Tue, 02 Jul 2013 20:42:12 -0400
7088419: Use x86 Hardware CRC32 Instruction with java.util.zip.CRC32
Summary: add intrinsics using new instruction to interpreter, C1, C2, for suitable x86; add test
Reviewed-by: kvn, twisti
1 /*
2 * Copyright (c) 1997, 2013, 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 "classfile/systemDictionary.hpp"
29 #include "code/exceptionHandlerTable.hpp"
30 #include "code/nmethod.hpp"
31 #include "compiler/compileLog.hpp"
32 #include "compiler/disassembler.hpp"
33 #include "compiler/oopMap.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/block.hpp"
36 #include "opto/c2compiler.hpp"
37 #include "opto/callGenerator.hpp"
38 #include "opto/callnode.hpp"
39 #include "opto/cfgnode.hpp"
40 #include "opto/chaitin.hpp"
41 #include "opto/compile.hpp"
42 #include "opto/connode.hpp"
43 #include "opto/divnode.hpp"
44 #include "opto/escape.hpp"
45 #include "opto/idealGraphPrinter.hpp"
46 #include "opto/loopnode.hpp"
47 #include "opto/machnode.hpp"
48 #include "opto/macro.hpp"
49 #include "opto/matcher.hpp"
50 #include "opto/memnode.hpp"
51 #include "opto/mulnode.hpp"
52 #include "opto/node.hpp"
53 #include "opto/opcodes.hpp"
54 #include "opto/output.hpp"
55 #include "opto/parse.hpp"
56 #include "opto/phaseX.hpp"
57 #include "opto/rootnode.hpp"
58 #include "opto/runtime.hpp"
59 #include "opto/stringopts.hpp"
60 #include "opto/type.hpp"
61 #include "opto/vectornode.hpp"
62 #include "runtime/arguments.hpp"
63 #include "runtime/signature.hpp"
64 #include "runtime/stubRoutines.hpp"
65 #include "runtime/timer.hpp"
66 #include "trace/tracing.hpp"
67 #include "utilities/copy.hpp"
68 #ifdef TARGET_ARCH_MODEL_x86_32
69 # include "adfiles/ad_x86_32.hpp"
70 #endif
71 #ifdef TARGET_ARCH_MODEL_x86_64
72 # include "adfiles/ad_x86_64.hpp"
73 #endif
74 #ifdef TARGET_ARCH_MODEL_sparc
75 # include "adfiles/ad_sparc.hpp"
76 #endif
77 #ifdef TARGET_ARCH_MODEL_zero
78 # include "adfiles/ad_zero.hpp"
79 #endif
80 #ifdef TARGET_ARCH_MODEL_arm
81 # include "adfiles/ad_arm.hpp"
82 #endif
83 #ifdef TARGET_ARCH_MODEL_ppc
84 # include "adfiles/ad_ppc.hpp"
85 #endif
88 // -------------------- Compile::mach_constant_base_node -----------------------
89 // Constant table base node singleton.
90 MachConstantBaseNode* Compile::mach_constant_base_node() {
91 if (_mach_constant_base_node == NULL) {
92 _mach_constant_base_node = new (C) MachConstantBaseNode();
93 _mach_constant_base_node->add_req(C->root());
94 }
95 return _mach_constant_base_node;
96 }
99 /// Support for intrinsics.
101 // Return the index at which m must be inserted (or already exists).
102 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
103 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
104 #ifdef ASSERT
105 for (int i = 1; i < _intrinsics->length(); i++) {
106 CallGenerator* cg1 = _intrinsics->at(i-1);
107 CallGenerator* cg2 = _intrinsics->at(i);
108 assert(cg1->method() != cg2->method()
109 ? cg1->method() < cg2->method()
110 : cg1->is_virtual() < cg2->is_virtual(),
111 "compiler intrinsics list must stay sorted");
112 }
113 #endif
114 // Binary search sorted list, in decreasing intervals [lo, hi].
115 int lo = 0, hi = _intrinsics->length()-1;
116 while (lo <= hi) {
117 int mid = (uint)(hi + lo) / 2;
118 ciMethod* mid_m = _intrinsics->at(mid)->method();
119 if (m < mid_m) {
120 hi = mid-1;
121 } else if (m > mid_m) {
122 lo = mid+1;
123 } else {
124 // look at minor sort key
125 bool mid_virt = _intrinsics->at(mid)->is_virtual();
126 if (is_virtual < mid_virt) {
127 hi = mid-1;
128 } else if (is_virtual > mid_virt) {
129 lo = mid+1;
130 } else {
131 return mid; // exact match
132 }
133 }
134 }
135 return lo; // inexact match
136 }
138 void Compile::register_intrinsic(CallGenerator* cg) {
139 if (_intrinsics == NULL) {
140 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
141 }
142 // This code is stolen from ciObjectFactory::insert.
143 // Really, GrowableArray should have methods for
144 // insert_at, remove_at, and binary_search.
145 int len = _intrinsics->length();
146 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
147 if (index == len) {
148 _intrinsics->append(cg);
149 } else {
150 #ifdef ASSERT
151 CallGenerator* oldcg = _intrinsics->at(index);
152 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
153 #endif
154 _intrinsics->append(_intrinsics->at(len-1));
155 int pos;
156 for (pos = len-2; pos >= index; pos--) {
157 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
158 }
159 _intrinsics->at_put(index, cg);
160 }
161 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
162 }
164 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
165 assert(m->is_loaded(), "don't try this on unloaded methods");
166 if (_intrinsics != NULL) {
167 int index = intrinsic_insertion_index(m, is_virtual);
168 if (index < _intrinsics->length()
169 && _intrinsics->at(index)->method() == m
170 && _intrinsics->at(index)->is_virtual() == is_virtual) {
171 return _intrinsics->at(index);
172 }
173 }
174 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
175 if (m->intrinsic_id() != vmIntrinsics::_none &&
176 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
177 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
178 if (cg != NULL) {
179 // Save it for next time:
180 register_intrinsic(cg);
181 return cg;
182 } else {
183 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
184 }
185 }
186 return NULL;
187 }
189 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
190 // in library_call.cpp.
193 #ifndef PRODUCT
194 // statistics gathering...
196 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
197 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
199 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
200 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
201 int oflags = _intrinsic_hist_flags[id];
202 assert(flags != 0, "what happened?");
203 if (is_virtual) {
204 flags |= _intrinsic_virtual;
205 }
206 bool changed = (flags != oflags);
207 if ((flags & _intrinsic_worked) != 0) {
208 juint count = (_intrinsic_hist_count[id] += 1);
209 if (count == 1) {
210 changed = true; // first time
211 }
212 // increment the overall count also:
213 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
214 }
215 if (changed) {
216 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
217 // Something changed about the intrinsic's virtuality.
218 if ((flags & _intrinsic_virtual) != 0) {
219 // This is the first use of this intrinsic as a virtual call.
220 if (oflags != 0) {
221 // We already saw it as a non-virtual, so note both cases.
222 flags |= _intrinsic_both;
223 }
224 } else if ((oflags & _intrinsic_both) == 0) {
225 // This is the first use of this intrinsic as a non-virtual
226 flags |= _intrinsic_both;
227 }
228 }
229 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
230 }
231 // update the overall flags also:
232 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
233 return changed;
234 }
236 static char* format_flags(int flags, char* buf) {
237 buf[0] = 0;
238 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
239 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
240 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
241 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
242 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
243 if (buf[0] == 0) strcat(buf, ",");
244 assert(buf[0] == ',', "must be");
245 return &buf[1];
246 }
248 void Compile::print_intrinsic_statistics() {
249 char flagsbuf[100];
250 ttyLocker ttyl;
251 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
252 tty->print_cr("Compiler intrinsic usage:");
253 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
254 if (total == 0) total = 1; // avoid div0 in case of no successes
255 #define PRINT_STAT_LINE(name, c, f) \
256 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
257 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
258 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
259 int flags = _intrinsic_hist_flags[id];
260 juint count = _intrinsic_hist_count[id];
261 if ((flags | count) != 0) {
262 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
263 }
264 }
265 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
266 if (xtty != NULL) xtty->tail("statistics");
267 }
269 void Compile::print_statistics() {
270 { ttyLocker ttyl;
271 if (xtty != NULL) xtty->head("statistics type='opto'");
272 Parse::print_statistics();
273 PhaseCCP::print_statistics();
274 PhaseRegAlloc::print_statistics();
275 Scheduling::print_statistics();
276 PhasePeephole::print_statistics();
277 PhaseIdealLoop::print_statistics();
278 if (xtty != NULL) xtty->tail("statistics");
279 }
280 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
281 // put this under its own <statistics> element.
282 print_intrinsic_statistics();
283 }
284 }
285 #endif //PRODUCT
287 // Support for bundling info
288 Bundle* Compile::node_bundling(const Node *n) {
289 assert(valid_bundle_info(n), "oob");
290 return &_node_bundling_base[n->_idx];
291 }
293 bool Compile::valid_bundle_info(const Node *n) {
294 return (_node_bundling_limit > n->_idx);
295 }
298 void Compile::gvn_replace_by(Node* n, Node* nn) {
299 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
300 Node* use = n->last_out(i);
301 bool is_in_table = initial_gvn()->hash_delete(use);
302 uint uses_found = 0;
303 for (uint j = 0; j < use->len(); j++) {
304 if (use->in(j) == n) {
305 if (j < use->req())
306 use->set_req(j, nn);
307 else
308 use->set_prec(j, nn);
309 uses_found++;
310 }
311 }
312 if (is_in_table) {
313 // reinsert into table
314 initial_gvn()->hash_find_insert(use);
315 }
316 record_for_igvn(use);
317 i -= uses_found; // we deleted 1 or more copies of this edge
318 }
319 }
322 static inline bool not_a_node(const Node* n) {
323 if (n == NULL) return true;
324 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
325 if (*(address*)n == badAddress) return true; // kill by Node::destruct
326 return false;
327 }
329 // Identify all nodes that are reachable from below, useful.
330 // Use breadth-first pass that records state in a Unique_Node_List,
331 // recursive traversal is slower.
332 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
333 int estimated_worklist_size = unique();
334 useful.map( estimated_worklist_size, NULL ); // preallocate space
336 // Initialize worklist
337 if (root() != NULL) { useful.push(root()); }
338 // If 'top' is cached, declare it useful to preserve cached node
339 if( cached_top_node() ) { useful.push(cached_top_node()); }
341 // Push all useful nodes onto the list, breadthfirst
342 for( uint next = 0; next < useful.size(); ++next ) {
343 assert( next < unique(), "Unique useful nodes < total nodes");
344 Node *n = useful.at(next);
345 uint max = n->len();
346 for( uint i = 0; i < max; ++i ) {
347 Node *m = n->in(i);
348 if (not_a_node(m)) continue;
349 useful.push(m);
350 }
351 }
352 }
354 // Update dead_node_list with any missing dead nodes using useful
355 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
356 void Compile::update_dead_node_list(Unique_Node_List &useful) {
357 uint max_idx = unique();
358 VectorSet& useful_node_set = useful.member_set();
360 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
361 // If node with index node_idx is not in useful set,
362 // mark it as dead in dead node list.
363 if (! useful_node_set.test(node_idx) ) {
364 record_dead_node(node_idx);
365 }
366 }
367 }
369 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
370 int shift = 0;
371 for (int i = 0; i < inlines->length(); i++) {
372 CallGenerator* cg = inlines->at(i);
373 CallNode* call = cg->call_node();
374 if (shift > 0) {
375 inlines->at_put(i-shift, cg);
376 }
377 if (!useful.member(call)) {
378 shift++;
379 }
380 }
381 inlines->trunc_to(inlines->length()-shift);
382 }
384 // Disconnect all useless nodes by disconnecting those at the boundary.
385 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
386 uint next = 0;
387 while (next < useful.size()) {
388 Node *n = useful.at(next++);
389 // Use raw traversal of out edges since this code removes out edges
390 int max = n->outcnt();
391 for (int j = 0; j < max; ++j) {
392 Node* child = n->raw_out(j);
393 if (! useful.member(child)) {
394 assert(!child->is_top() || child != top(),
395 "If top is cached in Compile object it is in useful list");
396 // Only need to remove this out-edge to the useless node
397 n->raw_del_out(j);
398 --j;
399 --max;
400 }
401 }
402 if (n->outcnt() == 1 && n->has_special_unique_user()) {
403 record_for_igvn(n->unique_out());
404 }
405 }
406 // Remove useless macro and predicate opaq nodes
407 for (int i = C->macro_count()-1; i >= 0; i--) {
408 Node* n = C->macro_node(i);
409 if (!useful.member(n)) {
410 remove_macro_node(n);
411 }
412 }
413 // Remove useless expensive node
414 for (int i = C->expensive_count()-1; i >= 0; i--) {
415 Node* n = C->expensive_node(i);
416 if (!useful.member(n)) {
417 remove_expensive_node(n);
418 }
419 }
420 // clean up the late inline lists
421 remove_useless_late_inlines(&_string_late_inlines, useful);
422 remove_useless_late_inlines(&_boxing_late_inlines, useful);
423 remove_useless_late_inlines(&_late_inlines, useful);
424 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
425 }
427 //------------------------------frame_size_in_words-----------------------------
428 // frame_slots in units of words
429 int Compile::frame_size_in_words() const {
430 // shift is 0 in LP32 and 1 in LP64
431 const int shift = (LogBytesPerWord - LogBytesPerInt);
432 int words = _frame_slots >> shift;
433 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
434 return words;
435 }
437 // ============================================================================
438 //------------------------------CompileWrapper---------------------------------
439 class CompileWrapper : public StackObj {
440 Compile *const _compile;
441 public:
442 CompileWrapper(Compile* compile);
444 ~CompileWrapper();
445 };
447 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
448 // the Compile* pointer is stored in the current ciEnv:
449 ciEnv* env = compile->env();
450 assert(env == ciEnv::current(), "must already be a ciEnv active");
451 assert(env->compiler_data() == NULL, "compile already active?");
452 env->set_compiler_data(compile);
453 assert(compile == Compile::current(), "sanity");
455 compile->set_type_dict(NULL);
456 compile->set_type_hwm(NULL);
457 compile->set_type_last_size(0);
458 compile->set_last_tf(NULL, NULL);
459 compile->set_indexSet_arena(NULL);
460 compile->set_indexSet_free_block_list(NULL);
461 compile->init_type_arena();
462 Type::Initialize(compile);
463 _compile->set_scratch_buffer_blob(NULL);
464 _compile->begin_method();
465 }
466 CompileWrapper::~CompileWrapper() {
467 _compile->end_method();
468 if (_compile->scratch_buffer_blob() != NULL)
469 BufferBlob::free(_compile->scratch_buffer_blob());
470 _compile->env()->set_compiler_data(NULL);
471 }
474 //----------------------------print_compile_messages---------------------------
475 void Compile::print_compile_messages() {
476 #ifndef PRODUCT
477 // Check if recompiling
478 if (_subsume_loads == false && PrintOpto) {
479 // Recompiling without allowing machine instructions to subsume loads
480 tty->print_cr("*********************************************************");
481 tty->print_cr("** Bailout: Recompile without subsuming loads **");
482 tty->print_cr("*********************************************************");
483 }
484 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
485 // Recompiling without escape analysis
486 tty->print_cr("*********************************************************");
487 tty->print_cr("** Bailout: Recompile without escape analysis **");
488 tty->print_cr("*********************************************************");
489 }
490 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
491 // Recompiling without boxing elimination
492 tty->print_cr("*********************************************************");
493 tty->print_cr("** Bailout: Recompile without boxing elimination **");
494 tty->print_cr("*********************************************************");
495 }
496 if (env()->break_at_compile()) {
497 // Open the debugger when compiling this method.
