235 if (failing()) {
2236 return;
2237 }
2239 // Perform instruction selection. You might think we could reclaim Matcher
2240 // memory PDQ, but actually the Matcher is used in generating spill code.
2241 // Internals of the Matcher (including some VectorSets) must remain live
2242 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2243 // set a bit in reclaimed memory.
2245 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2246 // nodes. Mapping is only valid at the root of each matched subtree.
2247 NOT_PRODUCT( verify_graph_edges(); )
2249 Matcher matcher;
2250 _matcher = &matcher;
2251 {
2252 TracePhase t2("matcher", &_t_matcher, true);
2253 matcher.match();
2254 }
2255 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2256 // nodes. Mapping is only valid at the root of each matched subtree.
2257 NOT_PRODUCT( verify_graph_edges(); )
2259 // If you have too many nodes, or if matching has failed, bail out
2260 check_node_count(0, "out of nodes matching instructions");
2261 if (failing()) {
2262 return;
2263 }
2265 // Build a proper-looking CFG
2266 PhaseCFG cfg(node_arena(), root(), matcher);
2267 _cfg = &cfg;
2268 {
2269 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2270 bool success = cfg.do_global_code_motion();
2271 if (!success) {
2272 return;
2273 }
2275 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2276 NOT_PRODUCT( verify_graph_edges(); )
2277 debug_only( cfg.verify(); )
2278 }
2280 PhaseChaitin regalloc(unique(), cfg, matcher);
2281 _regalloc = ®alloc;
2282 {
2283 TracePhase t2("regalloc", &_t_registerAllocation, true);
2284 // Perform register allocation. After Chaitin, use-def chains are
2285 // no longer accurate (at spill code) and so must be ignored.
2286 // Node->LRG->reg mappings are still accurate.
2287 _regalloc->Register_Allocate();
2289 // Bail out if the allocator builds too many nodes
2290 if (failing()) {
2291 return;
2292 }
2293 }
2295 // Prior to register allocation we kept empty basic blocks in case the
2296 // the allocator needed a place to spill. After register allocation we
2297 // are not adding any new instructions. If any basic block is empty, we
2298 // can now safely remove it.
2299 {
2300 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2301 cfg.remove_empty_blocks();
2302 if (do_freq_based_layout()) {
2303 PhaseBlockLayout layout(cfg);
2304 } else {
2305 cfg.set_loop_alignment();
2306 }
2307 cfg.fixup_flow();
2308 }
2310 // Apply peephole optimizations
2311 if( OptoPeephole ) {
2312 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2313 PhasePeephole peep( _regalloc, cfg);
2314 peep.do_transform();
2315 }
2317 // Do late expand if CPU requires this.
2318 if (Matcher::require_postalloc_expand) {
2319 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2320 cfg.postalloc_expand(_regalloc);
2321 }
2323 // Convert Nodes to instruction bits in a buffer
2324 {
2325 // %%%% workspace merge brought two timers together for one job
2326 TracePhase t2a("output", &_t_output, true);
2327 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2328 Output();
2329 }
2331 print_method(PHASE_FINAL_CODE);
2333 // He's dead, Jim.
2334 _cfg = (PhaseCFG*)0xdeadbeef;
2335 _regalloc = (PhaseChaitin*)0xdeadbeef;
2336 }
2339 //------------------------------dump_asm---------------------------------------
2340 // Dump formatted assembly
2341 #ifndef PRODUCT
2342 void Compile::dump_asm(int *pcs, uint pc_limit) {
2343 bool cut_short = false;
2344 tty->print_cr("#");
2345 tty->print("# "); _tf->dump(); tty->cr();
2346 tty->print_cr("#");
2348 // For all blocks
2349 int pc = 0x0; // Program counter
2350 char starts_bundle = ' ';
2351 _regalloc->dump_frame();
2353 Node *n = NULL;
2354 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2355 if (VMThread::should_terminate()) {
2356 cut_short = true;
2357 break;
2358 }
2359 Block* block = _cfg->get_block(i);
2360 if (block->is_connector() && !Verbose) {
2361 continue;
2362 }
2363 n = block->head();
2364 if (pcs && n->_idx < pc_limit) {
2365 tty->print("%3.3x ", pcs[n->_idx]);
2366 } else {
2367 tty->print(" ");
2368 }
2369 block->dump_head(_cfg);
2370 if (block->is_connector()) {
2371 tty->print_cr(" # Empty connector block");
2372 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2373 tty->print_cr(" # Block is sole successor of call");
2374 }
2376 // For all instructions
2377 Node *delay = NULL;
2378 for (uint j = 0; j < block->number_of_nodes(); j++) {
2379 if (VMThread::should_terminate()) {
2380 cut_short = true;
2381 break;
2382 }
2383 n = block->get_node(j);
2384 if (valid_bundle_info(n)) {
2385 Bundle* bundle = node_bundling(n);
2386 if (bundle->used_in_unconditional_delay()) {
2387 delay = n;
2388 continue;
2389 }
2390 if (bundle->starts_bundle()) {
2391 starts_bundle = '+';
2392 }
2393 }
2395 if (WizardMode) {
2396 n->dump();
2397 }
2399 if( !n->is_Region() && // Dont print in the Assembly
2400 !n->is_Phi() && // a few noisely useless nodes
2401 !n->is_Proj() &&
2402 !n->is_MachTemp() &&
2403 !n->is_SafePointScalarObject() &&
2404 !n->is_Catch() && // Would be nice to print exception table targets
2405 !n->is_MergeMem() && // Not very interesting
2406 !n->is_top() && // Debug info table constants
2407 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2408 ) {
2409 if (pcs && n->_idx < pc_limit)
2410 tty->print("%3.3x", pcs[n->_idx]);
2411 else
2412 tty->print(" ");
2413 tty->print(" %c ", starts_bundle);
2414 starts_bundle = ' ';
2415 tty->print("\t");
2416 n->format(_regalloc, tty);
2417 tty->cr();
2418 }
2420 // If we have an instruction with a delay slot, and have seen a delay,
2421 // then back up and print it
2422 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2423 assert(delay != NULL, "no unconditional delay instruction");
2424 if (WizardMode) delay->dump();
2426 if (node_bundling(delay)->starts_bundle())
2427 starts_bundle = '+';
2428 if (pcs && n->_idx < pc_limit)
2429 tty->print("%3.3x", pcs[n->_idx]);
2430 else
2431 tty->print(" ");
2432 tty->print(" %c ", starts_bundle);
2433 starts_bundle = ' ';
2434 tty->print("\t");
2435 delay->format(_regalloc, tty);
2436 tty->cr();
2437 delay = NULL;
2438 }
2440 // Dump the exception table as well
2441 if( n->is_Catch() && (Verbose || WizardMode) ) {
2442 // Print the exception table for this offset
2443 _handler_table.print_subtable_for(pc);
2444 }
2445 }
2447 if (pcs && n->_idx < pc_limit)
2448 tty->print_cr("%3.3x", pcs[n->_idx]);
2449 else
2450 tty->cr();
2452 assert(cut_short || delay == NULL, "no unconditional delay branch");
2454 } // End of per-block dump
2455 tty->cr();
2457 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2458 }
2459 #endif
2461 //------------------------------Final_Reshape_Counts---------------------------
2462 // This class defines counters to help identify when a method
2463 // may/must be executed using hardware with only 24-bit precision.
