Mon, 25 Apr 2016 21:03:53 +0000
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());
612 // Emitting into the scratch buffer should not fail
613 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason()));
615 if (is_branch) // Restore label.
616 n->as_MachBranch()->label_set(saveL, save_bnum);
618 // End scratch_emit_size section.
619 set_in_scratch_emit_size(false);
621 return buf.insts_size();
622 }
625 // ============================================================================
626 //------------------------------Compile standard-------------------------------
627 debug_only( int Compile::_debug_idx = 100000; )
629 // Compile a method. entry_bci is -1 for normal compilations and indicates
630 // the continuation bci for on stack replacement.
633 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
634 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
635 : Phase(Compiler),
636 _env(ci_env),
637 _log(ci_env->log()),
638 _compile_id(ci_env->compile_id()),
639 _save_argument_registers(false),
640 _stub_name(NULL),
641 _stub_function(NULL),
642 _stub_entry_point(NULL),
643 _method(target),
644 _entry_bci(osr_bci),
645 _initial_gvn(NULL),
646 _for_igvn(NULL),
647 _warm_calls(NULL),
648 _subsume_loads(subsume_loads),
649 _do_escape_analysis(do_escape_analysis),
650 _eliminate_boxing(eliminate_boxing),
651 _failure_reason(NULL),
652 _code_buffer("Compile::Fill_buffer"),
653 _orig_pc_slot(0),
654 _orig_pc_slot_offset_in_bytes(0),
655 _has_method_handle_invokes(false),
656 _mach_constant_base_node(NULL),
657 _node_bundling_limit(0),
658 _node_bundling_base(NULL),
659 _java_calls(0),
660 _inner_loops(0),
661 _scratch_const_size(-1),
662 _in_scratch_emit_size(false),
663 _dead_node_list(comp_arena()),
664 _dead_node_count(0),
665 #ifndef PRODUCT
666 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
667 _in_dump_cnt(0),
668 _printer(IdealGraphPrinter::printer()),
669 #endif
670 _congraph(NULL),
671 _comp_arena(mtCompiler),
672 _node_arena(mtCompiler),
673 _old_arena(mtCompiler),
674 _Compile_types(mtCompiler),
675 _replay_inline_data(NULL),
676 _late_inlines(comp_arena(), 2, 0, NULL),
677 _string_late_inlines(comp_arena(), 2, 0, NULL),
678 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
679 _late_inlines_pos(0),
680 _number_of_mh_late_inlines(0),
681 _inlining_progress(false),
682 _inlining_incrementally(false),
683 _print_inlining_list(NULL),
684 _print_inlining_idx(0),
685 _interpreter_frame_size(0),
686 _max_node_limit(MaxNodeLimit) {
687 C = this;
689 CompileWrapper cw(this);
690 #ifndef PRODUCT
691 if (TimeCompiler2) {
692 tty->print(" ");
693 target->holder()->name()->print();
694 tty->print(".");
695 target->print_short_name();
696 tty->print(" ");
697 }
698 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
699 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
700 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
701 if (!print_opto_assembly) {
702 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
703 if (print_assembly && !Disassembler::can_decode()) {
704 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
705 print_opto_assembly = true;
706 }
707 }
708 set_print_assembly(print_opto_assembly);
709 set_parsed_irreducible_loop(false);
711 if (method()->has_option("ReplayInline")) {
712 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
713 }
714 #endif
715 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
716 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
717 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
719 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
720 // Make sure the method being compiled gets its own MDO,
721 // so we can at least track the decompile_count().
722 // Need MDO to record RTM code generation state.
723 method()->ensure_method_data();
724 }
726 Init(::AliasLevel);
729 print_compile_messages();
731 _ilt = InlineTree::build_inline_tree_root();
733 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
734 assert(num_alias_types() >= AliasIdxRaw, "");
736 #define MINIMUM_NODE_HASH 1023
737 // Node list that Iterative GVN will start with
738 Unique_Node_List for_igvn(comp_arena());
739 set_for_igvn(&for_igvn);
741 // GVN that will be run immediately on new nodes
742 uint estimated_size = method()->code_size()*4+64;
743 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
744 PhaseGVN gvn(node_arena(), estimated_size);
745 set_initial_gvn(&gvn);
747 if (print_inlining() || print_intrinsics()) {
748 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
749 }
750 { // Scope for timing the parser
751 TracePhase t3("parse", &_t_parser, true);
753 // Put top into the hash table ASAP.
754 initial_gvn()->transform_no_reclaim(top());
756 // Set up tf(), start(), and find a CallGenerator.
757 CallGenerator* cg = NULL;
758 if (is_osr_compilation()) {
759 const TypeTuple *domain = StartOSRNode::osr_domain();
760 const TypeTuple *range = TypeTuple::make_range(method()->signature());
761 init_tf(TypeFunc::make(domain, range));
762 StartNode* s = new (this) StartOSRNode(root(), domain);
763 initial_gvn()->set_type_bottom(s);
764 init_start(s);
765 cg = CallGenerator::for_osr(method(), entry_bci());
766 } else {
767 // Normal case.
768 init_tf(TypeFunc::make(method()));
769 StartNode* s = new (this) StartNode(root(), tf()->domain());
770 initial_gvn()->set_type_bottom(s);
771 init_start(s);
772 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
773 // With java.lang.ref.reference.get() we must go through the
774 // intrinsic when G1 is enabled - even when get() is the root
775 // method of the compile - so that, if necessary, the value in
776 // the referent field of the reference object gets recorded by
777 // the pre-barrier code.
778 // Specifically, if G1 is enabled, the value in the referent
779 // field is recorded by the G1 SATB pre barrier. This will
780 // result in the referent being marked live and the reference
781 // object removed from the list of discovered references during
782 // reference processing.
783 cg = find_intrinsic(method(), false);
784 }
785 if (cg == NULL) {
786 float past_uses = method()->interpreter_invocation_count();
787 float expected_uses = past_uses;
788 cg = CallGenerator::for_inline(method(), expected_uses);
789 }
790 }
791 if (failing()) return;
792 if (cg == NULL) {
793 record_method_not_compilable_all_tiers("cannot parse method");
794 return;
795 }
796 JVMState* jvms = build_start_state(start(), tf());
797 if ((jvms = cg->generate(jvms)) == NULL) {
798 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
799 record_method_not_compilable("method parse failed");
800 }
801 return;
802 }
803 GraphKit kit(jvms);
805 if (!kit.stopped()) {
806 // Accept return values, and transfer control we know not where.