498 tty->print("### Breaking when compiling: ");
499 method()->print_short_name();
500 tty->cr();
501 BREAKPOINT;
502 }
504 if( PrintOpto ) {
505 if (is_osr_compilation()) {
506 tty->print("[OSR]%3d", _compile_id);
507 } else {
508 tty->print("%3d", _compile_id);
509 }
510 }
511 #endif
512 }
515 //-----------------------init_scratch_buffer_blob------------------------------
516 // Construct a temporary BufferBlob and cache it for this compile.
517 void Compile::init_scratch_buffer_blob(int const_size) {
518 // If there is already a scratch buffer blob allocated and the
519 // constant section is big enough, use it. Otherwise free the
520 // current and allocate a new one.
521 BufferBlob* blob = scratch_buffer_blob();
522 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
523 // Use the current blob.
524 } else {
525 if (blob != NULL) {
526 BufferBlob::free(blob);
527 }
529 ResourceMark rm;
530 _scratch_const_size = const_size;
531 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
532 blob = BufferBlob::create("Compile::scratch_buffer", size);
533 // Record the buffer blob for next time.
534 set_scratch_buffer_blob(blob);
535 // Have we run out of code space?
536 if (scratch_buffer_blob() == NULL) {
537 // Let CompilerBroker disable further compilations.
538 record_failure("Not enough space for scratch buffer in CodeCache");
539 return;
540 }
541 }
543 // Initialize the relocation buffers
544 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
545 set_scratch_locs_memory(locs_buf);
546 }
549 //-----------------------scratch_emit_size-------------------------------------
550 // Helper function that computes size by emitting code
551 uint Compile::scratch_emit_size(const Node* n) {
552 // Start scratch_emit_size section.
553 set_in_scratch_emit_size(true);
555 // Emit into a trash buffer and count bytes emitted.
556 // This is a pretty expensive way to compute a size,
557 // but it works well enough if seldom used.
558 // All common fixed-size instructions are given a size
559 // method by the AD file.
560 // Note that the scratch buffer blob and locs memory are
561 // allocated at the beginning of the compile task, and
562 // may be shared by several calls to scratch_emit_size.
563 // The allocation of the scratch buffer blob is particularly
564 // expensive, since it has to grab the code cache lock.
565 BufferBlob* blob = this->scratch_buffer_blob();
566 assert(blob != NULL, "Initialize BufferBlob at start");
567 assert(blob->size() > MAX_inst_size, "sanity");
568 relocInfo* locs_buf = scratch_locs_memory();
569 address blob_begin = blob->content_begin();
570 address blob_end = (address)locs_buf;
571 assert(blob->content_contains(blob_end), "sanity");
572 CodeBuffer buf(blob_begin, blob_end - blob_begin);
573 buf.initialize_consts_size(_scratch_const_size);
574 buf.initialize_stubs_size(MAX_stubs_size);
575 assert(locs_buf != NULL, "sanity");
576 int lsize = MAX_locs_size / 3;
577 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
578 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
579 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
581 // Do the emission.
583 Label fakeL; // Fake label for branch instructions.
584 Label* saveL = NULL;
585 uint save_bnum = 0;
586 bool is_branch = n->is_MachBranch();
587 if (is_branch) {
588 MacroAssembler masm(&buf);
589 masm.bind(fakeL);
590 n->as_MachBranch()->save_label(&saveL, &save_bnum);
591 n->as_MachBranch()->label_set(&fakeL, 0);
592 }
593 n->emit(buf, this->regalloc());
594 if (is_branch) // Restore label.
595 n->as_MachBranch()->label_set(saveL, save_bnum);
597 // End scratch_emit_size section.
598 set_in_scratch_emit_size(false);
600 return buf.insts_size();
601 }
604 // ============================================================================
605 //------------------------------Compile standard-------------------------------
606 debug_only( int Compile::_debug_idx = 100000; )
608 // Compile a method. entry_bci is -1 for normal compilations and indicates
609 // the continuation bci for on stack replacement.
612 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
613 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
614 : Phase(Compiler),
615 _env(ci_env),
616 _log(ci_env->log()),
617 _compile_id(ci_env->compile_id()),
618 _save_argument_registers(false),
619 _stub_name(NULL),
620 _stub_function(NULL),
621 _stub_entry_point(NULL),
622 _method(target),
623 _entry_bci(osr_bci),
624 _initial_gvn(NULL),
625 _for_igvn(NULL),
626 _warm_calls(NULL),
627 _subsume_loads(subsume_loads),
628 _do_escape_analysis(do_escape_analysis),
629 _eliminate_boxing(eliminate_boxing),
630 _failure_reason(NULL),
631 _code_buffer("Compile::Fill_buffer"),
632 _orig_pc_slot(0),
633 _orig_pc_slot_offset_in_bytes(0),
634 _has_method_handle_invokes(false),
635 _mach_constant_base_node(NULL),
636 _node_bundling_limit(0),
637 _node_bundling_base(NULL),
638 _java_calls(0),
639 _inner_loops(0),
640 _scratch_const_size(-1),
641 _in_scratch_emit_size(false),
642 _dead_node_list(comp_arena()),
643 _dead_node_count(0),
644 #ifndef PRODUCT
645 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
646 _printer(IdealGraphPrinter::printer()),
647 #endif
648 _congraph(NULL),
649 _late_inlines(comp_arena(), 2, 0, NULL),
650 _string_late_inlines(comp_arena(), 2, 0, NULL),
651 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
652 _late_inlines_pos(0),
653 _number_of_mh_late_inlines(0),
654 _inlining_progress(false),
655 _inlining_incrementally(false),
656 _print_inlining_list(NULL),
657 _print_inlining(0) {
658 C = this;
660 CompileWrapper cw(this);
661 #ifndef PRODUCT
662 if (TimeCompiler2) {
663 tty->print(" ");
664 target->holder()->name()->print();
665 tty->print(".");
666 target->print_short_name();
667 tty->print(" ");
668 }
669 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
670 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
671 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
672 if (!print_opto_assembly) {
673 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
674 if (print_assembly && !Disassembler::can_decode()) {
675 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
676 print_opto_assembly = true;
677 }
678 }
679 set_print_assembly(print_opto_assembly);
680 set_parsed_irreducible_loop(false);
681 #endif
683 if (ProfileTraps) {
684 // Make sure the method being compiled gets its own MDO,
685 // so we can at least track the decompile_count().
686 method()->ensure_method_data();
687 }
689 Init(::AliasLevel);
692 print_compile_messages();
694 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
695 _ilt = InlineTree::build_inline_tree_root();
696 else
697 _ilt = NULL;
699 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
700 assert(num_alias_types() >= AliasIdxRaw, "");
702 #define MINIMUM_NODE_HASH 1023
703 // Node list that Iterative GVN will start with
704 Unique_Node_List for_igvn(comp_arena());
705 set_for_igvn(&for_igvn);
707 // GVN that will be run immediately on new nodes
708 uint estimated_size = method()->code_size()*4+64;
709 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
710 PhaseGVN gvn(node_arena(), estimated_size);
711 set_initial_gvn(&gvn);
713 if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) {
714 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
715 }
716 { // Scope for timing the parser
717 TracePhase t3("parse", &_t_parser, true);
719 // Put top into the hash table ASAP.
720 initial_gvn()->transform_no_reclaim(top());
722 // Set up tf(), start(), and find a CallGenerator.
723 CallGenerator* cg = NULL;
724 if (is_osr_compilation()) {
725 const TypeTuple *domain = StartOSRNode::osr_domain();
726 const TypeTuple *range = TypeTuple::make_range(method()->signature());
727 init_tf(TypeFunc::make(domain, range));
728 StartNode* s = new (this) StartOSRNode(root(), domain);
729 initial_gvn()->set_type_bottom(s);
730 init_start(s);
731 cg = CallGenerator::for_osr(method(), entry_bci());
732 } else {
733 // Normal case.
734 init_tf(TypeFunc::make(method()));
735 StartNode* s = new (this) StartNode(root(), tf()->domain());
736 initial_gvn()->set_type_bottom(s);
737 init_start(s);
738 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
739 // With java.lang.ref.reference.get() we must go through the
740 // intrinsic when G1 is enabled - even when get() is the root
741 // method of the compile - so that, if necessary, the value in
742 // the referent field of the reference object gets recorded by
743 // the pre-barrier code.
744 // Specifically, if G1 is enabled, the value in the referent
745 // field is recorded by the G1 SATB pre barrier. This will
746 // result in the referent being marked live and the reference
747 // object removed from the list of discovered references during
748 // reference processing.
749 cg = find_intrinsic(method(), false);
750 }
751 if (cg == NULL) {
752 float past_uses = method()->interpreter_invocation_count();
753 float expected_uses = past_uses;
754 cg = CallGenerator::for_inline(method(), expected_uses);
755 }
756 }
757 if (failing()) return;
758 if (cg == NULL) {
759 record_method_not_compilable_all_tiers("cannot parse method");
760 return;
761 }
762 JVMState* jvms = build_start_state(start(), tf());
763 if ((jvms = cg->generate(jvms)) == NULL) {
764 record_method_not_compilable("method parse failed");
765 return;
766 }
767 GraphKit kit(jvms);
769 if (!kit.stopped()) {
770 // Accept return values, and transfer control we know not where.
771 // This is done by a special, unique ReturnNode bound to root.
772 return_values(kit.jvms());
773 }
775 if (kit.has_exceptions()) {
776 // Any exceptions that escape from this call must be rethrown
777 // to whatever caller is dynamically above us on the stack.
778 // This is done by a special, unique RethrowNode bound to root.
779 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
780 }
782 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
784 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
785 inline_string_calls(true);
786 }
788 if (failing()) return;
790 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
792 // Remove clutter produced by parsing.
793 if (!failing()) {
794 ResourceMark rm;
795 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
796 }
797 }
799 // Note: Large methods are capped off in do_one_bytecode().
800 if (failing()) return;
802 // After parsing, node notes are no longer automagic.
803 // They must be propagated by register_new_node_with_optimizer(),
804 // clone(), or the like.
805 set_default_node_notes(NULL);
807 for (;;) {
808 int successes = Inline_Warm();
809 if (failing()) return;
810 if (successes == 0) break;
811 }
813 // Drain the list.
814 Finish_Warm();
815 #ifndef PRODUCT
816 if (_printer) {
817 _printer->print_inlining(this);
818 }
819 #endif
821 if (failing()) return;
822 NOT_PRODUCT( verify_graph_edges(); )
824 // Now optimize
825 Optimize();
826 if (failing()) return;
827 NOT_PRODUCT( verify_graph_edges(); )
829 #ifndef PRODUCT
830 if (PrintIdeal) {
831 ttyLocker ttyl; // keep the following output all in one block
832 // This output goes directly to the tty, not the compiler log.
833 // To enable tools to match it up with the compilation activity,
834 // be sure to tag this tty output with the compile ID.
835 if (xtty != NULL) {
836 xtty->head("ideal compile_id='%d'%s", compile_id(),
837 is_osr_compilation() ? " compile_kind='osr'" :
838 "");
839 }
840 root()->dump(9999);
841 if (xtty != NULL) {
842 xtty->tail("ideal");
843 }
844 }
845 #endif
847 // Now that we know the size of all the monitors we can add a fixed slot
848 // for the original deopt pc.
850 _orig_pc_slot = fixed_slots();
851 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
852 set_fixed_slots(next_slot);
854 // Now generate code
855 Code_Gen();
856 if (failing()) return;
858 // Check if we want to skip execution of all compiled code.