2464 struct Final_Reshape_Counts : public StackObj {
2465 int _call_count; // count non-inlined 'common' calls
2466 int _float_count; // count float ops requiring 24-bit precision
2467 int _double_count; // count double ops requiring more precision
2468 int _java_call_count; // count non-inlined 'java' calls
2469 int _inner_loop_count; // count loops which need alignment
2470 VectorSet _visited; // Visitation flags
2471 Node_List _tests; // Set of IfNodes & PCTableNodes
2473 Final_Reshape_Counts() :
2474 _call_count(0), _float_count(0), _double_count(0),
2475 _java_call_count(0), _inner_loop_count(0),
2476 _visited( Thread::current()->resource_area() ) { }
2478 void inc_call_count () { _call_count ++; }
2479 void inc_float_count () { _float_count ++; }
2480 void inc_double_count() { _double_count++; }
2481 void inc_java_call_count() { _java_call_count++; }
2482 void inc_inner_loop_count() { _inner_loop_count++; }
2484 int get_call_count () const { return _call_count ; }
2485 int get_float_count () const { return _float_count ; }
2486 int get_double_count() const { return _double_count; }
2487 int get_java_call_count() const { return _java_call_count; }
2488 int get_inner_loop_count() const { return _inner_loop_count; }
2489 };
2491 #ifdef ASSERT
2492 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2493 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2494 // Make sure the offset goes inside the instance layout.
2495 return k->contains_field_offset(tp->offset());
2496 // Note that OffsetBot and OffsetTop are very negative.
2497 }
2498 #endif
2500 // Eliminate trivially redundant StoreCMs and accumulate their
2501 // precedence edges.
2502 void Compile::eliminate_redundant_card_marks(Node* n) {
2503 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2504 if (n->in(MemNode::Address)->outcnt() > 1) {
2505 // There are multiple users of the same address so it might be
2506 // possible to eliminate some of the StoreCMs
2507 Node* mem = n->in(MemNode::Memory);
2508 Node* adr = n->in(MemNode::Address);
2509 Node* val = n->in(MemNode::ValueIn);
2510 Node* prev = n;
2511 bool done = false;
2512 // Walk the chain of StoreCMs eliminating ones that match. As
2513 // long as it's a chain of single users then the optimization is
2514 // safe. Eliminating partially redundant StoreCMs would require
2515 // cloning copies down the other paths.
2516 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2517 if (adr == mem->in(MemNode::Address) &&
2518 val == mem->in(MemNode::ValueIn)) {
2519 // redundant StoreCM
2520 if (mem->req() > MemNode::OopStore) {
2521 // Hasn't been processed by this code yet.
2522 n->add_prec(mem->in(MemNode::OopStore));
2523 } else {
2524 // Already converted to precedence edge
2525 for (uint i = mem->req(); i < mem->len(); i++) {
2526 // Accumulate any precedence edges
2527 if (mem->in(i) != NULL) {
2528 n->add_prec(mem->in(i));
2529 }
2530 }
2531 // Everything above this point has been processed.
2532 done = true;
2533 }
2534 // Eliminate the previous StoreCM
2535 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2536 assert(mem->outcnt() == 0, "should be dead");
2537 mem->disconnect_inputs(NULL, this);
2538 } else {
2539 prev = mem;
2540 }
2541 mem = prev->in(MemNode::Memory);
2542 }
2543 }
2544 }
2546 //------------------------------final_graph_reshaping_impl----------------------
2547 // Implement items 1-5 from final_graph_reshaping below.
2548 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2550 if ( n->outcnt() == 0 ) return; // dead node
2551 uint nop = n->Opcode();
2553 // Check for 2-input instruction with "last use" on right input.
2554 // Swap to left input. Implements item (2).
2555 if( n->req() == 3 && // two-input instruction
2556 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2557 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2558 n->in(2)->outcnt() == 1 &&// right use IS a last use
2559 !n->in(2)->is_Con() ) { // right use is not a constant
2560 // Check for commutative opcode
2561 switch( nop ) {
2562 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2563 case Op_MaxI: case Op_MinI:
2564 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2565 case Op_AndL: case Op_XorL: case Op_OrL:
2566 case Op_AndI: case Op_XorI: case Op_OrI: {
2567 // Move "last use" input to left by swapping inputs
2568 n->swap_edges(1, 2);
2569 break;
2570 }
2571 default:
2572 break;
2573 }
2574 }
2576 #ifdef ASSERT
2577 if( n->is_Mem() ) {
2578 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2579 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2580 // oop will be recorded in oop map if load crosses safepoint
2581 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2582 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2583 "raw memory operations should have control edge");
2584 }
2585 #endif
2586 // Count FPU ops and common calls, implements item (3)
2587 switch( nop ) {
2588 // Count all float operations that may use FPU
2589 case Op_AddF:
2590 case Op_SubF:
2591 case Op_MulF:
2592 case Op_DivF:
2593 case Op_NegF:
2594 case Op_ModF:
2595 case Op_ConvI2F:
2596 case Op_ConF:
2597 case Op_CmpF:
2598 case Op_CmpF3:
2599 // case Op_ConvL2F: // longs are split into 32-bit halves
2600 frc.inc_float_count();
2601 break;
2603 case Op_ConvF2D:
2604 case Op_ConvD2F:
2605 frc.inc_float_count();
2606 frc.inc_double_count();
2607 break;
2609 // Count all double operations that may use FPU
2610 case Op_AddD:
2611 case Op_SubD:
2612 case Op_MulD:
2613 case Op_DivD:
2614 case Op_NegD:
2615 case Op_ModD:
2616 case Op_ConvI2D:
2617 case Op_ConvD2I:
2618 // case Op_ConvL2D: // handled by leaf call
2619 // case Op_ConvD2L: // handled by leaf call
2620 case Op_ConD:
2621 case Op_CmpD:
2622 case Op_CmpD3:
2623 frc.inc_double_count();
2624 break;
2625 case Op_Opaque1: // Remove Opaque Nodes before matching
2626 case Op_Opaque2: // Remove Opaque Nodes before matching
2627 case Op_Opaque3:
2628 n->subsume_by(n->in(1), this);
2629 break;
2630 case Op_CallStaticJava:
2631 case Op_CallJava:
2632 case Op_CallDynamicJava:
2633 frc.inc_java_call_count(); // Count java call site;
2634 case Op_CallRuntime:
2635 case Op_CallLeaf:
2636 case Op_CallLeafNoFP: {
2637 assert( n->is_Call(), "" );
2638 CallNode *call = n->as_Call();
2639 // Count call sites where the FP mode bit would have to be flipped.
2640 // Do not count uncommon runtime calls:
2641 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2642 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2643 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2644 frc.inc_call_count(); // Count the call site
2645 } else { // See if uncommon argument is shared
2646 Node *n = call->in(TypeFunc::Parms);
2647 int nop = n->Opcode();
2648 // Clone shared simple arguments to uncommon calls, item (1).