807 // This is done by a special, unique ReturnNode bound to root.
808 return_values(kit.jvms());
809 }
811 if (kit.has_exceptions()) {
812 // Any exceptions that escape from this call must be rethrown
813 // to whatever caller is dynamically above us on the stack.
814 // This is done by a special, unique RethrowNode bound to root.
815 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
816 }
818 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
820 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
821 inline_string_calls(true);
822 }
824 if (failing()) return;
826 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
828 // Remove clutter produced by parsing.
829 if (!failing()) {
830 ResourceMark rm;
831 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
832 }
833 }
835 // Note: Large methods are capped off in do_one_bytecode().
836 if (failing()) return;
838 // After parsing, node notes are no longer automagic.
839 // They must be propagated by register_new_node_with_optimizer(),
840 // clone(), or the like.
841 set_default_node_notes(NULL);
843 for (;;) {
844 int successes = Inline_Warm();
845 if (failing()) return;
846 if (successes == 0) break;
847 }
849 // Drain the list.
850 Finish_Warm();
851 #ifndef PRODUCT
852 if (_printer) {
853 _printer->print_inlining(this);
854 }
855 #endif
857 if (failing()) return;
858 NOT_PRODUCT( verify_graph_edges(); )
860 // Now optimize
861 Optimize();
862 if (failing()) return;
863 NOT_PRODUCT( verify_graph_edges(); )
865 #ifndef PRODUCT
866 if (PrintIdeal) {
867 ttyLocker ttyl; // keep the following output all in one block
868 // This output goes directly to the tty, not the compiler log.
869 // To enable tools to match it up with the compilation activity,
870 // be sure to tag this tty output with the compile ID.
871 if (xtty != NULL) {
872 xtty->head("ideal compile_id='%d'%s", compile_id(),
873 is_osr_compilation() ? " compile_kind='osr'" :
874 "");
875 }
876 root()->dump(9999);
877 if (xtty != NULL) {
878 xtty->tail("ideal");
879 }
880 }
881 #endif
883 NOT_PRODUCT( verify_barriers(); )
885 // Dump compilation data to replay it.
886 if (method()->has_option("DumpReplay")) {
887 env()->dump_replay_data(_compile_id);
888 }
889 if (method()->has_option("DumpInline") && (ilt() != NULL)) {
890 env()->dump_inline_data(_compile_id);
891 }
893 // Now that we know the size of all the monitors we can add a fixed slot
894 // for the original deopt pc.
896 _orig_pc_slot = fixed_slots();
897 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
898 set_fixed_slots(next_slot);
900 // Compute when to use implicit null checks. Used by matching trap based
901 // nodes and NullCheck optimization.
902 set_allowed_deopt_reasons();
904 // Now generate code
905 Code_Gen();
906 if (failing()) return;
908 // Check if we want to skip execution of all compiled code.
909 {
910 #ifndef PRODUCT
911 if (OptoNoExecute) {
912 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
913 return;
914 }
915 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
916 #endif
918 if (is_osr_compilation()) {
919 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
920 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
921 } else {
922 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
923 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
924 }
926 env()->register_method(_method, _entry_bci,
927 &_code_offsets,
928 _orig_pc_slot_offset_in_bytes,
929 code_buffer(),
930 frame_size_in_words(), _oop_map_set,
931 &_handler_table, &_inc_table,
932 compiler,
933 env()->comp_level(),
934 has_unsafe_access(),
935 SharedRuntime::is_wide_vector(max_vector_size()),
936 rtm_state()
937 );
939 if (log() != NULL) // Print code cache state into compiler log
940 log()->code_cache_state();
941 }
942 }
944 //------------------------------Compile----------------------------------------
945 // Compile a runtime stub
946 Compile::Compile( ciEnv* ci_env,
947 TypeFunc_generator generator,
948 address stub_function,
949 const char *stub_name,
950 int is_fancy_jump,
951 bool pass_tls,
952 bool save_arg_registers,
953 bool return_pc )
954 : Phase(Compiler),
955 _env(ci_env),
956 _log(ci_env->log()),
957 _compile_id(0),
958 _save_argument_registers(save_arg_registers),
959 _method(NULL),
960 _stub_name(stub_name),
961 _stub_function(stub_function),
962 _stub_entry_point(NULL),
963 _entry_bci(InvocationEntryBci),
964 _initial_gvn(NULL),
965 _for_igvn(NULL),
966 _warm_calls(NULL),
967 _orig_pc_slot(0),
968 _orig_pc_slot_offset_in_bytes(0),
969 _subsume_loads(true),
970 _do_escape_analysis(false),
971 _eliminate_boxing(false),
972 _failure_reason(NULL),
973 _code_buffer("Compile::Fill_buffer"),
974 _has_method_handle_invokes(false),
975 _mach_constant_base_node(NULL),
976 _node_bundling_limit(0),
977 _node_bundling_base(NULL),
978 _java_calls(0),
979 _inner_loops(0),
980 #ifndef PRODUCT
981 _trace_opto_output(TraceOptoOutput),
982 _in_dump_cnt(0),
983 _printer(NULL),
984 #endif
985 _comp_arena(mtCompiler),
986 _node_arena(mtCompiler),
987 _old_arena(mtCompiler),
988 _Compile_types(mtCompiler),
989 _dead_node_list(comp_arena()),
990 _dead_node_count(0),
991 _congraph(NULL),
992 _replay_inline_data(NULL),
993 _number_of_mh_late_inlines(0),
994 _inlining_progress(false),
995 _inlining_incrementally(false),
996 _print_inlining_list(NULL),
997 _print_inlining_idx(0),
998 _allowed_reasons(0),
999 _interpreter_frame_size(0),
1000 _max_node_limit(MaxNodeLimit) {
1001 C = this;
1003 #ifndef PRODUCT
1004 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
1005 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
1006 set_print_assembly(PrintFrameConverterAssembly);
1007 set_parsed_irreducible_loop(false);
1008 #endif
1009 set_has_irreducible_loop(false); // no loops
1011 CompileWrapper cw(this);
1012 Init(/*AliasLevel=*/ 0);
1013 init_tf((*generator)());
1015 {
1016 // The following is a dummy for the sake of GraphKit::gen_stub
1017 Unique_Node_List for_igvn(comp_arena());
1018 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1019 PhaseGVN gvn(Thread::current()->resource_area(),255);
1020 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1021 gvn.transform_no_reclaim(top());
1023 GraphKit kit;
1024 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1025 }
1027 NOT_PRODUCT( verify_graph_edges(); )
1028 Code_Gen();
1029 if (failing()) return;
1032 // Entry point will be accessed using compile->stub_entry_point();
1033 if (code_buffer() == NULL) {
1034 Matcher::soft_match_failure();
1035 } else {