859 {
860 #ifndef PRODUCT
861 if (OptoNoExecute) {
862 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
863 return;
864 }
865 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
866 #endif
868 if (is_osr_compilation()) {
869 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
870 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
871 } else {
872 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
873 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
874 }
876 env()->register_method(_method, _entry_bci,
877 &_code_offsets,
878 _orig_pc_slot_offset_in_bytes,
879 code_buffer(),
880 frame_size_in_words(), _oop_map_set,
881 &_handler_table, &_inc_table,
882 compiler,
883 env()->comp_level(),
884 has_unsafe_access(),
885 SharedRuntime::is_wide_vector(max_vector_size())
886 );
888 if (log() != NULL) // Print code cache state into compiler log
889 log()->code_cache_state();
890 }
891 }
893 //------------------------------Compile----------------------------------------
894 // Compile a runtime stub
895 Compile::Compile( ciEnv* ci_env,
896 TypeFunc_generator generator,
897 address stub_function,
898 const char *stub_name,
899 int is_fancy_jump,
900 bool pass_tls,
901 bool save_arg_registers,
902 bool return_pc )
903 : Phase(Compiler),
904 _env(ci_env),
905 _log(ci_env->log()),
906 _compile_id(0),
907 _save_argument_registers(save_arg_registers),
908 _method(NULL),
909 _stub_name(stub_name),
910 _stub_function(stub_function),
911 _stub_entry_point(NULL),
912 _entry_bci(InvocationEntryBci),
913 _initial_gvn(NULL),
914 _for_igvn(NULL),
915 _warm_calls(NULL),
916 _orig_pc_slot(0),
917 _orig_pc_slot_offset_in_bytes(0),
918 _subsume_loads(true),
919 _do_escape_analysis(false),
920 _eliminate_boxing(false),
921 _failure_reason(NULL),
922 _code_buffer("Compile::Fill_buffer"),
923 _has_method_handle_invokes(false),
924 _mach_constant_base_node(NULL),
925 _node_bundling_limit(0),
926 _node_bundling_base(NULL),
927 _java_calls(0),
928 _inner_loops(0),
929 #ifndef PRODUCT
930 _trace_opto_output(TraceOptoOutput),
931 _printer(NULL),
932 #endif
933 _dead_node_list(comp_arena()),
934 _dead_node_count(0),
935 _congraph(NULL),
936 _number_of_mh_late_inlines(0),
937 _inlining_progress(false),
938 _inlining_incrementally(false),
939 _print_inlining_list(NULL),
940 _print_inlining(0) {
941 C = this;
943 #ifndef PRODUCT
944 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
945 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
946 set_print_assembly(PrintFrameConverterAssembly);
947 set_parsed_irreducible_loop(false);
948 #endif
949 CompileWrapper cw(this);
950 Init(/*AliasLevel=*/ 0);
951 init_tf((*generator)());
953 {
954 // The following is a dummy for the sake of GraphKit::gen_stub
955 Unique_Node_List for_igvn(comp_arena());
956 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
957 PhaseGVN gvn(Thread::current()->resource_area(),255);
958 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
959 gvn.transform_no_reclaim(top());
961 GraphKit kit;
962 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
963 }
965 NOT_PRODUCT( verify_graph_edges(); )
966 Code_Gen();
967 if (failing()) return;
970 // Entry point will be accessed using compile->stub_entry_point();
971 if (code_buffer() == NULL) {
972 Matcher::soft_match_failure();
973 } else {
974 if (PrintAssembly && (WizardMode || Verbose))
975 tty->print_cr("### Stub::%s", stub_name);
977 if (!failing()) {
978 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
980 // Make the NMethod
981 // For now we mark the frame as never safe for profile stackwalking
982 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
983 code_buffer(),
984 CodeOffsets::frame_never_safe,
985 // _code_offsets.value(CodeOffsets::Frame_Complete),
986 frame_size_in_words(),
987 _oop_map_set,
988 save_arg_registers);
989 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
991 _stub_entry_point = rs->entry_point();
992 }
993 }
994 }
996 //------------------------------Init-------------------------------------------
997 // Prepare for a single compilation
998 void Compile::Init(int aliaslevel) {
999 _unique = 0;
1000 _regalloc = NULL;
1002 _tf = NULL; // filled in later
1003 _top = NULL; // cached later
1004 _matcher = NULL; // filled in later
1005 _cfg = NULL; // filled in later
1007 set_24_bit_selection_and_mode(Use24BitFP, false);
1009 _node_note_array = NULL;
1010 _default_node_notes = NULL;
1012 _immutable_memory = NULL; // filled in at first inquiry
1014 // Globally visible Nodes
1015 // First set TOP to NULL to give safe behavior during creation of RootNode
1016 set_cached_top_node(NULL);
1017 set_root(new (this) RootNode());
1018 // Now that you have a Root to point to, create the real TOP
1019 set_cached_top_node( new (this) ConNode(Type::TOP) );
1020 set_recent_alloc(NULL, NULL);
1022 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1023 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1024 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1025 env()->set_dependencies(new Dependencies(env()));
1027 _fixed_slots = 0;
1028 set_has_split_ifs(false);
1029 set_has_loops(has_method() && method()->has_loops()); // first approximation
1030 set_has_stringbuilder(false);
1031 set_has_boxed_value(false);
1032 _trap_can_recompile = false; // no traps emitted yet
1033 _major_progress = true; // start out assuming good things will happen
1034 set_has_unsafe_access(false);
1035 set_max_vector_size(0);
1036 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1037 set_decompile_count(0);
1039 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1040 set_num_loop_opts(LoopOptsCount);
1041 set_do_inlining(Inline);
1042 set_max_inline_size(MaxInlineSize);
1043 set_freq_inline_size(FreqInlineSize);
1044 set_do_scheduling(OptoScheduling);
1045 set_do_count_invocations(false);
1046 set_do_method_data_update(false);
1048 if (debug_info()->recording_non_safepoints()) {
1049 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1050 (comp_arena(), 8, 0, NULL));
1051 set_default_node_notes(Node_Notes::make(this));
1052 }
1054 // // -- Initialize types before each compile --
1055 // // Update cached type information
1056 // if( _method && _method->constants() )
1057 // Type::update_loaded_types(_method, _method->constants());
1059 // Init alias_type map.
1060 if (!_do_escape_analysis && aliaslevel == 3)
1061 aliaslevel = 2; // No unique types without escape analysis
1062 _AliasLevel = aliaslevel;
1063 const int grow_ats = 16;
1064 _max_alias_types = grow_ats;
1065 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1066 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1067 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1068 {
1069 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1070 }
1071 // Initialize the first few types.
1072 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1073 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1074 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1075 _num_alias_types = AliasIdxRaw+1;
1076 // Zero out the alias type cache.
1077 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1078 // A NULL adr_type hits in the cache right away. Preload the right answer.
1079 probe_alias_cache(NULL)->_index = AliasIdxTop;
1081 _intrinsics = NULL;
1082 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1083 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1084 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1085 register_library_intrinsics();
1086 }
1088 //---------------------------init_start----------------------------------------
1089 // Install the StartNode on this compile object.
1090 void Compile::init_start(StartNode* s) {
1091 if (failing())
1092 return; // already failing
1093 assert(s == start(), "");
1094 }
1096 StartNode* Compile::start() const {
1097 assert(!failing(), "");
1098 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1099 Node* start = root()->fast_out(i);
1100 if( start->is_Start() )
1101 return start->as_Start();
1102 }
1103 ShouldNotReachHere();
1104 return NULL;
1105 }
1107 //-------------------------------immutable_memory-------------------------------------
1108 // Access immutable memory
1109 Node* Compile::immutable_memory() {
1110 if (_immutable_memory != NULL) {
1111 return _immutable_memory;
1112 }
1113 StartNode* s = start();
1114 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1115 Node *p = s->fast_out(i);
1116 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1117 _immutable_memory = p;
1118 return _immutable_memory;
1119 }
1120 }
1121 ShouldNotReachHere();
1122 return NULL;
1123 }
1125 //----------------------set_cached_top_node------------------------------------
1126 // Install the cached top node, and make sure Node::is_top works correctly.
1127 void Compile::set_cached_top_node(Node* tn) {
1128 if (tn != NULL) verify_top(tn);
1129 Node* old_top = _top;
1130 _top = tn;
1131 // Calling Node::setup_is_top allows the nodes the chance to adjust
1132 // their _out arrays.
1133 if (_top != NULL) _top->setup_is_top();
1134 if (old_top != NULL) old_top->setup_is_top();
1135 assert(_top == NULL || top()->is_top(), "");
1136 }
1138 #ifdef ASSERT
1139 uint Compile::count_live_nodes_by_graph_walk() {
1140 Unique_Node_List useful(comp_arena());
1141 // Get useful node list by walking the graph.
1142 identify_useful_nodes(useful);
1143 return useful.size();
1144 }
1146 void Compile::print_missing_nodes() {
1148 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1149 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1150 return;
1151 }
1153 // This is an expensive function. It is executed only when the user
1154 // specifies VerifyIdealNodeCount option or otherwise knows the
1155 // additional work that needs to be done to identify reachable nodes
1156 // by walking the flow graph and find the missing ones using
1157 // _dead_node_list.
1159 Unique_Node_List useful(comp_arena());
1160 // Get useful node list by walking the graph.
1161 identify_useful_nodes(useful);
1163 uint l_nodes = C->live_nodes();
1164 uint l_nodes_by_walk = useful.size();
1166 if (l_nodes != l_nodes_by_walk) {
1167 if (_log != NULL) {
1168 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1169 _log->stamp();
1170 _log->end_head();
1171 }
1172 VectorSet& useful_member_set = useful.member_set();
1173 int last_idx = l_nodes_by_walk;
1174 for (int i = 0; i < last_idx; i++) {
1175 if (useful_member_set.test(i)) {
1176 if (_dead_node_list.test(i)) {
1177 if (_log != NULL) {
1178 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1179 }
1180 if (PrintIdealNodeCount) {
1181 // Print the log message to tty
1182 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1183 useful.at(i)->dump();
1184 }
1185 }
1186 }
1187 else if (! _dead_node_list.test(i)) {
1188 if (_log != NULL) {
1189 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1190 }
1191 if (PrintIdealNodeCount) {
1192 // Print the log message to tty
1193 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1194 }
1195 }
1196 }
1197 if (_log != NULL) {
1198 _log->tail("mismatched_nodes");
1199 }
1200 }
1201 }
1202 #endif
1204 #ifndef PRODUCT
1205 void Compile::verify_top(Node* tn) const {
1206 if (tn != NULL) {
1207 assert(tn->is_Con(), "top node must be a constant");
1208 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1209 assert(tn->in(0) != NULL, "must have live top node");
1210 }
1211 }
1212 #endif
1215 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1217 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1218 guarantee(arr != NULL, "");
1219 int num_blocks = arr->length();
1220 if (grow_by < num_blocks) grow_by = num_blocks;
1221 int num_notes = grow_by * _node_notes_block_size;
1222 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1223 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1224 while (num_notes > 0) {
1225 arr->append(notes);
1226 notes += _node_notes_block_size;
1227 num_notes -= _node_notes_block_size;
1228 }
1229 assert(num_notes == 0, "exact multiple, please");
1230 }
1232 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1233 if (source == NULL || dest == NULL) return false;
1235 if (dest->is_Con())
1236 return false; // Do not push debug info onto constants.
1238 #ifdef ASSERT
1239 // Leave a bread crumb trail pointing to the original node:
1240 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1241 dest->set_debug_orig(source);
1242 }
1243 #endif
1245 if (node_note_array() == NULL)
1246 return false; // Not collecting any notes now.
1248 // This is a copy onto a pre-existing node, which may already have notes.
1249 // If both nodes have notes, do not overwrite any pre-existing notes.
1250 Node_Notes* source_notes = node_notes_at(source->_idx);
1251 if (source_notes == NULL || source_notes->is_clear()) return false;
1252 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1253 if (dest_notes == NULL || dest_notes->is_clear()) {
1254 return set_node_notes_at(dest->_idx, source_notes);
1255 }
1257 Node_Notes merged_notes = (*source_notes);
1258 // The order of operations here ensures that dest notes will win...
1259 merged_notes.update_from(dest_notes);
1260 return set_node_notes_at(dest->_idx, &merged_notes);
1261 }
1264 //--------------------------allow_range_check_smearing-------------------------
1265 // Gating condition for coalescing similar range checks.
1266 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1267 // single covering check that is at least as strong as any of them.
1268 // If the optimization succeeds, the simplified (strengthened) range check
1269 // will always succeed. If it fails, we will deopt, and then give up
1270 // on the optimization.
1271 bool Compile::allow_range_check_smearing() const {
1272 // If this method has already thrown a range-check,
1273 // assume it was because we already tried range smearing
1274 // and it failed.
1275 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1276 return !already_trapped;
1277 }
1280 //------------------------------flatten_alias_type-----------------------------
1281 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1282 int offset = tj->offset();
1283 TypePtr::PTR ptr = tj->ptr();
1285 // Known instance (scalarizable allocation) alias only with itself.
1286 bool is_known_inst = tj->isa_oopptr() != NULL &&
1287 tj->is_oopptr()->is_known_instance();
1289 // Process weird unsafe references.
1290 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1291 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1292 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1293 tj = TypeOopPtr::BOTTOM;
1294 ptr = tj->ptr();
1295 offset = tj->offset();
1296 }
1298 // Array pointers need some flattening
1299 const TypeAryPtr *ta = tj->isa_aryptr();
1300 if( ta && is_known_inst ) {
1301 if ( offset != Type::OffsetBot &&
1302 offset > arrayOopDesc::length_offset_in_bytes() ) {
1303 offset = Type::OffsetBot; // Flatten constant access into array body only
1304 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1305 }
1306 } else if( ta && _AliasLevel >= 2 ) {
1307 // For arrays indexed by constant indices, we flatten the alias
1308 // space to include all of the array body. Only the header, klass
1309 // and array length can be accessed un-aliased.
1310 if( offset != Type::OffsetBot ) {
1311 if( ta->const_oop() ) { // MethodData* or Method*
1312 offset = Type::OffsetBot; // Flatten constant access into array body
1313 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1314 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1315 // range is OK as-is.
1316 tj = ta = TypeAryPtr::RANGE;
1317 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1318 tj = TypeInstPtr::KLASS; // all klass loads look alike
1319 ta = TypeAryPtr::RANGE; // generic ignored junk
1320 ptr = TypePtr::BotPTR;
1321 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1322 tj = TypeInstPtr::MARK;
1323 ta = TypeAryPtr::RANGE; // generic ignored junk
1324 ptr = TypePtr::BotPTR;
1325 } else { // Random constant offset into array body
1326 offset = Type::OffsetBot; // Flatten constant access into array body
1327 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1328 }
1329 }
1330 // Arrays of fixed size alias with arrays of unknown size.
1331 if (ta->size() != TypeInt::POS) {
1332 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1333 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1334 }
1335 // Arrays of known objects become arrays of unknown objects.
1336 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1337 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1338 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1339 }
1340 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1341 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1342 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1343 }
1344 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1345 // cannot be distinguished by bytecode alone.
1346 if (ta->elem() == TypeInt::BOOL) {
1347 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1348 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1349 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1350 }
1351 // During the 2nd round of IterGVN, NotNull castings are removed.
1352 // Make sure the Bottom and NotNull variants alias the same.
1353 // Also, make sure exact and non-exact variants alias the same.
1354 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1355 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1356 }
1357 }
1359 // Oop pointers need some flattening
1360 const TypeInstPtr *to = tj->isa_instptr();
1361 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1362 ciInstanceKlass *k = to->klass()->as_instance_klass();
1363 if( ptr == TypePtr::Constant ) {
1364 if (to->klass() != ciEnv::current()->Class_klass() ||
1365 offset < k->size_helper() * wordSize) {
1366 // No constant oop pointers (such as Strings); they alias with
1367 // unknown strings.
1368 assert(!is_known_inst, "not scalarizable allocation");
1369 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1370 }
1371 } else if( is_known_inst ) {
1372 tj = to; // Keep NotNull and klass_is_exact for instance type
1373 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1374 // During the 2nd round of IterGVN, NotNull castings are removed.
1375 // Make sure the Bottom and NotNull variants alias the same.
1376 // Also, make sure exact and non-exact variants alias the same.
1377 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1378 }
1379 // Canonicalize the holder of this field
1380 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1381 // First handle header references such as a LoadKlassNode, even if the
1382 // object's klass is unloaded at compile time (4965979).
1383 if (!is_known_inst) { // Do it only for non-instance types
1384 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1385 }
1386 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1387 // Static fields are in the space above the normal instance
1388 // fields in the java.lang.Class instance.
1389 if (to->klass() != ciEnv::current()->Class_klass()) {
1390 to = NULL;
1391 tj = TypeOopPtr::BOTTOM;
1392 offset = tj->offset();
1393 }
1394 } else {
1395 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1396 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1397 if( is_known_inst ) {
1398 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1399 } else {
1400 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1401 }
1402 }
1403 }
1404 }
1406 // Klass pointers to object array klasses need some flattening
1407 const TypeKlassPtr *tk = tj->isa_klassptr();
1408 if( tk ) {
1409 // If we are referencing a field within a Klass, we need
1410 // to assume the worst case of an Object. Both exact and
1411 // inexact types must flatten to the same alias class so
1412 // use NotNull as the PTR.
1413 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1415 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1416 TypeKlassPtr::OBJECT->klass(),
1417 offset);
1418 }
1420 ciKlass* klass = tk->klass();
1421 if( klass->is_obj_array_klass() ) {
1422 ciKlass* k = TypeAryPtr::OOPS->klass();
1423 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1424 k = TypeInstPtr::BOTTOM->klass();
1425 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1426 }
1428 // Check for precise loads from the primary supertype array and force them
1429 // to the supertype cache alias index. Check for generic array loads from
1430 // the primary supertype array and also force them to the supertype cache
1431 // alias index. Since the same load can reach both, we need to merge
1432 // these 2 disparate memories into the same alias class. Since the
1433 // primary supertype array is read-only, there's no chance of confusion
1434 // where we bypass an array load and an array store.