2649 if( n->outcnt() > 1 &&
2650 !n->is_Proj() &&
2651 nop != Op_CreateEx &&
2652 nop != Op_CheckCastPP &&
2653 nop != Op_DecodeN &&
2654 nop != Op_DecodeNKlass &&
2655 !n->is_Mem() ) {
2656 Node *x = n->clone();
2657 call->set_req( TypeFunc::Parms, x );
2658 }
2659 }
2660 break;
2661 }
2663 case Op_StoreD:
2664 case Op_LoadD:
2665 case Op_LoadD_unaligned:
2666 frc.inc_double_count();
2667 goto handle_mem;
2668 case Op_StoreF:
2669 case Op_LoadF:
2670 frc.inc_float_count();
2671 goto handle_mem;
2673 case Op_StoreCM:
2674 {
2675 // Convert OopStore dependence into precedence edge
2676 Node* prec = n->in(MemNode::OopStore);
2677 n->del_req(MemNode::OopStore);
2678 n->add_prec(prec);
2679 eliminate_redundant_card_marks(n);
2680 }
2682 // fall through
2684 case Op_StoreB:
2685 case Op_StoreC:
2686 case Op_StorePConditional:
2687 case Op_StoreI:
2688 case Op_StoreL:
2689 case Op_StoreIConditional:
2690 case Op_StoreLConditional:
2691 case Op_CompareAndSwapI:
2692 case Op_CompareAndSwapL:
2693 case Op_CompareAndSwapP:
2694 case Op_CompareAndSwapN:
2695 case Op_GetAndAddI:
2696 case Op_GetAndAddL:
2697 case Op_GetAndSetI:
2698 case Op_GetAndSetL:
2699 case Op_GetAndSetP:
2700 case Op_GetAndSetN:
2701 case Op_StoreP:
2702 case Op_StoreN:
2703 case Op_StoreNKlass:
2704 case Op_LoadB:
2705 case Op_LoadUB:
2706 case Op_LoadUS:
2707 case Op_LoadI:
2708 case Op_LoadKlass:
2709 case Op_LoadNKlass:
2710 case Op_LoadL:
2711 case Op_LoadL_unaligned:
2712 case Op_LoadPLocked:
2713 case Op_LoadP:
2714 case Op_LoadN:
2715 case Op_LoadRange:
2716 case Op_LoadS: {
2717 handle_mem:
2718 #ifdef ASSERT
2719 if( VerifyOptoOopOffsets ) {
2720 assert( n->is_Mem(), "" );
2721 MemNode *mem = (MemNode*)n;
2722 // Check to see if address types have grounded out somehow.
2723 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2724 assert( !tp || oop_offset_is_sane(tp), "" );
2725 }
2726 #endif
2727 break;
2728 }
2730 case Op_AddP: { // Assert sane base pointers
2731 Node *addp = n->in(AddPNode::Address);
2732 assert( !addp->is_AddP() ||
2733 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2734 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2735 "Base pointers must match" );
2736 #ifdef _LP64
2737 if ((UseCompressedOops || UseCompressedClassPointers) &&
2738 addp->Opcode() == Op_ConP &&
2739 addp == n->in(AddPNode::Base) &&
2740 n->in(AddPNode::Offset)->is_Con()) {
2741 // Use addressing with narrow klass to load with offset on x86.
2742 // On sparc loading 32-bits constant and decoding it have less
2743 // instructions (4) then load 64-bits constant (7).
2744 // Do this transformation here since IGVN will convert ConN back to ConP.
2745 const Type* t = addp->bottom_type();
2746 if (t->isa_oopptr() || t->isa_klassptr()) {
2747 Node* nn = NULL;
2749 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2751 // Look for existing ConN node of the same exact type.
2752 Node* r = root();
2753 uint cnt = r->outcnt();
2754 for (uint i = 0; i < cnt; i++) {
2755 Node* m = r->raw_out(i);
2756 if (m!= NULL && m->Opcode() == op &&
2757 m->bottom_type()->make_ptr() == t) {
2758 nn = m;
2759 break;
2760 }
2761 }
2762 if (nn != NULL) {
2763 // Decode a narrow oop to match address
2764 // [R12 + narrow_oop_reg<<3 + offset]
2765 if (t->isa_oopptr()) {
2766 nn = new (this) DecodeNNode(nn, t);
2767 } else {
2768 nn = new (this) DecodeNKlassNode(nn, t);
2769 }
2770 n->set_req(AddPNode::Base, nn);
2771 n->set_req(AddPNode::Address, nn);
2772 if (addp->outcnt() == 0) {
2773 addp->disconnect_inputs(NULL, this);
2774 }
2775 }
2776 }
2777 }
2778 #endif
2779 break;
2780 }
2782 #ifdef _LP64
2783 case Op_CastPP:
2784 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2785 Node* in1 = n->in(1);
2786 const Type* t = n->bottom_type();
2787 Node* new_in1 = in1->clone();
2788 new_in1->as_DecodeN()->set_type(t);
2790 if (!Matcher::narrow_oop_use_complex_address()) {
2791 //
2792 // x86, ARM and friends can handle 2 adds in addressing mode
2793 // and Matcher can fold a DecodeN node into address by using
2794 // a narrow oop directly and do implicit NULL check in address:
2795 //
2796 // [R12 + narrow_oop_reg<<3 + offset]
2797 // NullCheck narrow_oop_reg
2798 //
2799 // On other platforms (Sparc) we have to keep new DecodeN node and
2800 // use it to do implicit NULL check in address:
2801 //
2802 // decode_not_null narrow_oop_reg, base_reg
2803 // [base_reg + offset]
2804 // NullCheck base_reg
2805 //
2806 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2807 // to keep the information to which NULL check the new DecodeN node
2808 // corresponds to use it as value in implicit_null_check().
2809 //
2810 new_in1->set_req(0, n->in(0));
2811 }
2813 n->subsume_by(new_in1, this);
2814 if (in1->outcnt() == 0) {
2815 in1->disconnect_inputs(NULL, this);
2816 }
2817 }
2818 break;
2820 case Op_CmpP:
2821 // Do this transformation here to preserve CmpPNode::sub() and
2822 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2823 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2824 Node* in1 = n->in(1);
2825 Node* in2 = n->in(2);
2826 if (!in1->is_DecodeNarrowPtr()) {
2827 in2 = in1;
2828 in1 = n->in(2);
2829 }
2830 assert(in1->is_DecodeNarrowPtr(), "sanity");
2832 Node* new_in2 = NULL;
2833 if (in2->is_DecodeNarrowPtr()) {
2834 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2835 new_in2 = in2->in(1);
2836 } else if (in2->Opcode() == Op_ConP) {
2837 const Type* t = in2->bottom_type();
2838 if (t == TypePtr::NULL_PTR) {
2839 assert(in1->is_DecodeN(), "compare klass to null?");
2840 // Don't convert CmpP null check into CmpN if compressed
2841 // oops implicit null check is not generated.
2842 // This will allow to generate normal oop implicit null check.
2843 if (Matcher::gen_narrow_oop_implicit_null_checks())
2844 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2845 //
2846 // This transformation together with CastPP transformation above
2847 // will generated code for implicit NULL checks for compressed oops.
2848 //
2849 // The original code after Optimize()
2850 //
2851 // LoadN memory, narrow_oop_reg
2852 // decode narrow_oop_reg, base_reg
2853 // CmpP base_reg, NULL
2854 // CastPP base_reg // NotNull
2855 // Load [base_reg + offset], val_reg
2856 //
2857 // after these transformations will be
2858 //
2859 // LoadN memory, narrow_oop_reg
2860 // CmpN narrow_oop_reg, NULL
2861 // decode_not_null narrow_oop_reg, base_reg
2862 // Load [base_reg + offset], val_reg
2863 //
2864 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2865 // since narrow oops can be used in debug info now (see the code in
2866 // final_graph_reshaping_walk()).
2867 //
2868 // At the end the code will be matched to
2869 // on x86:
2870 //
2871 // Load_narrow_oop memory, narrow_oop_reg
2872 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2873 // NullCheck narrow_oop_reg
2874 //
2875 // and on sparc:
2876 //
2877 // Load_narrow_oop memory, narrow_oop_reg
2878 // decode_not_null narrow_oop_reg, base_reg
2879 // Load [base_reg + offset], val_reg
2880 // NullCheck base_reg
2881 //
2882 } else if (t->isa_oopptr()) {
2883 new_in2 = ConNode::make(this, t->make_narrowoop());
2884 } else if (t->isa_klassptr()) {
2885 new_in2 = ConNode::make(this, t->make_narrowklass());
2886 }
2887 }
2888 if (new_in2 != NULL) {
2889 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2890 n->subsume_by(cmpN, this);
2891 if (in1->outcnt() == 0) {
2892 in1->disconnect_inputs(NULL, this);
2893 }
2894 if (in2->outcnt() == 0) {
2895 in2->disconnect_inputs(NULL, this);
2896 }
2897 }
2898 }
2899 break;
2901 case Op_DecodeN:
2902 case Op_DecodeNKlass:
2903 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2904 // DecodeN could be pinned when it can't be fold into
2905 // an address expression, see the code for Op_CastPP above.