1036 if (PrintAssembly && (WizardMode || Verbose))
1037 tty->print_cr("### Stub::%s", stub_name);
1039 if (!failing()) {
1040 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1042 // Make the NMethod
1043 // For now we mark the frame as never safe for profile stackwalking
1044 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1045 code_buffer(),
1046 CodeOffsets::frame_never_safe,
1047 // _code_offsets.value(CodeOffsets::Frame_Complete),
1048 frame_size_in_words(),
1049 _oop_map_set,
1050 save_arg_registers);
1051 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1053 _stub_entry_point = rs->entry_point();
1054 }
1055 }
1056 }
1058 //------------------------------Init-------------------------------------------
1059 // Prepare for a single compilation
1060 void Compile::Init(int aliaslevel) {
1061 _unique = 0;
1062 _regalloc = NULL;
1064 _tf = NULL; // filled in later
1065 _top = NULL; // cached later
1066 _matcher = NULL; // filled in later
1067 _cfg = NULL; // filled in later
1069 set_24_bit_selection_and_mode(Use24BitFP, false);
1071 _node_note_array = NULL;
1072 _default_node_notes = NULL;
1074 _immutable_memory = NULL; // filled in at first inquiry
1076 // Globally visible Nodes
1077 // First set TOP to NULL to give safe behavior during creation of RootNode
1078 set_cached_top_node(NULL);
1079 set_root(new (this) RootNode());
1080 // Now that you have a Root to point to, create the real TOP
1081 set_cached_top_node( new (this) ConNode(Type::TOP) );
1082 set_recent_alloc(NULL, NULL);
1084 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1085 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1086 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1087 env()->set_dependencies(new Dependencies(env()));
1089 _fixed_slots = 0;
1090 set_has_split_ifs(false);
1091 set_has_loops(has_method() && method()->has_loops()); // first approximation
1092 set_has_stringbuilder(false);
1093 set_has_boxed_value(false);
1094 _trap_can_recompile = false; // no traps emitted yet
1095 _major_progress = true; // start out assuming good things will happen
1096 set_has_unsafe_access(false);
1097 set_max_vector_size(0);
1098 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1099 set_decompile_count(0);
1101 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1102 set_num_loop_opts(LoopOptsCount);
1103 set_do_inlining(Inline);
1104 set_max_inline_size(MaxInlineSize);
1105 set_freq_inline_size(FreqInlineSize);
1106 set_do_scheduling(OptoScheduling);
1107 set_do_count_invocations(false);
1108 set_do_method_data_update(false);
1109 set_rtm_state(NoRTM); // No RTM lock eliding by default
1110 method_has_option_value("MaxNodeLimit", _max_node_limit);
1111 #if INCLUDE_RTM_OPT
1112 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1113 int rtm_state = method()->method_data()->rtm_state();
1114 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1115 // Don't generate RTM lock eliding code.
1116 set_rtm_state(NoRTM);
1117 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1118 // Generate RTM lock eliding code without abort ratio calculation code.
1119 set_rtm_state(UseRTM);
1120 } else if (UseRTMDeopt) {
1121 // Generate RTM lock eliding code and include abort ratio calculation
1122 // code if UseRTMDeopt is on.
1123 set_rtm_state(ProfileRTM);
1124 }
1125 }
1126 #endif
1127 if (debug_info()->recording_non_safepoints()) {
1128 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1129 (comp_arena(), 8, 0, NULL));
1130 set_default_node_notes(Node_Notes::make(this));
1131 }
1133 // // -- Initialize types before each compile --
1134 // // Update cached type information
1135 // if( _method && _method->constants() )
1136 // Type::update_loaded_types(_method, _method->constants());
1138 // Init alias_type map.
1139 if (!_do_escape_analysis && aliaslevel == 3)
1140 aliaslevel = 2; // No unique types without escape analysis
1141 _AliasLevel = aliaslevel;
1142 const int grow_ats = 16;
1143 _max_alias_types = grow_ats;
1144 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1145 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1146 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1147 {
1148 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1149 }
1150 // Initialize the first few types.
1151 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1152 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1153 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1154 _num_alias_types = AliasIdxRaw+1;
1155 // Zero out the alias type cache.
1156 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1157 // A NULL adr_type hits in the cache right away. Preload the right answer.
1158 probe_alias_cache(NULL)->_index = AliasIdxTop;
1160 _intrinsics = NULL;
1161 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1162 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1163 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1164 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1165 register_library_intrinsics();
1166 }
1168 //---------------------------init_start----------------------------------------
1169 // Install the StartNode on this compile object.
1170 void Compile::init_start(StartNode* s) {
1171 if (failing())
1172 return; // already failing
1173 assert(s == start(), "");
1174 }
1176 StartNode* Compile::start() const {
1177 assert(!failing(), "");
1178 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1179 Node* start = root()->fast_out(i);
1180 if( start->is_Start() )
1181 return start->as_Start();
1182 }
1183 fatal("Did not find Start node!");
1184 return NULL;
1185 }
1187 //-------------------------------immutable_memory-------------------------------------
1188 // Access immutable memory
1189 Node* Compile::immutable_memory() {
1190 if (_immutable_memory != NULL) {
1191 return _immutable_memory;
1192 }
1193 StartNode* s = start();
1194 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1195 Node *p = s->fast_out(i);
1196 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1197 _immutable_memory = p;
1198 return _immutable_memory;
1199 }
1200 }
1201 ShouldNotReachHere();
1202 return NULL;
1203 }
1205 //----------------------set_cached_top_node------------------------------------
1206 // Install the cached top node, and make sure Node::is_top works correctly.
1207 void Compile::set_cached_top_node(Node* tn) {
1208 if (tn != NULL) verify_top(tn);
1209 Node* old_top = _top;
1210 _top = tn;
1211 // Calling Node::setup_is_top allows the nodes the chance to adjust
1212 // their _out arrays.