1435 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1436 if (offset == Type::OffsetBot ||
1437 (offset >= primary_supers_offset &&
1438 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1439 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1440 offset = in_bytes(Klass::secondary_super_cache_offset());
1441 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1442 }
1443 }
1445 // Flatten all Raw pointers together.
1446 if (tj->base() == Type::RawPtr)
1447 tj = TypeRawPtr::BOTTOM;
1449 if (tj->base() == Type::AnyPtr)
1450 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1452 // Flatten all to bottom for now
1453 switch( _AliasLevel ) {
1454 case 0:
1455 tj = TypePtr::BOTTOM;
1456 break;
1457 case 1: // Flatten to: oop, static, field or array
1458 switch (tj->base()) {
1459 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1460 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1461 case Type::AryPtr: // do not distinguish arrays at all
1462 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1463 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1464 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1465 default: ShouldNotReachHere();
1466 }
1467 break;
1468 case 2: // No collapsing at level 2; keep all splits
1469 case 3: // No collapsing at level 3; keep all splits
1470 break;
1471 default:
1472 Unimplemented();
1473 }
1475 offset = tj->offset();
1476 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1478 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1479 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1480 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1481 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1482 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1483 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1484 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1485 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1486 assert( tj->ptr() != TypePtr::TopPTR &&
1487 tj->ptr() != TypePtr::AnyNull &&
1488 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1489 // assert( tj->ptr() != TypePtr::Constant ||
1490 // tj->base() == Type::RawPtr ||
1491 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1493 return tj;
1494 }
1496 void Compile::AliasType::Init(int i, const TypePtr* at) {
1497 _index = i;
1498 _adr_type = at;
1499 _field = NULL;
1500 _is_rewritable = true; // default
1501 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1502 if (atoop != NULL && atoop->is_known_instance()) {
1503 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1504 _general_index = Compile::current()->get_alias_index(gt);
1505 } else {
1506 _general_index = 0;
1507 }
1508 }
1510 //---------------------------------print_on------------------------------------
1511 #ifndef PRODUCT
1512 void Compile::AliasType::print_on(outputStream* st) {
1513 if (index() < 10)
1514 st->print("@ <%d> ", index());
1515 else st->print("@ <%d>", index());
1516 st->print(is_rewritable() ? " " : " RO");
1517 int offset = adr_type()->offset();
1518 if (offset == Type::OffsetBot)
1519 st->print(" +any");
1520 else st->print(" +%-3d", offset);
1521 st->print(" in ");
1522 adr_type()->dump_on(st);
1523 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1524 if (field() != NULL && tjp) {
1525 if (tjp->klass() != field()->holder() ||
1526 tjp->offset() != field()->offset_in_bytes()) {
1527 st->print(" != ");
1528 field()->print();
1529 st->print(" ***");
1530 }
1531 }
1532 }
1534 void print_alias_types() {
1535 Compile* C = Compile::current();
1536 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1537 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1538 C->alias_type(idx)->print_on(tty);
1539 tty->cr();
1540 }
1541 }
1542 #endif
1545 //----------------------------probe_alias_cache--------------------------------
1546 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1547 intptr_t key = (intptr_t) adr_type;
1548 key ^= key >> logAliasCacheSize;
1549 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1550 }
1553 //-----------------------------grow_alias_types--------------------------------
1554 void Compile::grow_alias_types() {
1555 const int old_ats = _max_alias_types; // how many before?
1556 const int new_ats = old_ats; // how many more?
1557 const int grow_ats = old_ats+new_ats; // how many now?
1558 _max_alias_types = grow_ats;
1559 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1560 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1561 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1562 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1563 }
1566 //--------------------------------find_alias_type------------------------------
1567 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1568 if (_AliasLevel == 0)
1569 return alias_type(AliasIdxBot);
1571 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1572 if (ace->_adr_type == adr_type) {
1573 return alias_type(ace->_index);
1574 }
1576 // Handle special cases.
1577 if (adr_type == NULL) return alias_type(AliasIdxTop);
1578 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1580 // Do it the slow way.
1581 const TypePtr* flat = flatten_alias_type(adr_type);
1583 #ifdef ASSERT
1584 assert(flat == flatten_alias_type(flat), "idempotent");
1585 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1586 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1587 const TypeOopPtr* foop = flat->is_oopptr();
1588 // Scalarizable allocations have exact klass always.
1589 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1590 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1591 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1592 }
1593 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1594 #endif
1596 int idx = AliasIdxTop;
1597 for (int i = 0; i < num_alias_types(); i++) {
1598 if (alias_type(i)->adr_type() == flat) {
1599 idx = i;
1600 break;
1601 }
1602 }
1604 if (idx == AliasIdxTop) {
1605 if (no_create) return NULL;
1606 // Grow the array if necessary.
1607 if (_num_alias_types == _max_alias_types) grow_alias_types();
1608 // Add a new alias type.
1609 idx = _num_alias_types++;
1610 _alias_types[idx]->Init(idx, flat);
1611 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1612 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1613 if (flat->isa_instptr()) {
1614 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1615 && flat->is_instptr()->klass() == env()->Class_klass())
1616 alias_type(idx)->set_rewritable(false);
1617 }
1618 if (flat->isa_klassptr()) {
1619 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1620 alias_type(idx)->set_rewritable(false);
1621 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1622 alias_type(idx)->set_rewritable(false);
1623 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1624 alias_type(idx)->set_rewritable(false);
1625 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1626 alias_type(idx)->set_rewritable(false);
1627 }
1628 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1629 // but the base pointer type is not distinctive enough to identify
1630 // references into JavaThread.)
1632 // Check for final fields.
1633 const TypeInstPtr* tinst = flat->isa_instptr();
1634 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1635 ciField* field;
1636 if (tinst->const_oop() != NULL &&
1637 tinst->klass() == ciEnv::current()->Class_klass() &&
1638 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1639 // static field
1640 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1641 field = k->get_field_by_offset(tinst->offset(), true);
1642 } else {
1643 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1644 field = k->get_field_by_offset(tinst->offset(), false);
1645 }
1646 assert(field == NULL ||
1647 original_field == NULL ||
1648 (field->holder() == original_field->holder() &&
1649 field->offset() == original_field->offset() &&
1650 field->is_static() == original_field->is_static()), "wrong field?");
1651 // Set field() and is_rewritable() attributes.
1652 if (field != NULL) alias_type(idx)->set_field(field);
1653 }
1654 }
1656 // Fill the cache for next time.
1657 ace->_adr_type = adr_type;
1658 ace->_index = idx;
1659 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1661 // Might as well try to fill the cache for the flattened version, too.
1662 AliasCacheEntry* face = probe_alias_cache(flat);
1663 if (face->_adr_type == NULL) {
1664 face->_adr_type = flat;
1665 face->_index = idx;
1666 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1667 }
1669 return alias_type(idx);
1670 }
1673 Compile::AliasType* Compile::alias_type(ciField* field) {
1674 const TypeOopPtr* t;
1675 if (field->is_static())
1676 t = TypeInstPtr::make(field->holder()->java_mirror());
1677 else
1678 t = TypeOopPtr::make_from_klass_raw(field->holder());
1679 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1680 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1681 return atp;
1682 }
1685 //------------------------------have_alias_type--------------------------------
1686 bool Compile::have_alias_type(const TypePtr* adr_type) {
1687 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1688 if (ace->_adr_type == adr_type) {
1689 return true;
1690 }
1692 // Handle special cases.
1693 if (adr_type == NULL) return true;
1694 if (adr_type == TypePtr::BOTTOM) return true;
1696 return find_alias_type(adr_type, true, NULL) != NULL;
1697 }
1699 //-----------------------------must_alias--------------------------------------
1700 // True if all values of the given address type are in the given alias category.
1701 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1702 if (alias_idx == AliasIdxBot) return true; // the universal category
1703 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1704 if (alias_idx == AliasIdxTop) return false; // the empty category
1705 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1707 // the only remaining possible overlap is identity
1708 int adr_idx = get_alias_index(adr_type);
1709 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1710 assert(adr_idx == alias_idx ||
1711 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1712 && adr_type != TypeOopPtr::BOTTOM),
1713 "should not be testing for overlap with an unsafe pointer");
1714 return adr_idx == alias_idx;
1715 }
1717 //------------------------------can_alias--------------------------------------
1718 // True if any values of the given address type are in the given alias category.
1719 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1720 if (alias_idx == AliasIdxTop) return false; // the empty category
1721 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1722 if (alias_idx == AliasIdxBot) return true; // the universal category
1723 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1725 // the only remaining possible overlap is identity
1726 int adr_idx = get_alias_index(adr_type);
1727 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1728 return adr_idx == alias_idx;
1729 }
1733 //---------------------------pop_warm_call-------------------------------------
1734 WarmCallInfo* Compile::pop_warm_call() {
1735 WarmCallInfo* wci = _warm_calls;
1736 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1737 return wci;
1738 }
1740 //----------------------------Inline_Warm--------------------------------------
1741 int Compile::Inline_Warm() {
1742 // If there is room, try to inline some more warm call sites.
1743 // %%% Do a graph index compaction pass when we think we're out of space?
1744 if (!InlineWarmCalls) return 0;
1746 int calls_made_hot = 0;
1747 int room_to_grow = NodeCountInliningCutoff - unique();
1748 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1749 int amount_grown = 0;
1750 WarmCallInfo* call;
1751 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1752 int est_size = (int)call->size();
1753 if (est_size > (room_to_grow - amount_grown)) {
1754 // This one won't fit anyway. Get rid of it.
1755 call->make_cold();
1756 continue;
1757 }
1758 call->make_hot();
1759 calls_made_hot++;
1760 amount_grown += est_size;
1761 amount_to_grow -= est_size;
1762 }
1764 if (calls_made_hot > 0) set_major_progress();
1765 return calls_made_hot;
1766 }
1769 //----------------------------Finish_Warm--------------------------------------
1770 void Compile::Finish_Warm() {
1771 if (!InlineWarmCalls) return;
1772 if (failing()) return;
1773 if (warm_calls() == NULL) return;
1775 // Clean up loose ends, if we are out of space for inlining.
1776 WarmCallInfo* call;
1777 while ((call = pop_warm_call()) != NULL) {
1778 call->make_cold();
1779 }
1780 }
1782 //---------------------cleanup_loop_predicates-----------------------
1783 // Remove the opaque nodes that protect the predicates so that all unused
1784 // checks and uncommon_traps will be eliminated from the ideal graph
1785 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1786 if (predicate_count()==0) return;
1787 for (int i = predicate_count(); i > 0; i--) {
1788 Node * n = predicate_opaque1_node(i-1);
1789 assert(n->Opcode() == Op_Opaque1, "must be");
1790 igvn.replace_node(n, n->in(1));
1791 }
1792 assert(predicate_count()==0, "should be clean!");
1793 }
1795 // StringOpts and late inlining of string methods
1796 void Compile::inline_string_calls(bool parse_time) {
1797 {
1798 // remove useless nodes to make the usage analysis simpler
1799 ResourceMark rm;
1800 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1801 }
1803 {
1804 ResourceMark rm;
1805 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1806 PhaseStringOpts pso(initial_gvn(), for_igvn());
1807 print_method(PHASE_AFTER_STRINGOPTS, 3);
1808 }
1810 // now inline anything that we skipped the first time around
1811 if (!parse_time) {
1812 _late_inlines_pos = _late_inlines.length();
1813 }
1815 while (_string_late_inlines.length() > 0) {
1816 CallGenerator* cg = _string_late_inlines.pop();
1817 cg->do_late_inline();
1818 if (failing()) return;
1819 }
1820 _string_late_inlines.trunc_to(0);
1821 }
1823 // Late inlining of boxing methods
1824 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1825 if (_boxing_late_inlines.length() > 0) {
1826 assert(has_boxed_value(), "inconsistent");
1828 PhaseGVN* gvn = initial_gvn();
1829 set_inlining_incrementally(true);
1831 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1832 for_igvn()->clear();
1833 gvn->replace_with(&igvn);
1835 while (_boxing_late_inlines.length() > 0) {
1836 CallGenerator* cg = _boxing_late_inlines.pop();
1837 cg->do_late_inline();
1838 if (failing()) return;
1839 }
1840 _boxing_late_inlines.trunc_to(0);
1842 {
1843 ResourceMark rm;
1844 PhaseRemoveUseless pru(gvn, for_igvn());
1845 }
1847 igvn = PhaseIterGVN(gvn);
1848 igvn.optimize();
1850 set_inlining_progress(false);
1851 set_inlining_incrementally(false);
1852 }
1853 }
1855 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1856 assert(IncrementalInline, "incremental inlining should be on");
1857 PhaseGVN* gvn = initial_gvn();
1859 set_inlining_progress(false);
1860 for_igvn()->clear();
1861 gvn->replace_with(&igvn);
1863 int i = 0;
1865 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1866 CallGenerator* cg = _late_inlines.at(i);
1867 _late_inlines_pos = i+1;
1868 cg->do_late_inline();
1869 if (failing()) return;
1870 }
1871 int j = 0;
1872 for (; i < _late_inlines.length(); i++, j++) {
1873 _late_inlines.at_put(j, _late_inlines.at(i));
1874 }
1875 _late_inlines.trunc_to(j);
1877 {
1878 ResourceMark rm;
1879 PhaseRemoveUseless pru(gvn, for_igvn());
1880 }
1882 igvn = PhaseIterGVN(gvn);
1883 }
1885 // Perform incremental inlining until bound on number of live nodes is reached
1886 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1887 PhaseGVN* gvn = initial_gvn();
1889 set_inlining_incrementally(true);
1890 set_inlining_progress(true);
1891 uint low_live_nodes = 0;
1893 while(inlining_progress() && _late_inlines.length() > 0) {
1895 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1896 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1897 // PhaseIdealLoop is expensive so we only try it once we are
1898 // out of loop and we only try it again if the previous helped
1899 // got the number of nodes down significantly
1900 PhaseIdealLoop ideal_loop( igvn, false, true );
1901 if (failing()) return;
1902 low_live_nodes = live_nodes();
1903 _major_progress = true;
1904 }
1906 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1907 break;
1908 }
1909 }
1911 inline_incrementally_one(igvn);
1913 if (failing()) return;
1915 igvn.optimize();
1917 if (failing()) return;
1918 }
1920 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1922 if (_string_late_inlines.length() > 0) {
1923 assert(has_stringbuilder(), "inconsistent");
1924 for_igvn()->clear();
1925 initial_gvn()->replace_with(&igvn);
1927 inline_string_calls(false);
1929 if (failing()) return;
1931 {
1932 ResourceMark rm;
1933 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1934 }
1936 igvn = PhaseIterGVN(gvn);
1938 igvn.optimize();
1939 }
1941 set_inlining_incrementally(false);
1942 }
1945 //------------------------------Optimize---------------------------------------
1946 // Given a graph, optimize it.