2906 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2907 break;
2909 case Op_EncodeP:
2910 case Op_EncodePKlass: {
2911 Node* in1 = n->in(1);
2912 if (in1->is_DecodeNarrowPtr()) {
2913 n->subsume_by(in1->in(1), this);
2914 } else if (in1->Opcode() == Op_ConP) {
2915 const Type* t = in1->bottom_type();
2916 if (t == TypePtr::NULL_PTR) {
2917 assert(t->isa_oopptr(), "null klass?");
2918 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2919 } else if (t->isa_oopptr()) {
2920 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2921 } else if (t->isa_klassptr()) {
2922 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2923 }
2924 }
2925 if (in1->outcnt() == 0) {
2926 in1->disconnect_inputs(NULL, this);
2927 }
2928 break;
2929 }
2931 case Op_Proj: {
2932 if (OptimizeStringConcat) {
2933 ProjNode* p = n->as_Proj();
2934 if (p->_is_io_use) {
2935 // Separate projections were used for the exception path which
2936 // are normally removed by a late inline. If it wasn't inlined
2937 // then they will hang around and should just be replaced with
2938 // the original one.
2939 Node* proj = NULL;
2940 // Replace with just one
2941 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2942 Node *use = i.get();
2943 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2944 proj = use;
2945 break;
2946 }
2947 }
2948 assert(proj != NULL, "must be found");
2949 p->subsume_by(proj, this);
2950 }
2951 }
2952 break;
2953 }
2955 case Op_Phi:
2956 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
2957 // The EncodeP optimization may create Phi with the same edges
2958 // for all paths. It is not handled well by Register Allocator.
2959 Node* unique_in = n->in(1);
2960 assert(unique_in != NULL, "");
2961 uint cnt = n->req();
2962 for (uint i = 2; i < cnt; i++) {
2963 Node* m = n->in(i);
2964 assert(m != NULL, "");
2965 if (unique_in != m)
2966 unique_in = NULL;
2967 }
2968 if (unique_in != NULL) {
2969 n->subsume_by(unique_in, this);
2970 }
2971 }
2972 break;
2974 #endif
2976 case Op_ModI:
2977 if (UseDivMod) {
2978 // Check if a%b and a/b both exist
2979 Node* d = n->find_similar(Op_DivI);
2980 if (d) {
2981 // Replace them with a fused divmod if supported
2982 if (Matcher::has_match_rule(Op_DivModI)) {
2983 DivModINode* divmod = DivModINode::make(this, n);
2984 d->subsume_by(divmod->div_proj(), this);
2985 n->subsume_by(divmod->mod_proj(), this);
2986 } else {
2987 // replace a%b with a-((a/b)*b)
2988 Node* mult = new (this) MulINode(d, d->in(2));
2989 Node* sub = new (this) SubINode(d->in(1), mult);
2990 n->subsume_by(sub, this);
2991 }
2992 }
2993 }
2994 break;
2996 case Op_ModL:
2997 if (UseDivMod) {
2998 // Check if a%b and a/b both exist
2999 Node* d = n->find_similar(Op_DivL);
3000 if (d) {
3001 // Replace them with a fused divmod if supported
3002 if (Matcher::has_match_rule(Op_DivModL)) {
3003 DivModLNode* divmod = DivModLNode::make(this, n);
3004 d->subsume_by(divmod->div_proj(), this);
3005 n->subsume_by(divmod->mod_proj(), this);
3006 } else {
3007 // replace a%b with a-((a/b)*b)
3008 Node* mult = new (this) MulLNode(d, d->in(2));
3009 Node* sub = new (this) SubLNode(d->in(1), mult);
3010 n->subsume_by(sub, this);
3011 }
3012 }
3013 }
3014 break;
3016 case Op_LoadVector:
3017 case Op_StoreVector:
3018 break;
3020 case Op_PackB:
3021 case Op_PackS:
3022 case Op_PackI:
3023 case Op_PackF:
3024 case Op_PackL:
3025 case Op_PackD:
3026 if (n->req()-1 > 2) {
3027 // Replace many operand PackNodes with a binary tree for matching
3028 PackNode* p = (PackNode*) n;
3029 Node* btp = p->binary_tree_pack(this, 1, n->req());
3030 n->subsume_by(btp, this);
3031 }
3032 break;
3033 case Op_Loop:
3034 case Op_CountedLoop:
3035 if (n->as_Loop()->is_inner_loop()) {
3036 frc.inc_inner_loop_count();
3037 }
3038 break;
3039 case Op_LShiftI:
3040 case Op_RShiftI:
3041 case Op_URShiftI:
3042 case Op_LShiftL:
3043 case Op_RShiftL:
3044 case Op_URShiftL:
3045 if (Matcher::need_masked_shift_count) {
3046 // The cpu's shift instructions don't restrict the count to the
3047 // lower 5/6 bits. We need to do the masking ourselves.
3048 Node* in2 = n->in(2);
3049 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3050 const TypeInt* t = in2->find_int_type();
3051 if (t != NULL && t->is_con()) {
3052 juint shift = t->get_con();
3053 if (shift > mask) { // Unsigned cmp
3054 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
3055 }
3056 } else {
3057 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3058 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3059 n->set_req(2, shift);
3060 }
3061 }
3062 if (in2->outcnt() == 0) { // Remove dead node
3063 in2->disconnect_inputs(NULL, this);
3064 }
3065 }
3066 break;
3067 case Op_MemBarStoreStore:
3068 case Op_MemBarRelease:
3069 // Break the link with AllocateNode: it is no longer useful and
3070 // confuses register allocation.
3071 if (n->req() > MemBarNode::Precedent) {
3072 n->set_req(MemBarNode::Precedent, top());
3073 }
3074 break;
3075 default:
3076 assert( !n->is_Call(), "" );
3077 assert( !n->is_Mem(), "" );
3078 break;
3079 }
3081 // Collect CFG split points
3082 if (n->is_MultiBranch())
3083 frc._tests.push(n);
3084 }
3086 //------------------------------final_graph_reshaping_walk---------------------
3087 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3088 // requires that the walk visits a node's inputs before visiting the node.
3089 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3090 ResourceArea *area = Thread::current()->resource_area();
3091 Unique_Node_List sfpt(area);
3093 frc._visited.set(root->_idx); // first, mark node as visited
3094 uint cnt = root->req();
3095 Node *n = root;
3096 uint i = 0;
3097 while (true) {
3098 if (i < cnt) {
3099 // Place all non-visited non-null inputs onto stack
3100 Node* m = n->in(i);
3101 ++i;
3102 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3103 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3104 // compute worst case interpreter size in case of a deoptimization
3105 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3107 sfpt.push(m);
3108 }
3109 cnt = m->req();
3110 nstack.push(n, i); // put on stack parent and next input's index
3111 n = m;
3112 i = 0;
3113 }
3114 } else {
3115 // Now do post-visit work
3116 final_graph_reshaping_impl( n, frc );
3117 if (nstack.is_empty())
3118 break; // finished
3119 n = nstack.node(); // Get node from stack
3120 cnt = n->req();
3121 i = nstack.index();
3122 nstack.pop(); // Shift to the next node on stack
3123 }
3124 }
3126 // Skip next transformation if compressed oops are not used.