1213 if (_top != NULL) _top->setup_is_top();
1214 if (old_top != NULL) old_top->setup_is_top();
1215 assert(_top == NULL || top()->is_top(), "");
1216 }
1218 #ifdef ASSERT
1219 uint Compile::count_live_nodes_by_graph_walk() {
1220 Unique_Node_List useful(comp_arena());
1221 // Get useful node list by walking the graph.
1222 identify_useful_nodes(useful);
1223 return useful.size();
1224 }
1226 void Compile::print_missing_nodes() {
1228 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1229 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1230 return;
1231 }
1233 // This is an expensive function. It is executed only when the user
1234 // specifies VerifyIdealNodeCount option or otherwise knows the
1235 // additional work that needs to be done to identify reachable nodes
1236 // by walking the flow graph and find the missing ones using
1237 // _dead_node_list.
1239 Unique_Node_List useful(comp_arena());
1240 // Get useful node list by walking the graph.
1241 identify_useful_nodes(useful);
1243 uint l_nodes = C->live_nodes();
1244 uint l_nodes_by_walk = useful.size();
1246 if (l_nodes != l_nodes_by_walk) {
1247 if (_log != NULL) {
1248 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1249 _log->stamp();
1250 _log->end_head();
1251 }
1252 VectorSet& useful_member_set = useful.member_set();
1253 int last_idx = l_nodes_by_walk;
1254 for (int i = 0; i < last_idx; i++) {
1255 if (useful_member_set.test(i)) {
1256 if (_dead_node_list.test(i)) {
1257 if (_log != NULL) {
1258 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1259 }
1260 if (PrintIdealNodeCount) {
1261 // Print the log message to tty
1262 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1263 useful.at(i)->dump();
1264 }
1265 }
1266 }
1267 else if (! _dead_node_list.test(i)) {
1268 if (_log != NULL) {
1269 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1270 }
1271 if (PrintIdealNodeCount) {
1272 // Print the log message to tty
1273 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1274 }
1275 }
1276 }
1277 if (_log != NULL) {
1278 _log->tail("mismatched_nodes");
1279 }
1280 }
1281 }
1282 #endif
1284 #ifndef PRODUCT
1285 void Compile::verify_top(Node* tn) const {
1286 if (tn != NULL) {
1287 assert(tn->is_Con(), "top node must be a constant");
1288 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1289 assert(tn->in(0) != NULL, "must have live top node");
1290 }
1291 }
1292 #endif
1295 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1297 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1298 guarantee(arr != NULL, "");
1299 int num_blocks = arr->length();
1300 if (grow_by < num_blocks) grow_by = num_blocks;
1301 int num_notes = grow_by * _node_notes_block_size;
1302 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1303 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1304 while (num_notes > 0) {
1305 arr->append(notes);
1306 notes += _node_notes_block_size;
1307 num_notes -= _node_notes_block_size;
1308 }
1309 assert(num_notes == 0, "exact multiple, please");
1310 }
1312 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1313 if (source == NULL || dest == NULL) return false;
1315 if (dest->is_Con())
1316 return false; // Do not push debug info onto constants.
1318 #ifdef ASSERT
1319 // Leave a bread crumb trail pointing to the original node:
1320 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1321 dest->set_debug_orig(source);
1322 }
1323 #endif
1325 if (node_note_array() == NULL)
1326 return false; // Not collecting any notes now.
1328 // This is a copy onto a pre-existing node, which may already have notes.
1329 // If both nodes have notes, do not overwrite any pre-existing notes.
1330 Node_Notes* source_notes = node_notes_at(source->_idx);
1331 if (source_notes == NULL || source_notes->is_clear()) return false;
1332 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1333 if (dest_notes == NULL || dest_notes->is_clear()) {
1334 return set_node_notes_at(dest->_idx, source_notes);
1335 }
1337 Node_Notes merged_notes = (*source_notes);
1338 // The order of operations here ensures that dest notes will win...
1339 merged_notes.update_from(dest_notes);
1340 return set_node_notes_at(dest->_idx, &merged_notes);
1341 }
1344 //--------------------------allow_range_check_smearing-------------------------
1345 // Gating condition for coalescing similar range checks.
1346 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1347 // single covering check that is at least as strong as any of them.
1348 // If the optimization succeeds, the simplified (strengthened) range check
1349 // will always succeed. If it fails, we will deopt, and then give up
1350 // on the optimization.
1351 bool Compile::allow_range_check_smearing() const {
1352 // If this method has already thrown a range-check,
1353 // assume it was because we already tried range smearing
1354 // and it failed.
1355 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1356 return !already_trapped;
1357 }
1360 //------------------------------flatten_alias_type-----------------------------
1361 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1362 int offset = tj->offset();
1363 TypePtr::PTR ptr = tj->ptr();
1365 // Known instance (scalarizable allocation) alias only with itself.
1366 bool is_known_inst = tj->isa_oopptr() != NULL &&
1367 tj->is_oopptr()->is_known_instance();
1369 // Process weird unsafe references.
1370 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1371 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1372 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1373 tj = TypeOopPtr::BOTTOM;
1374 ptr = tj->ptr();
1375 offset = tj->offset();
1376 }
1378 // Array pointers need some flattening
1379 const TypeAryPtr *ta = tj->isa_aryptr();
1380 if (ta && ta->is_stable()) {
1381 // Erase stability property for alias analysis.
1382 tj = ta = ta->cast_to_stable(false);
1383 }
1384 if( ta && is_known_inst ) {
1385 if ( offset != Type::OffsetBot &&
1386 offset > arrayOopDesc::length_offset_in_bytes() ) {
1387 offset = Type::OffsetBot; // Flatten constant access into array body only
1388 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1389 }
1390 } else if( ta && _AliasLevel >= 2 ) {
1391 // For arrays indexed by constant indices, we flatten the alias
1392 // space to include all of the array body. Only the header, klass
1393 // and array length can be accessed un-aliased.
1394 if( offset != Type::OffsetBot ) {
1395 if( ta->const_oop() ) { // MethodData* or Method*
1396 offset = Type::OffsetBot; // Flatten constant access into array body
1397 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1398 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1399 // range is OK as-is.