1947 void Compile::Optimize() {
1948 TracePhase t1("optimizer", &_t_optimizer, true);
1950 #ifndef PRODUCT
1951 if (env()->break_at_compile()) {
1952 BREAKPOINT;
1953 }
1955 #endif
1957 ResourceMark rm;
1958 int loop_opts_cnt;
1960 NOT_PRODUCT( verify_graph_edges(); )
1962 print_method(PHASE_AFTER_PARSING);
1964 {
1965 // Iterative Global Value Numbering, including ideal transforms
1966 // Initialize IterGVN with types and values from parse-time GVN
1967 PhaseIterGVN igvn(initial_gvn());
1968 {
1969 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1970 igvn.optimize();
1971 }
1973 print_method(PHASE_ITER_GVN1, 2);
1975 if (failing()) return;
1977 {
1978 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
1979 inline_incrementally(igvn);
1980 }
1982 print_method(PHASE_INCREMENTAL_INLINE, 2);
1984 if (failing()) return;
1986 if (eliminate_boxing()) {
1987 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
1988 // Inline valueOf() methods now.
1989 inline_boxing_calls(igvn);
1991 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
1993 if (failing()) return;
1994 }
1996 // No more new expensive nodes will be added to the list from here
1997 // so keep only the actual candidates for optimizations.
1998 cleanup_expensive_nodes(igvn);
2000 // Perform escape analysis
2001 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2002 if (has_loops()) {
2003 // Cleanup graph (remove dead nodes).
2004 TracePhase t2("idealLoop", &_t_idealLoop, true);
2005 PhaseIdealLoop ideal_loop( igvn, false, true );
2006 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2007 if (failing()) return;
2008 }
2009 ConnectionGraph::do_analysis(this, &igvn);
2011 if (failing()) return;
2013 // Optimize out fields loads from scalar replaceable allocations.
2014 igvn.optimize();
2015 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2017 if (failing()) return;
2019 if (congraph() != NULL && macro_count() > 0) {
2020 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2021 PhaseMacroExpand mexp(igvn);
2022 mexp.eliminate_macro_nodes();
2023 igvn.set_delay_transform(false);
2025 igvn.optimize();
2026 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2028 if (failing()) return;
2029 }
2030 }
2032 // Loop transforms on the ideal graph. Range Check Elimination,
2033 // peeling, unrolling, etc.
2035 // Set loop opts counter
2036 loop_opts_cnt = num_loop_opts();
2037 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2038 {
2039 TracePhase t2("idealLoop", &_t_idealLoop, true);
2040 PhaseIdealLoop ideal_loop( igvn, true );
2041 loop_opts_cnt--;
2042 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2043 if (failing()) return;
2044 }
2045 // Loop opts pass if partial peeling occurred in previous pass
2046 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2047 TracePhase t3("idealLoop", &_t_idealLoop, true);
2048 PhaseIdealLoop ideal_loop( igvn, false );
2049 loop_opts_cnt--;
2050 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2051 if (failing()) return;
2052 }
2053 // Loop opts pass for loop-unrolling before CCP
2054 if(major_progress() && (loop_opts_cnt > 0)) {
2055 TracePhase t4("idealLoop", &_t_idealLoop, true);
2056 PhaseIdealLoop ideal_loop( igvn, false );
2057 loop_opts_cnt--;
2058 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2059 }
2060 if (!failing()) {
2061 // Verify that last round of loop opts produced a valid graph
2062 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2063 PhaseIdealLoop::verify(igvn);
2064 }
2065 }
2066 if (failing()) return;
2068 // Conditional Constant Propagation;
2069 PhaseCCP ccp( &igvn );
2070 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2071 {
2072 TracePhase t2("ccp", &_t_ccp, true);
2073 ccp.do_transform();
2074 }
2075 print_method(PHASE_CPP1, 2);
2077 assert( true, "Break here to ccp.dump_old2new_map()");
2079 // Iterative Global Value Numbering, including ideal transforms
2080 {
2081 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2082 igvn = ccp;
2083 igvn.optimize();
2084 }
2086 print_method(PHASE_ITER_GVN2, 2);
2088 if (failing()) return;
2090 // Loop transforms on the ideal graph. Range Check Elimination,
2091 // peeling, unrolling, etc.
2092 if(loop_opts_cnt > 0) {
2093 debug_only( int cnt = 0; );
2094 while(major_progress() && (loop_opts_cnt > 0)) {
2095 TracePhase t2("idealLoop", &_t_idealLoop, true);
2096 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2097 PhaseIdealLoop ideal_loop( igvn, true);
2098 loop_opts_cnt--;
2099 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2100 if (failing()) return;
2101 }
2102 }
2104 {
2105 // Verify that all previous optimizations produced a valid graph
2106 // at least to this point, even if no loop optimizations were done.
2107 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2108 PhaseIdealLoop::verify(igvn);
2109 }
2111 {
2112 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2113 PhaseMacroExpand mex(igvn);
2114 if (mex.expand_macro_nodes()) {
2115 assert(failing(), "must bail out w/ explicit message");
2116 return;
2117 }
2118 }
2120 } // (End scope of igvn; run destructor if necessary for asserts.)
2122 dump_inlining();
2123 // A method with only infinite loops has no edges entering loops from root
2124 {
2125 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2126 if (final_graph_reshaping()) {
2127 assert(failing(), "must bail out w/ explicit message");
2128 return;
2129 }
2130 }
2132 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2133 }
2136 //------------------------------Code_Gen---------------------------------------
2137 // Given a graph, generate code for it
2138 void Compile::Code_Gen() {
2139 if (failing()) return;
2141 // Perform instruction selection. You might think we could reclaim Matcher
2142 // memory PDQ, but actually the Matcher is used in generating spill code.
2143 // Internals of the Matcher (including some VectorSets) must remain live
2144 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2145 // set a bit in reclaimed memory.
2147 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2148 // nodes. Mapping is only valid at the root of each matched subtree.
2149 NOT_PRODUCT( verify_graph_edges(); )
2151 Node_List proj_list;
2152 Matcher m(proj_list);
2153 _matcher = &m;
2154 {
2155 TracePhase t2("matcher", &_t_matcher, true);
2156 m.match();
2157 }
2158 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2159 // nodes. Mapping is only valid at the root of each matched subtree.
2160 NOT_PRODUCT( verify_graph_edges(); )
2162 // If you have too many nodes, or if matching has failed, bail out
2163 check_node_count(0, "out of nodes matching instructions");
2164 if (failing()) return;
2166 // Build a proper-looking CFG
2167 PhaseCFG cfg(node_arena(), root(), m);
2168 _cfg = &cfg;
2169 {
2170 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2171 cfg.Dominators();
2172 if (failing()) return;
2174 NOT_PRODUCT( verify_graph_edges(); )
2176 cfg.Estimate_Block_Frequency();
2177 cfg.GlobalCodeMotion(m,unique(),proj_list);
2178 if (failing()) return;
2180 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2182 NOT_PRODUCT( verify_graph_edges(); )
2184 debug_only( cfg.verify(); )
2185 }
2186 NOT_PRODUCT( verify_graph_edges(); )
2188 PhaseChaitin regalloc(unique(), cfg, m);
2189 _regalloc = ®alloc;
2190 {
2191 TracePhase t2("regalloc", &_t_registerAllocation, true);
2192 // Perform register allocation. After Chaitin, use-def chains are
2193 // no longer accurate (at spill code) and so must be ignored.
2194 // Node->LRG->reg mappings are still accurate.
2195 _regalloc->Register_Allocate();
2197 // Bail out if the allocator builds too many nodes
2198 if (failing()) {
2199 return;
2200 }
2201 }
2203 // Prior to register allocation we kept empty basic blocks in case the
2204 // the allocator needed a place to spill. After register allocation we
2205 // are not adding any new instructions. If any basic block is empty, we
2206 // can now safely remove it.
2207 {
2208 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2209 cfg.remove_empty();
2210 if (do_freq_based_layout()) {
2211 PhaseBlockLayout layout(cfg);
2212 } else {
2213 cfg.set_loop_alignment();
2214 }
2215 cfg.fixup_flow();
2216 }
2218 // Apply peephole optimizations
2219 if( OptoPeephole ) {
2220 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2221 PhasePeephole peep( _regalloc, cfg);
2222 peep.do_transform();
2223 }
2225 // Convert Nodes to instruction bits in a buffer
2226 {
2227 // %%%% workspace merge brought two timers together for one job
2228 TracePhase t2a("output", &_t_output, true);
2229 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2230 Output();
2231 }
2233 print_method(PHASE_FINAL_CODE);
2235 // He's dead, Jim.
2236 _cfg = (PhaseCFG*)0xdeadbeef;
2237 _regalloc = (PhaseChaitin*)0xdeadbeef;
2238 }
2241 //------------------------------dump_asm---------------------------------------
2242 // Dump formatted assembly
2243 #ifndef PRODUCT
2244 void Compile::dump_asm(int *pcs, uint pc_limit) {
2245 bool cut_short = false;
2246 tty->print_cr("#");
2247 tty->print("# "); _tf->dump(); tty->cr();
2248 tty->print_cr("#");
2250 // For all blocks
2251 int pc = 0x0; // Program counter
2252 char starts_bundle = ' ';
2253 _regalloc->dump_frame();
2255 Node *n = NULL;
2256 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
2257 if (VMThread::should_terminate()) { cut_short = true; break; }
2258 Block *b = _cfg->_blocks[i];
2259 if (b->is_connector() && !Verbose) continue;
2260 n = b->_nodes[0];
2261 if (pcs && n->_idx < pc_limit)
2262 tty->print("%3.3x ", pcs[n->_idx]);
2263 else
2264 tty->print(" ");
2265 b->dump_head( &_cfg->_bbs );
2266 if (b->is_connector()) {
2267 tty->print_cr(" # Empty connector block");
2268 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2269 tty->print_cr(" # Block is sole successor of call");
2270 }
2272 // For all instructions
2273 Node *delay = NULL;
2274 for( uint j = 0; j<b->_nodes.size(); j++ ) {
2275 if (VMThread::should_terminate()) { cut_short = true; break; }
2276 n = b->_nodes[j];
2277 if (valid_bundle_info(n)) {
2278 Bundle *bundle = node_bundling(n);
2279 if (bundle->used_in_unconditional_delay()) {
2280 delay = n;
2281 continue;
2282 }
2283 if (bundle->starts_bundle())
2284 starts_bundle = '+';
2285 }
2287 if (WizardMode) n->dump();
2289 if( !n->is_Region() && // Dont print in the Assembly
2290 !n->is_Phi() && // a few noisely useless nodes
2291 !n->is_Proj() &&
2292 !n->is_MachTemp() &&
2293 !n->is_SafePointScalarObject() &&
2294 !n->is_Catch() && // Would be nice to print exception table targets
2295 !n->is_MergeMem() && // Not very interesting
2296 !n->is_top() && // Debug info table constants
2297 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2298 ) {
2299 if (pcs && n->_idx < pc_limit)
2300 tty->print("%3.3x", pcs[n->_idx]);
2301 else
2302 tty->print(" ");
2303 tty->print(" %c ", starts_bundle);
2304 starts_bundle = ' ';
2305 tty->print("\t");
2306 n->format(_regalloc, tty);
2307 tty->cr();
2308 }
2310 // If we have an instruction with a delay slot, and have seen a delay,
2311 // then back up and print it
2312 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2313 assert(delay != NULL, "no unconditional delay instruction");
2314 if (WizardMode) delay->dump();
2316 if (node_bundling(delay)->starts_bundle())
2317 starts_bundle = '+';
2318 if (pcs && n->_idx < pc_limit)
2319 tty->print("%3.3x", pcs[n->_idx]);
2320 else
2321 tty->print(" ");
2322 tty->print(" %c ", starts_bundle);
2323 starts_bundle = ' ';
2324 tty->print("\t");
2325 delay->format(_regalloc, tty);
2326 tty->print_cr("");
2327 delay = NULL;
2328 }
2330 // Dump the exception table as well
2331 if( n->is_Catch() && (Verbose || WizardMode) ) {
2332 // Print the exception table for this offset
2333 _handler_table.print_subtable_for(pc);
2334 }
2335 }
2337 if (pcs && n->_idx < pc_limit)
2338 tty->print_cr("%3.3x", pcs[n->_idx]);
2339 else
2340 tty->print_cr("");
2342 assert(cut_short || delay == NULL, "no unconditional delay branch");
2344 } // End of per-block dump
2345 tty->print_cr("");
2347 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2348 }
2349 #endif
2351 //------------------------------Final_Reshape_Counts---------------------------
2352 // This class defines counters to help identify when a method
2353 // may/must be executed using hardware with only 24-bit precision.
2354 struct Final_Reshape_Counts : public StackObj {
2355 int _call_count; // count non-inlined 'common' calls
2356 int _float_count; // count float ops requiring 24-bit precision
2357 int _double_count; // count double ops requiring more precision
2358 int _java_call_count; // count non-inlined 'java' calls
2359 int _inner_loop_count; // count loops which need alignment
2360 VectorSet _visited; // Visitation flags
2361 Node_List _tests; // Set of IfNodes & PCTableNodes
2363 Final_Reshape_Counts() :
2364 _call_count(0), _float_count(0), _double_count(0),
2365 _java_call_count(0), _inner_loop_count(0),
2366 _visited( Thread::current()->resource_area() ) { }
2368 void inc_call_count () { _call_count ++; }
2369 void inc_float_count () { _float_count ++; }
2370 void inc_double_count() { _double_count++; }
2371 void inc_java_call_count() { _java_call_count++; }
2372 void inc_inner_loop_count() { _inner_loop_count++; }
2374 int get_call_count () const { return _call_count ; }
2375 int get_float_count () const { return _float_count ; }
2376 int get_double_count() const { return _double_count; }
2377 int get_java_call_count() const { return _java_call_count; }
2378 int get_inner_loop_count() const { return _inner_loop_count; }
2379 };
2381 #ifdef ASSERT
2382 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2383 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2384 // Make sure the offset goes inside the instance layout.
2385 return k->contains_field_offset(tp->offset());
2386 // Note that OffsetBot and OffsetTop are very negative.
2387 }
2388 #endif
2390 // Eliminate trivially redundant StoreCMs and accumulate their
2391 // precedence edges.
2392 void Compile::eliminate_redundant_card_marks(Node* n) {
2393 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2394 if (n->in(MemNode::Address)->outcnt() > 1) {
2395 // There are multiple users of the same address so it might be
2396 // possible to eliminate some of the StoreCMs
2397 Node* mem = n->in(MemNode::Memory);
2398 Node* adr = n->in(MemNode::Address);
2399 Node* val = n->in(MemNode::ValueIn);
2400 Node* prev = n;
2401 bool done = false;
2402 // Walk the chain of StoreCMs eliminating ones that match. As
2403 // long as it's a chain of single users then the optimization is
2404 // safe. Eliminating partially redundant StoreCMs would require
2405 // cloning copies down the other paths.
2406 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2407 if (adr == mem->in(MemNode::Address) &&
2408 val == mem->in(MemNode::ValueIn)) {
2409 // redundant StoreCM
2410 if (mem->req() > MemNode::OopStore) {
2411 // Hasn't been processed by this code yet.
2412 n->add_prec(mem->in(MemNode::OopStore));
2413 } else {
2414 // Already converted to precedence edge
2415 for (uint i = mem->req(); i < mem->len(); i++) {
2416 // Accumulate any precedence edges
2417 if (mem->in(i) != NULL) {
2418 n->add_prec(mem->in(i));
2419 }
2420 }
2421 // Everything above this point has been processed.