3127 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3128 (!UseCompressedOops && !UseCompressedClassPointers))
3129 return;
3131 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3132 // It could be done for an uncommon traps or any safepoints/calls
3133 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3134 while (sfpt.size() > 0) {
3135 n = sfpt.pop();
3136 JVMState *jvms = n->as_SafePoint()->jvms();
3137 assert(jvms != NULL, "sanity");
3138 int start = jvms->debug_start();
3139 int end = n->req();
3140 bool is_uncommon = (n->is_CallStaticJava() &&
3141 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3142 for (int j = start; j < end; j++) {
3143 Node* in = n->in(j);
3144 if (in->is_DecodeNarrowPtr()) {
3145 bool safe_to_skip = true;
3146 if (!is_uncommon ) {
3147 // Is it safe to skip?
3148 for (uint i = 0; i < in->outcnt(); i++) {
3149 Node* u = in->raw_out(i);
3150 if (!u->is_SafePoint() ||
3151 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3152 safe_to_skip = false;
3153 }
3154 }
3155 }
3156 if (safe_to_skip) {
3157 n->set_req(j, in->in(1));
3158 }
3159 if (in->outcnt() == 0) {
3160 in->disconnect_inputs(NULL, this);
3161 }
3162 }
3163 }
3164 }
3165 }
3167 //------------------------------final_graph_reshaping--------------------------
3168 // Final Graph Reshaping.
3169 //
3170 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3171 // and not commoned up and forced early. Must come after regular
3172 // optimizations to avoid GVN undoing the cloning. Clone constant
3173 // inputs to Loop Phis; these will be split by the allocator anyways.
3174 // Remove Opaque nodes.
3175 // (2) Move last-uses by commutative operations to the left input to encourage
3176 // Intel update-in-place two-address operations and better register usage
3177 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3178 // calls canonicalizing them back.
3179 // (3) Count the number of double-precision FP ops, single-precision FP ops
3180 // and call sites. On Intel, we can get correct rounding either by
3181 // forcing singles to memory (requires extra stores and loads after each
3182 // FP bytecode) or we can set a rounding mode bit (requires setting and
3183 // clearing the mode bit around call sites). The mode bit is only used
3184 // if the relative frequency of single FP ops to calls is low enough.
3185 // This is a key transform for SPEC mpeg_audio.
3186 // (4) Detect infinite loops; blobs of code reachable from above but not
3187 // below. Several of the Code_Gen algorithms fail on such code shapes,
3188 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3189 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3190 // Detection is by looking for IfNodes where only 1 projection is
3191 // reachable from below or CatchNodes missing some targets.
3192 // (5) Assert for insane oop offsets in debug mode.
3194 bool Compile::final_graph_reshaping() {
3195 // an infinite loop may have been eliminated by the optimizer,
3196 // in which case the graph will be empty.
3197 if (root()->req() == 1) {
3198 record_method_not_compilable("trivial infinite loop");
3199 return true;
3200 }
3202 // Expensive nodes have their control input set to prevent the GVN
3203 // from freely commoning them. There's no GVN beyond this point so
3204 // no need to keep the control input. We want the expensive nodes to
3205 // be freely moved to the least frequent code path by gcm.
3206 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3207 for (int i = 0; i < expensive_count(); i++) {
3208 _expensive_nodes->at(i)->set_req(0, NULL);
3209 }
3211 Final_Reshape_Counts frc;
3213 // Visit everybody reachable!
3214 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3215 Node_Stack nstack(unique() >> 1);
3216 final_graph_reshaping_walk(nstack, root(), frc);
3218 // Check for unreachable (from below) code (i.e., infinite loops).
3219 for( uint i = 0; i < frc._tests.size(); i++ ) {
3220 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3221 // Get number of CFG targets.
3222 // Note that PCTables include exception targets after calls.
3223 uint required_outcnt = n->required_outcnt();
3224 if (n->outcnt() != required_outcnt) {
3225 // Check for a few special cases. Rethrow Nodes never take the
3226 // 'fall-thru' path, so expected kids is 1 less.
3227 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3228 if (n->in(0)->in(0)->is_Call()) {
3229 CallNode *call = n->in(0)->in(0)->as_Call();
3230 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3231 required_outcnt--; // Rethrow always has 1 less kid
3232 } else if (call->req() > TypeFunc::Parms &&
3233 call->is_CallDynamicJava()) {
3234 // Check for null receiver. In such case, the optimizer has
3235 // detected that the virtual call will always result in a null
3236 // pointer exception. The fall-through projection of this CatchNode
3237 // will not be populated.
3238 Node *arg0 = call->in(TypeFunc::Parms);
3239 if (arg0->is_Type() &&
3240 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3241 required_outcnt--;
3242 }
3243 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3244 call->req() > TypeFunc::Parms+1 &&
3245 call->is_CallStaticJava()) {
3246 // Check for negative array length. In such case, the optimizer has
3247 // detected that the allocation attempt will always result in an
3248 // exception. There is no fall-through projection of this CatchNode .
3249 Node *arg1 = call->in(TypeFunc::Parms+1);
3250 if (arg1->is_Type() &&
3251 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3252 required_outcnt--;
3253 }
3254 }
3255 }
3256 }
3257 // Recheck with a better notion of 'required_outcnt'
3258 if (n->outcnt() != required_outcnt) {
3259 record_method_not_compilable("malformed control flow");
3260 return true; // Not all targets reachable!
3261 }
3262 }
3263 // Check that I actually visited all kids. Unreached kids
3264 // must be infinite loops.
3265 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3266 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3267 record_method_not_compilable("infinite loop");
3268 return true; // Found unvisited kid; must be unreach
3269 }
3270 }
3272 // If original bytecodes contained a mixture of floats and doubles
3273 // check if the optimizer has made it homogenous, item (3).
3274 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3275 frc.get_float_count() > 32 &&
3276 frc.get_double_count() == 0 &&
3277 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3278 set_24_bit_selection_and_mode( false, true );
3279 }
3281 set_java_calls(frc.get_java_call_count());
3282 set_inner_loops(frc.get_inner_loop_count());
3284 // No infinite loops, no reason to bail out.
3285 return false;
3286 }
3288 //-----------------------------too_many_traps----------------------------------
3289 // Report if there are too many traps at the current method and bci.
3290 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3291 bool Compile::too_many_traps(ciMethod* method,
3292 int bci,
3293 Deoptimization::DeoptReason reason) {
3294 ciMethodData* md = method->method_data();
3295 if (md->is_empty()) {
3296 // Assume the trap has not occurred, or that it occurred only
3297 // because of a transient condition during start-up in the interpreter.
3298 return false;
3299 }
3300 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3301 if (md->has_trap_at(bci, m, reason) != 0) {
3302 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3303 // Also, if there are multiple reasons, or if there is no per-BCI record,
3304 // assume the worst.
3305 if (log())
3306 log()->elem("observe trap='%s' count='%d'",
3307 Deoptimization::trap_reason_name(reason),
3308 md->trap_count(reason));
3309 return true;
3310 } else {
3311 // Ignore method/bci and see if there have been too many globally.
3312 return too_many_traps(reason, md);
3313 }
3314 }
3316 // Less-accurate variant which does not require a method and bci.
3317 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3318 ciMethodData* logmd) {
3319 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3320 // Too many traps globally.
3321 // Note that we use cumulative trap_count, not just md->trap_count.
3322 if (log()) {
3323 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3324 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3325 Deoptimization::trap_reason_name(reason),
3326 mcount, trap_count(reason));
3327 }
3328 return true;
3329 } else {
3330 // The coast is clear.
