Mon, 25 Apr 2016 11:01:13 -0700
Merge
1 /*
2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "compiler/disassembler.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/block.hpp"
37 #include "opto/c2compiler.hpp"
38 #include "opto/callGenerator.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/cfgnode.hpp"
41 #include "opto/chaitin.hpp"
42 #include "opto/compile.hpp"
43 #include "opto/connode.hpp"
44 #include "opto/divnode.hpp"
45 #include "opto/escape.hpp"
46 #include "opto/idealGraphPrinter.hpp"
47 #include "opto/loopnode.hpp"
48 #include "opto/machnode.hpp"
49 #include "opto/macro.hpp"
50 #include "opto/matcher.hpp"
51 #include "opto/mathexactnode.hpp"
52 #include "opto/memnode.hpp"
53 #include "opto/mulnode.hpp"
54 #include "opto/node.hpp"
55 #include "opto/opcodes.hpp"
56 #include "opto/output.hpp"
57 #include "opto/parse.hpp"
58 #include "opto/phaseX.hpp"
59 #include "opto/rootnode.hpp"
60 #include "opto/runtime.hpp"
61 #include "opto/stringopts.hpp"
62 #include "opto/type.hpp"
63 #include "opto/vectornode.hpp"
64 #include "runtime/arguments.hpp"
65 #include "runtime/signature.hpp"
66 #include "runtime/stubRoutines.hpp"
67 #include "runtime/timer.hpp"
68 #include "trace/tracing.hpp"
69 #include "utilities/copy.hpp"
70 #if defined AD_MD_HPP
71 # include AD_MD_HPP
72 #elif defined TARGET_ARCH_MODEL_x86_32
73 # include "adfiles/ad_x86_32.hpp"
74 #elif defined TARGET_ARCH_MODEL_x86_64
75 # include "adfiles/ad_x86_64.hpp"
76 #elif defined TARGET_ARCH_MODEL_sparc
77 # include "adfiles/ad_sparc.hpp"
78 #elif defined TARGET_ARCH_MODEL_zero
79 # include "adfiles/ad_zero.hpp"
80 #elif defined TARGET_ARCH_MODEL_ppc_64
81 # include "adfiles/ad_ppc_64.hpp"
82 #endif
85 // -------------------- Compile::mach_constant_base_node -----------------------
86 // Constant table base node singleton.
87 MachConstantBaseNode* Compile::mach_constant_base_node() {
88 if (_mach_constant_base_node == NULL) {
89 _mach_constant_base_node = new (C) MachConstantBaseNode();
90 _mach_constant_base_node->add_req(C->root());
91 }
92 return _mach_constant_base_node;
93 }
96 /// Support for intrinsics.
98 // Return the index at which m must be inserted (or already exists).
99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
100 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
101 #ifdef ASSERT
102 for (int i = 1; i < _intrinsics->length(); i++) {
103 CallGenerator* cg1 = _intrinsics->at(i-1);
104 CallGenerator* cg2 = _intrinsics->at(i);
105 assert(cg1->method() != cg2->method()
106 ? cg1->method() < cg2->method()
107 : cg1->is_virtual() < cg2->is_virtual(),
108 "compiler intrinsics list must stay sorted");
109 }
110 #endif
111 // Binary search sorted list, in decreasing intervals [lo, hi].
112 int lo = 0, hi = _intrinsics->length()-1;
113 while (lo <= hi) {
114 int mid = (uint)(hi + lo) / 2;
115 ciMethod* mid_m = _intrinsics->at(mid)->method();
116 if (m < mid_m) {
117 hi = mid-1;
118 } else if (m > mid_m) {
119 lo = mid+1;
120 } else {
121 // look at minor sort key
122 bool mid_virt = _intrinsics->at(mid)->is_virtual();
123 if (is_virtual < mid_virt) {
124 hi = mid-1;
125 } else if (is_virtual > mid_virt) {
126 lo = mid+1;
127 } else {
128 return mid; // exact match
129 }
130 }
131 }
132 return lo; // inexact match
133 }
135 void Compile::register_intrinsic(CallGenerator* cg) {
136 if (_intrinsics == NULL) {
137 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
138 }
139 // This code is stolen from ciObjectFactory::insert.
140 // Really, GrowableArray should have methods for
141 // insert_at, remove_at, and binary_search.
142 int len = _intrinsics->length();
143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
144 if (index == len) {
145 _intrinsics->append(cg);
146 } else {
147 #ifdef ASSERT
148 CallGenerator* oldcg = _intrinsics->at(index);
149 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
150 #endif
151 _intrinsics->append(_intrinsics->at(len-1));
152 int pos;
153 for (pos = len-2; pos >= index; pos--) {
154 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
155 }
156 _intrinsics->at_put(index, cg);
157 }
158 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
159 }
161 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
162 assert(m->is_loaded(), "don't try this on unloaded methods");
163 if (_intrinsics != NULL) {
164 int index = intrinsic_insertion_index(m, is_virtual);
165 if (index < _intrinsics->length()
166 && _intrinsics->at(index)->method() == m
167 && _intrinsics->at(index)->is_virtual() == is_virtual) {
168 return _intrinsics->at(index);
169 }
170 }
171 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
172 if (m->intrinsic_id() != vmIntrinsics::_none &&
173 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
174 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
175 if (cg != NULL) {
176 // Save it for next time:
177 register_intrinsic(cg);
178 return cg;
179 } else {
180 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
181 }
182 }
183 return NULL;
184 }
186 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
187 // in library_call.cpp.
190 #ifndef PRODUCT
191 // statistics gathering...
193 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
194 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
196 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
197 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
198 int oflags = _intrinsic_hist_flags[id];
199 assert(flags != 0, "what happened?");
200 if (is_virtual) {
201 flags |= _intrinsic_virtual;
202 }
203 bool changed = (flags != oflags);
204 if ((flags & _intrinsic_worked) != 0) {
205 juint count = (_intrinsic_hist_count[id] += 1);
206 if (count == 1) {
207 changed = true; // first time
208 }
209 // increment the overall count also:
210 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
211 }
212 if (changed) {
213 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
214 // Something changed about the intrinsic's virtuality.
215 if ((flags & _intrinsic_virtual) != 0) {
216 // This is the first use of this intrinsic as a virtual call.
217 if (oflags != 0) {
218 // We already saw it as a non-virtual, so note both cases.
219 flags |= _intrinsic_both;
220 }
221 } else if ((oflags & _intrinsic_both) == 0) {
222 // This is the first use of this intrinsic as a non-virtual
223 flags |= _intrinsic_both;
224 }
225 }
226 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
227 }
228 // update the overall flags also:
229 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
230 return changed;
231 }
233 static char* format_flags(int flags, char* buf) {
234 buf[0] = 0;
235 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
236 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
237 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
238 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
239 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
240 if (buf[0] == 0) strcat(buf, ",");
241 assert(buf[0] == ',', "must be");
242 return &buf[1];
243 }
245 void Compile::print_intrinsic_statistics() {
246 char flagsbuf[100];
247 ttyLocker ttyl;
248 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
249 tty->print_cr("Compiler intrinsic usage:");
250 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
251 if (total == 0) total = 1; // avoid div0 in case of no successes
252 #define PRINT_STAT_LINE(name, c, f) \
253 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
254 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
255 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
256 int flags = _intrinsic_hist_flags[id];
257 juint count = _intrinsic_hist_count[id];
258 if ((flags | count) != 0) {
259 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
260 }
261 }
262 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
263 if (xtty != NULL) xtty->tail("statistics");
264 }
266 void Compile::print_statistics() {
267 { ttyLocker ttyl;
268 if (xtty != NULL) xtty->head("statistics type='opto'");
269 Parse::print_statistics();
270 PhaseCCP::print_statistics();
271 PhaseRegAlloc::print_statistics();
272 Scheduling::print_statistics();
273 PhasePeephole::print_statistics();
274 PhaseIdealLoop::print_statistics();
275 if (xtty != NULL) xtty->tail("statistics");
276 }
277 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
278 // put this under its own <statistics> element.
279 print_intrinsic_statistics();
280 }
281 }
282 #endif //PRODUCT
284 // Support for bundling info
285 Bundle* Compile::node_bundling(const Node *n) {
286 assert(valid_bundle_info(n), "oob");
287 return &_node_bundling_base[n->_idx];
288 }
290 bool Compile::valid_bundle_info(const Node *n) {
291 return (_node_bundling_limit > n->_idx);
292 }
295 void Compile::gvn_replace_by(Node* n, Node* nn) {
296 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
297 Node* use = n->last_out(i);
298 bool is_in_table = initial_gvn()->hash_delete(use);
299 uint uses_found = 0;
300 for (uint j = 0; j < use->len(); j++) {
301 if (use->in(j) == n) {
302 if (j < use->req())
303 use->set_req(j, nn);
304 else
305 use->set_prec(j, nn);
306 uses_found++;
307 }
308 }
309 if (is_in_table) {
310 // reinsert into table
311 initial_gvn()->hash_find_insert(use);
312 }
313 record_for_igvn(use);
314 i -= uses_found; // we deleted 1 or more copies of this edge
315 }
316 }
319 static inline bool not_a_node(const Node* n) {
320 if (n == NULL) return true;
321 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
322 if (*(address*)n == badAddress) return true; // kill by Node::destruct
323 return false;
324 }
326 // Identify all nodes that are reachable from below, useful.
327 // Use breadth-first pass that records state in a Unique_Node_List,
328 // recursive traversal is slower.
329 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
330 int estimated_worklist_size = live_nodes();
331 useful.map( estimated_worklist_size, NULL ); // preallocate space
333 // Initialize worklist
334 if (root() != NULL) { useful.push(root()); }
335 // If 'top' is cached, declare it useful to preserve cached node
336 if( cached_top_node() ) { useful.push(cached_top_node()); }
338 // Push all useful nodes onto the list, breadthfirst
339 for( uint next = 0; next < useful.size(); ++next ) {
340 assert( next < unique(), "Unique useful nodes < total nodes");
341 Node *n = useful.at(next);
342 uint max = n->len();
343 for( uint i = 0; i < max; ++i ) {
344 Node *m = n->in(i);
345 if (not_a_node(m)) continue;
346 useful.push(m);
347 }
348 }
349 }
351 // Update dead_node_list with any missing dead nodes using useful
352 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
353 void Compile::update_dead_node_list(Unique_Node_List &useful) {
354 uint max_idx = unique();
355 VectorSet& useful_node_set = useful.member_set();
357 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
358 // If node with index node_idx is not in useful set,
359 // mark it as dead in dead node list.
360 if (! useful_node_set.test(node_idx) ) {
361 record_dead_node(node_idx);
362 }
363 }
364 }
366 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
367 int shift = 0;
368 for (int i = 0; i < inlines->length(); i++) {
369 CallGenerator* cg = inlines->at(i);
370 CallNode* call = cg->call_node();
371 if (shift > 0) {
372 inlines->at_put(i-shift, cg);
373 }
374 if (!useful.member(call)) {
375 shift++;
376 }
377 }
378 inlines->trunc_to(inlines->length()-shift);
379 }
381 // Disconnect all useless nodes by disconnecting those at the boundary.
382 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
383 uint next = 0;
384 while (next < useful.size()) {
385 Node *n = useful.at(next++);
386 if (n->is_SafePoint()) {
387 // We're done with a parsing phase. Replaced nodes are not valid
388 // beyond that point.
389 n->as_SafePoint()->delete_replaced_nodes();
390 }
391 // Use raw traversal of out edges since this code removes out edges
392 int max = n->outcnt();
393 for (int j = 0; j < max; ++j) {
394 Node* child = n->raw_out(j);
395 if (! useful.member(child)) {
396 assert(!child->is_top() || child != top(),
397 "If top is cached in Compile object it is in useful list");
398 // Only need to remove this out-edge to the useless node
399 n->raw_del_out(j);
400 --j;
401 --max;
402 }
403 }
404 if (n->outcnt() == 1 && n->has_special_unique_user()) {
405 record_for_igvn(n->unique_out());
406 }
407 }
408 // Remove useless macro and predicate opaq nodes
409 for (int i = C->macro_count()-1; i >= 0; i--) {
410 Node* n = C->macro_node(i);
411 if (!useful.member(n)) {
412 remove_macro_node(n);
413 }
414 }
415 // Remove useless CastII nodes with range check dependency
416 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
417 Node* cast = range_check_cast_node(i);
418 if (!useful.member(cast)) {
419 remove_range_check_cast(cast);
420 }
421 }
422 // Remove useless expensive node
423 for (int i = C->expensive_count()-1; i >= 0; i--) {
424 Node* n = C->expensive_node(i);
425 if (!useful.member(n)) {
426 remove_expensive_node(n);
427 }
428 }
429 // clean up the late inline lists
430 remove_useless_late_inlines(&_string_late_inlines, useful);
431 remove_useless_late_inlines(&_boxing_late_inlines, useful);
432 remove_useless_late_inlines(&_late_inlines, useful);
433 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
434 }
436 //------------------------------frame_size_in_words-----------------------------
437 // frame_slots in units of words
438 int Compile::frame_size_in_words() const {
439 // shift is 0 in LP32 and 1 in LP64
440 const int shift = (LogBytesPerWord - LogBytesPerInt);
441 int words = _frame_slots >> shift;
442 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
443 return words;
444 }
446 // To bang the stack of this compiled method we use the stack size
447 // that the interpreter would need in case of a deoptimization. This
448 // removes the need to bang the stack in the deoptimization blob which
449 // in turn simplifies stack overflow handling.
450 int Compile::bang_size_in_bytes() const {
451 return MAX2(_interpreter_frame_size, frame_size_in_bytes());
452 }
454 // ============================================================================
455 //------------------------------CompileWrapper---------------------------------
456 class CompileWrapper : public StackObj {
457 Compile *const _compile;
458 public:
459 CompileWrapper(Compile* compile);
461 ~CompileWrapper();
462 };
464 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
465 // the Compile* pointer is stored in the current ciEnv:
466 ciEnv* env = compile->env();
467 assert(env == ciEnv::current(), "must already be a ciEnv active");
468 assert(env->compiler_data() == NULL, "compile already active?");
469 env->set_compiler_data(compile);
470 assert(compile == Compile::current(), "sanity");
472 compile->set_type_dict(NULL);
473 compile->set_type_hwm(NULL);
474 compile->set_type_last_size(0);
475 compile->set_last_tf(NULL, NULL);
476 compile->set_indexSet_arena(NULL);
477 compile->set_indexSet_free_block_list(NULL);
478 compile->init_type_arena();
479 Type::Initialize(compile);
480 _compile->set_scratch_buffer_blob(NULL);
481 _compile->begin_method();
482 }
483 CompileWrapper::~CompileWrapper() {
484 _compile->end_method();
485 if (_compile->scratch_buffer_blob() != NULL)
486 BufferBlob::free(_compile->scratch_buffer_blob());
487 _compile->env()->set_compiler_data(NULL);
488 }
491 //----------------------------print_compile_messages---------------------------
492 void Compile::print_compile_messages() {
493 #ifndef PRODUCT
494 // Check if recompiling
495 if (_subsume_loads == false && PrintOpto) {
496 // Recompiling without allowing machine instructions to subsume loads
497 tty->print_cr("*********************************************************");
498 tty->print_cr("** Bailout: Recompile without subsuming loads **");
499 tty->print_cr("*********************************************************");
500 }
501 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
502 // Recompiling without escape analysis
503 tty->print_cr("*********************************************************");
504 tty->print_cr("** Bailout: Recompile without escape analysis **");
505 tty->print_cr("*********************************************************");
506 }
507 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
508 // Recompiling without boxing elimination
509 tty->print_cr("*********************************************************");
510 tty->print_cr("** Bailout: Recompile without boxing elimination **");
511 tty->print_cr("*********************************************************");
512 }
513 if (env()->break_at_compile()) {
514 // Open the debugger when compiling this method.
515 tty->print("### Breaking when compiling: ");
516 method()->print_short_name();
517 tty->cr();
518 BREAKPOINT;
519 }
521 if( PrintOpto ) {
522 if (is_osr_compilation()) {
523 tty->print("[OSR]%3d", _compile_id);
524 } else {
525 tty->print("%3d", _compile_id);
526 }
527 }
528 #endif
529 }
532 //-----------------------init_scratch_buffer_blob------------------------------
533 // Construct a temporary BufferBlob and cache it for this compile.