1400 tj = ta = TypeAryPtr::RANGE;
1401 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1402 tj = TypeInstPtr::KLASS; // all klass loads look alike
1403 ta = TypeAryPtr::RANGE; // generic ignored junk
1404 ptr = TypePtr::BotPTR;
1405 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1406 tj = TypeInstPtr::MARK;
1407 ta = TypeAryPtr::RANGE; // generic ignored junk
1408 ptr = TypePtr::BotPTR;
1409 } else { // Random constant offset into array body
1410 offset = Type::OffsetBot; // Flatten constant access into array body
1411 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1412 }
1413 }
1414 // Arrays of fixed size alias with arrays of unknown size.
1415 if (ta->size() != TypeInt::POS) {
1416 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1417 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1418 }
1419 // Arrays of known objects become arrays of unknown objects.
1420 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1421 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1422 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1423 }
1424 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1425 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1426 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1427 }
1428 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1429 // cannot be distinguished by bytecode alone.
1430 if (ta->elem() == TypeInt::BOOL) {
1431 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1432 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1433 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1434 }
1435 // During the 2nd round of IterGVN, NotNull castings are removed.
1436 // Make sure the Bottom and NotNull variants alias the same.
1437 // Also, make sure exact and non-exact variants alias the same.
1438 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1439 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1440 }
1441 }
1443 // Oop pointers need some flattening
1444 const TypeInstPtr *to = tj->isa_instptr();
1445 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1446 ciInstanceKlass *k = to->klass()->as_instance_klass();
1447 if( ptr == TypePtr::Constant ) {
1448 if (to->klass() != ciEnv::current()->Class_klass() ||
1449 offset < k->size_helper() * wordSize) {
1450 // No constant oop pointers (such as Strings); they alias with
1451 // unknown strings.
1452 assert(!is_known_inst, "not scalarizable allocation");
1453 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1454 }
1455 } else if( is_known_inst ) {
1456 tj = to; // Keep NotNull and klass_is_exact for instance type
1457 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1458 // During the 2nd round of IterGVN, NotNull castings are removed.
1459 // Make sure the Bottom and NotNull variants alias the same.
1460 // Also, make sure exact and non-exact variants alias the same.
1461 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1462 }
1463 if (to->speculative() != NULL) {
1464 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1465 }
1466 // Canonicalize the holder of this field
1467 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1468 // First handle header references such as a LoadKlassNode, even if the
1469 // object's klass is unloaded at compile time (4965979).
1470 if (!is_known_inst) { // Do it only for non-instance types
1471 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1472 }
1473 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1474 // Static fields are in the space above the normal instance
1475 // fields in the java.lang.Class instance.
1476 if (to->klass() != ciEnv::current()->Class_klass()) {
1477 to = NULL;
1478 tj = TypeOopPtr::BOTTOM;
1479 offset = tj->offset();
1480 }
1481 } else {
1482 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1483 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1484 if( is_known_inst ) {
1485 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1486 } else {
1487 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1488 }
1489 }
1490 }
1491 }
1493 // Klass pointers to object array klasses need some flattening
1494 const TypeKlassPtr *tk = tj->isa_klassptr();
1495 if( tk ) {
1496 // If we are referencing a field within a Klass, we need
1497 // to assume the worst case of an Object. Both exact and
1498 // inexact types must flatten to the same alias class so
1499 // use NotNull as the PTR.
1500 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1502 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1503 TypeKlassPtr::OBJECT->klass(),
1504 offset);
1505 }
1507 ciKlass* klass = tk->klass();
1508 if( klass->is_obj_array_klass() ) {
1509 ciKlass* k = TypeAryPtr::OOPS->klass();
1510 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1511 k = TypeInstPtr::BOTTOM->klass();
1512 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1513 }
1515 // Check for precise loads from the primary supertype array and force them
1516 // to the supertype cache alias index. Check for generic array loads from
1517 // the primary supertype array and also force them to the supertype cache
1518 // alias index. Since the same load can reach both, we need to merge
1519 // these 2 disparate memories into the same alias class. Since the
1520 // primary supertype array is read-only, there's no chance of confusion
1521 // where we bypass an array load and an array store.
1522 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1523 if (offset == Type::OffsetBot ||
1524 (offset >= primary_supers_offset &&
1525 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1526 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1527 offset = in_bytes(Klass::secondary_super_cache_offset());
1528 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1529 }
1530 }
1532 // Flatten all Raw pointers together.
1533 if (tj->base() == Type::RawPtr)
1534 tj = TypeRawPtr::BOTTOM;
1536 if (tj->base() == Type::AnyPtr)
1537 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1539 // Flatten all to bottom for now
1540 switch( _AliasLevel ) {
1541 case 0:
1542 tj = TypePtr::BOTTOM;
1543 break;
1544 case 1: // Flatten to: oop, static, field or array
1545 switch (tj->base()) {
1546 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1547 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1548 case Type::AryPtr: // do not distinguish arrays at all
1549 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1550 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1551 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1552 default: ShouldNotReachHere();
1553 }
1554 break;
1555 case 2: // No collapsing at level 2; keep all splits
1556 case 3: // No collapsing at level 3; keep all splits
1557 break;
1558 default:
1559 Unimplemented();
1560 }
1562 offset = tj->offset();
1563 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1565 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1566 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1567 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1568 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1569 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1570 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1571 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1572 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1573 assert( tj->ptr() != TypePtr::TopPTR &&
1574 tj->ptr() != TypePtr::AnyNull &&
1575 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1576 // assert( tj->ptr() != TypePtr::Constant ||
1577 // tj->base() == Type::RawPtr ||
1578 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1580 return tj;
1581 }
1583 void Compile::AliasType::Init(int i, const TypePtr* at) {
1584 _index = i;
1585 _adr_type = at;
1586 _field = NULL;
1587 _element = NULL;
1588 _is_rewritable = true; // default
1589 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1590 if (atoop != NULL && atoop->is_known_instance()) {
1591 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1592 _general_index = Compile::current()->get_alias_index(gt);
1593 } else {
1594 _general_index = 0;
1595 }
1596 }
1598 //---------------------------------print_on------------------------------------
1599 #ifndef PRODUCT
1600 void Compile::AliasType::print_on(outputStream* st) {
1601 if (index() < 10)
1602 st->print("@ <%d> ", index());
1603 else st->print("@ <%d>", index());
1604 st->print(is_rewritable() ? " " : " RO");
1605 int offset = adr_type()->offset();
1606 if (offset == Type::OffsetBot)
1607 st->print(" +any");
1608 else st->print(" +%-3d", offset);
1609 st->print(" in ");
1610 adr_type()->dump_on(st);
1611 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1612 if (field() != NULL && tjp) {
1613 if (tjp->klass() != field()->holder() ||
1614 tjp->offset() != field()->offset_in_bytes()) {
1615 st->print(" != ");
1616 field()->print();
1617 st->print(" ***");
1618 }
1619 }
1620 }
1622 void print_alias_types() {
1623 Compile* C = Compile::current();
1624 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1625 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1626 C->alias_type(idx)->print_on(tty);
1627 tty->cr();
1628 }
1629 }
1630 #endif
1633 //----------------------------probe_alias_cache--------------------------------
1634 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1635 intptr_t key = (intptr_t) adr_type;
1636 key ^= key >> logAliasCacheSize;
1637 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1638 }
1641 //-----------------------------grow_alias_types--------------------------------
1642 void Compile::grow_alias_types() {
1643 const int old_ats = _max_alias_types; // how many before?