3331 return false;
3332 }
3333 }
3335 //--------------------------too_many_recompiles--------------------------------
3336 // Report if there are too many recompiles at the current method and bci.
3337 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3338 // Is not eager to return true, since this will cause the compiler to use
3339 // Action_none for a trap point, to avoid too many recompilations.
3340 bool Compile::too_many_recompiles(ciMethod* method,
3341 int bci,
3342 Deoptimization::DeoptReason reason) {
3343 ciMethodData* md = method->method_data();
3344 if (md->is_empty()) {
3345 // Assume the trap has not occurred, or that it occurred only
3346 // because of a transient condition during start-up in the interpreter.
3347 return false;
3348 }
3349 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3350 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3351 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3352 Deoptimization::DeoptReason per_bc_reason
3353 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3354 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3355 if ((per_bc_reason == Deoptimization::Reason_none
3356 || md->has_trap_at(bci, m, reason) != 0)
3357 // The trap frequency measure we care about is the recompile count:
3358 && md->trap_recompiled_at(bci, m)
3359 && md->overflow_recompile_count() >= bc_cutoff) {
3360 // Do not emit a trap here if it has already caused recompilations.
3361 // Also, if there are multiple reasons, or if there is no per-BCI record,
3362 // assume the worst.
3363 if (log())
3364 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3365 Deoptimization::trap_reason_name(reason),
3366 md->trap_count(reason),
3367 md->overflow_recompile_count());
3368 return true;
3369 } else if (trap_count(reason) != 0
3370 && decompile_count() >= m_cutoff) {
3371 // Too many recompiles globally, and we have seen this sort of trap.
3372 // Use cumulative decompile_count, not just md->decompile_count.
3373 if (log())
3374 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3375 Deoptimization::trap_reason_name(reason),
3376 md->trap_count(reason), trap_count(reason),
3377 md->decompile_count(), decompile_count());
3378 return true;
3379 } else {
3380 // The coast is clear.
3381 return false;
3382 }
3383 }
3385 // Compute when not to trap. Used by matching trap based nodes and
3386 // NullCheck optimization.
3387 void Compile::set_allowed_deopt_reasons() {
3388 _allowed_reasons = 0;
3389 if (is_method_compilation()) {
3390 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3391 assert(rs < BitsPerInt, "recode bit map");
3392 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3393 _allowed_reasons |= nth_bit(rs);
3394 }
3395 }
3396 }
3397 }
3399 #ifndef PRODUCT
3400 //------------------------------verify_graph_edges---------------------------
3401 // Walk the Graph and verify that there is a one-to-one correspondence
3402 // between Use-Def edges and Def-Use edges in the graph.
3403 void Compile::verify_graph_edges(bool no_dead_code) {
3404 if (VerifyGraphEdges) {
3405 ResourceArea *area = Thread::current()->resource_area();
3406 Unique_Node_List visited(area);
3407 // Call recursive graph walk to check edges
3408 _root->verify_edges(visited);
3409 if (no_dead_code) {
3410 // Now make sure that no visited node is used by an unvisited node.
3411 bool dead_nodes = 0;
3412 Unique_Node_List checked(area);
3413 while (visited.size() > 0) {
3414 Node* n = visited.pop();
3415 checked.push(n);
3416 for (uint i = 0; i < n->outcnt(); i++) {
3417 Node* use = n->raw_out(i);
3418 if (checked.member(use)) continue; // already checked
3419 if (visited.member(use)) continue; // already in the graph
3420 if (use->is_Con()) continue; // a dead ConNode is OK
3421 // At this point, we have found a dead node which is DU-reachable.
3422 if (dead_nodes++ == 0)
3423 tty->print_cr("*** Dead nodes reachable via DU edges:");
3424 use->dump(2);
3425 tty->print_cr("---");
3426 checked.push(use); // No repeats; pretend it is now checked.
3427 }
3428 }
3429 assert(dead_nodes == 0, "using nodes must be reachable from root");
3430 }
3431 }
3432 }
3434 // Verify GC barriers consistency
3435 // Currently supported:
3436 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3437 void Compile::verify_barriers() {
3438 if (UseG1GC) {
3439 // Verify G1 pre-barriers
3440 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3442 ResourceArea *area = Thread::current()->resource_area();
3443 Unique_Node_List visited(area);
3444 Node_List worklist(area);
3445 // We're going to walk control flow backwards starting from the Root
3446 worklist.push(_root);
3447 while (worklist.size() > 0) {
3448 Node* x = worklist.pop();
3449 if (x == NULL || x == top()) continue;
3450 if (visited.member(x)) {
3451 continue;
3452 } else {
3453 visited.push(x);
3454 }
3456 if (x->is_Region()) {
3457 for (uint i = 1; i < x->req(); i++) {
3458 worklist.push(x->in(i));
3459 }
3460 } else {
3461 worklist.push(x->in(0));
3462 // We are looking for the pattern:
3463 // /->ThreadLocal
3464 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3465 // \->ConI(0)
3466 // We want to verify that the If and the LoadB have the same control
3467 // See GraphKit::g1_write_barrier_pre()
3468 if (x->is_If()) {
3469 IfNode *iff = x->as_If();
3470 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3471 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3472 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3473 && cmp->in(1)->is_Load()) {
3474 LoadNode* load = cmp->in(1)->as_Load();
3475 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3476 && load->in(2)->in(3)->is_Con()
3477 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3479 Node* if_ctrl = iff->in(0);
3480 Node* load_ctrl = load->in(0);
3482 if (if_ctrl != load_ctrl) {
3483 // Skip possible CProj->NeverBranch in infinite loops
3484 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3485 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3486 if_ctrl = if_ctrl->in(0)->in(0);
3487 }
3488 }
3489 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3490 }
3491 }
3492 }
3493 }
3494 }
3495 }
3496 }
3497 }
3499 #endif
3501 // The Compile object keeps track of failure reasons separately from the ciEnv.
3502 // This is required because there is not quite a 1-1 relation between the
3503 // ciEnv and its compilation task and the Compile object. Note that one
3504 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3505 // to backtrack and retry without subsuming loads. Other than this backtracking
3506 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3507 // by the logic in C2Compiler.
3508 void Compile::record_failure(const char* reason) {
3509 if (log() != NULL) {
3510 log()->elem("failure reason='%s' phase='compile'", reason);
3511 }
3512 if (_failure_reason == NULL) {
3513 // Record the first failure reason.
3514 _failure_reason = reason;
3515 }
3517 EventCompilerFailure event;
3518 if (event.should_commit()) {
3519 event.set_compileID(Compile::compile_id());
3520 event.set_failure(reason);
3521 event.commit();
3522 }
3524 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3525 C->print_method(PHASE_FAILURE);
3526 }
3527 _root = NULL; // flush the graph, too
3528 }
3530 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3531 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3532 _phase_name(name), _dolog(dolog)
3533 {
3534 if (dolog) {
3535 C = Compile::current();
3536 _log = C->log();
3537 } else {
3538 C = NULL;
3539 _log = NULL;
3540 }
3541 if (_log != NULL) {
3542 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3543 _log->stamp();
3544 _log->end_head();
3545 }
3546 }
3548 Compile::TracePhase::~TracePhase() {
3550 C = Compile::current();
3551 if (_dolog) {
3552 _log = C->log();
3553 } else {
3554 _log = NULL;
3555 }
3557 #ifdef ASSERT
3558 if (PrintIdealNodeCount) {
3559 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3560 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3561 }
3563 if (VerifyIdealNodeCount) {
3564 Compile::current()->print_missing_nodes();
3565 }
3566 #endif
3568 if (_log != NULL) {
3569 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3570 }
3571 }
3573 //=============================================================================
3574 // Two Constant's are equal when the type and the value are equal.