534 void Compile::init_scratch_buffer_blob(int const_size) {
535 // If there is already a scratch buffer blob allocated and the
536 // constant section is big enough, use it. Otherwise free the
537 // current and allocate a new one.
538 BufferBlob* blob = scratch_buffer_blob();
539 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
540 // Use the current blob.
541 } else {
542 if (blob != NULL) {
543 BufferBlob::free(blob);
544 }
546 ResourceMark rm;
547 _scratch_const_size = const_size;
548 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
549 blob = BufferBlob::create("Compile::scratch_buffer", size);
550 // Record the buffer blob for next time.
551 set_scratch_buffer_blob(blob);
552 // Have we run out of code space?
553 if (scratch_buffer_blob() == NULL) {
554 // Let CompilerBroker disable further compilations.
555 record_failure("Not enough space for scratch buffer in CodeCache");
556 return;
557 }
558 }
560 // Initialize the relocation buffers
561 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
562 set_scratch_locs_memory(locs_buf);
563 }
566 //-----------------------scratch_emit_size-------------------------------------
567 // Helper function that computes size by emitting code
568 uint Compile::scratch_emit_size(const Node* n) {
569 // Start scratch_emit_size section.
570 set_in_scratch_emit_size(true);
572 // Emit into a trash buffer and count bytes emitted.
573 // This is a pretty expensive way to compute a size,
574 // but it works well enough if seldom used.
575 // All common fixed-size instructions are given a size
576 // method by the AD file.
577 // Note that the scratch buffer blob and locs memory are
578 // allocated at the beginning of the compile task, and
579 // may be shared by several calls to scratch_emit_size.
580 // The allocation of the scratch buffer blob is particularly
581 // expensive, since it has to grab the code cache lock.
582 BufferBlob* blob = this->scratch_buffer_blob();
583 assert(blob != NULL, "Initialize BufferBlob at start");
584 assert(blob->size() > MAX_inst_size, "sanity");
585 relocInfo* locs_buf = scratch_locs_memory();
586 address blob_begin = blob->content_begin();
587 address blob_end = (address)locs_buf;
588 assert(blob->content_contains(blob_end), "sanity");
589 CodeBuffer buf(blob_begin, blob_end - blob_begin);
590 buf.initialize_consts_size(_scratch_const_size);
591 buf.initialize_stubs_size(MAX_stubs_size);
592 assert(locs_buf != NULL, "sanity");
593 int lsize = MAX_locs_size / 3;
594 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
595 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
596 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
598 // Do the emission.
600 Label fakeL; // Fake label for branch instructions.
601 Label* saveL = NULL;
602 uint save_bnum = 0;
603 bool is_branch = n->is_MachBranch();
604 if (is_branch) {
605 MacroAssembler masm(&buf);
606 masm.bind(fakeL);
607 n->as_MachBranch()->save_label(&saveL, &save_bnum);
608 n->as_MachBranch()->label_set(&fakeL, 0);
609 }
610 n->emit(buf, this->regalloc());
611 if (is_branch) // Restore label.
612 n->as_MachBranch()->label_set(saveL, save_bnum);
614 // End scratch_emit_size section.
615 set_in_scratch_emit_size(false);
617 return buf.insts_size();
618 }
621 // ============================================================================
622 //------------------------------Compile standard-------------------------------
623 debug_only( int Compile::_debug_idx = 100000; )
625 // Compile a method. entry_bci is -1 for normal compilations and indicates
626 // the continuation bci for on stack replacement.
629 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
630 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
631 : Phase(Compiler),
632 _env(ci_env),
633 _log(ci_env->log()),
634 _compile_id(ci_env->compile_id()),
635 _save_argument_registers(false),
636 _stub_name(NULL),
637 _stub_function(NULL),
638 _stub_entry_point(NULL),
639 _method(target),
640 _entry_bci(osr_bci),
641 _initial_gvn(NULL),
642 _for_igvn(NULL),
643 _warm_calls(NULL),
644 _subsume_loads(subsume_loads),
645 _do_escape_analysis(do_escape_analysis),
646 _eliminate_boxing(eliminate_boxing),
647 _failure_reason(NULL),
648 _code_buffer("Compile::Fill_buffer"),
649 _orig_pc_slot(0),
650 _orig_pc_slot_offset_in_bytes(0),
651 _has_method_handle_invokes(false),
652 _mach_constant_base_node(NULL),
653 _node_bundling_limit(0),
654 _node_bundling_base(NULL),
655 _java_calls(0),
656 _inner_loops(0),
657 _scratch_const_size(-1),
658 _in_scratch_emit_size(false),
659 _dead_node_list(comp_arena()),
660 _dead_node_count(0),
661 #ifndef PRODUCT
662 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
663 _in_dump_cnt(0),
664 _printer(IdealGraphPrinter::printer()),
665 #endif
666 _congraph(NULL),
667 _comp_arena(mtCompiler),
668 _node_arena(mtCompiler),
669 _old_arena(mtCompiler),
670 _Compile_types(mtCompiler),
671 _replay_inline_data(NULL),
672 _late_inlines(comp_arena(), 2, 0, NULL),
673 _string_late_inlines(comp_arena(), 2, 0, NULL),
674 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
675 _late_inlines_pos(0),
676 _number_of_mh_late_inlines(0),
677 _inlining_progress(false),
678 _inlining_incrementally(false),
679 _print_inlining_list(NULL),
680 _print_inlining_idx(0),
681 _interpreter_frame_size(0),
682 _max_node_limit(MaxNodeLimit) {
683 C = this;
685 CompileWrapper cw(this);
686 #ifndef PRODUCT
687 if (TimeCompiler2) {
688 tty->print(" ");
689 target->holder()->name()->print();
690 tty->print(".");
691 target->print_short_name();
692 tty->print(" ");
693 }
694 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
695 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
696 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
697 if (!print_opto_assembly) {
698 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
699 if (print_assembly && !Disassembler::can_decode()) {
700 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
701 print_opto_assembly = true;
702 }
703 }
704 set_print_assembly(print_opto_assembly);
705 set_parsed_irreducible_loop(false);
707 if (method()->has_option("ReplayInline")) {
708 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
709 }
710 #endif
711 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
712 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
713 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
715 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
716 // Make sure the method being compiled gets its own MDO,
717 // so we can at least track the decompile_count().
718 // Need MDO to record RTM code generation state.
719 method()->ensure_method_data();
720 }
722 Init(::AliasLevel);
725 print_compile_messages();
727 _ilt = InlineTree::build_inline_tree_root();
729 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
730 assert(num_alias_types() >= AliasIdxRaw, "");
732 #define MINIMUM_NODE_HASH 1023
733 // Node list that Iterative GVN will start with
734 Unique_Node_List for_igvn(comp_arena());
735 set_for_igvn(&for_igvn);
737 // GVN that will be run immediately on new nodes
738 uint estimated_size = method()->code_size()*4+64;
739 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
740 PhaseGVN gvn(node_arena(), estimated_size);
741 set_initial_gvn(&gvn);
743 if (print_inlining() || print_intrinsics()) {
744 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
745 }
746 { // Scope for timing the parser
747 TracePhase t3("parse", &_t_parser, true);
749 // Put top into the hash table ASAP.
750 initial_gvn()->transform_no_reclaim(top());
752 // Set up tf(), start(), and find a CallGenerator.
753 CallGenerator* cg = NULL;
754 if (is_osr_compilation()) {
755 const TypeTuple *domain = StartOSRNode::osr_domain();
756 const TypeTuple *range = TypeTuple::make_range(method()->signature());
757 init_tf(TypeFunc::make(domain, range));
758 StartNode* s = new (this) StartOSRNode(root(), domain);
759 initial_gvn()->set_type_bottom(s);
760 init_start(s);
761 cg = CallGenerator::for_osr(method(), entry_bci());
762 } else {
763 // Normal case.
764 init_tf(TypeFunc::make(method()));
765 StartNode* s = new (this) StartNode(root(), tf()->domain());
766 initial_gvn()->set_type_bottom(s);
767 init_start(s);
768 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
769 // With java.lang.ref.reference.get() we must go through the
770 // intrinsic when G1 is enabled - even when get() is the root
771 // method of the compile - so that, if necessary, the value in
772 // the referent field of the reference object gets recorded by
773 // the pre-barrier code.
774 // Specifically, if G1 is enabled, the value in the referent
775 // field is recorded by the G1 SATB pre barrier. This will
776 // result in the referent being marked live and the reference
777 // object removed from the list of discovered references during
778 // reference processing.
779 cg = find_intrinsic(method(), false);
780 }
781 if (cg == NULL) {
782 float past_uses = method()->interpreter_invocation_count();
783 float expected_uses = past_uses;
784 cg = CallGenerator::for_inline(method(), expected_uses);
785 }
786 }
787 if (failing()) return;
788 if (cg == NULL) {
789 record_method_not_compilable_all_tiers("cannot parse method");
790 return;
791 }
792 JVMState* jvms = build_start_state(start(), tf());
793 if ((jvms = cg->generate(jvms)) == NULL) {
794 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
795 record_method_not_compilable("method parse failed");
796 }
797 return;
798 }
799 GraphKit kit(jvms);
801 if (!kit.stopped()) {
802 // Accept return values, and transfer control we know not where.
803 // This is done by a special, unique ReturnNode bound to root.
804 return_values(kit.jvms());
805 }
807 if (kit.has_exceptions()) {
808 // Any exceptions that escape from this call must be rethrown
809 // to whatever caller is dynamically above us on the stack.
810 // This is done by a special, unique RethrowNode bound to root.
811 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
812 }
814 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
816 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
817 inline_string_calls(true);
818 }
820 if (failing()) return;
822 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
824 // Remove clutter produced by parsing.
825 if (!failing()) {
826 ResourceMark rm;
827 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
828 }
829 }
831 // Note: Large methods are capped off in do_one_bytecode().
832 if (failing()) return;
834 // After parsing, node notes are no longer automagic.
835 // They must be propagated by register_new_node_with_optimizer(),
836 // clone(), or the like.
837 set_default_node_notes(NULL);
839 for (;;) {
840 int successes = Inline_Warm();
841 if (failing()) return;
842 if (successes == 0) break;
843 }
845 // Drain the list.
846 Finish_Warm();
847 #ifndef PRODUCT
848 if (_printer) {
849 _printer->print_inlining(this);
850 }
851 #endif
853 if (failing()) return;
854 NOT_PRODUCT( verify_graph_edges(); )
856 // Now optimize
857 Optimize();
858 if (failing()) return;
859 NOT_PRODUCT( verify_graph_edges(); )
861 #ifndef PRODUCT
862 if (PrintIdeal) {
863 ttyLocker ttyl; // keep the following output all in one block
864 // This output goes directly to the tty, not the compiler log.
865 // To enable tools to match it up with the compilation activity,
866 // be sure to tag this tty output with the compile ID.
867 if (xtty != NULL) {
868 xtty->head("ideal compile_id='%d'%s", compile_id(),
869 is_osr_compilation() ? " compile_kind='osr'" :
870 "");
871 }
872 root()->dump(9999);
873 if (xtty != NULL) {
874 xtty->tail("ideal");
875 }
876 }
877 #endif
879 NOT_PRODUCT( verify_barriers(); )
881 // Dump compilation data to replay it.
882 if (method()->has_option("DumpReplay")) {
883 env()->dump_replay_data(_compile_id);
884 }
885 if (method()->has_option("DumpInline") && (ilt() != NULL)) {
886 env()->dump_inline_data(_compile_id);
887 }
889 // Now that we know the size of all the monitors we can add a fixed slot
890 // for the original deopt pc.
892 _orig_pc_slot = fixed_slots();
893 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
894 set_fixed_slots(next_slot);
896 // Compute when to use implicit null checks. Used by matching trap based
897 // nodes and NullCheck optimization.
898 set_allowed_deopt_reasons();
900 // Now generate code
901 Code_Gen();
902 if (failing()) return;
904 // Check if we want to skip execution of all compiled code.
905 {
906 #ifndef PRODUCT
907 if (OptoNoExecute) {
908 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
909 return;
910 }
911 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
912 #endif
914 if (is_osr_compilation()) {
915 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
916 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
917 } else {
918 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
919 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
920 }
922 env()->register_method(_method, _entry_bci,
923 &_code_offsets,
924 _orig_pc_slot_offset_in_bytes,
925 code_buffer(),
926 frame_size_in_words(), _oop_map_set,
927 &_handler_table, &_inc_table,
928 compiler,
929 env()->comp_level(),
930 has_unsafe_access(),
931 SharedRuntime::is_wide_vector(max_vector_size()),
932 rtm_state()
933 );
935 if (log() != NULL) // Print code cache state into compiler log
936 log()->code_cache_state();
937 }
938 }
940 //------------------------------Compile----------------------------------------
941 // Compile a runtime stub
942 Compile::Compile( ciEnv* ci_env,
943 TypeFunc_generator generator,
944 address stub_function,
945 const char *stub_name,
946 int is_fancy_jump,
947 bool pass_tls,
948 bool save_arg_registers,
949 bool return_pc )
950 : Phase(Compiler),
951 _env(ci_env),
952 _log(ci_env->log()),
953 _compile_id(0),
954 _save_argument_registers(save_arg_registers),
955 _method(NULL),
956 _stub_name(stub_name),
957 _stub_function(stub_function),
958 _stub_entry_point(NULL),
959 _entry_bci(InvocationEntryBci),
960 _initial_gvn(NULL),
961 _for_igvn(NULL),
962 _warm_calls(NULL),
963 _orig_pc_slot(0),
964 _orig_pc_slot_offset_in_bytes(0),
965 _subsume_loads(true),
966 _do_escape_analysis(false),
967 _eliminate_boxing(false),
968 _failure_reason(NULL),
969 _code_buffer("Compile::Fill_buffer"),
970 _has_method_handle_invokes(false),
971 _mach_constant_base_node(NULL),
972 _node_bundling_limit(0),
973 _node_bundling_base(NULL),
974 _java_calls(0),
975 _inner_loops(0),
976 #ifndef PRODUCT
977 _trace_opto_output(TraceOptoOutput),
978 _in_dump_cnt(0),
979 _printer(NULL),
980 #endif
981 _comp_arena(mtCompiler),
982 _node_arena(mtCompiler),
983 _old_arena(mtCompiler),
984 _Compile_types(mtCompiler),
985 _dead_node_list(comp_arena()),
986 _dead_node_count(0),
987 _congraph(NULL),
988 _replay_inline_data(NULL),
989 _number_of_mh_late_inlines(0),
990 _inlining_progress(false),
991 _inlining_incrementally(false),
992 _print_inlining_list(NULL),
993 _print_inlining_idx(0),
994 _allowed_reasons(0),
995 _interpreter_frame_size(0),
996 _max_node_limit(MaxNodeLimit) {
997 C = this;
999 #ifndef PRODUCT
1000 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
1001 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
1002 set_print_assembly(PrintFrameConverterAssembly);
1003 set_parsed_irreducible_loop(false);
1004 #endif
1005 set_has_irreducible_loop(false); // no loops
1007 CompileWrapper cw(this);
1008 Init(/*AliasLevel=*/ 0);
1009 init_tf((*generator)());
1011 {
1012 // The following is a dummy for the sake of GraphKit::gen_stub
1013 Unique_Node_List for_igvn(comp_arena());
1014 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1015 PhaseGVN gvn(Thread::current()->resource_area(),255);
1016 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1017 gvn.transform_no_reclaim(top());
1019 GraphKit kit;
1020 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1021 }
1023 NOT_PRODUCT( verify_graph_edges(); )
1024 Code_Gen();
1025 if (failing()) return;
1028 // Entry point will be accessed using compile->stub_entry_point();
1029 if (code_buffer() == NULL) {
1030 Matcher::soft_match_failure();
1031 } else {
1032 if (PrintAssembly && (WizardMode || Verbose))
1033 tty->print_cr("### Stub::%s", stub_name);
1035 if (!failing()) {
1036 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1038 // Make the NMethod
1039 // For now we mark the frame as never safe for profile stackwalking
1040 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1041 code_buffer(),
1042 CodeOffsets::frame_never_safe,
1043 // _code_offsets.value(CodeOffsets::Frame_Complete),
1044 frame_size_in_words(),
1045 _oop_map_set,
1046 save_arg_registers);
1047 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1049 _stub_entry_point = rs->entry_point();
1050 }
1051 }
1052 }
1054 //------------------------------Init-------------------------------------------
1055 // Prepare for a single compilation
1056 void Compile::Init(int aliaslevel) {
1057 _unique = 0;
1058 _regalloc = NULL;
1060 _tf = NULL; // filled in later
1061 _top = NULL; // cached later
1062 _matcher = NULL; // filled in later
1063 _cfg = NULL; // filled in later
1065 set_24_bit_selection_and_mode(Use24BitFP, false);
1067 _node_note_array = NULL;
1068 _default_node_notes = NULL;
1070 _immutable_memory = NULL; // filled in at first inquiry
1072 // Globally visible Nodes
1073 // First set TOP to NULL to give safe behavior during creation of RootNode
1074 set_cached_top_node(NULL);
1075 set_root(new (this) RootNode());
1076 // Now that you have a Root to point to, create the real TOP
1077 set_cached_top_node( new (this) ConNode(Type::TOP) );
1078 set_recent_alloc(NULL, NULL);
1080 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1081 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1082 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1083 env()->set_dependencies(new Dependencies(env()));
1085 _fixed_slots = 0;
1086 set_has_split_ifs(false);
1087 set_has_loops(has_method() && method()->has_loops()); // first approximation
1088 set_has_stringbuilder(false);
1089 set_has_boxed_value(false);
1090 _trap_can_recompile = false; // no traps emitted yet
1091 _major_progress = true; // start out assuming good things will happen
1092 set_has_unsafe_access(false);
1093 set_max_vector_size(0);
1094 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1095 set_decompile_count(0);
1097 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1098 set_num_loop_opts(LoopOptsCount);
1099 set_do_inlining(Inline);
1100 set_max_inline_size(MaxInlineSize);
1101 set_freq_inline_size(FreqInlineSize);
1102 set_do_scheduling(OptoScheduling);
1103 set_do_count_invocations(false);
1104 set_do_method_data_update(false);
1105 set_rtm_state(NoRTM); // No RTM lock eliding by default
1106 method_has_option_value("MaxNodeLimit", _max_node_limit);
1107 #if INCLUDE_RTM_OPT
1108 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1109 int rtm_state = method()->method_data()->rtm_state();
1110 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1111 // Don't generate RTM lock eliding code.