1644 const int new_ats = old_ats; // how many more?
1645 const int grow_ats = old_ats+new_ats; // how many now?
1646 _max_alias_types = grow_ats;
1647 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1648 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1649 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1650 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1651 }
1654 //--------------------------------find_alias_type------------------------------
1655 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1656 if (_AliasLevel == 0)
1657 return alias_type(AliasIdxBot);
1659 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1660 if (ace->_adr_type == adr_type) {
1661 return alias_type(ace->_index);
1662 }
1664 // Handle special cases.
1665 if (adr_type == NULL) return alias_type(AliasIdxTop);
1666 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1668 // Do it the slow way.
1669 const TypePtr* flat = flatten_alias_type(adr_type);
1671 #ifdef ASSERT
1672 assert(flat == flatten_alias_type(flat), "idempotent");
1673 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1674 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1675 const TypeOopPtr* foop = flat->is_oopptr();
1676 // Scalarizable allocations have exact klass always.
1677 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1678 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1679 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1680 }
1681 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1682 #endif
1684 int idx = AliasIdxTop;
1685 for (int i = 0; i < num_alias_types(); i++) {
1686 if (alias_type(i)->adr_type() == flat) {
1687 idx = i;
1688 break;
1689 }
1690 }
1692 if (idx == AliasIdxTop) {
1693 if (no_create) return NULL;
1694 // Grow the array if necessary.
1695 if (_num_alias_types == _max_alias_types) grow_alias_types();
1696 // Add a new alias type.
1697 idx = _num_alias_types++;
1698 _alias_types[idx]->Init(idx, flat);
1699 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1700 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1701 if (flat->isa_instptr()) {
1702 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1703 && flat->is_instptr()->klass() == env()->Class_klass())
1704 alias_type(idx)->set_rewritable(false);
1705 }
1706 if (flat->isa_aryptr()) {
1707 #ifdef ASSERT
1708 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1709 // (T_BYTE has the weakest alignment and size restrictions...)
1710 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1711 #endif
1712 if (flat->offset() == TypePtr::OffsetBot) {
1713 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1714 }
1715 }
1716 if (flat->isa_klassptr()) {
1717 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1718 alias_type(idx)->set_rewritable(false);
1719 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1720 alias_type(idx)->set_rewritable(false);
1721 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1722 alias_type(idx)->set_rewritable(false);
1723 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1724 alias_type(idx)->set_rewritable(false);
1725 }
1726 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1727 // but the base pointer type is not distinctive enough to identify
1728 // references into JavaThread.)
1730 // Check for final fields.
1731 const TypeInstPtr* tinst = flat->isa_instptr();
1732 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1733 ciField* field;
1734 if (tinst->const_oop() != NULL &&
1735 tinst->klass() == ciEnv::current()->Class_klass() &&
1736 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1737 // static field
1738 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1739 field = k->get_field_by_offset(tinst->offset(), true);
1740 } else {
1741 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1742 field = k->get_field_by_offset(tinst->offset(), false);
1743 }
1744 assert(field == NULL ||
1745 original_field == NULL ||
1746 (field->holder() == original_field->holder() &&
1747 field->offset() == original_field->offset() &&
1748 field->is_static() == original_field->is_static()), "wrong field?");
1749 // Set field() and is_rewritable() attributes.
1750 if (field != NULL) alias_type(idx)->set_field(field);
1751 }
1752 }
1754 // Fill the cache for next time.
1755 ace->_adr_type = adr_type;
1756 ace->_index = idx;
1757 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1759 // Might as well try to fill the cache for the flattened version, too.
1760 AliasCacheEntry* face = probe_alias_cache(flat);
1761 if (face->_adr_type == NULL) {
1762 face->_adr_type = flat;
1763 face->_index = idx;
1764 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1765 }
1767 return alias_type(idx);
1768 }
1771 Compile::AliasType* Compile::alias_type(ciField* field) {
1772 const TypeOopPtr* t;
1773 if (field->is_static())
1774 t = TypeInstPtr::make(field->holder()->java_mirror());
1775 else
1776 t = TypeOopPtr::make_from_klass_raw(field->holder());
1777 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1778 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1779 return atp;
1780 }
1783 //------------------------------have_alias_type--------------------------------
1784 bool Compile::have_alias_type(const TypePtr* adr_type) {
1785 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1786 if (ace->_adr_type == adr_type) {
1787 return true;
1788 }
1790 // Handle special cases.
1791 if (adr_type == NULL) return true;
1792 if (adr_type == TypePtr::BOTTOM) return true;
1794 return find_alias_type(adr_type, true, NULL) != NULL;
1795 }
1797 //-----------------------------must_alias--------------------------------------
1798 // True if all values of the given address type are in the given alias category.
1799 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1800 if (alias_idx == AliasIdxBot) return true; // the universal category
1801 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1802 if (alias_idx == AliasIdxTop) return false; // the empty category
1803 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1805 // the only remaining possible overlap is identity
1806 int adr_idx = get_alias_index(adr_type);
1807 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1808 assert(adr_idx == alias_idx ||
1809 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1810 && adr_type != TypeOopPtr::BOTTOM),
1811 "should not be testing for overlap with an unsafe pointer");
1812 return adr_idx == alias_idx;
1813 }
1815 //------------------------------can_alias--------------------------------------
1816 // True if any values of the given address type are in the given alias category.