2422 done = true;
2423 }
2424 // Eliminate the previous StoreCM
2425 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2426 assert(mem->outcnt() == 0, "should be dead");
2427 mem->disconnect_inputs(NULL, this);
2428 } else {
2429 prev = mem;
2430 }
2431 mem = prev->in(MemNode::Memory);
2432 }
2433 }
2434 }
2436 //------------------------------final_graph_reshaping_impl----------------------
2437 // Implement items 1-5 from final_graph_reshaping below.
2438 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2440 if ( n->outcnt() == 0 ) return; // dead node
2441 uint nop = n->Opcode();
2443 // Check for 2-input instruction with "last use" on right input.
2444 // Swap to left input. Implements item (2).
2445 if( n->req() == 3 && // two-input instruction
2446 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2447 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2448 n->in(2)->outcnt() == 1 &&// right use IS a last use
2449 !n->in(2)->is_Con() ) { // right use is not a constant
2450 // Check for commutative opcode
2451 switch( nop ) {
2452 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2453 case Op_MaxI: case Op_MinI:
2454 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2455 case Op_AndL: case Op_XorL: case Op_OrL:
2456 case Op_AndI: case Op_XorI: case Op_OrI: {
2457 // Move "last use" input to left by swapping inputs
2458 n->swap_edges(1, 2);
2459 break;
2460 }
2461 default:
2462 break;
2463 }
2464 }
2466 #ifdef ASSERT
2467 if( n->is_Mem() ) {
2468 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2469 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2470 // oop will be recorded in oop map if load crosses safepoint
2471 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2472 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2473 "raw memory operations should have control edge");
2474 }
2475 #endif
2476 // Count FPU ops and common calls, implements item (3)
2477 switch( nop ) {
2478 // Count all float operations that may use FPU
2479 case Op_AddF:
2480 case Op_SubF:
2481 case Op_MulF:
2482 case Op_DivF:
2483 case Op_NegF:
2484 case Op_ModF:
2485 case Op_ConvI2F:
2486 case Op_ConF:
2487 case Op_CmpF:
2488 case Op_CmpF3:
2489 // case Op_ConvL2F: // longs are split into 32-bit halves
2490 frc.inc_float_count();
2491 break;
2493 case Op_ConvF2D:
2494 case Op_ConvD2F:
2495 frc.inc_float_count();
2496 frc.inc_double_count();
2497 break;
2499 // Count all double operations that may use FPU
2500 case Op_AddD:
2501 case Op_SubD:
2502 case Op_MulD:
2503 case Op_DivD:
2504 case Op_NegD:
2505 case Op_ModD:
2506 case Op_ConvI2D:
2507 case Op_ConvD2I:
2508 // case Op_ConvL2D: // handled by leaf call
2509 // case Op_ConvD2L: // handled by leaf call
2510 case Op_ConD:
2511 case Op_CmpD:
2512 case Op_CmpD3:
2513 frc.inc_double_count();
2514 break;
2515 case Op_Opaque1: // Remove Opaque Nodes before matching
2516 case Op_Opaque2: // Remove Opaque Nodes before matching
2517 n->subsume_by(n->in(1), this);
2518 break;
2519 case Op_CallStaticJava:
2520 case Op_CallJava:
2521 case Op_CallDynamicJava:
2522 frc.inc_java_call_count(); // Count java call site;
2523 case Op_CallRuntime:
2524 case Op_CallLeaf:
2525 case Op_CallLeafNoFP: {
2526 assert( n->is_Call(), "" );
2527 CallNode *call = n->as_Call();
2528 // Count call sites where the FP mode bit would have to be flipped.
2529 // Do not count uncommon runtime calls:
2530 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2531 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2532 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2533 frc.inc_call_count(); // Count the call site
2534 } else { // See if uncommon argument is shared
2535 Node *n = call->in(TypeFunc::Parms);
2536 int nop = n->Opcode();
2537 // Clone shared simple arguments to uncommon calls, item (1).
2538 if( n->outcnt() > 1 &&
2539 !n->is_Proj() &&
2540 nop != Op_CreateEx &&
2541 nop != Op_CheckCastPP &&
2542 nop != Op_DecodeN &&
2543 nop != Op_DecodeNKlass &&
2544 !n->is_Mem() ) {
2545 Node *x = n->clone();
2546 call->set_req( TypeFunc::Parms, x );
2547 }
2548 }
2549 break;
2550 }
2552 case Op_StoreD:
2553 case Op_LoadD:
2554 case Op_LoadD_unaligned:
2555 frc.inc_double_count();
2556 goto handle_mem;
2557 case Op_StoreF:
2558 case Op_LoadF:
2559 frc.inc_float_count();
2560 goto handle_mem;
2562 case Op_StoreCM:
2563 {
2564 // Convert OopStore dependence into precedence edge
2565 Node* prec = n->in(MemNode::OopStore);
2566 n->del_req(MemNode::OopStore);
2567 n->add_prec(prec);
2568 eliminate_redundant_card_marks(n);
2569 }
2571 // fall through
2573 case Op_StoreB:
2574 case Op_StoreC:
2575 case Op_StorePConditional:
2576 case Op_StoreI:
2577 case Op_StoreL:
2578 case Op_StoreIConditional:
2579 case Op_StoreLConditional:
2580 case Op_CompareAndSwapI:
2581 case Op_CompareAndSwapL:
2582 case Op_CompareAndSwapP:
2583 case Op_CompareAndSwapN:
2584 case Op_GetAndAddI:
2585 case Op_GetAndAddL:
2586 case Op_GetAndSetI:
2587 case Op_GetAndSetL:
2588 case Op_GetAndSetP:
2589 case Op_GetAndSetN:
2590 case Op_StoreP:
2591 case Op_StoreN:
2592 case Op_StoreNKlass:
2593 case Op_LoadB:
2594 case Op_LoadUB:
2595 case Op_LoadUS:
2596 case Op_LoadI:
2597 case Op_LoadKlass:
2598 case Op_LoadNKlass:
2599 case Op_LoadL:
2600 case Op_LoadL_unaligned:
2601 case Op_LoadPLocked:
2602 case Op_LoadP:
2603 case Op_LoadN:
2604 case Op_LoadRange:
2605 case Op_LoadS: {
2606 handle_mem:
2607 #ifdef ASSERT
2608 if( VerifyOptoOopOffsets ) {
2609 assert( n->is_Mem(), "" );
2610 MemNode *mem = (MemNode*)n;
2611 // Check to see if address types have grounded out somehow.
2612 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2613 assert( !tp || oop_offset_is_sane(tp), "" );
2614 }
2615 #endif
2616 break;
2617 }
2619 case Op_AddP: { // Assert sane base pointers
2620 Node *addp = n->in(AddPNode::Address);
2621 assert( !addp->is_AddP() ||
2622 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2623 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2624 "Base pointers must match" );
2625 #ifdef _LP64
2626 if ((UseCompressedOops || UseCompressedKlassPointers) &&
2627 addp->Opcode() == Op_ConP &&
2628 addp == n->in(AddPNode::Base) &&
2629 n->in(AddPNode::Offset)->is_Con()) {
2630 // Use addressing with narrow klass to load with offset on x86.
2631 // On sparc loading 32-bits constant and decoding it have less
2632 // instructions (4) then load 64-bits constant (7).
2633 // Do this transformation here since IGVN will convert ConN back to ConP.
2634 const Type* t = addp->bottom_type();
2635 if (t->isa_oopptr() || t->isa_klassptr()) {
2636 Node* nn = NULL;
2638 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2640 // Look for existing ConN node of the same exact type.
2641 Node* r = root();
2642 uint cnt = r->outcnt();
2643 for (uint i = 0; i < cnt; i++) {
2644 Node* m = r->raw_out(i);
2645 if (m!= NULL && m->Opcode() == op &&
2646 m->bottom_type()->make_ptr() == t) {
2647 nn = m;
2648 break;
2649 }
2650 }
2651 if (nn != NULL) {
2652 // Decode a narrow oop to match address
2653 // [R12 + narrow_oop_reg<<3 + offset]
2654 if (t->isa_oopptr()) {
2655 nn = new (this) DecodeNNode(nn, t);
2656 } else {
2657 nn = new (this) DecodeNKlassNode(nn, t);
2658 }
2659 n->set_req(AddPNode::Base, nn);
2660 n->set_req(AddPNode::Address, nn);
2661 if (addp->outcnt() == 0) {
2662 addp->disconnect_inputs(NULL, this);
2663 }
2664 }
2665 }
2666 }
2667 #endif
2668 break;
2669 }
2671 #ifdef _LP64
2672 case Op_CastPP:
2673 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2674 Node* in1 = n->in(1);
2675 const Type* t = n->bottom_type();
2676 Node* new_in1 = in1->clone();
2677 new_in1->as_DecodeN()->set_type(t);
2679 if (!Matcher::narrow_oop_use_complex_address()) {
2680 //
2681 // x86, ARM and friends can handle 2 adds in addressing mode
2682 // and Matcher can fold a DecodeN node into address by using
2683 // a narrow oop directly and do implicit NULL check in address:
2684 //
2685 // [R12 + narrow_oop_reg<<3 + offset]
2686 // NullCheck narrow_oop_reg
2687 //
2688 // On other platforms (Sparc) we have to keep new DecodeN node and
2689 // use it to do implicit NULL check in address:
2690 //
2691 // decode_not_null narrow_oop_reg, base_reg
2692 // [base_reg + offset]
2693 // NullCheck base_reg
2694 //
2695 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2696 // to keep the information to which NULL check the new DecodeN node
2697 // corresponds to use it as value in implicit_null_check().
2698 //
2699 new_in1->set_req(0, n->in(0));
2700 }
2702 n->subsume_by(new_in1, this);
2703 if (in1->outcnt() == 0) {
2704 in1->disconnect_inputs(NULL, this);
2705 }
2706 }
2707 break;
2709 case Op_CmpP:
2710 // Do this transformation here to preserve CmpPNode::sub() and
2711 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2712 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2713 Node* in1 = n->in(1);
2714 Node* in2 = n->in(2);
2715 if (!in1->is_DecodeNarrowPtr()) {
2716 in2 = in1;
2717 in1 = n->in(2);
2718 }
2719 assert(in1->is_DecodeNarrowPtr(), "sanity");
2721 Node* new_in2 = NULL;
2722 if (in2->is_DecodeNarrowPtr()) {
2723 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2724 new_in2 = in2->in(1);
2725 } else if (in2->Opcode() == Op_ConP) {
2726 const Type* t = in2->bottom_type();
2727 if (t == TypePtr::NULL_PTR) {
2728 assert(in1->is_DecodeN(), "compare klass to null?");
2729 // Don't convert CmpP null check into CmpN if compressed
2730 // oops implicit null check is not generated.
2731 // This will allow to generate normal oop implicit null check.
2732 if (Matcher::gen_narrow_oop_implicit_null_checks())
2733 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2734 //
2735 // This transformation together with CastPP transformation above
2736 // will generated code for implicit NULL checks for compressed oops.
2737 //
2738 // The original code after Optimize()
2739 //
2740 // LoadN memory, narrow_oop_reg
2741 // decode narrow_oop_reg, base_reg
2742 // CmpP base_reg, NULL
2743 // CastPP base_reg // NotNull
2744 // Load [base_reg + offset], val_reg
2745 //
2746 // after these transformations will be
2747 //
2748 // LoadN memory, narrow_oop_reg
2749 // CmpN narrow_oop_reg, NULL
2750 // decode_not_null narrow_oop_reg, base_reg
2751 // Load [base_reg + offset], val_reg
2752 //
2753 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2754 // since narrow oops can be used in debug info now (see the code in
2755 // final_graph_reshaping_walk()).
2756 //
2757 // At the end the code will be matched to
2758 // on x86:
2759 //
2760 // Load_narrow_oop memory, narrow_oop_reg
2761 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2762 // NullCheck narrow_oop_reg
2763 //
2764 // and on sparc:
2765 //
2766 // Load_narrow_oop memory, narrow_oop_reg
2767 // decode_not_null narrow_oop_reg, base_reg
2768 // Load [base_reg + offset], val_reg
2769 // NullCheck base_reg
2770 //
2771 } else if (t->isa_oopptr()) {
2772 new_in2 = ConNode::make(this, t->make_narrowoop());
2773 } else if (t->isa_klassptr()) {
2774 new_in2 = ConNode::make(this, t->make_narrowklass());
2775 }
2776 }
2777 if (new_in2 != NULL) {
2778 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2779 n->subsume_by(cmpN, this);
2780 if (in1->outcnt() == 0) {
2781 in1->disconnect_inputs(NULL, this);
2782 }
2783 if (in2->outcnt() == 0) {
2784 in2->disconnect_inputs(NULL, this);
2785 }
2786 }
2787 }
2788 break;
2790 case Op_DecodeN:
2791 case Op_DecodeNKlass:
2792 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2793 // DecodeN could be pinned when it can't be fold into
2794 // an address expression, see the code for Op_CastPP above.
2795 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2796 break;
2798 case Op_EncodeP:
2799 case Op_EncodePKlass: {
2800 Node* in1 = n->in(1);
2801 if (in1->is_DecodeNarrowPtr()) {
2802 n->subsume_by(in1->in(1), this);
2803 } else if (in1->Opcode() == Op_ConP) {
2804 const Type* t = in1->bottom_type();
2805 if (t == TypePtr::NULL_PTR) {
2806 assert(t->isa_oopptr(), "null klass?");
2807 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2808 } else if (t->isa_oopptr()) {
2809 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2810 } else if (t->isa_klassptr()) {
2811 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2812 }
2813 }
2814 if (in1->outcnt() == 0) {
2815 in1->disconnect_inputs(NULL, this);
2816 }
2817 break;
2818 }
2820 case Op_Proj: {
2821 if (OptimizeStringConcat) {
2822 ProjNode* p = n->as_Proj();
2823 if (p->_is_io_use) {
2824 // Separate projections were used for the exception path which
2825 // are normally removed by a late inline. If it wasn't inlined
2826 // then they will hang around and should just be replaced with
2827 // the original one.
2828 Node* proj = NULL;
2829 // Replace with just one
2830 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2831 Node *use = i.get();
2832 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2833 proj = use;
2834 break;
2835 }
2836 }
2837 assert(proj != NULL, "must be found");
2838 p->subsume_by(proj, this);
2839 }
2840 }
2841 break;
2842 }
2844 case Op_Phi:
2845 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
2846 // The EncodeP optimization may create Phi with the same edges
2847 // for all paths. It is not handled well by Register Allocator.