1112 set_rtm_state(NoRTM);
1113 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1114 // Generate RTM lock eliding code without abort ratio calculation code.
1115 set_rtm_state(UseRTM);
1116 } else if (UseRTMDeopt) {
1117 // Generate RTM lock eliding code and include abort ratio calculation
1118 // code if UseRTMDeopt is on.
1119 set_rtm_state(ProfileRTM);
1120 }
1121 }
1122 #endif
1123 if (debug_info()->recording_non_safepoints()) {
1124 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1125 (comp_arena(), 8, 0, NULL));
1126 set_default_node_notes(Node_Notes::make(this));
1127 }
1129 // // -- Initialize types before each compile --
1130 // // Update cached type information
1131 // if( _method && _method->constants() )
1132 // Type::update_loaded_types(_method, _method->constants());
1134 // Init alias_type map.
1135 if (!_do_escape_analysis && aliaslevel == 3)
1136 aliaslevel = 2; // No unique types without escape analysis
1137 _AliasLevel = aliaslevel;
1138 const int grow_ats = 16;
1139 _max_alias_types = grow_ats;
1140 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1141 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1142 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1143 {
1144 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1145 }
1146 // Initialize the first few types.
1147 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1148 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1149 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1150 _num_alias_types = AliasIdxRaw+1;
1151 // Zero out the alias type cache.
1152 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1153 // A NULL adr_type hits in the cache right away. Preload the right answer.
1154 probe_alias_cache(NULL)->_index = AliasIdxTop;
1156 _intrinsics = NULL;
1157 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1158 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1159 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1160 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1161 register_library_intrinsics();
1162 }
1164 //---------------------------init_start----------------------------------------
1165 // Install the StartNode on this compile object.
1166 void Compile::init_start(StartNode* s) {
1167 if (failing())
1168 return; // already failing
1169 assert(s == start(), "");
1170 }
1172 StartNode* Compile::start() const {
1173 assert(!failing(), "");
1174 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1175 Node* start = root()->fast_out(i);
1176 if( start->is_Start() )
1177 return start->as_Start();
1178 }
1179 fatal("Did not find Start node!");
1180 return NULL;
1181 }
1183 //-------------------------------immutable_memory-------------------------------------
1184 // Access immutable memory
1185 Node* Compile::immutable_memory() {
1186 if (_immutable_memory != NULL) {
1187 return _immutable_memory;
1188 }
1189 StartNode* s = start();
1190 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1191 Node *p = s->fast_out(i);
1192 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1193 _immutable_memory = p;
1194 return _immutable_memory;
1195 }
1196 }
1197 ShouldNotReachHere();
1198 return NULL;
1199 }
1201 //----------------------set_cached_top_node------------------------------------
1202 // Install the cached top node, and make sure Node::is_top works correctly.
1203 void Compile::set_cached_top_node(Node* tn) {
1204 if (tn != NULL) verify_top(tn);
1205 Node* old_top = _top;
1206 _top = tn;
1207 // Calling Node::setup_is_top allows the nodes the chance to adjust
1208 // their _out arrays.
1209 if (_top != NULL) _top->setup_is_top();
1210 if (old_top != NULL) old_top->setup_is_top();
1211 assert(_top == NULL || top()->is_top(), "");
1212 }
1214 #ifdef ASSERT
1215 uint Compile::count_live_nodes_by_graph_walk() {
1216 Unique_Node_List useful(comp_arena());
1217 // Get useful node list by walking the graph.
1218 identify_useful_nodes(useful);
1219 return useful.size();
1220 }
1222 void Compile::print_missing_nodes() {
1224 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1225 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1226 return;
1227 }
1229 // This is an expensive function. It is executed only when the user
1230 // specifies VerifyIdealNodeCount option or otherwise knows the
1231 // additional work that needs to be done to identify reachable nodes
1232 // by walking the flow graph and find the missing ones using
1233 // _dead_node_list.
1235 Unique_Node_List useful(comp_arena());
1236 // Get useful node list by walking the graph.
1237 identify_useful_nodes(useful);
1239 uint l_nodes = C->live_nodes();
1240 uint l_nodes_by_walk = useful.size();
1242 if (l_nodes != l_nodes_by_walk) {
1243 if (_log != NULL) {
1244 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1245 _log->stamp();
1246 _log->end_head();
1247 }
1248 VectorSet& useful_member_set = useful.member_set();
1249 int last_idx = l_nodes_by_walk;
1250 for (int i = 0; i < last_idx; i++) {
1251 if (useful_member_set.test(i)) {
1252 if (_dead_node_list.test(i)) {
1253 if (_log != NULL) {
1254 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1255 }
1256 if (PrintIdealNodeCount) {
1257 // Print the log message to tty
1258 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1259 useful.at(i)->dump();
1260 }
1261 }
1262 }
1263 else if (! _dead_node_list.test(i)) {
1264 if (_log != NULL) {
1265 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1266 }
1267 if (PrintIdealNodeCount) {
1268 // Print the log message to tty
1269 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1270 }
1271 }
1272 }
1273 if (_log != NULL) {
1274 _log->tail("mismatched_nodes");
1275 }
1276 }
1277 }
1278 #endif
1280 #ifndef PRODUCT
1281 void Compile::verify_top(Node* tn) const {
1282 if (tn != NULL) {
1283 assert(tn->is_Con(), "top node must be a constant");
1284 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1285 assert(tn->in(0) != NULL, "must have live top node");
1286 }
1287 }
1288 #endif
1291 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1293 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1294 guarantee(arr != NULL, "");
1295 int num_blocks = arr->length();
1296 if (grow_by < num_blocks) grow_by = num_blocks;
1297 int num_notes = grow_by * _node_notes_block_size;
1298 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1299 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1300 while (num_notes > 0) {
1301 arr->append(notes);
1302 notes += _node_notes_block_size;
1303 num_notes -= _node_notes_block_size;
1304 }
1305 assert(num_notes == 0, "exact multiple, please");
1306 }
1308 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1309 if (source == NULL || dest == NULL) return false;
1311 if (dest->is_Con())
1312 return false; // Do not push debug info onto constants.
1314 #ifdef ASSERT
1315 // Leave a bread crumb trail pointing to the original node:
1316 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1317 dest->set_debug_orig(source);
1318 }
1319 #endif
1321 if (node_note_array() == NULL)
1322 return false; // Not collecting any notes now.
1324 // This is a copy onto a pre-existing node, which may already have notes.
1325 // If both nodes have notes, do not overwrite any pre-existing notes.
1326 Node_Notes* source_notes = node_notes_at(source->_idx);
1327 if (source_notes == NULL || source_notes->is_clear()) return false;
1328 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1329 if (dest_notes == NULL || dest_notes->is_clear()) {
1330 return set_node_notes_at(dest->_idx, source_notes);
1331 }
1333 Node_Notes merged_notes = (*source_notes);
1334 // The order of operations here ensures that dest notes will win...
1335 merged_notes.update_from(dest_notes);
1336 return set_node_notes_at(dest->_idx, &merged_notes);
1337 }
1340 //--------------------------allow_range_check_smearing-------------------------
1341 // Gating condition for coalescing similar range checks.
1342 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1343 // single covering check that is at least as strong as any of them.
1344 // If the optimization succeeds, the simplified (strengthened) range check
1345 // will always succeed. If it fails, we will deopt, and then give up
1346 // on the optimization.
1347 bool Compile::allow_range_check_smearing() const {
1348 // If this method has already thrown a range-check,
1349 // assume it was because we already tried range smearing
1350 // and it failed.
1351 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1352 return !already_trapped;
1353 }
1356 //------------------------------flatten_alias_type-----------------------------
1357 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1358 int offset = tj->offset();
1359 TypePtr::PTR ptr = tj->ptr();
1361 // Known instance (scalarizable allocation) alias only with itself.
1362 bool is_known_inst = tj->isa_oopptr() != NULL &&
1363 tj->is_oopptr()->is_known_instance();
1365 // Process weird unsafe references.
1366 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1367 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1368 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1369 tj = TypeOopPtr::BOTTOM;
1370 ptr = tj->ptr();
1371 offset = tj->offset();
1372 }
1374 // Array pointers need some flattening
1375 const TypeAryPtr *ta = tj->isa_aryptr();
1376 if (ta && ta->is_stable()) {
1377 // Erase stability property for alias analysis.
1378 tj = ta = ta->cast_to_stable(false);
1379 }
1380 if( ta && is_known_inst ) {
1381 if ( offset != Type::OffsetBot &&
1382 offset > arrayOopDesc::length_offset_in_bytes() ) {
1383 offset = Type::OffsetBot; // Flatten constant access into array body only
1384 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1385 }
1386 } else if( ta && _AliasLevel >= 2 ) {
1387 // For arrays indexed by constant indices, we flatten the alias
1388 // space to include all of the array body. Only the header, klass
1389 // and array length can be accessed un-aliased.
1390 if( offset != Type::OffsetBot ) {
1391 if( ta->const_oop() ) { // MethodData* or Method*
1392 offset = Type::OffsetBot; // Flatten constant access into array body
1393 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1394 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1395 // range is OK as-is.
1396 tj = ta = TypeAryPtr::RANGE;
1397 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1398 tj = TypeInstPtr::KLASS; // all klass loads look alike
1399 ta = TypeAryPtr::RANGE; // generic ignored junk
1400 ptr = TypePtr::BotPTR;
1401 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1402 tj = TypeInstPtr::MARK;
1403 ta = TypeAryPtr::RANGE; // generic ignored junk
1404 ptr = TypePtr::BotPTR;
1405 } else { // Random constant offset into array body
1406 offset = Type::OffsetBot; // Flatten constant access into array body
1407 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1408 }
1409 }
1410 // Arrays of fixed size alias with arrays of unknown size.
1411 if (ta->size() != TypeInt::POS) {
1412 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1413 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1414 }
1415 // Arrays of known objects become arrays of unknown objects.
1416 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1417 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1418 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1419 }
1420 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1421 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1422 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1423 }
1424 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1425 // cannot be distinguished by bytecode alone.
1426 if (ta->elem() == TypeInt::BOOL) {
1427 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1428 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1429 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1430 }
1431 // During the 2nd round of IterGVN, NotNull castings are removed.
1432 // Make sure the Bottom and NotNull variants alias the same.
1433 // Also, make sure exact and non-exact variants alias the same.
1434 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1435 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1436 }
1437 }
1439 // Oop pointers need some flattening
1440 const TypeInstPtr *to = tj->isa_instptr();
1441 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1442 ciInstanceKlass *k = to->klass()->as_instance_klass();
1443 if( ptr == TypePtr::Constant ) {
1444 if (to->klass() != ciEnv::current()->Class_klass() ||
1445 offset < k->size_helper() * wordSize) {
1446 // No constant oop pointers (such as Strings); they alias with
1447 // unknown strings.
1448 assert(!is_known_inst, "not scalarizable allocation");
1449 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1450 }
1451 } else if( is_known_inst ) {
1452 tj = to; // Keep NotNull and klass_is_exact for instance type
1453 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1454 // During the 2nd round of IterGVN, NotNull castings are removed.
1455 // Make sure the Bottom and NotNull variants alias the same.
1456 // Also, make sure exact and non-exact variants alias the same.
1457 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1458 }
1459 if (to->speculative() != NULL) {
1460 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1461 }
1462 // Canonicalize the holder of this field
1463 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1464 // First handle header references such as a LoadKlassNode, even if the
1465 // object's klass is unloaded at compile time (4965979).
1466 if (!is_known_inst) { // Do it only for non-instance types
1467 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1468 }
1469 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1470 // Static fields are in the space above the normal instance
1471 // fields in the java.lang.Class instance.
1472 if (to->klass() != ciEnv::current()->Class_klass()) {
1473 to = NULL;
1474 tj = TypeOopPtr::BOTTOM;
1475 offset = tj->offset();
1476 }
1477 } else {
1478 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1479 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1480 if( is_known_inst ) {
1481 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1482 } else {
1483 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1484 }
1485 }
1486 }
1487 }
1489 // Klass pointers to object array klasses need some flattening
1490 const TypeKlassPtr *tk = tj->isa_klassptr();
1491 if( tk ) {
1492 // If we are referencing a field within a Klass, we need
1493 // to assume the worst case of an Object. Both exact and
1494 // inexact types must flatten to the same alias class so
1495 // use NotNull as the PTR.
1496 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1498 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1499 TypeKlassPtr::OBJECT->klass(),
1500 offset);
1501 }
1503 ciKlass* klass = tk->klass();
1504 if( klass->is_obj_array_klass() ) {
1505 ciKlass* k = TypeAryPtr::OOPS->klass();
1506 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1507 k = TypeInstPtr::BOTTOM->klass();
1508 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1509 }
1511 // Check for precise loads from the primary supertype array and force them
1512 // to the supertype cache alias index. Check for generic array loads from
1513 // the primary supertype array and also force them to the supertype cache
1514 // alias index. Since the same load can reach both, we need to merge
1515 // these 2 disparate memories into the same alias class. Since the
1516 // primary supertype array is read-only, there's no chance of confusion
1517 // where we bypass an array load and an array store.
1518 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1519 if (offset == Type::OffsetBot ||
1520 (offset >= primary_supers_offset &&
1521 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1522 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1523 offset = in_bytes(Klass::secondary_super_cache_offset());
1524 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1525 }
1526 }
1528 // Flatten all Raw pointers together.