1817 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1818 if (alias_idx == AliasIdxTop) return false; // the empty category
1819 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1820 if (alias_idx == AliasIdxBot) return true; // the universal category
1821 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1823 // the only remaining possible overlap is identity
1824 int adr_idx = get_alias_index(adr_type);
1825 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1826 return adr_idx == alias_idx;
1827 }
1831 //---------------------------pop_warm_call-------------------------------------
1832 WarmCallInfo* Compile::pop_warm_call() {
1833 WarmCallInfo* wci = _warm_calls;
1834 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1835 return wci;
1836 }
1838 //----------------------------Inline_Warm--------------------------------------
1839 int Compile::Inline_Warm() {
1840 // If there is room, try to inline some more warm call sites.
1841 // %%% Do a graph index compaction pass when we think we're out of space?
1842 if (!InlineWarmCalls) return 0;
1844 int calls_made_hot = 0;
1845 int room_to_grow = NodeCountInliningCutoff - unique();
1846 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1847 int amount_grown = 0;
1848 WarmCallInfo* call;
1849 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1850 int est_size = (int)call->size();
1851 if (est_size > (room_to_grow - amount_grown)) {
1852 // This one won't fit anyway. Get rid of it.
1853 call->make_cold();
1854 continue;
1855 }
1856 call->make_hot();
1857 calls_made_hot++;
1858 amount_grown += est_size;
1859 amount_to_grow -= est_size;
1860 }
1862 if (calls_made_hot > 0) set_major_progress();
1863 return calls_made_hot;
1864 }
1867 //----------------------------Finish_Warm--------------------------------------
1868 void Compile::Finish_Warm() {
1869 if (!InlineWarmCalls) return;
1870 if (failing()) return;
1871 if (warm_calls() == NULL) return;
1873 // Clean up loose ends, if we are out of space for inlining.
1874 WarmCallInfo* call;
1875 while ((call = pop_warm_call()) != NULL) {
1876 call->make_cold();
1877 }
1878 }
1880 //---------------------cleanup_loop_predicates-----------------------
1881 // Remove the opaque nodes that protect the predicates so that all unused
1882 // checks and uncommon_traps will be eliminated from the ideal graph
1883 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1884 if (predicate_count()==0) return;
1885 for (int i = predicate_count(); i > 0; i--) {
1886 Node * n = predicate_opaque1_node(i-1);
1887 assert(n->Opcode() == Op_Opaque1, "must be");
1888 igvn.replace_node(n, n->in(1));
1889 }
1890 assert(predicate_count()==0, "should be clean!");
1891 }
1893 void Compile::add_range_check_cast(Node* n) {
1894 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1895 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1896 _range_check_casts->append(n);
1897 }
1899 // Remove all range check dependent CastIINodes.
1900 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1901 for (int i = range_check_cast_count(); i > 0; i--) {
1902 Node* cast = range_check_cast_node(i-1);
1903 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1904 igvn.replace_node(cast, cast->in(1));
1905 }
1906 assert(range_check_cast_count() == 0, "should be empty");
1907 }
1909 // StringOpts and late inlining of string methods
1910 void Compile::inline_string_calls(bool parse_time) {
1911 {
1912 // remove useless nodes to make the usage analysis simpler
1913 ResourceMark rm;
1914 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1915 }
1917 {
1918 ResourceMark rm;
1919 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1920 PhaseStringOpts pso(initial_gvn(), for_igvn());
1921 print_method(PHASE_AFTER_STRINGOPTS, 3);
1922 }
1924 // now inline anything that we skipped the first time around
1925 if (!parse_time) {
1926 _late_inlines_pos = _late_inlines.length();
1927 }
1929 while (_string_late_inlines.length() > 0) {
1930 CallGenerator* cg = _string_late_inlines.pop();
1931 cg->do_late_inline();
1932 if (failing()) return;
1933 }
1934 _string_late_inlines.trunc_to(0);
1935 }
1937 // Late inlining of boxing methods
1938 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1939 if (_boxing_late_inlines.length() > 0) {
1940 assert(has_boxed_value(), "inconsistent");
1942 PhaseGVN* gvn = initial_gvn();
1943 set_inlining_incrementally(true);
1945 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1946 for_igvn()->clear();
1947 gvn->replace_with(&igvn);
1949 _late_inlines_pos = _late_inlines.length();
1951 while (_boxing_late_inlines.length() > 0) {
1952 CallGenerator* cg = _boxing_late_inlines.pop();
1953 cg->do_late_inline();
1954 if (failing()) return;
1955 }
1956 _boxing_late_inlines.trunc_to(0);
1958 {
1959 ResourceMark rm;
1960 PhaseRemoveUseless pru(gvn, for_igvn());
1961 }
1963 igvn = PhaseIterGVN(gvn);
1964 igvn.optimize();
1966 set_inlining_progress(false);
1967 set_inlining_incrementally(false);
1968 }
1969 }
1971 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1972 assert(IncrementalInline, "incremental inlining should be on");
1973 PhaseGVN* gvn = initial_gvn();
1975 set_inlining_progress(false);
1976 for_igvn()->clear();
1977 gvn->replace_with(&igvn);
1979 int i = 0;
1981 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1982 CallGenerator* cg = _late_inlines.at(i);
1983 _late_inlines_pos = i+1;
1984 cg->do_late_inline();
1985 if (failing()) return;
1986 }
1987 int j = 0;
1988 for (; i < _late_inlines.length(); i++, j++) {
1989 _late_inlines.at_put(j, _late_inlines.at(i));
1990 }
1991 _late_inlines.trunc_to(j);
1993 {
1994 ResourceMark rm;
1995 PhaseRemoveUseless pru(gvn, for_igvn());
1996 }
1998 igvn = PhaseIterGVN(gvn);
1999 }
2001 // Perform incremental inlining until bound on number of live nodes is reached
2002 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2003 PhaseGVN* gvn = initial_gvn();
2005 set_inlining_incrementally(true);
2006 set_inlining_progress(true);
2007 uint low_live_nodes = 0;
2009 while(inlining_progress() && _late_inlines.length() > 0) {
2011 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2012 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2013 // PhaseIdealLoop is expensive so we only try it once we are
2014 // out of live nodes and we only try it again if the previous
2015 // helped got the number of nodes down significantly
2016 PhaseIdealLoop ideal_loop( igvn, false, true );
2017 if (failing()) return;
2018 low_live_nodes = live_nodes();
2019 _major_progress = true;
2020 }
2022 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2023 break;
2024 }
2025 }
2027 inline_incrementally_one(igvn);
2029 if (failing()) return;
2031 igvn.optimize();
2033 if (failing()) return;
2034 }
2036 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2038 if (_string_late_inlines.length() > 0) {
2039 assert(has_stringbuilder(), "inconsistent");
2040 for_igvn()->clear();
2041 initial_gvn()->replace_with(&igvn);
2043 inline_string_calls(false);
2045 if (failing()) return;
2047 {
2048 ResourceMark rm;
2049 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2050 }
2052 igvn = PhaseIterGVN(gvn);
2054 igvn.optimize();
2055 }
2057 set_inlining_incrementally(false);
2058 }
2061 //------------------------------Optimize---------------------------------------
2062 // Given a graph, optimize it.