2848 Node* unique_in = n->in(1);
2849 assert(unique_in != NULL, "");
2850 uint cnt = n->req();
2851 for (uint i = 2; i < cnt; i++) {
2852 Node* m = n->in(i);
2853 assert(m != NULL, "");
2854 if (unique_in != m)
2855 unique_in = NULL;
2856 }
2857 if (unique_in != NULL) {
2858 n->subsume_by(unique_in, this);
2859 }
2860 }
2861 break;
2863 #endif
2865 case Op_ModI:
2866 if (UseDivMod) {
2867 // Check if a%b and a/b both exist
2868 Node* d = n->find_similar(Op_DivI);
2869 if (d) {
2870 // Replace them with a fused divmod if supported
2871 if (Matcher::has_match_rule(Op_DivModI)) {
2872 DivModINode* divmod = DivModINode::make(this, n);
2873 d->subsume_by(divmod->div_proj(), this);
2874 n->subsume_by(divmod->mod_proj(), this);
2875 } else {
2876 // replace a%b with a-((a/b)*b)
2877 Node* mult = new (this) MulINode(d, d->in(2));
2878 Node* sub = new (this) SubINode(d->in(1), mult);
2879 n->subsume_by(sub, this);
2880 }
2881 }
2882 }
2883 break;
2885 case Op_ModL:
2886 if (UseDivMod) {
2887 // Check if a%b and a/b both exist
2888 Node* d = n->find_similar(Op_DivL);
2889 if (d) {
2890 // Replace them with a fused divmod if supported
2891 if (Matcher::has_match_rule(Op_DivModL)) {
2892 DivModLNode* divmod = DivModLNode::make(this, n);
2893 d->subsume_by(divmod->div_proj(), this);
2894 n->subsume_by(divmod->mod_proj(), this);
2895 } else {
2896 // replace a%b with a-((a/b)*b)
2897 Node* mult = new (this) MulLNode(d, d->in(2));
2898 Node* sub = new (this) SubLNode(d->in(1), mult);
2899 n->subsume_by(sub, this);
2900 }
2901 }
2902 }
2903 break;
2905 case Op_LoadVector:
2906 case Op_StoreVector:
2907 break;
2909 case Op_PackB:
2910 case Op_PackS:
2911 case Op_PackI:
2912 case Op_PackF:
2913 case Op_PackL:
2914 case Op_PackD:
2915 if (n->req()-1 > 2) {
2916 // Replace many operand PackNodes with a binary tree for matching
2917 PackNode* p = (PackNode*) n;
2918 Node* btp = p->binary_tree_pack(this, 1, n->req());
2919 n->subsume_by(btp, this);
2920 }
2921 break;
2922 case Op_Loop:
2923 case Op_CountedLoop:
2924 if (n->as_Loop()->is_inner_loop()) {
2925 frc.inc_inner_loop_count();
2926 }
2927 break;
2928 case Op_LShiftI:
2929 case Op_RShiftI:
2930 case Op_URShiftI:
2931 case Op_LShiftL:
2932 case Op_RShiftL:
2933 case Op_URShiftL:
2934 if (Matcher::need_masked_shift_count) {
2935 // The cpu's shift instructions don't restrict the count to the
2936 // lower 5/6 bits. We need to do the masking ourselves.
2937 Node* in2 = n->in(2);
2938 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2939 const TypeInt* t = in2->find_int_type();
2940 if (t != NULL && t->is_con()) {
2941 juint shift = t->get_con();
2942 if (shift > mask) { // Unsigned cmp
2943 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
2944 }
2945 } else {
2946 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2947 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
2948 n->set_req(2, shift);
2949 }
2950 }
2951 if (in2->outcnt() == 0) { // Remove dead node
2952 in2->disconnect_inputs(NULL, this);
2953 }
2954 }
2955 break;
2956 case Op_MemBarStoreStore:
2957 case Op_MemBarRelease:
2958 // Break the link with AllocateNode: it is no longer useful and
2959 // confuses register allocation.
2960 if (n->req() > MemBarNode::Precedent) {
2961 n->set_req(MemBarNode::Precedent, top());
2962 }
2963 break;
2964 default:
2965 assert( !n->is_Call(), "" );
2966 assert( !n->is_Mem(), "" );
2967 break;
2968 }
2970 // Collect CFG split points
2971 if (n->is_MultiBranch())
2972 frc._tests.push(n);
2973 }
2975 //------------------------------final_graph_reshaping_walk---------------------
2976 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2977 // requires that the walk visits a node's inputs before visiting the node.
2978 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
2979 ResourceArea *area = Thread::current()->resource_area();
2980 Unique_Node_List sfpt(area);
2982 frc._visited.set(root->_idx); // first, mark node as visited
2983 uint cnt = root->req();
2984 Node *n = root;
2985 uint i = 0;
2986 while (true) {
2987 if (i < cnt) {
2988 // Place all non-visited non-null inputs onto stack
2989 Node* m = n->in(i);
2990 ++i;
2991 if (m != NULL && !frc._visited.test_set(m->_idx)) {
2992 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
2993 sfpt.push(m);
2994 cnt = m->req();
2995 nstack.push(n, i); // put on stack parent and next input's index
2996 n = m;
2997 i = 0;
2998 }
2999 } else {
3000 // Now do post-visit work
3001 final_graph_reshaping_impl( n, frc );
3002 if (nstack.is_empty())
3003 break; // finished
3004 n = nstack.node(); // Get node from stack
3005 cnt = n->req();
3006 i = nstack.index();
3007 nstack.pop(); // Shift to the next node on stack
3008 }
3009 }
3011 // Skip next transformation if compressed oops are not used.
3012 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3013 (!UseCompressedOops && !UseCompressedKlassPointers))
3014 return;
3016 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3017 // It could be done for an uncommon traps or any safepoints/calls
3018 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3019 while (sfpt.size() > 0) {
3020 n = sfpt.pop();
3021 JVMState *jvms = n->as_SafePoint()->jvms();
3022 assert(jvms != NULL, "sanity");
3023 int start = jvms->debug_start();
3024 int end = n->req();
3025 bool is_uncommon = (n->is_CallStaticJava() &&
3026 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3027 for (int j = start; j < end; j++) {
3028 Node* in = n->in(j);
3029 if (in->is_DecodeNarrowPtr()) {
3030 bool safe_to_skip = true;
3031 if (!is_uncommon ) {
3032 // Is it safe to skip?
3033 for (uint i = 0; i < in->outcnt(); i++) {
3034 Node* u = in->raw_out(i);
3035 if (!u->is_SafePoint() ||
3036 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3037 safe_to_skip = false;
3038 }
3039 }
3040 }
3041 if (safe_to_skip) {
3042 n->set_req(j, in->in(1));
3043 }
3044 if (in->outcnt() == 0) {
3045 in->disconnect_inputs(NULL, this);
3046 }
3047 }
3048 }
3049 }
3050 }
3052 //------------------------------final_graph_reshaping--------------------------
3053 // Final Graph Reshaping.
3054 //
3055 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3056 // and not commoned up and forced early. Must come after regular
3057 // optimizations to avoid GVN undoing the cloning. Clone constant
3058 // inputs to Loop Phis; these will be split by the allocator anyways.
3059 // Remove Opaque nodes.
3060 // (2) Move last-uses by commutative operations to the left input to encourage
3061 // Intel update-in-place two-address operations and better register usage
3062 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3063 // calls canonicalizing them back.
3064 // (3) Count the number of double-precision FP ops, single-precision FP ops
3065 // and call sites. On Intel, we can get correct rounding either by
3066 // forcing singles to memory (requires extra stores and loads after each
3067 // FP bytecode) or we can set a rounding mode bit (requires setting and
3068 // clearing the mode bit around call sites). The mode bit is only used
3069 // if the relative frequency of single FP ops to calls is low enough.
3070 // This is a key transform for SPEC mpeg_audio.
3071 // (4) Detect infinite loops; blobs of code reachable from above but not
3072 // below. Several of the Code_Gen algorithms fail on such code shapes,
3073 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3074 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3075 // Detection is by looking for IfNodes where only 1 projection is
3076 // reachable from below or CatchNodes missing some targets.
3077 // (5) Assert for insane oop offsets in debug mode.
3079 bool Compile::final_graph_reshaping() {
3080 // an infinite loop may have been eliminated by the optimizer,
3081 // in which case the graph will be empty.
3082 if (root()->req() == 1) {
3083 record_method_not_compilable("trivial infinite loop");
3084 return true;
3085 }
3087 // Expensive nodes have their control input set to prevent the GVN
3088 // from freely commoning them. There's no GVN beyond this point so
3089 // no need to keep the control input. We want the expensive nodes to
3090 // be freely moved to the least frequent code path by gcm.
3091 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3092 for (int i = 0; i < expensive_count(); i++) {
3093 _expensive_nodes->at(i)->set_req(0, NULL);
3094 }
3096 Final_Reshape_Counts frc;
3098 // Visit everybody reachable!
3099 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3100 Node_Stack nstack(unique() >> 1);
3101 final_graph_reshaping_walk(nstack, root(), frc);
3103 // Check for unreachable (from below) code (i.e., infinite loops).
3104 for( uint i = 0; i < frc._tests.size(); i++ ) {
3105 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3106 // Get number of CFG targets.
3107 // Note that PCTables include exception targets after calls.
3108 uint required_outcnt = n->required_outcnt();
3109 if (n->outcnt() != required_outcnt) {
3110 // Check for a few special cases. Rethrow Nodes never take the
3111 // 'fall-thru' path, so expected kids is 1 less.
3112 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3113 if (n->in(0)->in(0)->is_Call()) {
3114 CallNode *call = n->in(0)->in(0)->as_Call();
3115 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3116 required_outcnt--; // Rethrow always has 1 less kid
3117 } else if (call->req() > TypeFunc::Parms &&
3118 call->is_CallDynamicJava()) {
3119 // Check for null receiver. In such case, the optimizer has
3120 // detected that the virtual call will always result in a null
3121 // pointer exception. The fall-through projection of this CatchNode
3122 // will not be populated.
3123 Node *arg0 = call->in(TypeFunc::Parms);
3124 if (arg0->is_Type() &&
3125 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3126 required_outcnt--;
3127 }
3128 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3129 call->req() > TypeFunc::Parms+1 &&
3130 call->is_CallStaticJava()) {
3131 // Check for negative array length. In such case, the optimizer has
3132 // detected that the allocation attempt will always result in an
3133 // exception. There is no fall-through projection of this CatchNode .
3134 Node *arg1 = call->in(TypeFunc::Parms+1);
3135 if (arg1->is_Type() &&
3136 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3137 required_outcnt--;
3138 }
3139 }
3140 }
3141 }
3142 // Recheck with a better notion of 'required_outcnt'
3143 if (n->outcnt() != required_outcnt) {
3144 record_method_not_compilable("malformed control flow");
3145 return true; // Not all targets reachable!
3146 }
3147 }
3148 // Check that I actually visited all kids. Unreached kids
3149 // must be infinite loops.
3150 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3151 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3152 record_method_not_compilable("infinite loop");
3153 return true; // Found unvisited kid; must be unreach
3154 }
3155 }
3157 // If original bytecodes contained a mixture of floats and doubles
3158 // check if the optimizer has made it homogenous, item (3).
3159 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3160 frc.get_float_count() > 32 &&
3161 frc.get_double_count() == 0 &&
3162 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3163 set_24_bit_selection_and_mode( false, true );
3164 }
3166 set_java_calls(frc.get_java_call_count());
3167 set_inner_loops(frc.get_inner_loop_count());
3169 // No infinite loops, no reason to bail out.
3170 return false;
3171 }
3173 //-----------------------------too_many_traps----------------------------------
3174 // Report if there are too many traps at the current method and bci.
3175 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3176 bool Compile::too_many_traps(ciMethod* method,
3177 int bci,
3178 Deoptimization::DeoptReason reason) {
3179 ciMethodData* md = method->method_data();
3180 if (md->is_empty()) {
3181 // Assume the trap has not occurred, or that it occurred only
3182 // because of a transient condition during start-up in the interpreter.
3183 return false;
3184 }
3185 if (md->has_trap_at(bci, reason) != 0) {
3186 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3187 // Also, if there are multiple reasons, or if there is no per-BCI record,
3188 // assume the worst.
3189 if (log())
3190 log()->elem("observe trap='%s' count='%d'",
3191 Deoptimization::trap_reason_name(reason),
3192 md->trap_count(reason));
3193 return true;
3194 } else {
3195 // Ignore method/bci and see if there have been too many globally.
3196 return too_many_traps(reason, md);
3197 }
3198 }
3200 // Less-accurate variant which does not require a method and bci.
3201 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3202 ciMethodData* logmd) {
3203 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
3204 // Too many traps globally.
3205 // Note that we use cumulative trap_count, not just md->trap_count.
3206 if (log()) {
3207 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3208 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3209 Deoptimization::trap_reason_name(reason),
3210 mcount, trap_count(reason));
3211 }
3212 return true;
3213 } else {
3214 // The coast is clear.
3215 return false;
3216 }
3217 }
3219 //--------------------------too_many_recompiles--------------------------------
3220 // Report if there are too many recompiles at the current method and bci.
3221 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3222 // Is not eager to return true, since this will cause the compiler to use
3223 // Action_none for a trap point, to avoid too many recompilations.
3224 bool Compile::too_many_recompiles(ciMethod* method,
3225 int bci,
3226 Deoptimization::DeoptReason reason) {
3227 ciMethodData* md = method->method_data();
3228 if (md->is_empty()) {
3229 // Assume the trap has not occurred, or that it occurred only
3230 // because of a transient condition during start-up in the interpreter.
3231 return false;
3232 }
3233 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3234 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3235 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3236 Deoptimization::DeoptReason per_bc_reason
3237 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3238 if ((per_bc_reason == Deoptimization::Reason_none
3239 || md->has_trap_at(bci, reason) != 0)
3240 // The trap frequency measure we care about is the recompile count:
3241 && md->trap_recompiled_at(bci)
3242 && md->overflow_recompile_count() >= bc_cutoff) {
3243 // Do not emit a trap here if it has already caused recompilations.
3244 // Also, if there are multiple reasons, or if there is no per-BCI record,
3245 // assume the worst.
3246 if (log())
3247 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3248 Deoptimization::trap_reason_name(reason),
3249 md->trap_count(reason),
3250 md->overflow_recompile_count());
3251 return true;
3252 } else if (trap_count(reason) != 0
3253 && decompile_count() >= m_cutoff) {
3254 // Too many recompiles globally, and we have seen this sort of trap.
3255 // Use cumulative decompile_count, not just md->decompile_count.
3256 if (log())
3257 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3258 Deoptimization::trap_reason_name(reason),
3259 md->trap_count(reason), trap_count(reason),
3260 md->decompile_count(), decompile_count());
3261 return true;
3262 } else {
3263 // The coast is clear.
3264 return false;
3265 }
3266 }
3269 #ifndef PRODUCT
3270 //------------------------------verify_graph_edges---------------------------
3271 // Walk the Graph and verify that there is a one-to-one correspondence
3272 // between Use-Def edges and Def-Use edges in the graph.
3273 void Compile::verify_graph_edges(bool no_dead_code) {
3274 if (VerifyGraphEdges) {
3275 ResourceArea *area = Thread::current()->resource_area();
3276 Unique_Node_List visited(area);
3277 // Call recursive graph walk to check edges
3278 _root->verify_edges(visited);
3279 if (no_dead_code) {
3280 // Now make sure that no visited node is used by an unvisited node.