1529 if (tj->base() == Type::RawPtr)
1530 tj = TypeRawPtr::BOTTOM;
1532 if (tj->base() == Type::AnyPtr)
1533 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1535 // Flatten all to bottom for now
1536 switch( _AliasLevel ) {
1537 case 0:
1538 tj = TypePtr::BOTTOM;
1539 break;
1540 case 1: // Flatten to: oop, static, field or array
1541 switch (tj->base()) {
1542 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1543 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1544 case Type::AryPtr: // do not distinguish arrays at all
1545 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1546 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1547 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1548 default: ShouldNotReachHere();
1549 }
1550 break;
1551 case 2: // No collapsing at level 2; keep all splits
1552 case 3: // No collapsing at level 3; keep all splits
1553 break;
1554 default:
1555 Unimplemented();
1556 }
1558 offset = tj->offset();
1559 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1561 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1562 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1563 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1564 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1565 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1566 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1567 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1568 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1569 assert( tj->ptr() != TypePtr::TopPTR &&
1570 tj->ptr() != TypePtr::AnyNull &&
1571 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1572 // assert( tj->ptr() != TypePtr::Constant ||
1573 // tj->base() == Type::RawPtr ||
1574 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1576 return tj;
1577 }
1579 void Compile::AliasType::Init(int i, const TypePtr* at) {
1580 _index = i;
1581 _adr_type = at;
1582 _field = NULL;
1583 _element = NULL;
1584 _is_rewritable = true; // default
1585 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1586 if (atoop != NULL && atoop->is_known_instance()) {
1587 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1588 _general_index = Compile::current()->get_alias_index(gt);
1589 } else {
1590 _general_index = 0;
1591 }
1592 }
1594 //---------------------------------print_on------------------------------------
1595 #ifndef PRODUCT
1596 void Compile::AliasType::print_on(outputStream* st) {
1597 if (index() < 10)
1598 st->print("@ <%d> ", index());
1599 else st->print("@ <%d>", index());
1600 st->print(is_rewritable() ? " " : " RO");
1601 int offset = adr_type()->offset();
1602 if (offset == Type::OffsetBot)
1603 st->print(" +any");
1604 else st->print(" +%-3d", offset);
1605 st->print(" in ");
1606 adr_type()->dump_on(st);
1607 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1608 if (field() != NULL && tjp) {
1609 if (tjp->klass() != field()->holder() ||
1610 tjp->offset() != field()->offset_in_bytes()) {
1611 st->print(" != ");
1612 field()->print();
1613 st->print(" ***");
1614 }
1615 }
1616 }
1618 void print_alias_types() {
1619 Compile* C = Compile::current();
1620 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1621 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1622 C->alias_type(idx)->print_on(tty);
1623 tty->cr();
1624 }
1625 }
1626 #endif
1629 //----------------------------probe_alias_cache--------------------------------
1630 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1631 intptr_t key = (intptr_t) adr_type;
1632 key ^= key >> logAliasCacheSize;
1633 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1634 }
1637 //-----------------------------grow_alias_types--------------------------------
1638 void Compile::grow_alias_types() {
1639 const int old_ats = _max_alias_types; // how many before?
1640 const int new_ats = old_ats; // how many more?
1641 const int grow_ats = old_ats+new_ats; // how many now?
1642 _max_alias_types = grow_ats;
1643 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1644 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1645 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1646 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1647 }
1650 //--------------------------------find_alias_type------------------------------
1651 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1652 if (_AliasLevel == 0)
1653 return alias_type(AliasIdxBot);
1655 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1656 if (ace->_adr_type == adr_type) {
1657 return alias_type(ace->_index);
1658 }
1660 // Handle special cases.
1661 if (adr_type == NULL) return alias_type(AliasIdxTop);
1662 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1664 // Do it the slow way.
1665 const TypePtr* flat = flatten_alias_type(adr_type);
1667 #ifdef ASSERT
1668 assert(flat == flatten_alias_type(flat), "idempotent");
1669 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1670 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1671 const TypeOopPtr* foop = flat->is_oopptr();
1672 // Scalarizable allocations have exact klass always.
1673 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1674 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1675 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1676 }
1677 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1678 #endif
1680 int idx = AliasIdxTop;
1681 for (int i = 0; i < num_alias_types(); i++) {
1682 if (alias_type(i)->adr_type() == flat) {
1683 idx = i;
1684 break;
1685 }
1686 }
1688 if (idx == AliasIdxTop) {
1689 if (no_create) return NULL;
1690 // Grow the array if necessary.
1691 if (_num_alias_types == _max_alias_types) grow_alias_types();
1692 // Add a new alias type.
1693 idx = _num_alias_types++;
1694 _alias_types[idx]->Init(idx, flat);
1695 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1696 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1697 if (flat->isa_instptr()) {
1698 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1699 && flat->is_instptr()->klass() == env()->Class_klass())
1700 alias_type(idx)->set_rewritable(false);
1701 }
1702 if (flat->isa_aryptr()) {
1703 #ifdef ASSERT
1704 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1705 // (T_BYTE has the weakest alignment and size restrictions...)
1706 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1707 #endif
1708 if (flat->offset() == TypePtr::OffsetBot) {
1709 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1710 }
1711 }
1712 if (flat->isa_klassptr()) {
1713 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1714 alias_type(idx)->set_rewritable(false);
1715 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1716 alias_type(idx)->set_rewritable(false);
1717 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1718 alias_type(idx)->set_rewritable(false);
1719 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1720 alias_type(idx)->set_rewritable(false);
1721 }
1722 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1723 // but the base pointer type is not distinctive enough to identify
1724 // references into JavaThread.)
1726 // Check for final fields.
1727 const TypeInstPtr* tinst = flat->isa_instptr();
1728 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1729 ciField* field;
1730 if (tinst->const_oop() != NULL &&
1731 tinst->klass() == ciEnv::current()->Class_klass() &&
1732 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1733 // static field
1734 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1735 field = k->get_field_by_offset(tinst->offset(), true);
1736 } else {
1737 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1738 field = k->get_field_by_offset(tinst->offset(), false);
1739 }
1740 assert(field == NULL ||
1741 original_field == NULL ||
1742 (field->holder() == original_field->holder() &&
1743 field->offset() == original_field->offset() &&
1744 field->is_static() == original_field->is_static()), "wrong field?");
1745 // Set field() and is_rewritable() attributes.
1746 if (field != NULL) alias_type(idx)->set_field(field);
1747 }
1748 }
1750 // Fill the cache for next time.
1751 ace->_adr_type = adr_type;
1752 ace->_index = idx;
1753 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1755 // Might as well try to fill the cache for the flattened version, too.
1756 AliasCacheEntry* face = probe_alias_cache(flat);
1757 if (face->_adr_type == NULL) {
1758 face->_adr_type = flat;
1759 face->_index = idx;
1760 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1761 }
1763 return alias_type(idx);
1764 }
1767 Compile::AliasType* Compile::alias_type(ciField* field) {
1768 const TypeOopPtr* t;
1769 if (field->is_static())
1770 t = TypeInstPtr::make(field->holder()->java_mirror());
1771 else
1772 t = TypeOopPtr::make_from_klass_raw(field->holder());
1773 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1774 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1775 return atp;
1776 }
1779 //------------------------------have_alias_type--------------------------------
1780 bool Compile::have_alias_type(const TypePtr* adr_type) {
1781 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1782 if (ace->_adr_type == adr_type) {
1783 return true;
1784 }
1786 // Handle special cases.
1787 if (adr_type == NULL) return true;
1788 if (adr_type == TypePtr::BOTTOM) return true;
1790 return find_alias_type(adr_type, true, NULL) != NULL;
1791 }
1793 //-----------------------------must_alias--------------------------------------
1794 // True if all values of the given address type are in the given alias category.
1795 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1796 if (alias_idx == AliasIdxBot) return true; // the universal category
1797 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1798 if (alias_idx == AliasIdxTop) return false; // the empty category
1799 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1801 // the only remaining possible overlap is identity
1802 int adr_idx = get_alias_index(adr_type);
1803 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1804 assert(adr_idx == alias_idx ||
1805 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1806 && adr_type != TypeOopPtr::BOTTOM),
1807 "should not be testing for overlap with an unsafe pointer");
1808 return adr_idx == alias_idx;
1809 }
1811 //------------------------------can_alias--------------------------------------
1812 // True if any values of the given address type are in the given alias category.
1813 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1814 if (alias_idx == AliasIdxTop) return false; // the empty category
1815 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1816 if (alias_idx == AliasIdxBot) return true; // the universal category
1817 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1819 // the only remaining possible overlap is identity
1820 int adr_idx = get_alias_index(adr_type);
1821 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1822 return adr_idx == alias_idx;
1823 }
1827 //---------------------------pop_warm_call-------------------------------------
1828 WarmCallInfo* Compile::pop_warm_call() {
1829 WarmCallInfo* wci = _warm_calls;
1830 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1831 return wci;
1832 }
1834 //----------------------------Inline_Warm--------------------------------------
1835 int Compile::Inline_Warm() {
1836 // If there is room, try to inline some more warm call sites.
1837 // %%% Do a graph index compaction pass when we think we're out of space?
1838 if (!InlineWarmCalls) return 0;
1840 int calls_made_hot = 0;
1841 int room_to_grow = NodeCountInliningCutoff - unique();
1842 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1843 int amount_grown = 0;
1844 WarmCallInfo* call;
1845 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1846 int est_size = (int)call->size();
1847 if (est_size > (room_to_grow - amount_grown)) {
1848 // This one won't fit anyway. Get rid of it.
1849 call->make_cold();
1850 continue;
1851 }
1852 call->make_hot();
1853 calls_made_hot++;
1854 amount_grown += est_size;
1855 amount_to_grow -= est_size;
1856 }
1858 if (calls_made_hot > 0) set_major_progress();
1859 return calls_made_hot;
1860 }
1863 //----------------------------Finish_Warm--------------------------------------
1864 void Compile::Finish_Warm() {
1865 if (!InlineWarmCalls) return;
1866 if (failing()) return;
1867 if (warm_calls() == NULL) return;
1869 // Clean up loose ends, if we are out of space for inlining.
1870 WarmCallInfo* call;
1871 while ((call = pop_warm_call()) != NULL) {
1872 call->make_cold();
1873 }
1874 }
1876 //---------------------cleanup_loop_predicates-----------------------
1877 // Remove the opaque nodes that protect the predicates so that all unused
1878 // checks and uncommon_traps will be eliminated from the ideal graph
1879 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1880 if (predicate_count()==0) return;
1881 for (int i = predicate_count(); i > 0; i--) {
1882 Node * n = predicate_opaque1_node(i-1);
1883 assert(n->Opcode() == Op_Opaque1, "must be");
1884 igvn.replace_node(n, n->in(1));
1885 }
1886 assert(predicate_count()==0, "should be clean!");
1887 }
1889 void Compile::add_range_check_cast(Node* n) {
1890 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1891 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1892 _range_check_casts->append(n);
1893 }
1895 // Remove all range check dependent CastIINodes.
1896 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1897 for (int i = range_check_cast_count(); i > 0; i--) {
1898 Node* cast = range_check_cast_node(i-1);
1899 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1900 igvn.replace_node(cast, cast->in(1));
1901 }
1902 assert(range_check_cast_count() == 0, "should be empty");
1903 }
1905 // StringOpts and late inlining of string methods
1906 void Compile::inline_string_calls(bool parse_time) {
1907 {
1908 // remove useless nodes to make the usage analysis simpler
1909 ResourceMark rm;
1910 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1911 }
1913 {
1914 ResourceMark rm;
1915 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1916 PhaseStringOpts pso(initial_gvn(), for_igvn());
1917 print_method(PHASE_AFTER_STRINGOPTS, 3);
1918 }
1920 // now inline anything that we skipped the first time around
1921 if (!parse_time) {
1922 _late_inlines_pos = _late_inlines.length();
1923 }
1925 while (_string_late_inlines.length() > 0) {
1926 CallGenerator* cg = _string_late_inlines.pop();
1927 cg->do_late_inline();
1928 if (failing()) return;
1929 }
1930 _string_late_inlines.trunc_to(0);
1931 }
1933 // Late inlining of boxing methods
1934 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1935 if (_boxing_late_inlines.length() > 0) {
1936 assert(has_boxed_value(), "inconsistent");
1938 PhaseGVN* gvn = initial_gvn();
1939 set_inlining_incrementally(true);
1941 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1942 for_igvn()->clear();
1943 gvn->replace_with(&igvn);
1945 _late_inlines_pos = _late_inlines.length();
1947 while (_boxing_late_inlines.length() > 0) {
1948 CallGenerator* cg = _boxing_late_inlines.pop();
1949 cg->do_late_inline();
1950 if (failing()) return;
1951 }
1952 _boxing_late_inlines.trunc_to(0);
1954 {
1955 ResourceMark rm;
1956 PhaseRemoveUseless pru(gvn, for_igvn());
1957 }
1959 igvn = PhaseIterGVN(gvn);
1960 igvn.optimize();
1962 set_inlining_progress(false);
1963 set_inlining_incrementally(false);
1964 }
1965 }
1967 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1968 assert(IncrementalInline, "incremental inlining should be on");
1969 PhaseGVN* gvn = initial_gvn();
1971 set_inlining_progress(false);
1972 for_igvn()->clear();
1973 gvn->replace_with(&igvn);
1975 int i = 0;
1977 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1978 CallGenerator* cg = _late_inlines.at(i);
1979 _late_inlines_pos = i+1;
1980 cg->do_late_inline();
1981 if (failing()) return;
1982 }
1983 int j = 0;
1984 for (; i < _late_inlines.length(); i++, j++) {
1985 _late_inlines.at_put(j, _late_inlines.at(i));
1986 }
1987 _late_inlines.trunc_to(j);
1989 {
1990 ResourceMark rm;
1991 PhaseRemoveUseless pru(gvn, for_igvn());
1992 }
1994 igvn = PhaseIterGVN(gvn);
1995 }
1997 // Perform incremental inlining until bound on number of live nodes is reached
1998 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1999 PhaseGVN* gvn = initial_gvn();
2001 set_inlining_incrementally(true);
2002 set_inlining_progress(true);
2003 uint low_live_nodes = 0;
2005 while(inlining_progress() && _late_inlines.length() > 0) {
2007 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2008 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2009 // PhaseIdealLoop is expensive so we only try it once we are
2010 // out of live nodes and we only try it again if the previous
2011 // helped got the number of nodes down significantly
2012 PhaseIdealLoop ideal_loop( igvn, false, true );
2013 if (failing()) return;
2014 low_live_nodes = live_nodes();
2015 _major_progress = true;
2016 }
2018 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2019 break;
2020 }
2021 }
2023 inline_incrementally_one(igvn);
2025 if (failing()) return;
2027 igvn.optimize();
2029 if (failing()) return;
2030 }
2032 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2034 if (_string_late_inlines.length() > 0) {
2035 assert(has_stringbuilder(), "inconsistent");
2036 for_igvn()->clear();
2037 initial_gvn()->replace_with(&igvn);
2039 inline_string_calls(false);
2041 if (failing()) return;
2043 {
2044 ResourceMark rm;
2045 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2046 }
2048 igvn = PhaseIterGVN(gvn);
2050 igvn.optimize();
2051 }
2053 set_inlining_incrementally(false);
2054 }
2057 //------------------------------Optimize---------------------------------------
2058 // Given a graph, optimize it.
2059 void Compile::Optimize() {
2060 TracePhase t1("optimizer", &_t_optimizer, true);
2062 #ifndef PRODUCT
2063 if (env()->break_at_compile()) {
2064 BREAKPOINT;
2065 }
2067 #endif
2069 ResourceMark rm;
2070 int loop_opts_cnt;
2072 NOT_PRODUCT( verify_graph_edges(); )
2074 print_method(PHASE_AFTER_PARSING);
2076 {
2077 // Iterative Global Value Numbering, including ideal transforms
2078 // Initialize IterGVN with types and values from parse-time GVN
2079 PhaseIterGVN igvn(initial_gvn());
2080 {
2081 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2082 igvn.optimize();
2083 }
2085 print_method(PHASE_ITER_GVN1, 2);
2087 if (failing()) return;
2089 {
2090 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2091 inline_incrementally(igvn);
2092 }
2094 print_method(PHASE_INCREMENTAL_INLINE, 2);
2096 if (failing()) return;
2098 if (eliminate_boxing()) {
2099 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2100 // Inline valueOf() methods now.
2101 inline_boxing_calls(igvn);
2103 if (AlwaysIncrementalInline) {
2104 inline_incrementally(igvn);
2105 }
2107 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2109 if (failing()) return;
2110 }
2112 // Remove the speculative part of types and clean up the graph from
2113 // the extra CastPP nodes whose only purpose is to carry them. Do
2114 // that early so that optimizations are not disrupted by the extra
2115 // CastPP nodes.
2116 remove_speculative_types(igvn);
2118 // No more new expensive nodes will be added to the list from here
2119 // so keep only the actual candidates for optimizations.
2120 cleanup_expensive_nodes(igvn);
2122 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2123 NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);)
2124 initial_gvn()->replace_with(&igvn);
2125 for_igvn()->clear();
2126 Unique_Node_List new_worklist(C->comp_arena());
2127 {
2128 ResourceMark rm;
2129 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2130 }
2131 set_for_igvn(&new_worklist);
2132 igvn = PhaseIterGVN(initial_gvn());
2133 igvn.optimize();
2134 }
2136 // Perform escape analysis
2137 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2138 if (has_loops()) {
2139 // Cleanup graph (remove dead nodes).