3575 bool Compile::Constant::operator==(const Constant& other) {
3576 if (type() != other.type() ) return false;
3577 if (can_be_reused() != other.can_be_reused()) return false;
3578 // For floating point values we compare the bit pattern.
3579 switch (type()) {
3580 case T_FLOAT: return (_v._value.i == other._v._value.i);
3581 case T_LONG:
3582 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3583 case T_OBJECT:
3584 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3585 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3586 case T_METADATA: return (_v._metadata == other._v._metadata);
3587 default: ShouldNotReachHere();
3588 }
3589 return false;
3590 }
3592 static int type_to_size_in_bytes(BasicType t) {
3593 switch (t) {
3594 case T_LONG: return sizeof(jlong );
3595 case T_FLOAT: return sizeof(jfloat );
3596 case T_DOUBLE: return sizeof(jdouble);
3597 case T_METADATA: return sizeof(Metadata*);
3598 // We use T_VOID as marker for jump-table entries (labels) which
3599 // need an internal word relocation.
3600 case T_VOID:
3601 case T_ADDRESS:
3602 case T_OBJECT: return sizeof(jobject);
3603 }
3605 ShouldNotReachHere();
3606 return -1;
3607 }
3609 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3610 // sort descending
3611 if (a->freq() > b->freq()) return -1;
3612 if (a->freq() < b->freq()) return 1;
3613 return 0;
3614 }
3616 void Compile::ConstantTable::calculate_offsets_and_size() {
3617 // First, sort the array by frequencies.
3618 _constants.sort(qsort_comparator);
3620 #ifdef ASSERT
3621 // Make sure all jump-table entries were sorted to the end of the
3622 // array (they have a negative frequency).
3623 bool found_void = false;
3624 for (int i = 0; i < _constants.length(); i++) {
3625 Constant con = _constants.at(i);
3626 if (con.type() == T_VOID)
3627 found_void = true; // jump-tables
3628 else
3629 assert(!found_void, "wrong sorting");
3630 }
3631 #endif
3633 int offset = 0;
3634 for (int i = 0; i < _constants.length(); i++) {
3635 Constant* con = _constants.adr_at(i);
3637 // Align offset for type.
3638 int typesize = type_to_size_in_bytes(con->type());
3639 offset = align_size_up(offset, typesize);
3640 con->set_offset(offset); // set constant's offset
3642 if (con->type() == T_VOID) {
3643 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3644 offset = offset + typesize * n->outcnt(); // expand jump-table
3645 } else {
3646 offset = offset + typesize;
3647 }
3648 }
3650 // Align size up to the next section start (which is insts; see
3651 // CodeBuffer::align_at_start).
3652 assert(_size == -1, "already set?");
3653 _size = align_size_up(offset, CodeEntryAlignment);
3654 }
3656 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3657 MacroAssembler _masm(&cb);
3658 for (int i = 0; i < _constants.length(); i++) {
3659 Constant con = _constants.at(i);
3660 address constant_addr;
3661 switch (con.type()) {
3662 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3663 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3664 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3665 case T_OBJECT: {
3666 jobject obj = con.get_jobject();
3667 int oop_index = _masm.oop_recorder()->find_index(obj);
3668 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3669 break;
3670 }
3671 case T_ADDRESS: {
3672 address addr = (address) con.get_jobject();
3673 constant_addr = _masm.address_constant(addr);
3674 break;
3675 }
3676 // We use T_VOID as marker for jump-table entries (labels) which
3677 // need an internal word relocation.
3678 case T_VOID: {
3679 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3680 // Fill the jump-table with a dummy word. The real value is
3681 // filled in later in fill_jump_table.
3682 address dummy = (address) n;
3683 constant_addr = _masm.address_constant(dummy);
3684 // Expand jump-table
3685 for (uint i = 1; i < n->outcnt(); i++) {
3686 address temp_addr = _masm.address_constant(dummy + i);
3687 assert(temp_addr, "consts section too small");
3688 }
3689 break;
3690 }
3691 case T_METADATA: {
3692 Metadata* obj = con.get_metadata();
3693 int metadata_index = _masm.oop_recorder()->find_index(obj);
3694 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3695 break;
3696 }
3697 default: ShouldNotReachHere();
3698 }
3699 assert(constant_addr, "consts section too small");
3700 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3701 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())));
3702 }
3703 }
3705 int Compile::ConstantTable::find_offset(Constant& con) const {
3706 int idx = _constants.find(con);
3707 assert(idx != -1, "constant must be in constant table");
3708 int offset = _constants.at(idx).offset();
3709 assert(offset != -1, "constant table not emitted yet?");
3710 return offset;
3711 }
3713 void Compile::ConstantTable::add(Constant& con) {
3714 if (con.can_be_reused()) {
3715 int idx = _constants.find(con);
3716 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3717 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3718 return;
3719 }
3720 }
3721 (void) _constants.append(con);
3722 }
3724 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3725 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3726 Constant con(type, value, b->_freq);
3727 add(con);
3728 return con;
3729 }
3731 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3732 Constant con(metadata);
3733 add(con);
3734 return con;
3735 }
3737 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3738 jvalue value;
3739 BasicType type = oper->type()->basic_type();
3740 switch (type) {
3741 case T_LONG: value.j = oper->constantL(); break;
3742 case T_FLOAT: value.f = oper->constantF(); break;
3743 case T_DOUBLE: value.d = oper->constantD(); break;
3744 case T_OBJECT:
3745 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3746 case T_METADATA: return add((Metadata*)oper->constant()); break;
3747 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3748 }
3749 return add(n, type, value);
3750 }
3752 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3753 jvalue value;
3754 // We can use the node pointer here to identify the right jump-table
3755 // as this method is called from Compile::Fill_buffer right before
3756 // the MachNodes are emitted and the jump-table is filled (means the
3757 // MachNode pointers do not change anymore).
3758 value.l = (jobject) n;
3759 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3760 add(con);
3761 return con;
3762 }
3764 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3765 // If called from Compile::scratch_emit_size do nothing.
3766 if (Compile::current()->in_scratch_emit_size()) return;
3768 assert(labels.is_nonempty(), "must be");
3769 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3771 // Since MachConstantNode::constant_offset() also contains
3772 // table_base_offset() we need to subtract the table_base_offset()
3773 // to get the plain offset into the constant table.
3774 int offset = n->constant_offset() - table_base_offset();
3776 MacroAssembler _masm(&cb);
3777 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3779 for (uint i = 0; i < n->outcnt(); i++) {
3780 address* constant_addr = &jump_table_base[i];
3781 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i)));
3782 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3783 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3784 }
3785 }
3787 void Compile::dump_inlining() {
3788 if (print_inlining() || print_intrinsics()) {
3789 // Print inlining message for candidates that we couldn't inline
3790 // for lack of space or non constant receiver
3791 for (int i = 0; i < _late_inlines.length(); i++) {
3792 CallGenerator* cg = _late_inlines.at(i);
3793 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3794 }
3795 Unique_Node_List useful;
3796 useful.push(root());
3797 for (uint next = 0; next < useful.size(); ++next) {
3798 Node* n = useful.at(next);
3799 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3800 CallNode* call = n->as_Call();
3801 CallGenerator* cg = call->generator();
3802 cg->print_inlining_late("receiver not constant");
3803 }
3804 uint max = n->len();
3805 for ( uint i = 0; i < max; ++i ) {
3806 Node *m = n->in(i);
3807 if ( m == NULL ) continue;
3808 useful.push(m);
3809 }
3810 }
3811 for (int i = 0; i < _print_inlining_list->length(); i++) {
3812 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
3813 }
3814 }
3815 }
3817 // Dump inlining replay data to the stream.