3281 bool dead_nodes = 0;
3282 Unique_Node_List checked(area);
3283 while (visited.size() > 0) {
3284 Node* n = visited.pop();
3285 checked.push(n);
3286 for (uint i = 0; i < n->outcnt(); i++) {
3287 Node* use = n->raw_out(i);
3288 if (checked.member(use)) continue; // already checked
3289 if (visited.member(use)) continue; // already in the graph
3290 if (use->is_Con()) continue; // a dead ConNode is OK
3291 // At this point, we have found a dead node which is DU-reachable.
3292 if (dead_nodes++ == 0)
3293 tty->print_cr("*** Dead nodes reachable via DU edges:");
3294 use->dump(2);
3295 tty->print_cr("---");
3296 checked.push(use); // No repeats; pretend it is now checked.
3297 }
3298 }
3299 assert(dead_nodes == 0, "using nodes must be reachable from root");
3300 }
3301 }
3302 }
3303 #endif
3305 // The Compile object keeps track of failure reasons separately from the ciEnv.
3306 // This is required because there is not quite a 1-1 relation between the
3307 // ciEnv and its compilation task and the Compile object. Note that one
3308 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3309 // to backtrack and retry without subsuming loads. Other than this backtracking
3310 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3311 // by the logic in C2Compiler.
3312 void Compile::record_failure(const char* reason) {
3313 if (log() != NULL) {
3314 log()->elem("failure reason='%s' phase='compile'", reason);
3315 }
3316 if (_failure_reason == NULL) {
3317 // Record the first failure reason.
3318 _failure_reason = reason;
3319 }
3321 EventCompilerFailure event;
3322 if (event.should_commit()) {
3323 event.set_compileID(Compile::compile_id());
3324 event.set_failure(reason);
3325 event.commit();
3326 }
3328 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3329 C->print_method(PHASE_FAILURE);
3330 }
3331 _root = NULL; // flush the graph, too
3332 }
3334 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3335 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3336 _phase_name(name), _dolog(dolog)
3337 {
3338 if (dolog) {
3339 C = Compile::current();
3340 _log = C->log();
3341 } else {
3342 C = NULL;
3343 _log = NULL;
3344 }
3345 if (_log != NULL) {
3346 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3347 _log->stamp();
3348 _log->end_head();
3349 }
3350 }
3352 Compile::TracePhase::~TracePhase() {
3354 C = Compile::current();
3355 if (_dolog) {
3356 _log = C->log();
3357 } else {
3358 _log = NULL;
3359 }
3361 #ifdef ASSERT
3362 if (PrintIdealNodeCount) {
3363 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3364 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3365 }
3367 if (VerifyIdealNodeCount) {
3368 Compile::current()->print_missing_nodes();
3369 }
3370 #endif
3372 if (_log != NULL) {
3373 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3374 }
3375 }
3377 //=============================================================================
3378 // Two Constant's are equal when the type and the value are equal.
3379 bool Compile::Constant::operator==(const Constant& other) {
3380 if (type() != other.type() ) return false;
3381 if (can_be_reused() != other.can_be_reused()) return false;
3382 // For floating point values we compare the bit pattern.
3383 switch (type()) {
3384 case T_FLOAT: return (_v._value.i == other._v._value.i);
3385 case T_LONG:
3386 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3387 case T_OBJECT:
3388 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3389 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3390 case T_METADATA: return (_v._metadata == other._v._metadata);
3391 default: ShouldNotReachHere();
3392 }
3393 return false;
3394 }
3396 static int type_to_size_in_bytes(BasicType t) {
3397 switch (t) {
3398 case T_LONG: return sizeof(jlong );
3399 case T_FLOAT: return sizeof(jfloat );
3400 case T_DOUBLE: return sizeof(jdouble);
3401 case T_METADATA: return sizeof(Metadata*);
3402 // We use T_VOID as marker for jump-table entries (labels) which
3403 // need an internal word relocation.
3404 case T_VOID:
3405 case T_ADDRESS:
3406 case T_OBJECT: return sizeof(jobject);
3407 }
3409 ShouldNotReachHere();
3410 return -1;
3411 }
3413 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3414 // sort descending
3415 if (a->freq() > b->freq()) return -1;
3416 if (a->freq() < b->freq()) return 1;
3417 return 0;
3418 }
3420 void Compile::ConstantTable::calculate_offsets_and_size() {
3421 // First, sort the array by frequencies.
3422 _constants.sort(qsort_comparator);
3424 #ifdef ASSERT
3425 // Make sure all jump-table entries were sorted to the end of the
3426 // array (they have a negative frequency).
3427 bool found_void = false;
3428 for (int i = 0; i < _constants.length(); i++) {
3429 Constant con = _constants.at(i);
3430 if (con.type() == T_VOID)
3431 found_void = true; // jump-tables
3432 else
3433 assert(!found_void, "wrong sorting");
3434 }
3435 #endif
3437 int offset = 0;
3438 for (int i = 0; i < _constants.length(); i++) {
3439 Constant* con = _constants.adr_at(i);
3441 // Align offset for type.
3442 int typesize = type_to_size_in_bytes(con->type());
3443 offset = align_size_up(offset, typesize);
3444 con->set_offset(offset); // set constant's offset
3446 if (con->type() == T_VOID) {
3447 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3448 offset = offset + typesize * n->outcnt(); // expand jump-table
3449 } else {
3450 offset = offset + typesize;
3451 }
3452 }
3454 // Align size up to the next section start (which is insts; see
3455 // CodeBuffer::align_at_start).
3456 assert(_size == -1, "already set?");
3457 _size = align_size_up(offset, CodeEntryAlignment);
3458 }
3460 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3461 MacroAssembler _masm(&cb);
3462 for (int i = 0; i < _constants.length(); i++) {
3463 Constant con = _constants.at(i);
3464 address constant_addr;
3465 switch (con.type()) {
3466 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3467 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3468 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3469 case T_OBJECT: {
3470 jobject obj = con.get_jobject();
3471 int oop_index = _masm.oop_recorder()->find_index(obj);
3472 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3473 break;
3474 }
3475 case T_ADDRESS: {
3476 address addr = (address) con.get_jobject();
3477 constant_addr = _masm.address_constant(addr);
3478 break;
3479 }
3480 // We use T_VOID as marker for jump-table entries (labels) which
3481 // need an internal word relocation.
3482 case T_VOID: {
3483 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3484 // Fill the jump-table with a dummy word. The real value is
3485 // filled in later in fill_jump_table.
3486 address dummy = (address) n;
3487 constant_addr = _masm.address_constant(dummy);
3488 // Expand jump-table
3489 for (uint i = 1; i < n->outcnt(); i++) {
3490 address temp_addr = _masm.address_constant(dummy + i);
3491 assert(temp_addr, "consts section too small");
3492 }
3493 break;
3494 }
3495 case T_METADATA: {
3496 Metadata* obj = con.get_metadata();
3497 int metadata_index = _masm.oop_recorder()->find_index(obj);
3498 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3499 break;
3500 }
3501 default: ShouldNotReachHere();
3502 }
3503 assert(constant_addr, "consts section too small");
3504 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3505 }
3506 }
3508 int Compile::ConstantTable::find_offset(Constant& con) const {
3509 int idx = _constants.find(con);
3510 assert(idx != -1, "constant must be in constant table");
3511 int offset = _constants.at(idx).offset();
3512 assert(offset != -1, "constant table not emitted yet?");
3513 return offset;
3514 }
3516 void Compile::ConstantTable::add(Constant& con) {
3517 if (con.can_be_reused()) {
3518 int idx = _constants.find(con);
3519 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3520 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3521 return;
3522 }
3523 }
3524 (void) _constants.append(con);
3525 }
3527 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3528 Block* b = Compile::current()->cfg()->_bbs[n->_idx];
3529 Constant con(type, value, b->_freq);
3530 add(con);
3531 return con;
3532 }
3534 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3535 Constant con(metadata);
3536 add(con);
3537 return con;
3538 }
3540 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3541 jvalue value;
3542 BasicType type = oper->type()->basic_type();
3543 switch (type) {
3544 case T_LONG: value.j = oper->constantL(); break;
3545 case T_FLOAT: value.f = oper->constantF(); break;
3546 case T_DOUBLE: value.d = oper->constantD(); break;
3547 case T_OBJECT:
3548 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3549 case T_METADATA: return add((Metadata*)oper->constant()); break;
3550 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3551 }
3552 return add(n, type, value);
3553 }
3555 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3556 jvalue value;
3557 // We can use the node pointer here to identify the right jump-table
3558 // as this method is called from Compile::Fill_buffer right before
3559 // the MachNodes are emitted and the jump-table is filled (means the
3560 // MachNode pointers do not change anymore).
3561 value.l = (jobject) n;
3562 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3563 add(con);
3564 return con;
3565 }
3567 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3568 // If called from Compile::scratch_emit_size do nothing.
3569 if (Compile::current()->in_scratch_emit_size()) return;
3571 assert(labels.is_nonempty(), "must be");
3572 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3574 // Since MachConstantNode::constant_offset() also contains
3575 // table_base_offset() we need to subtract the table_base_offset()
3576 // to get the plain offset into the constant table.
3577 int offset = n->constant_offset() - table_base_offset();
3579 MacroAssembler _masm(&cb);
3580 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3582 for (uint i = 0; i < n->outcnt(); i++) {
3583 address* constant_addr = &jump_table_base[i];
3584 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i)));
3585 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3586 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3587 }
3588 }
3590 void Compile::dump_inlining() {
3591 if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) {
3592 // Print inlining message for candidates that we couldn't inline
3593 // for lack of space or non constant receiver
3594 for (int i = 0; i < _late_inlines.length(); i++) {
3595 CallGenerator* cg = _late_inlines.at(i);
3596 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3597 }
3598 Unique_Node_List useful;
3599 useful.push(root());
3600 for (uint next = 0; next < useful.size(); ++next) {
3601 Node* n = useful.at(next);
3602 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3603 CallNode* call = n->as_Call();
3604 CallGenerator* cg = call->generator();
3605 cg->print_inlining_late("receiver not constant");
3606 }
3607 uint max = n->len();
3608 for ( uint i = 0; i < max; ++i ) {
3609 Node *m = n->in(i);
3610 if ( m == NULL ) continue;
3611 useful.push(m);
3612 }
3613 }
3614 for (int i = 0; i < _print_inlining_list->length(); i++) {
3615 tty->print(_print_inlining_list->at(i).ss()->as_string());
3616 }
3617 }
3618 }
3620 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3621 if (n1->Opcode() < n2->Opcode()) return -1;
3622 else if (n1->Opcode() > n2->Opcode()) return 1;
3624 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()));
3625 for (uint i = 1; i < n1->req(); i++) {
3626 if (n1->in(i) < n2->in(i)) return -1;
3627 else if (n1->in(i) > n2->in(i)) return 1;
3628 }
3630 return 0;
3631 }
3633 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3634 Node* n1 = *n1p;
3635 Node* n2 = *n2p;
3637 return cmp_expensive_nodes(n1, n2);
3638 }
3640 void Compile::sort_expensive_nodes() {
3641 if (!expensive_nodes_sorted()) {
3642 _expensive_nodes->sort(cmp_expensive_nodes);
3643 }
3644 }
3646 bool Compile::expensive_nodes_sorted() const {
3647 for (int i = 1; i < _expensive_nodes->length(); i++) {
3648 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3649 return false;
3650 }
3651 }
3652 return true;
3653 }
3655 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3656 if (_expensive_nodes->length() == 0) {
3657 return false;
3658 }
3660 assert(OptimizeExpensiveOps, "optimization off?");
3662 // Take this opportunity to remove dead nodes from the list
3663 int j = 0;
3664 for (int i = 0; i < _expensive_nodes->length(); i++) {
3665 Node* n = _expensive_nodes->at(i);
3666 if (!n->is_unreachable(igvn)) {
3667 assert(n->is_expensive(), "should be expensive");
3668 _expensive_nodes->at_put(j, n);
3669 j++;
3670 }
3671 }
3672 _expensive_nodes->trunc_to(j);
3674 // Then sort the list so that similar nodes are next to each other
3675 // and check for at least two nodes of identical kind with same data
3676 // inputs.
3677 sort_expensive_nodes();
3679 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3680 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3681 return true;
3682 }
3683 }
3685 return false;
3686 }
3688 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3689 if (_expensive_nodes->length() == 0) {
3690 return;
3691 }
3693 assert(OptimizeExpensiveOps, "optimization off?");
3695 // Sort to bring similar nodes next to each other and clear the
3696 // control input of nodes for which there's only a single copy.
3697 sort_expensive_nodes();
3699 int j = 0;
3700 int identical = 0;
3701 int i = 0;
3702 for (; i < _expensive_nodes->length()-1; i++) {
3703 assert(j <= i, "can't write beyond current index");
3704 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3705 identical++;
3706 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3707 continue;
3708 }
3709 if (identical > 0) {
3710 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3711 identical = 0;
3712 } else {
3713 Node* n = _expensive_nodes->at(i);
3714 igvn.hash_delete(n);
3715 n->set_req(0, NULL);
3716 igvn.hash_insert(n);
3717 }
3718 }
3719 if (identical > 0) {
3720 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3721 } else if (_expensive_nodes->length() >= 1) {
3722 Node* n = _expensive_nodes->at(i);
3723 igvn.hash_delete(n);
3724 n->set_req(0, NULL);
3725 igvn.hash_insert(n);
3726 }
3727 _expensive_nodes->trunc_to(j);
3728 }
3730 void Compile::add_expensive_node(Node * n) {
3731 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3732 assert(n->is_expensive(), "expensive nodes with non-null control here only");
3733 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
3734 if (OptimizeExpensiveOps) {
3735 _expensive_nodes->append(n);
3736 } else {
3737 // Clear control input and let IGVN optimize expensive nodes if
3738 // OptimizeExpensiveOps is off.
3739 n->set_req(0, NULL);
3740 }
3741 }
3743 // Auxiliary method to support randomized stressing/fuzzing.
3744 //
3745 // This method can be called the arbitrary number of times, with current count
3746 // as the argument. The logic allows selecting a single candidate from the
3747 // running list of candidates as follows:
3748 // int count = 0;
3749 // Cand* selected = null;
3750 // while(cand = cand->next()) {
3751 // if (randomized_select(++count)) {
3752 // selected = cand;
3753 // }
3754 // }
3755 //
3756 // Including count equalizes the chances any candidate is "selected".
3757 // This is useful when we don't have the complete list of candidates to choose
3758 // from uniformly. In this case, we need to adjust the randomicity of the
3759 // selection, or else we will end up biasing the selection towards the latter
3760 // candidates.
3761 //
3762 // Quick back-envelope calculation shows that for the list of n candidates
3763 // the equal probability for the candidate to persist as "best" can be
3764 // achieved by replacing it with "next" k-th candidate with the probability
3765 // of 1/k. It can be easily shown that by the end of the run, the
3766 // probability for any candidate is converged to 1/n, thus giving the
3767 // uniform distribution among all the candidates.
3768 //
3769 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
3770 #define RANDOMIZED_DOMAIN_POW 29
3771 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
3772 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
3773 bool Compile::randomized_select(int count) {
3774 assert(count > 0, "only positive");
3775 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
3776 }