2140 TracePhase t2("idealLoop", &_t_idealLoop, true);
2141 PhaseIdealLoop ideal_loop( igvn, false, true );
2142 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2143 if (failing()) return;
2144 }
2145 ConnectionGraph::do_analysis(this, &igvn);
2147 if (failing()) return;
2149 // Optimize out fields loads from scalar replaceable allocations.
2150 igvn.optimize();
2151 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2153 if (failing()) return;
2155 if (congraph() != NULL && macro_count() > 0) {
2156 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2157 PhaseMacroExpand mexp(igvn);
2158 mexp.eliminate_macro_nodes();
2159 igvn.set_delay_transform(false);
2161 igvn.optimize();
2162 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2164 if (failing()) return;
2165 }
2166 }
2168 // Loop transforms on the ideal graph. Range Check Elimination,
2169 // peeling, unrolling, etc.
2171 // Set loop opts counter
2172 loop_opts_cnt = num_loop_opts();
2173 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2174 {
2175 TracePhase t2("idealLoop", &_t_idealLoop, true);
2176 PhaseIdealLoop ideal_loop( igvn, true );
2177 loop_opts_cnt--;
2178 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2179 if (failing()) return;
2180 }
2181 // Loop opts pass if partial peeling occurred in previous pass
2182 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2183 TracePhase t3("idealLoop", &_t_idealLoop, true);
2184 PhaseIdealLoop ideal_loop( igvn, false );
2185 loop_opts_cnt--;
2186 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2187 if (failing()) return;
2188 }
2189 // Loop opts pass for loop-unrolling before CCP
2190 if(major_progress() && (loop_opts_cnt > 0)) {
2191 TracePhase t4("idealLoop", &_t_idealLoop, true);
2192 PhaseIdealLoop ideal_loop( igvn, false );
2193 loop_opts_cnt--;
2194 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2195 }
2196 if (!failing()) {
2197 // Verify that last round of loop opts produced a valid graph
2198 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2199 PhaseIdealLoop::verify(igvn);
2200 }
2201 }
2202 if (failing()) return;
2204 // Conditional Constant Propagation;
2205 PhaseCCP ccp( &igvn );
2206 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2207 {
2208 TracePhase t2("ccp", &_t_ccp, true);
2209 ccp.do_transform();
2210 }
2211 print_method(PHASE_CPP1, 2);
2213 assert( true, "Break here to ccp.dump_old2new_map()");
2215 // Iterative Global Value Numbering, including ideal transforms
2216 {
2217 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2218 igvn = ccp;
2219 igvn.optimize();
2220 }
2222 print_method(PHASE_ITER_GVN2, 2);
2224 if (failing()) return;
2226 // Loop transforms on the ideal graph. Range Check Elimination,
2227 // peeling, unrolling, etc.
2228 if(loop_opts_cnt > 0) {
2229 debug_only( int cnt = 0; );
2230 while(major_progress() && (loop_opts_cnt > 0)) {
2231 TracePhase t2("idealLoop", &_t_idealLoop, true);
2232 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2233 PhaseIdealLoop ideal_loop( igvn, true);
2234 loop_opts_cnt--;
2235 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2236 if (failing()) return;
2237 }
2238 }
2240 {
2241 // Verify that all previous optimizations produced a valid graph
2242 // at least to this point, even if no loop optimizations were done.
2243 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2244 PhaseIdealLoop::verify(igvn);
2245 }
2247 if (range_check_cast_count() > 0) {
2248 // No more loop optimizations. Remove all range check dependent CastIINodes.
2249 C->remove_range_check_casts(igvn);
2250 igvn.optimize();
2251 }
2253 {
2254 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2255 PhaseMacroExpand mex(igvn);
2256 if (mex.expand_macro_nodes()) {
2257 assert(failing(), "must bail out w/ explicit message");
2258 return;
2259 }
2260 }
2262 } // (End scope of igvn; run destructor if necessary for asserts.)
2264 dump_inlining();
2265 // A method with only infinite loops has no edges entering loops from root
2266 {
2267 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2268 if (final_graph_reshaping()) {
2269 assert(failing(), "must bail out w/ explicit message");
2270 return;
2271 }
2272 }
2274 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2275 }
2278 //------------------------------Code_Gen---------------------------------------
2279 // Given a graph, generate code for it
2280 void Compile::Code_Gen() {
2281 if (failing()) {
2282 return;
2283 }
2285 // Perform instruction selection. You might think we could reclaim Matcher
2286 // memory PDQ, but actually the Matcher is used in generating spill code.
2287 // Internals of the Matcher (including some VectorSets) must remain live
2288 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2289 // set a bit in reclaimed memory.
2291 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2292 // nodes. Mapping is only valid at the root of each matched subtree.
2293 NOT_PRODUCT( verify_graph_edges(); )
2295 Matcher matcher;
2296 _matcher = &matcher;
2297 {
2298 TracePhase t2("matcher", &_t_matcher, true);
2299 matcher.match();
2300 }
2301 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2302 // nodes. Mapping is only valid at the root of each matched subtree.
2303 NOT_PRODUCT( verify_graph_edges(); )
2305 // If you have too many nodes, or if matching has failed, bail out
2306 check_node_count(0, "out of nodes matching instructions");
2307 if (failing()) {
2308 return;
2309 }
2311 // Build a proper-looking CFG
2312 PhaseCFG cfg(node_arena(), root(), matcher);
2313 _cfg = &cfg;
2314 {
2315 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2316 bool success = cfg.do_global_code_motion();
2317 if (!success) {
2318 return;
2319 }
2321 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2322 NOT_PRODUCT( verify_graph_edges(); )
2323 debug_only( cfg.verify(); )
2324 }
2326 PhaseChaitin regalloc(unique(), cfg, matcher);
2327 _regalloc = ®alloc;
2328 {
2329 TracePhase t2("regalloc", &_t_registerAllocation, true);
2330 // Perform register allocation. After Chaitin, use-def chains are
2331 // no longer accurate (at spill code) and so must be ignored.
2332 // Node->LRG->reg mappings are still accurate.
2333 _regalloc->Register_Allocate();
2335 // Bail out if the allocator builds too many nodes
2336 if (failing()) {
2337 return;
2338 }
2339 }
2341 // Prior to register allocation we kept empty basic blocks in case the
2342 // the allocator needed a place to spill. After register allocation we
2343 // are not adding any new instructions. If any basic block is empty, we
2344 // can now safely remove it.
2345 {
2346 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2347 cfg.remove_empty_blocks();
2348 if (do_freq_based_layout()) {
2349 PhaseBlockLayout layout(cfg);
2350 } else {
2351 cfg.set_loop_alignment();
2352 }
2353 cfg.fixup_flow();
2354 }
2356 // Apply peephole optimizations
2357 if( OptoPeephole ) {
2358 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2359 PhasePeephole peep( _regalloc, cfg);
2360 peep.do_transform();
2361 }
2363 // Do late expand if CPU requires this.
2364 if (Matcher::require_postalloc_expand) {
2365 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2366 cfg.postalloc_expand(_regalloc);
2367 }
2369 // Convert Nodes to instruction bits in a buffer
2370 {
2371 // %%%% workspace merge brought two timers together for one job
2372 TracePhase t2a("output", &_t_output, true);
2373 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2374 Output();
2375 }
2377 print_method(PHASE_FINAL_CODE);
2379 // He's dead, Jim.
2380 _cfg = (PhaseCFG*)0xdeadbeef;
2381 _regalloc = (PhaseChaitin*)0xdeadbeef;
2382 }
2385 //------------------------------dump_asm---------------------------------------
2386 // Dump formatted assembly
2387 #ifndef PRODUCT
2388 void Compile::dump_asm(int *pcs, uint pc_limit) {
2389 bool cut_short = false;
2390 tty->print_cr("#");
2391 tty->print("# "); _tf->dump(); tty->cr();
2392 tty->print_cr("#");
2394 // For all blocks
2395 int pc = 0x0; // Program counter
2396 char starts_bundle = ' ';
2397 _regalloc->dump_frame();
2399 Node *n = NULL;
2400 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2401 if (VMThread::should_terminate()) {
2402 cut_short = true;
2403 break;
2404 }
2405 Block* block = _cfg->get_block(i);
2406 if (block->is_connector() && !Verbose) {
2407 continue;
2408 }
2409 n = block->head();
2410 if (pcs && n->_idx < pc_limit) {
2411 tty->print("%3.3x ", pcs[n->_idx]);
2412 } else {
2413 tty->print(" ");
2414 }
2415 block->dump_head(_cfg);
2416 if (block->is_connector()) {
2417 tty->print_cr(" # Empty connector block");
2418 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2419 tty->print_cr(" # Block is sole successor of call");
2420 }
2422 // For all instructions
2423 Node *delay = NULL;
2424 for (uint j = 0; j < block->number_of_nodes(); j++) {
2425 if (VMThread::should_terminate()) {
2426 cut_short = true;
2427 break;
2428 }
2429 n = block->get_node(j);
2430 if (valid_bundle_info(n)) {
2431 Bundle* bundle = node_bundling(n);
2432 if (bundle->used_in_unconditional_delay()) {
2433 delay = n;
2434 continue;
2435 }
2436 if (bundle->starts_bundle()) {
2437 starts_bundle = '+';
2438 }
2439 }
2441 if (WizardMode) {
2442 n->dump();
2443 }
2445 if( !n->is_Region() && // Dont print in the Assembly
2446 !n->is_Phi() && // a few noisely useless nodes
2447 !n->is_Proj() &&
2448 !n->is_MachTemp() &&
2449 !n->is_SafePointScalarObject() &&
2450 !n->is_Catch() && // Would be nice to print exception table targets
2451 !n->is_MergeMem() && // Not very interesting
2452 !n->is_top() && // Debug info table constants
2453 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2454 ) {
2455 if (pcs && n->_idx < pc_limit)
2456 tty->print("%3.3x", pcs[n->_idx]);
2457 else
2458 tty->print(" ");
2459 tty->print(" %c ", starts_bundle);
2460 starts_bundle = ' ';
2461 tty->print("\t");
2462 n->format(_regalloc, tty);
2463 tty->cr();
2464 }
2466 // If we have an instruction with a delay slot, and have seen a delay,
2467 // then back up and print it
2468 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2469 assert(delay != NULL, "no unconditional delay instruction");
2470 if (WizardMode) delay->dump();
2472 if (node_bundling(delay)->starts_bundle())
2473 starts_bundle = '+';
2474 if (pcs && n->_idx < pc_limit)
2475 tty->print("%3.3x", pcs[n->_idx]);
2476 else
2477 tty->print(" ");
2478 tty->print(" %c ", starts_bundle);
2479 starts_bundle = ' ';
2480 tty->print("\t");
2481 delay->format(_regalloc, tty);
2482 tty->cr();
2483 delay = NULL;
2484 }
2486 // Dump the exception table as well
2487 if( n->is_Catch() && (Verbose || WizardMode) ) {
2488 // Print the exception table for this offset
2489 _handler_table.print_subtable_for(pc);
2490 }
2491 }
2493 if (pcs && n->_idx < pc_limit)
2494 tty->print_cr("%3.3x", pcs[n->_idx]);
2495 else
2496 tty->cr();
2498 assert(cut_short || delay == NULL, "no unconditional delay branch");
2500 } // End of per-block dump
2501 tty->cr();
2503 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2504 }
2505 #endif
2507 //------------------------------Final_Reshape_Counts---------------------------
2508 // This class defines counters to help identify when a method
2509 // may/must be executed using hardware with only 24-bit precision.
2510 struct Final_Reshape_Counts : public StackObj {
2511 int _call_count; // count non-inlined 'common' calls
2512 int _float_count; // count float ops requiring 24-bit precision
2513 int _double_count; // count double ops requiring more precision
2514 int _java_call_count; // count non-inlined 'java' calls
2515 int _inner_loop_count; // count loops which need alignment
2516 VectorSet _visited; // Visitation flags
2517 Node_List _tests; // Set of IfNodes & PCTableNodes
2519 Final_Reshape_Counts() :
2520 _call_count(0), _float_count(0), _double_count(0),
2521 _java_call_count(0), _inner_loop_count(0),
2522 _visited( Thread::current()->resource_area() ) { }
2524 void inc_call_count () { _call_count ++; }
2525 void inc_float_count () { _float_count ++; }
2526 void inc_double_count() { _double_count++; }
2527 void inc_java_call_count() { _java_call_count++; }
2528 void inc_inner_loop_count() { _inner_loop_count++; }
2530 int get_call_count () const { return _call_count ; }
2531 int get_float_count () const { return _float_count ; }
2532 int get_double_count() const { return _double_count; }
2533 int get_java_call_count() const { return _java_call_count; }
2534 int get_inner_loop_count() const { return _inner_loop_count; }
2535 };
2537 #ifdef ASSERT
2538 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2539 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2540 // Make sure the offset goes inside the instance layout.
2541 return k->contains_field_offset(tp->offset());
2542 // Note that OffsetBot and OffsetTop are very negative.
2543 }
2544 #endif
2546 // Eliminate trivially redundant StoreCMs and accumulate their
2547 // precedence edges.
2548 void Compile::eliminate_redundant_card_marks(Node* n) {
2549 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2550 if (n->in(MemNode::Address)->outcnt() > 1) {
2551 // There are multiple users of the same address so it might be
2552 // possible to eliminate some of the StoreCMs
2553 Node* mem = n->in(MemNode::Memory);
2554 Node* adr = n->in(MemNode::Address);
2555 Node* val = n->in(MemNode::ValueIn);
2556 Node* prev = n;
2557 bool done = false;
2558 // Walk the chain of StoreCMs eliminating ones that match. As
2559 // long as it's a chain of single users then the optimization is
2560 // safe. Eliminating partially redundant StoreCMs would require
2561 // cloning copies down the other paths.
2562 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2563 if (adr == mem->in(MemNode::Address) &&
2564 val == mem->in(MemNode::ValueIn)) {
2565 // redundant StoreCM
2566 if (mem->req() > MemNode::OopStore) {
2567 // Hasn't been processed by this code yet.
2568 n->add_prec(mem->in(MemNode::OopStore));
2569 } else {
2570 // Already converted to precedence edge
2571 for (uint i = mem->req(); i < mem->len(); i++) {
2572 // Accumulate any precedence edges
2573 if (mem->in(i) != NULL) {
2574 n->add_prec(mem->in(i));
2575 }
2576 }
2577 // Everything above this point has been processed.
2578 done = true;
2579 }
2580 // Eliminate the previous StoreCM
2581 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2582 assert(mem->outcnt() == 0, "should be dead");
2583 mem->disconnect_inputs(NULL, this);
2584 } else {
2585 prev = mem;
2586 }
2587 mem = prev->in(MemNode::Memory);
2588 }
2589 }
2590 }
2592 //------------------------------final_graph_reshaping_impl----------------------
2593 // Implement items 1-5 from final_graph_reshaping below.
2594 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2596 if ( n->outcnt() == 0 ) return; // dead node
2597 uint nop = n->Opcode();
2599 // Check for 2-input instruction with "last use" on right input.
2600 // Swap to left input. Implements item (2).