3818 // Don't change thread state and acquire any locks.
3819 void Compile::dump_inline_data(outputStream* out) {
3820 InlineTree* inl_tree = ilt();
3821 if (inl_tree != NULL) {
3822 out->print(" inline %d", inl_tree->count());
3823 inl_tree->dump_replay_data(out);
3824 }
3825 }
3827 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3828 if (n1->Opcode() < n2->Opcode()) return -1;
3829 else if (n1->Opcode() > n2->Opcode()) return 1;
3831 assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()));
3832 for (uint i = 1; i < n1->req(); i++) {
3833 if (n1->in(i) < n2->in(i)) return -1;
3834 else if (n1->in(i) > n2->in(i)) return 1;
3835 }
3837 return 0;
3838 }
3840 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3841 Node* n1 = *n1p;
3842 Node* n2 = *n2p;
3844 return cmp_expensive_nodes(n1, n2);
3845 }
3847 void Compile::sort_expensive_nodes() {
3848 if (!expensive_nodes_sorted()) {
3849 _expensive_nodes->sort(cmp_expensive_nodes);
3850 }
3851 }
3853 bool Compile::expensive_nodes_sorted() const {
3854 for (int i = 1; i < _expensive_nodes->length(); i++) {
3855 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3856 return false;
3857 }
3858 }
3859 return true;
3860 }
3862 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3863 if (_expensive_nodes->length() == 0) {
3864 return false;
3865 }
3867 assert(OptimizeExpensiveOps, "optimization off?");
3869 // Take this opportunity to remove dead nodes from the list
3870 int j = 0;
3871 for (int i = 0; i < _expensive_nodes->length(); i++) {
3872 Node* n = _expensive_nodes->at(i);
3873 if (!n->is_unreachable(igvn)) {
3874 assert(n->is_expensive(), "should be expensive");
3875 _expensive_nodes->at_put(j, n);
3876 j++;
3877 }
3878 }
3879 _expensive_nodes->trunc_to(j);
3881 // Then sort the list so that similar nodes are next to each other
3882 // and check for at least two nodes of identical kind with same data
3883 // inputs.
3884 sort_expensive_nodes();
3886 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3887 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3888 return true;
3889 }
3890 }
3892 return false;
3893 }
3895 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3896 if (_expensive_nodes->length() == 0) {
3897 return;
3898 }
3900 assert(OptimizeExpensiveOps, "optimization off?");
3902 // Sort to bring similar nodes next to each other and clear the
3903 // control input of nodes for which there's only a single copy.
3904 sort_expensive_nodes();
3906 int j = 0;
3907 int identical = 0;
3908 int i = 0;
3909 for (; i < _expensive_nodes->length()-1; i++) {
3910 assert(j <= i, "can't write beyond current index");
3911 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3912 identical++;
3913 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3914 continue;
3915 }
3916 if (identical > 0) {
3917 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3918 identical = 0;
3919 } else {
3920 Node* n = _expensive_nodes->at(i);
3921 igvn.hash_delete(n);
3922 n->set_req(0, NULL);
3923 igvn.hash_insert(n);
3924 }
3925 }
3926 if (identical > 0) {
3927 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3928 } else if (_expensive_nodes->length() >= 1) {
3929 Node* n = _expensive_nodes->at(i);
3930 igvn.hash_delete(n);
3931 n->set_req(0, NULL);
3932 igvn.hash_insert(n);
3933 }
3934 _expensive_nodes->trunc_to(j);
3935 }
3937 void Compile::add_expensive_node(Node * n) {
3938 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3939 assert(n->is_expensive(), "expensive nodes with non-null control here only");
3940 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
3941 if (OptimizeExpensiveOps) {
3942 _expensive_nodes->append(n);
3943 } else {
3944 // Clear control input and let IGVN optimize expensive nodes if
3945 // OptimizeExpensiveOps is off.
3946 n->set_req(0, NULL);
3947 }
3948 }
3950 /**
3951 * Remove the speculative part of types and clean up the graph
3952 */
3953 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
3954 if (UseTypeSpeculation) {
3955 Unique_Node_List worklist;
3956 worklist.push(root());
3957 int modified = 0;
3958 // Go over all type nodes that carry a speculative type, drop the
3959 // speculative part of the type and enqueue the node for an igvn
3960 // which may optimize it out.
3961 for (uint next = 0; next < worklist.size(); ++next) {
3962 Node *n = worklist.at(next);
3963 if (n->is_Type()) {
3964 TypeNode* tn = n->as_Type();
3965 const Type* t = tn->type();
3966 const Type* t_no_spec = t->remove_speculative();
3967 if (t_no_spec != t) {
3968 bool in_hash = igvn.hash_delete(n);
3969 assert(in_hash, "node should be in igvn hash table");
3970 tn->set_type(t_no_spec);
3971 igvn.hash_insert(n);
3972 igvn._worklist.push(n); // give it a chance to go away
3973 modified++;
3974 }
3975 }
3976 uint max = n->len();
3977 for( uint i = 0; i < max; ++i ) {
3978 Node *m = n->in(i);
3979 if (not_a_node(m)) continue;
3980 worklist.push(m);
3981 }
3982 }
3983 // Drop the speculative part of all types in the igvn's type table
3984 igvn.remove_speculative_types();
3985 if (modified > 0) {
3986 igvn.optimize();
3987 }
3988 #ifdef ASSERT
3989 // Verify that after the IGVN is over no speculative type has resurfaced
3990 worklist.clear();
3991 worklist.push(root());
3992 for (uint next = 0; next < worklist.size(); ++next) {
3993 Node *n = worklist.at(next);
3994 const Type* t = igvn.type_or_null(n);
3995 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
3996 if (n->is_Type()) {
3997 t = n->as_Type()->type();
3998 assert(t == t->remove_speculative(), "no more speculative types");
3999 }
4000 uint max = n->len();
4001 for( uint i = 0; i < max; ++i ) {
4002 Node *m = n->in(i);
4003 if (not_a_node(m)) continue;
4004 worklist.push(m);
4005 }
4006 }
4007 igvn.check_no_speculative_types();
4008 #endif
4009 }
4010 }
4012 // Auxiliary method to support randomized stressing/fuzzing.
4013 //
4014 // This method can be called the arbitrary number of times, with current count
4015 // as the argument. The logic allows selecting a single candidate from the
4016 // running list of candidates as follows:
4017 // int count = 0;
4018 // Cand* selected = null;
4019 // while(cand = cand->next()) {
4020 // if (randomized_select(++count)) {
4021 // selected = cand;
4022 // }
4023 // }
4024 //
4025 // Including count equalizes the chances any candidate is "selected".
4026 // This is useful when we don't have the complete list of candidates to choose
4027 // from uniformly. In this case, we need to adjust the randomicity of the
4028 // selection, or else we will end up biasing the selection towards the latter
4029 // candidates.
4030 //
4031 // Quick back-envelope calculation shows that for the list of n candidates
4032 // the equal probability for the candidate to persist as "best" can be
4033 // achieved by replacing it with "next" k-th candidate with the probability
4034 // of 1/k. It can be easily shown that by the end of the run, the
4035 // probability for any candidate is converged to 1/n, thus giving the
4036 // uniform distribution among all the candidates.
4037 //
4038 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4039 #define RANDOMIZED_DOMAIN_POW 29
4040 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4041 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4042 bool Compile::randomized_select(int count) {
4043 assert(count > 0, "only positive");
4044 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4045 }