2601 if( n->req() == 3 && // two-input instruction
2602 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2603 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2604 n->in(2)->outcnt() == 1 &&// right use IS a last use
2605 !n->in(2)->is_Con() ) { // right use is not a constant
2606 // Check for commutative opcode
2607 switch( nop ) {
2608 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2609 case Op_MaxI: case Op_MinI:
2610 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2611 case Op_AndL: case Op_XorL: case Op_OrL:
2612 case Op_AndI: case Op_XorI: case Op_OrI: {
2613 // Move "last use" input to left by swapping inputs
2614 n->swap_edges(1, 2);
2615 break;
2616 }
2617 default:
2618 break;
2619 }
2620 }
2622 #ifdef ASSERT
2623 if( n->is_Mem() ) {
2624 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2625 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2626 // oop will be recorded in oop map if load crosses safepoint
2627 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2628 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2629 "raw memory operations should have control edge");
2630 }
2631 #endif
2632 // Count FPU ops and common calls, implements item (3)
2633 switch( nop ) {
2634 // Count all float operations that may use FPU
2635 case Op_AddF:
2636 case Op_SubF:
2637 case Op_MulF:
2638 case Op_DivF:
2639 case Op_NegF:
2640 case Op_ModF:
2641 case Op_ConvI2F:
2642 case Op_ConF:
2643 case Op_CmpF:
2644 case Op_CmpF3:
2645 // case Op_ConvL2F: // longs are split into 32-bit halves
2646 frc.inc_float_count();
2647 break;
2649 case Op_ConvF2D:
2650 case Op_ConvD2F:
2651 frc.inc_float_count();
2652 frc.inc_double_count();
2653 break;
2655 // Count all double operations that may use FPU
2656 case Op_AddD:
2657 case Op_SubD:
2658 case Op_MulD:
2659 case Op_DivD:
2660 case Op_NegD:
2661 case Op_ModD:
2662 case Op_ConvI2D:
2663 case Op_ConvD2I:
2664 // case Op_ConvL2D: // handled by leaf call
2665 // case Op_ConvD2L: // handled by leaf call
2666 case Op_ConD:
2667 case Op_CmpD:
2668 case Op_CmpD3:
2669 frc.inc_double_count();
2670 break;
2671 case Op_Opaque1: // Remove Opaque Nodes before matching
2672 case Op_Opaque2: // Remove Opaque Nodes before matching
2673 case Op_Opaque3:
2674 n->subsume_by(n->in(1), this);
2675 break;
2676 case Op_CallStaticJava:
2677 case Op_CallJava:
2678 case Op_CallDynamicJava:
2679 frc.inc_java_call_count(); // Count java call site;
2680 case Op_CallRuntime:
2681 case Op_CallLeaf:
2682 case Op_CallLeafNoFP: {
2683 assert( n->is_Call(), "" );
2684 CallNode *call = n->as_Call();
2685 // Count call sites where the FP mode bit would have to be flipped.
2686 // Do not count uncommon runtime calls:
2687 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2688 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2689 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2690 frc.inc_call_count(); // Count the call site
2691 } else { // See if uncommon argument is shared
2692 Node *n = call->in(TypeFunc::Parms);
2693 int nop = n->Opcode();
2694 // Clone shared simple arguments to uncommon calls, item (1).
2695 if( n->outcnt() > 1 &&
2696 !n->is_Proj() &&
2697 nop != Op_CreateEx &&
2698 nop != Op_CheckCastPP &&
2699 nop != Op_DecodeN &&
2700 nop != Op_DecodeNKlass &&
2701 !n->is_Mem() ) {
2702 Node *x = n->clone();
2703 call->set_req( TypeFunc::Parms, x );
2704 }
2705 }
2706 break;
2707 }
2709 case Op_StoreD:
2710 case Op_LoadD:
2711 case Op_LoadD_unaligned:
2712 frc.inc_double_count();
2713 goto handle_mem;
2714 case Op_StoreF:
2715 case Op_LoadF:
2716 frc.inc_float_count();
2717 goto handle_mem;
2719 case Op_StoreCM:
2720 {
2721 // Convert OopStore dependence into precedence edge
2722 Node* prec = n->in(MemNode::OopStore);
2723 n->del_req(MemNode::OopStore);
2724 n->add_prec(prec);
2725 eliminate_redundant_card_marks(n);
2726 }
2728 // fall through
2730 case Op_StoreB:
2731 case Op_StoreC:
2732 case Op_StorePConditional:
2733 case Op_StoreI:
2734 case Op_StoreL:
2735 case Op_StoreIConditional:
2736 case Op_StoreLConditional:
2737 case Op_CompareAndSwapI:
2738 case Op_CompareAndSwapL:
2739 case Op_CompareAndSwapP:
2740 case Op_CompareAndSwapN:
2741 case Op_GetAndAddI:
2742 case Op_GetAndAddL:
2743 case Op_GetAndSetI:
2744 case Op_GetAndSetL:
2745 case Op_GetAndSetP:
2746 case Op_GetAndSetN:
2747 case Op_StoreP:
2748 case Op_StoreN:
2749 case Op_StoreNKlass:
2750 case Op_LoadB:
2751 case Op_LoadUB:
2752 case Op_LoadUS:
2753 case Op_LoadI:
2754 case Op_LoadKlass:
2755 case Op_LoadNKlass:
2756 case Op_LoadL:
2757 case Op_LoadL_unaligned:
2758 case Op_LoadPLocked:
2759 case Op_LoadP:
2760 case Op_LoadN:
2761 case Op_LoadRange:
2762 case Op_LoadS: {
2763 handle_mem:
2764 #ifdef ASSERT
2765 if( VerifyOptoOopOffsets ) {
2766 assert( n->is_Mem(), "" );
2767 MemNode *mem = (MemNode*)n;
2768 // Check to see if address types have grounded out somehow.
2769 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2770 assert( !tp || oop_offset_is_sane(tp), "" );
2771 }
2772 #endif
2773 break;
2774 }
2776 case Op_AddP: { // Assert sane base pointers
2777 Node *addp = n->in(AddPNode::Address);
2778 assert( !addp->is_AddP() ||
2779 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2780 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2781 "Base pointers must match" );
2782 #ifdef _LP64
2783 if ((UseCompressedOops || UseCompressedClassPointers) &&
2784 addp->Opcode() == Op_ConP &&
2785 addp == n->in(AddPNode::Base) &&
2786 n->in(AddPNode::Offset)->is_Con()) {
2787 // Use addressing with narrow klass to load with offset on x86.
2788 // On sparc loading 32-bits constant and decoding it have less
2789 // instructions (4) then load 64-bits constant (7).
2790 // Do this transformation here since IGVN will convert ConN back to ConP.
2791 const Type* t = addp->bottom_type();
2792 if (t->isa_oopptr() || t->isa_klassptr()) {
2793 Node* nn = NULL;
2795 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2797 // Look for existing ConN node of the same exact type.
2798 Node* r = root();
2799 uint cnt = r->outcnt();
2800 for (uint i = 0; i < cnt; i++) {
2801 Node* m = r->raw_out(i);
2802 if (m!= NULL && m->Opcode() == op &&
2803 m->bottom_type()->make_ptr() == t) {
2804 nn = m;
2805 break;
2806 }
2807 }
2808 if (nn != NULL) {
2809 // Decode a narrow oop to match address
2810 // [R12 + narrow_oop_reg<<3 + offset]
2811 if (t->isa_oopptr()) {
2812 nn = new (this) DecodeNNode(nn, t);
2813 } else {
2814 nn = new (this) DecodeNKlassNode(nn, t);
2815 }
2816 n->set_req(AddPNode::Base, nn);
2817 n->set_req(AddPNode::Address, nn);
2818 if (addp->outcnt() == 0) {
2819 addp->disconnect_inputs(NULL, this);
2820 }
2821 }
2822 }
2823 }
2824 #endif
2825 break;
2826 }
2828 #ifdef _LP64
2829 case Op_CastPP:
2830 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2831 Node* in1 = n->in(1);
2832 const Type* t = n->bottom_type();
2833 Node* new_in1 = in1->clone();
2834 new_in1->as_DecodeN()->set_type(t);
2836 if (!Matcher::narrow_oop_use_complex_address()) {
2837 //
2838 // x86, ARM and friends can handle 2 adds in addressing mode
2839 // and Matcher can fold a DecodeN node into address by using
2840 // a narrow oop directly and do implicit NULL check in address:
2841 //
2842 // [R12 + narrow_oop_reg<<3 + offset]
2843 // NullCheck narrow_oop_reg
2844 //
2845 // On other platforms (Sparc) we have to keep new DecodeN node and
2846 // use it to do implicit NULL check in address:
2847 //
2848 // decode_not_null narrow_oop_reg, base_reg
2849 // [base_reg + offset]
2850 // NullCheck base_reg
2851 //
2852 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2853 // to keep the information to which NULL check the new DecodeN node
2854 // corresponds to use it as value in implicit_null_check().
2855 //
2856 new_in1->set_req(0, n->in(0));
2857 }
2859 n->subsume_by(new_in1, this);
2860 if (in1->outcnt() == 0) {
2861 in1->disconnect_inputs(NULL, this);
2862 }
2863 }
2864 break;
2866 case Op_CmpP:
2867 // Do this transformation here to preserve CmpPNode::sub() and
2868 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2869 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2870 Node* in1 = n->in(1);
2871 Node* in2 = n->in(2);
2872 if (!in1->is_DecodeNarrowPtr()) {
2873 in2 = in1;
2874 in1 = n->in(2);
2875 }
2876 assert(in1->is_DecodeNarrowPtr(), "sanity");
2878 Node* new_in2 = NULL;
2879 if (in2->is_DecodeNarrowPtr()) {
2880 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2881 new_in2 = in2->in(1);
2882 } else if (in2->Opcode() == Op_ConP) {
2883 const Type* t = in2->bottom_type();
2884 if (t == TypePtr::NULL_PTR) {
2885 assert(in1->is_DecodeN(), "compare klass to null?");
2886 // Don't convert CmpP null check into CmpN if compressed
2887 // oops implicit null check is not generated.
2888 // This will allow to generate normal oop implicit null check.
2889 if (Matcher::gen_narrow_oop_implicit_null_checks())
2890 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2891 //
2892 // This transformation together with CastPP transformation above
2893 // will generated code for implicit NULL checks for compressed oops.
2894 //
2895 // The original code after Optimize()
2896 //
2897 // LoadN memory, narrow_oop_reg
2898 // decode narrow_oop_reg, base_reg
2899 // CmpP base_reg, NULL
2900 // CastPP base_reg // NotNull
2901 // Load [base_reg + offset], val_reg
2902 //
2903 // after these transformations will be
2904 //
2905 // LoadN memory, narrow_oop_reg
2906 // CmpN narrow_oop_reg, NULL
2907 // decode_not_null narrow_oop_reg, base_reg
2908 // Load [base_reg + offset], val_reg
2909 //
2910 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2911 // since narrow oops can be used in debug info now (see the code in
2912 // final_graph_reshaping_walk()).
2913 //
2914 // At the end the code will be matched to
2915 // on x86:
2916 //
2917 // Load_narrow_oop memory, narrow_oop_reg
2918 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2919 // NullCheck narrow_oop_reg
2920 //
2921 // and on sparc:
2922 //
2923 // Load_narrow_oop memory, narrow_oop_reg
2924 // decode_not_null narrow_oop_reg, base_reg
2925 // Load [base_reg + offset], val_reg
2926 // NullCheck base_reg
2927 //
2928 } else if (t->isa_oopptr()) {
2929 new_in2 = ConNode::make(this, t->make_narrowoop());
2930 } else if (t->isa_klassptr()) {
2931 new_in2 = ConNode::make(this, t->make_narrowklass());
2932 }
2933 }
2934 if (new_in2 != NULL) {
2935 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2936 n->subsume_by(cmpN, this);
2937 if (in1->outcnt() == 0) {
2938 in1->disconnect_inputs(NULL, this);
2939 }
2940 if (in2->outcnt() == 0) {
2941 in2->disconnect_inputs(NULL, this);
2942 }
2943 }
2944 }
2945 break;
2947 case Op_DecodeN:
2948 case Op_DecodeNKlass:
2949 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2950 // DecodeN could be pinned when it can't be fold into
2951 // an address expression, see the code for Op_CastPP above.
2952 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2953 break;
2955 case Op_EncodeP:
2956 case Op_EncodePKlass: {
2957 Node* in1 = n->in(1);
2958 if (in1->is_DecodeNarrowPtr()) {
2959 n->subsume_by(in1->in(1), this);
2960 } else if (in1->Opcode() == Op_ConP) {
2961 const Type* t = in1->bottom_type();
2962 if (t == TypePtr::NULL_PTR) {
2963 assert(t->isa_oopptr(), "null klass?");
2964 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2965 } else if (t->isa_oopptr()) {
2966 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2967 } else if (t->isa_klassptr()) {
2968 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2969 }
2970 }
2971 if (in1->outcnt() == 0) {
2972 in1->disconnect_inputs(NULL, this);
2973 }
2974 break;
2975 }
2977 case Op_Proj: {
2978 if (OptimizeStringConcat) {
2979 ProjNode* p = n->as_Proj();
2980 if (p->_is_io_use) {
2981 // Separate projections were used for the exception path which
2982 // are normally removed by a late inline. If it wasn't inlined
2983 // then they will hang around and should just be replaced with
2984 // the original one.
2985 Node* proj = NULL;
2986 // Replace with just one
2987 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2988 Node *use = i.get();
2989 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2990 proj = use;
2991 break;
2992 }
2993 }
2994 assert(proj != NULL, "must be found");
2995 p->subsume_by(proj, this);
2996 }
2997 }
2998 break;
2999 }
3001 case Op_Phi:
3002 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3003 // The EncodeP optimization may create Phi with the same edges
3004 // for all paths. It is not handled well by Register Allocator.
3005 Node* unique_in = n->in(1);
3006 assert(unique_in != NULL, "");
3007 uint cnt = n->req();
3008 for (uint i = 2; i < cnt; i++) {
3009 Node* m = n->in(i);
3010 assert(m != NULL, "");
3011 if (unique_in != m)
3012 unique_in = NULL;
3013 }
3014 if (unique_in != NULL) {
3015 n->subsume_by(unique_in, this);
3016 }
3017 }
3018 break;
3020 #endif
3022 #ifdef ASSERT
3023 case Op_CastII:
3024 // Verify that all range check dependent CastII nodes were removed.
3025 if (n->isa_CastII()->has_range_check()) {
3026 n->dump(3);
3027 assert(false, "Range check dependent CastII node was not removed");
3028 }
3029 break;
3030 #endif
3032 case Op_ModI:
3033 if (UseDivMod) {
3034 // Check if a%b and a/b both exist
3035 Node* d = n->find_similar(Op_DivI);
3036 if (d) {
3037 // Replace them with a fused divmod if supported
3038 if (Matcher::has_match_rule(Op_DivModI)) {
3039 DivModINode* divmod = DivModINode::make(this, n);
3040 d->subsume_by(divmod->div_proj(), this);
3041 n->subsume_by(divmod->mod_proj(), this);
3042 } else {
3043 // replace a%b with a-((a/b)*b)
3044 Node* mult = new (this) MulINode(d, d->in(2));
3045 Node* sub = new (this) SubINode(d->in(1), mult);
3046 n->subsume_by(sub, this);
3047 }
3048 }
3049 }
3050 break;
3052 case Op_ModL:
3053 if (UseDivMod) {
3054 // Check if a%b and a/b both exist
3055 Node* d = n->find_similar(Op_DivL);
3056 if (d) {
3057 // Replace them with a fused divmod if supported
3058 if (Matcher::has_match_rule(Op_DivModL)) {
3059 DivModLNode* divmod = DivModLNode::make(this, n);
3060 d->subsume_by(divmod->div_proj(), this);
3061 n->subsume_by(divmod->mod_proj(), this);
3062 } else {
3063 // replace a%b with a-((a/b)*b)
3064 Node* mult = new (this) MulLNode(d, d->in(2));
3065 Node* sub = new (this) SubLNode(d->in(1), mult);
3066 n->subsume_by(sub, this);
3067 }
3068 }
3069 }
3070 break;
3072 case Op_LoadVector:
3073 case Op_StoreVector:
3074 break;
3076 case Op_PackB:
3077 case Op_PackS:
3078 case Op_PackI:
3079 case Op_PackF:
3080 case Op_PackL:
3081 case Op_PackD:
3082 if (n->req()-1 > 2) {
3083 // Replace many operand PackNodes with a binary tree for matching
3084 PackNode* p = (PackNode*) n;
3085 Node* btp = p->binary_tree_pack(this, 1, n->req());
3086 n->subsume_by(btp, this);
3087 }
3088 break;
3089 case Op_Loop:
3090 case Op_CountedLoop:
3091 if (n->as_Loop()->is_inner_loop()) {
3092 frc.inc_inner_loop_count();
3093 }
3094 break;
3095 case Op_LShiftI:
3096 case Op_RShiftI:
3097 case Op_URShiftI:
3098 case Op_LShiftL:
3099 case Op_RShiftL:
3100 case Op_URShiftL:
3101 if (Matcher::need_masked_shift_count) {
3102 // The cpu's shift instructions don't restrict the count to the
3103 // lower 5/6 bits. We need to do the masking ourselves.
3104 Node* in2 = n->in(2);
3105 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3106 const TypeInt* t = in2->find_int_type();
3107 if (t != NULL && t->is_con()) {
3108 juint shift = t->get_con();
3109 if (shift > mask) { // Unsigned cmp
3110 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
3111 }
3112 } else {
3113 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3114 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3115 n->set_req(2, shift);
3116 }
3117 }
3118 if (in2->outcnt() == 0) { // Remove dead node
3119 in2->disconnect_inputs(NULL, this);
3120 }
3121 }
3122 break;
3123 case Op_MemBarStoreStore:
3124 case Op_MemBarRelease:
3125 // Break the link with AllocateNode: it is no longer useful and
3126 // confuses register allocation.
3127 if (n->req() > MemBarNode::Precedent) {
3128 n->set_req(MemBarNode::Precedent, top());
3129 }
3130 break;
3131 default:
3132 assert( !n->is_Call(), "" );
3133 assert( !n->is_Mem(), "" );
3134 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3135 break;
3136 }
3138 // Collect CFG split points
3139 if (n->is_MultiBranch())
3140 frc._tests.push(n);
3141 }
3143 //------------------------------final_graph_reshaping_walk---------------------
3144 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3145 // requires that the walk visits a node's inputs before visiting the node.
3146 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3147 ResourceArea *area = Thread::current()->resource_area();
3148 Unique_Node_List sfpt(area);
3150 frc._visited.set(root->_idx); // first, mark node as visited
3151 uint cnt = root->req();
3152 Node *n = root;
3153 uint i = 0;
3154 while (true) {
3155 if (i < cnt) {
3156 // Place all non-visited non-null inputs onto stack
3157 Node* m = n->in(i);
3158 ++i;
3159 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3160 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3161 // compute worst case interpreter size in case of a deoptimization
3162 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3164 sfpt.push(m);
3165 }
3166 cnt = m->req();
3167 nstack.push(n, i); // put on stack parent and next input's index
3168 n = m;
3169 i = 0;
3170 }
3171 } else {
3172 // Now do post-visit work
3173 final_graph_reshaping_impl( n, frc );
3174 if (nstack.is_empty())
3175 break; // finished
3176 n = nstack.node(); // Get node from stack
3177 cnt = n->req();
3178 i = nstack.index();
3179 nstack.pop(); // Shift to the next node on stack
3180 }
3181 }
3183 // Skip next transformation if compressed oops are not used.
3184 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3185 (!UseCompressedOops && !UseCompressedClassPointers))
3186 return;
3188 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3189 // It could be done for an uncommon traps or any safepoints/calls
3190 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3191 while (sfpt.size() > 0) {
3192 n = sfpt.pop();
3193 JVMState *jvms = n->as_SafePoint()->jvms();
3194 assert(jvms != NULL, "sanity");
3195 int start = jvms->debug_start();
3196 int end = n->req();
3197 bool is_uncommon = (n->is_CallStaticJava() &&
3198 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3199 for (int j = start; j < end; j++) {
3200 Node* in = n->in(j);
3201 if (in->is_DecodeNarrowPtr()) {
3202 bool safe_to_skip = true;
3203 if (!is_uncommon ) {
3204 // Is it safe to skip?
3205 for (uint i = 0; i < in->outcnt(); i++) {
3206 Node* u = in->raw_out(i);
3207 if (!u->is_SafePoint() ||
3208 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3209 safe_to_skip = false;
3210 }
3211 }
3212 }
3213 if (safe_to_skip) {
3214 n->set_req(j, in->in(1));
3215 }
3216 if (in->outcnt() == 0) {
3217 in->disconnect_inputs(NULL, this);
3218 }
3219 }
3220 }
3221 }
3222 }
3224 //------------------------------final_graph_reshaping--------------------------
3225 // Final Graph Reshaping.
3226 //
3227 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3228 // and not commoned up and forced early. Must come after regular
3229 // optimizations to avoid GVN undoing the cloning. Clone constant
3230 // inputs to Loop Phis; these will be split by the allocator anyways.
3231 // Remove Opaque nodes.
3232 // (2) Move last-uses by commutative operations to the left input to encourage
3233 // Intel update-in-place two-address operations and better register usage
3234 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3235 // calls canonicalizing them back.
3236 // (3) Count the number of double-precision FP ops, single-precision FP ops
3237 // and call sites. On Intel, we can get correct rounding either by
3238 // forcing singles to memory (requires extra stores and loads after each
3239 // FP bytecode) or we can set a rounding mode bit (requires setting and
3240 // clearing the mode bit around call sites). The mode bit is only used
3241 // if the relative frequency of single FP ops to calls is low enough.
3242 // This is a key transform for SPEC mpeg_audio.
3243 // (4) Detect infinite loops; blobs of code reachable from above but not
3244 // below. Several of the Code_Gen algorithms fail on such code shapes,
3245 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3246 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3247 // Detection is by looking for IfNodes where only 1 projection is
3248 // reachable from below or CatchNodes missing some targets.
3249 // (5) Assert for insane oop offsets in debug mode.
3251 bool Compile::final_graph_reshaping() {
3252 // an infinite loop may have been eliminated by the optimizer,
3253 // in which case the graph will be empty.
3254 if (root()->req() == 1) {
3255 record_method_not_compilable("trivial infinite loop");
3256 return true;
3257 }
3259 // Expensive nodes have their control input set to prevent the GVN
3260 // from freely commoning them. There's no GVN beyond this point so
3261 // no need to keep the control input. We want the expensive nodes to
3262 // be freely moved to the least frequent code path by gcm.
3263 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3264 for (int i = 0; i < expensive_count(); i++) {
3265 _expensive_nodes->at(i)->set_req(0, NULL);
3266 }
3268 Final_Reshape_Counts frc;
3270 // Visit everybody reachable!
3271 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3272 Node_Stack nstack(live_nodes() >> 1);
3273 final_graph_reshaping_walk(nstack, root(), frc);
3275 // Check for unreachable (from below) code (i.e., infinite loops).
3276 for( uint i = 0; i < frc._tests.size(); i++ ) {
3277 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3278 // Get number of CFG targets.
3279 // Note that PCTables include exception targets after calls.
3280 uint required_outcnt = n->required_outcnt();
3281 if (n->outcnt() != required_outcnt) {
3282 // Check for a few special cases. Rethrow Nodes never take the
3283 // 'fall-thru' path, so expected kids is 1 less.
3284 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3285 if (n->in(0)->in(0)->is_Call()) {
3286 CallNode *call = n->in(0)->in(0)->as_Call();
3287 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3288 required_outcnt--; // Rethrow always has 1 less kid
3289 } else if (call->req() > TypeFunc::Parms &&
3290 call->is_CallDynamicJava()) {
3291 // Check for null receiver. In such case, the optimizer has
3292 // detected that the virtual call will always result in a null
3293 // pointer exception. The fall-through projection of this CatchNode
3294 // will not be populated.
3295 Node *arg0 = call->in(TypeFunc::Parms);
3296 if (arg0->is_Type() &&
3297 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3298 required_outcnt--;
3299 }
3300 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3301 call->req() > TypeFunc::Parms+1 &&
3302 call->is_CallStaticJava()) {
3303 // Check for negative array length. In such case, the optimizer has
3304 // detected that the allocation attempt will always result in an
3305 // exception. There is no fall-through projection of this CatchNode .
3306 Node *arg1 = call->in(TypeFunc::Parms+1);
3307 if (arg1->is_Type() &&
3308 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3309 required_outcnt--;
3310 }
3311 }
3312 }
3313 }
3314 // Recheck with a better notion of 'required_outcnt'
3315 if (n->outcnt() != required_outcnt) {
3316 record_method_not_compilable("malformed control flow");
3317 return true; // Not all targets reachable!
3318 }
3319 }
3320 // Check that I actually visited all kids. Unreached kids
3321 // must be infinite loops.
3322 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3323 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3324 record_method_not_compilable("infinite loop");
3325 return true; // Found unvisited kid; must be unreach
3326 }
3327 }
3329 // If original bytecodes contained a mixture of floats and doubles
3330 // check if the optimizer has made it homogenous, item (3).
3331 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3332 frc.get_float_count() > 32 &&
3333 frc.get_double_count() == 0 &&
3334 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3335 set_24_bit_selection_and_mode( false, true );
3336 }
3338 set_java_calls(frc.get_java_call_count());
3339 set_inner_loops(frc.get_inner_loop_count());
3341 // No infinite loops, no reason to bail out.
3342 return false;
3343 }
3345 //-----------------------------too_many_traps----------------------------------
3346 // Report if there are too many traps at the current method and bci.
3347 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3348 bool Compile::too_many_traps(ciMethod* method,
3349 int bci,
3350 Deoptimization::DeoptReason reason) {
3351 ciMethodData* md = method->method_data();
3352 if (md->is_empty()) {
3353 // Assume the trap has not occurred, or that it occurred only
3354 // because of a transient condition during start-up in the interpreter.
3355 return false;
3356 }
3357 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3358 if (md->has_trap_at(bci, m, reason) != 0) {
3359 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3360 // Also, if there are multiple reasons, or if there is no per-BCI record,
3361 // assume the worst.
3362 if (log())
3363 log()->elem("observe trap='%s' count='%d'",
3364 Deoptimization::trap_reason_name(reason),
3365 md->trap_count(reason));
3366 return true;
3367 } else {
3368 // Ignore method/bci and see if there have been too many globally.
3369 return too_many_traps(reason, md);
3370 }
3371 }
3373 // Less-accurate variant which does not require a method and bci.
3374 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3375 ciMethodData* logmd) {
3376 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3377 // Too many traps globally.
3378 // Note that we use cumulative trap_count, not just md->trap_count.
3379 if (log()) {
3380 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3381 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3382 Deoptimization::trap_reason_name(reason),
3383 mcount, trap_count(reason));
3384 }
3385 return true;
3386 } else {
3387 // The coast is clear.
3388 return false;
3389 }
3390 }
3392 //--------------------------too_many_recompiles--------------------------------
3393 // Report if there are too many recompiles at the current method and bci.
3394 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3395 // Is not eager to return true, since this will cause the compiler to use
3396 // Action_none for a trap point, to avoid too many recompilations.
3397 bool Compile::too_many_recompiles(ciMethod* method,
3398 int bci,
3399 Deoptimization::DeoptReason reason) {
3400 ciMethodData* md = method->method_data();
3401 if (md->is_empty()) {
3402 // Assume the trap has not occurred, or that it occurred only
3403 // because of a transient condition during start-up in the interpreter.
3404 return false;
3405 }
3406 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3407 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3408 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3409 Deoptimization::DeoptReason per_bc_reason
3410 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3411 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3412 if ((per_bc_reason == Deoptimization::Reason_none
3413 || md->has_trap_at(bci, m, reason) != 0)
3414 // The trap frequency measure we care about is the recompile count:
3415 && md->trap_recompiled_at(bci, m)
3416 && md->overflow_recompile_count() >= bc_cutoff) {
3417 // Do not emit a trap here if it has already caused recompilations.
3418 // Also, if there are multiple reasons, or if there is no per-BCI record,
3419 // assume the worst.
3420 if (log())
3421 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3422 Deoptimization::trap_reason_name(reason),
3423 md->trap_count(reason),
3424 md->overflow_recompile_count());
3425 return true;
3426 } else if (trap_count(reason) != 0
3427 && decompile_count() >= m_cutoff) {
3428 // Too many recompiles globally, and we have seen this sort of trap.
3429 // Use cumulative decompile_count, not just md->decompile_count.
3430 if (log())
3431 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3432 Deoptimization::trap_reason_name(reason),
3433 md->trap_count(reason), trap_count(reason),
3434 md->decompile_count(), decompile_count());
3435 return true;
3436 } else {
3437 // The coast is clear.
3438 return false;
3439 }
3440 }
3442 // Compute when not to trap. Used by matching trap based nodes and
3443 // NullCheck optimization.
3444 void Compile::set_allowed_deopt_reasons() {
3445 _allowed_reasons = 0;
3446 if (is_method_compilation()) {
3447 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3448 assert(rs < BitsPerInt, "recode bit map");
3449 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3450 _allowed_reasons |= nth_bit(rs);
3451 }
3452 }
3453 }
3454 }
3456 #ifndef PRODUCT
3457 //------------------------------verify_graph_edges---------------------------
3458 // Walk the Graph and verify that there is a one-to-one correspondence
3459 // between Use-Def edges and Def-Use edges in the graph.
3460 void Compile::verify_graph_edges(bool no_dead_code) {
3461 if (VerifyGraphEdges) {
3462 ResourceArea *area = Thread::current()->resource_area();
3463 Unique_Node_List visited(area);
3464 // Call recursive graph walk to check edges
3465 _root->verify_edges(visited);
3466 if (no_dead_code) {
3467 // Now make sure that no visited node is used by an unvisited node.
3468 bool dead_nodes = 0;
3469 Unique_Node_List checked(area);
3470 while (visited.size() > 0) {
3471 Node* n = visited.pop();
3472 checked.push(n);
3473 for (uint i = 0; i < n->outcnt(); i++) {
3474 Node* use = n->raw_out(i);
3475 if (checked.member(use)) continue; // already checked
3476 if (visited.member(use)) continue; // already in the graph
3477 if (use->is_Con()) continue; // a dead ConNode is OK
3478 // At this point, we have found a dead node which is DU-reachable.
3479 if (dead_nodes++ == 0)
3480 tty->print_cr("*** Dead nodes reachable via DU edges:");
3481 use->dump(2);
3482 tty->print_cr("---");
3483 checked.push(use); // No repeats; pretend it is now checked.
3484 }
3485 }
3486 assert(dead_nodes == 0, "using nodes must be reachable from root");
3487 }
3488 }
3489 }
3491 // Verify GC barriers consistency
3492 // Currently supported:
3493 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3494 void Compile::verify_barriers() {
3495 if (UseG1GC) {
3496 // Verify G1 pre-barriers
3497 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3499 ResourceArea *area = Thread::current()->resource_area();
3500 Unique_Node_List visited(area);
3501 Node_List worklist(area);
3502 // We're going to walk control flow backwards starting from the Root
3503 worklist.push(_root);
3504 while (worklist.size() > 0) {
3505 Node* x = worklist.pop();
3506 if (x == NULL || x == top()) continue;
3507 if (visited.member(x)) {
3508 continue;
3509 } else {
3510 visited.push(x);
3511 }
3513 if (x->is_Region()) {
3514 for (uint i = 1; i < x->req(); i++) {
3515 worklist.push(x->in(i));
3516 }
3517 } else {
3518 worklist.push(x->in(0));
3519 // We are looking for the pattern:
3520 // /->ThreadLocal
3521 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3522 // \->ConI(0)
3523 // We want to verify that the If and the LoadB have the same control
3524 // See GraphKit::g1_write_barrier_pre()
3525 if (x->is_If()) {
3526 IfNode *iff = x->as_If();
3527 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3528 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3529 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3530 && cmp->in(1)->is_Load()) {
3531 LoadNode* load = cmp->in(1)->as_Load();
3532 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3533 && load->in(2)->in(3)->is_Con()
3534 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3536 Node* if_ctrl = iff->in(0);
3537 Node* load_ctrl = load->in(0);
3539 if (if_ctrl != load_ctrl) {
3540 // Skip possible CProj->NeverBranch in infinite loops
3541 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3542 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3543 if_ctrl = if_ctrl->in(0)->in(0);
3544 }
3545 }
3546 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3547 }
3548 }
3549 }
3550 }
3551 }
3552 }
3553 }
3554 }
3556 #endif
3558 // The Compile object keeps track of failure reasons separately from the ciEnv.
3559 // This is required because there is not quite a 1-1 relation between the
3560 // ciEnv and its compilation task and the Compile object. Note that one
3561 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3562 // to backtrack and retry without subsuming loads. Other than this backtracking
3563 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3564 // by the logic in C2Compiler.
3565 void Compile::record_failure(const char* reason) {
3566 if (log() != NULL) {
3567 log()->elem("failure reason='%s' phase='compile'", reason);
3568 }
3569 if (_failure_reason == NULL) {
3570 // Record the first failure reason.
3571 _failure_reason = reason;
3572 }
3574 EventCompilerFailure event;
3575 if (event.should_commit()) {
3576 event.set_compileID(Compile::compile_id());
3577 event.set_failure(reason);
3578 event.commit();
3579 }
3581 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3582 C->print_method(PHASE_FAILURE);
3583 }
3584 _root = NULL; // flush the graph, too
3585 }
3587 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3588 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3589 _phase_name(name), _dolog(dolog)
3590 {
3591 if (dolog) {
3592 C = Compile::current();
3593 _log = C->log();
3594 } else {
3595 C = NULL;
3596 _log = NULL;
3597 }
3598 if (_log != NULL) {
3599 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3600 _log->stamp();
3601 _log->end_head();
3602 }
3603 }
3605 Compile::TracePhase::~TracePhase() {
3607 C = Compile::current();
3608 if (_dolog) {
3609 _log = C->log();
3610 } else {
3611 _log = NULL;
3612 }
3614 #ifdef ASSERT
3615 if (PrintIdealNodeCount) {
3616 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3617 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3618 }
3620 if (VerifyIdealNodeCount) {
3621 Compile::current()->print_missing_nodes();
3622 }
3623 #endif
3625 if (_log != NULL) {
3626 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3627 }
3628 }
3630 //=============================================================================
3631 // Two Constant's are equal when the type and the value are equal.
3632 bool Compile::Constant::operator==(const Constant& other) {
3633 if (type() != other.type() ) return false;
3634 if (can_be_reused() != other.can_be_reused()) return false;
3635 // For floating point values we compare the bit pattern.
3636 switch (type()) {
3637 case T_FLOAT: return (_v._value.i == other._v._value.i);
3638 case T_LONG:
3639 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3640 case T_OBJECT:
3641 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3642 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3643 case T_METADATA: return (_v._metadata == other._v._metadata);
3644 default: ShouldNotReachHere();
3645 }
3646 return false;
3647 }
3649 static int type_to_size_in_bytes(BasicType t) {
3650 switch (t) {
3651 case T_LONG: return sizeof(jlong );
3652 case T_FLOAT: return sizeof(jfloat );
3653 case T_DOUBLE: return sizeof(jdouble);
3654 case T_METADATA: return sizeof(Metadata*);
3655 // We use T_VOID as marker for jump-table entries (labels) which
3656 // need an internal word relocation.
3657 case T_VOID:
3658 case T_ADDRESS:
3659 case T_OBJECT: return sizeof(jobject);
3660 }
3662 ShouldNotReachHere();
3663 return -1;
3664 }
3666 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3667 // sort descending
3668 if (a->freq() > b->freq()) return -1;
3669 if (a->freq() < b->freq()) return 1;
3670 return 0;
3671 }
3673 void Compile::ConstantTable::calculate_offsets_and_size() {
3674 // First, sort the array by frequencies.
3675 _constants.sort(qsort_comparator);
3677 #ifdef ASSERT
3678 // Make sure all jump-table entries were sorted to the end of the
3679 // array (they have a negative frequency).
3680 bool found_void = false;
3681 for (int i = 0; i < _constants.length(); i++) {
3682 Constant con = _constants.at(i);
3683 if (con.type() == T_VOID)
3684 found_void = true; // jump-tables
3685 else
3686 assert(!found_void, "wrong sorting");
3687 }
3688 #endif
3690 int offset = 0;
3691 for (int i = 0; i < _constants.length(); i++) {
3692 Constant* con = _constants.adr_at(i);
3694 // Align offset for type.
3695 int typesize = type_to_size_in_bytes(con->type());
3696 offset = align_size_up(offset, typesize);
3697 con->set_offset(offset); // set constant's offset
3699 if (con->type() == T_VOID) {
3700 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3701 offset = offset + typesize * n->outcnt(); // expand jump-table
3702 } else {
3703 offset = offset + typesize;
3704 }
3705 }
3707 // Align size up to the next section start (which is insts; see
3708 // CodeBuffer::align_at_start).
3709 assert(_size == -1, "already set?");
3710 _size = align_size_up(offset, CodeEntryAlignment);
3711 }
3713 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3714 MacroAssembler _masm(&cb);
3715 for (int i = 0; i < _constants.length(); i++) {
3716 Constant con = _constants.at(i);
3717 address constant_addr = NULL;
3718 switch (con.type()) {
3719 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3720 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3721 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3722 case T_OBJECT: {
3723 jobject obj = con.get_jobject();
3724 int oop_index = _masm.oop_recorder()->find_index(obj);
3725 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3726 break;
3727 }
3728 case T_ADDRESS: {
3729 address addr = (address) con.get_jobject();
3730 constant_addr = _masm.address_constant(addr);
3731 break;
3732 }
3733 // We use T_VOID as marker for jump-table entries (labels) which
3734 // need an internal word relocation.
3735 case T_VOID: {
3736 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3737 // Fill the jump-table with a dummy word. The real value is
3738 // filled in later in fill_jump_table.
3739 address dummy = (address) n;
3740 constant_addr = _masm.address_constant(dummy);
3741 // Expand jump-table
3742 for (uint i = 1; i < n->outcnt(); i++) {
3743 address temp_addr = _masm.address_constant(dummy + i);
3744 assert(temp_addr, "consts section too small");
3745 }
3746 break;
3747 }
3748 case T_METADATA: {
3749 Metadata* obj = con.get_metadata();
3750 int metadata_index = _masm.oop_recorder()->find_index(obj);
3751 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3752 break;
3753 }
3754 default: ShouldNotReachHere();
3755 }
3756 assert(constant_addr, "consts section too small");
3757 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3758 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())));
3759 }
3760 }
3762 int Compile::ConstantTable::find_offset(Constant& con) const {
3763 int idx = _constants.find(con);
3764 assert(idx != -1, "constant must be in constant table");
3765 int offset = _constants.at(idx).offset();
3766 assert(offset != -1, "constant table not emitted yet?");
3767 return offset;
3768 }
3770 void Compile::ConstantTable::add(Constant& con) {
3771 if (con.can_be_reused()) {
3772 int idx = _constants.find(con);
3773 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3774 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3775 return;
3776 }
3777 }
3778 (void) _constants.append(con);
3779 }
3781 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3782 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3783 Constant con(type, value, b->_freq);
3784 add(con);
3785 return con;
3786 }
3788 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3789 Constant con(metadata);
3790 add(con);
3791 return con;
3792 }
3794 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3795 jvalue value;
3796 BasicType type = oper->type()->basic_type();
3797 switch (type) {
3798 case T_LONG: value.j = oper->constantL(); break;
3799 case T_FLOAT: value.f = oper->constantF(); break;
3800 case T_DOUBLE: value.d = oper->constantD(); break;
3801 case T_OBJECT:
3802 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3803 case T_METADATA: return add((Metadata*)oper->constant()); break;
3804 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3805 }
3806 return add(n, type, value);
3807 }
3809 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3810 jvalue value;
3811 // We can use the node pointer here to identify the right jump-table
3812 // as this method is called from Compile::Fill_buffer right before
3813 // the MachNodes are emitted and the jump-table is filled (means the
3814 // MachNode pointers do not change anymore).
3815 value.l = (jobject) n;
3816 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3817 add(con);
3818 return con;
3819 }
3821 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3822 // If called from Compile::scratch_emit_size do nothing.
3823 if (Compile::current()->in_scratch_emit_size()) return;
3825 assert(labels.is_nonempty(), "must be");
3826 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3828 // Since MachConstantNode::constant_offset() also contains
3829 // table_base_offset() we need to subtract the table_base_offset()
3830 // to get the plain offset into the constant table.
3831 int offset = n->constant_offset() - table_base_offset();
3833 MacroAssembler _masm(&cb);
3834 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3836 for (uint i = 0; i < n->outcnt(); i++) {
3837 address* constant_addr = &jump_table_base[i];
3838 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i)));
3839 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3840 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3841 }
3842 }
3844 void Compile::dump_inlining() {
3845 if (print_inlining() || print_intrinsics()) {
3846 // Print inlining message for candidates that we couldn't inline
3847 // for lack of space or non constant receiver
3848 for (int i = 0; i < _late_inlines.length(); i++) {
3849 CallGenerator* cg = _late_inlines.at(i);
3850 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3851 }
3852 Unique_Node_List useful;
3853 useful.push(root());
3854 for (uint next = 0; next < useful.size(); ++next) {
3855 Node* n = useful.at(next);
3856 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3857 CallNode* call = n->as_Call();
3858 CallGenerator* cg = call->generator();
3859 cg->print_inlining_late("receiver not constant");
3860 }
3861 uint max = n->len();
3862 for ( uint i = 0; i < max; ++i ) {
3863 Node *m = n->in(i);
3864 if ( m == NULL ) continue;
3865 useful.push(m);
3866 }
3867 }
3868 for (int i = 0; i < _print_inlining_list->length(); i++) {
3869 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
3870 }
3871 }
3872 }
3874 // Dump inlining replay data to the stream.
3875 // Don't change thread state and acquire any locks.
3876 void Compile::dump_inline_data(outputStream* out) {
3877 InlineTree* inl_tree = ilt();
3878 if (inl_tree != NULL) {
3879 out->print(" inline %d", inl_tree->count());
3880 inl_tree->dump_replay_data(out);
3881 }
3882 }
3884 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3885 if (n1->Opcode() < n2->Opcode()) return -1;
3886 else if (n1->Opcode() > n2->Opcode()) return 1;
3888 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()));
3889 for (uint i = 1; i < n1->req(); i++) {
3890 if (n1->in(i) < n2->in(i)) return -1;
3891 else if (n1->in(i) > n2->in(i)) return 1;
3892 }
3894 return 0;
3895 }
3897 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3898 Node* n1 = *n1p;
3899 Node* n2 = *n2p;
3901 return cmp_expensive_nodes(n1, n2);
3902 }
3904 void Compile::sort_expensive_nodes() {
3905 if (!expensive_nodes_sorted()) {
3906 _expensive_nodes->sort(cmp_expensive_nodes);
3907 }
3908 }
3910 bool Compile::expensive_nodes_sorted() const {
3911 for (int i = 1; i < _expensive_nodes->length(); i++) {
3912 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3913 return false;
3914 }
3915 }
3916 return true;
3917 }
3919 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3920 if (_expensive_nodes->length() == 0) {
3921 return false;
3922 }
3924 assert(OptimizeExpensiveOps, "optimization off?");
3926 // Take this opportunity to remove dead nodes from the list
3927 int j = 0;
3928 for (int i = 0; i < _expensive_nodes->length(); i++) {
3929 Node* n = _expensive_nodes->at(i);
3930 if (!n->is_unreachable(igvn)) {
3931 assert(n->is_expensive(), "should be expensive");
3932 _expensive_nodes->at_put(j, n);
3933 j++;
3934 }
3935 }
3936 _expensive_nodes->trunc_to(j);
3938 // Then sort the list so that similar nodes are next to each other
3939 // and check for at least two nodes of identical kind with same data
3940 // inputs.
3941 sort_expensive_nodes();
3943 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3944 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3945 return true;
3946 }
3947 }
3949 return false;
3950 }
3952 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3953 if (_expensive_nodes->length() == 0) {
3954 return;
3955 }
3957 assert(OptimizeExpensiveOps, "optimization off?");
3959 // Sort to bring similar nodes next to each other and clear the
3960 // control input of nodes for which there's only a single copy.
3961 sort_expensive_nodes();
3963 int j = 0;
3964 int identical = 0;
3965 int i = 0;
3966 for (; i < _expensive_nodes->length()-1; i++) {
3967 assert(j <= i, "can't write beyond current index");
3968 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3969 identical++;
3970 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3971 continue;
3972 }
3973 if (identical > 0) {
3974 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3975 identical = 0;
3976 } else {
3977 Node* n = _expensive_nodes->at(i);
3978 igvn.hash_delete(n);
3979 n->set_req(0, NULL);
3980 igvn.hash_insert(n);
3981 }
3982 }
3983 if (identical > 0) {
3984 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3985 } else if (_expensive_nodes->length() >= 1) {
3986 Node* n = _expensive_nodes->at(i);
3987 igvn.hash_delete(n);
3988 n->set_req(0, NULL);
3989 igvn.hash_insert(n);
3990 }
3991 _expensive_nodes->trunc_to(j);
3992 }
3994 void Compile::add_expensive_node(Node * n) {
3995 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3996 assert(n->is_expensive(), "expensive nodes with non-null control here only");
3997 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
3998 if (OptimizeExpensiveOps) {
3999 _expensive_nodes->append(n);
4000 } else {
4001 // Clear control input and let IGVN optimize expensive nodes if
4002 // OptimizeExpensiveOps is off.
4003 n->set_req(0, NULL);
4004 }
4005 }
4007 /**
4008 * Remove the speculative part of types and clean up the graph
4009 */
4010 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4011 if (UseTypeSpeculation) {
4012 Unique_Node_List worklist;
4013 worklist.push(root());
4014 int modified = 0;
4015 // Go over all type nodes that carry a speculative type, drop the
4016 // speculative part of the type and enqueue the node for an igvn
4017 // which may optimize it out.
4018 for (uint next = 0; next < worklist.size(); ++next) {
4019 Node *n = worklist.at(next);
4020 if (n->is_Type()) {
4021 TypeNode* tn = n->as_Type();
4022 const Type* t = tn->type();
4023 const Type* t_no_spec = t->remove_speculative();
4024 if (t_no_spec != t) {
4025 bool in_hash = igvn.hash_delete(n);
4026 assert(in_hash, "node should be in igvn hash table");
4027 tn->set_type(t_no_spec);
4028 igvn.hash_insert(n);
4029 igvn._worklist.push(n); // give it a chance to go away
4030 modified++;
4031 }
4032 }
4033 uint max = n->len();
4034 for( uint i = 0; i < max; ++i ) {
4035 Node *m = n->in(i);
4036 if (not_a_node(m)) continue;
4037 worklist.push(m);
4038 }
4039 }
4040 // Drop the speculative part of all types in the igvn's type table
4041 igvn.remove_speculative_types();
4042 if (modified > 0) {
4043 igvn.optimize();
4044 }
4045 #ifdef ASSERT
4046 // Verify that after the IGVN is over no speculative type has resurfaced
4047 worklist.clear();
4048 worklist.push(root());
4049 for (uint next = 0; next < worklist.size(); ++next) {
4050 Node *n = worklist.at(next);
4051 const Type* t = igvn.type_or_null(n);
4052 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4053 if (n->is_Type()) {
4054 t = n->as_Type()->type();
4055 assert(t == t->remove_speculative(), "no more speculative types");
4056 }
4057 uint max = n->len();
4058 for( uint i = 0; i < max; ++i ) {
4059 Node *m = n->in(i);
4060 if (not_a_node(m)) continue;
4061 worklist.push(m);
4062 }
4063 }
4064 igvn.check_no_speculative_types();
4065 #endif
4066 }
4067 }
4069 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4070 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4071 if (ctrl != NULL) {
4072 // Express control dependency by a CastII node with a narrow type.
4073 value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */);
4074 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4075 // node from floating above the range check during loop optimizations. Otherwise, the
4076 // ConvI2L node may be eliminated independently of the range check, causing the data path
4077 // to become TOP while the control path is still there (although it's unreachable).
4078 value->set_req(0, ctrl);
4079 // Save CastII node to remove it after loop optimizations.
4080 phase->C->add_range_check_cast(value);
4081 value = phase->transform(value);
4082 }
4083 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4084 return phase->transform(new (phase->C) ConvI2LNode(value, ltype));
4085 }
4087 // Auxiliary method to support randomized stressing/fuzzing.
4088 //
4089 // This method can be called the arbitrary number of times, with current count
4090 // as the argument. The logic allows selecting a single candidate from the
4091 // running list of candidates as follows:
4092 // int count = 0;
4093 // Cand* selected = null;
4094 // while(cand = cand->next()) {
4095 // if (randomized_select(++count)) {
4096 // selected = cand;
4097 // }
4098 // }
4099 //
4100 // Including count equalizes the chances any candidate is "selected".
4101 // This is useful when we don't have the complete list of candidates to choose
4102 // from uniformly. In this case, we need to adjust the randomicity of the
4103 // selection, or else we will end up biasing the selection towards the latter
4104 // candidates.
4105 //
4106 // Quick back-envelope calculation shows that for the list of n candidates
4107 // the equal probability for the candidate to persist as "best" can be
4108 // achieved by replacing it with "next" k-th candidate with the probability
4109 // of 1/k. It can be easily shown that by the end of the run, the
4110 // probability for any candidate is converged to 1/n, thus giving the
4111 // uniform distribution among all the candidates.
4112 //
4113 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4114 #define RANDOMIZED_DOMAIN_POW 29
4115 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4116 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4117 bool Compile::randomized_select(int count) {
4118 assert(count > 0, "only positive");
4119 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4120 }