2063 void Compile::Optimize() {
2064 TracePhase t1("optimizer", &_t_optimizer, true);
2066 #ifndef PRODUCT
2067 if (env()->break_at_compile()) {
2068 BREAKPOINT;
2069 }
2071 #endif
2073 ResourceMark rm;
2074 int loop_opts_cnt;
2076 NOT_PRODUCT( verify_graph_edges(); )
2078 print_method(PHASE_AFTER_PARSING);
2080 {
2081 // Iterative Global Value Numbering, including ideal transforms
2082 // Initialize IterGVN with types and values from parse-time GVN
2083 PhaseIterGVN igvn(initial_gvn());
2084 {
2085 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2086 igvn.optimize();
2087 }
2089 print_method(PHASE_ITER_GVN1, 2);
2091 if (failing()) return;
2093 {
2094 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2095 inline_incrementally(igvn);
2096 }
2098 print_method(PHASE_INCREMENTAL_INLINE, 2);
2100 if (failing()) return;
2102 if (eliminate_boxing()) {
2103 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2104 // Inline valueOf() methods now.
2105 inline_boxing_calls(igvn);
2107 if (AlwaysIncrementalInline) {
2108 inline_incrementally(igvn);
2109 }
2111 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2113 if (failing()) return;
2114 }
2116 // Remove the speculative part of types and clean up the graph from
2117 // the extra CastPP nodes whose only purpose is to carry them. Do
2118 // that early so that optimizations are not disrupted by the extra
2119 // CastPP nodes.
2120 remove_speculative_types(igvn);
2122 // No more new expensive nodes will be added to the list from here
2123 // so keep only the actual candidates for optimizations.
2124 cleanup_expensive_nodes(igvn);
2126 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2127 NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);)
2128 initial_gvn()->replace_with(&igvn);
2129 for_igvn()->clear();
2130 Unique_Node_List new_worklist(C->comp_arena());
2131 {
2132 ResourceMark rm;
2133 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2134 }
2135 set_for_igvn(&new_worklist);
2136 igvn = PhaseIterGVN(initial_gvn());
2137 igvn.optimize();
2138 }
2140 // Perform escape analysis
2141 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2142 if (has_loops()) {
2143 // Cleanup graph (remove dead nodes).
2144 TracePhase t2("idealLoop", &_t_idealLoop, true);
2145 PhaseIdealLoop ideal_loop( igvn, false, true );
2146 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2147 if (failing()) return;
2148 }
2149 ConnectionGraph::do_analysis(this, &igvn);
2151 if (failing()) return;
2153 // Optimize out fields loads from scalar replaceable allocations.
2154 igvn.optimize();
2155 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2157 if (failing()) return;
2159 if (congraph() != NULL && macro_count() > 0) {
2160 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2161 PhaseMacroExpand mexp(igvn);
2162 mexp.eliminate_macro_nodes();
2163 igvn.set_delay_transform(false);
2165 igvn.optimize();
2166 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2168 if (failing()) return;
2169 }
2170 }
2172 // Loop transforms on the ideal graph. Range Check Elimination,
2173 // peeling, unrolling, etc.
2175 // Set loop opts counter
2176 loop_opts_cnt = num_loop_opts();
2177 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2178 {
2179 TracePhase t2("idealLoop", &_t_idealLoop, true);
2180 PhaseIdealLoop ideal_loop( igvn, true );
2181 loop_opts_cnt--;
2182 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2183 if (failing()) return;
2184 }
2185 // Loop opts pass if partial peeling occurred in previous pass
2186 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2187 TracePhase t3("idealLoop", &_t_idealLoop, true);
2188 PhaseIdealLoop ideal_loop( igvn, false );
2189 loop_opts_cnt--;
2190 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2191 if (failing()) return;
2192 }
2193 // Loop opts pass for loop-unrolling before CCP
2194 if(major_progress() && (loop_opts_cnt > 0)) {
2195 TracePhase t4("idealLoop", &_t_idealLoop, true);
2196 PhaseIdealLoop ideal_loop( igvn, false );
2197 loop_opts_cnt--;
2198 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2199 }
2200 if (!failing()) {
2201 // Verify that last round of loop opts produced a valid graph
2202 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2203 PhaseIdealLoop::verify(igvn);
2204 }
2205 }
2206 if (failing()) return;
2208 // Conditional Constant Propagation;
2209 PhaseCCP ccp( &igvn );
2210 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2211 {
2212 TracePhase t2("ccp", &_t_ccp, true);
2213 ccp.do_transform();
2214 }
2215 print_method(PHASE_CPP1, 2);
2217 assert( true, "Break here to ccp.dump_old2new_map()");
2219 // Iterative Global Value Numbering, including ideal transforms
2220 {
2221 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2222 igvn = ccp;
2223 igvn.optimize();
2224 }
2226 print_method(PHASE_ITER_GVN2, 2);
2228 if (failing()) return;
2230 // Loop transforms on the ideal graph. Range Check Elimination,
2231 // peeling, unrolling, etc.
2232 if(loop_opts_cnt > 0) {
2233 debug_only( int cnt = 0; );
2234 while(major_progress() && (loop_opts_cnt > 0)) {
2235 TracePhase t2("idealLoop", &_t_idealLoop, true);
2236 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2237 PhaseIdealLoop ideal_loop( igvn, true);
2238 loop_opts_cnt--;
2239 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2240 if (failing()) return;
2241 }
2242 }
2244 {
2245 // Verify that all previous optimizations produced a valid graph
2246 // at least to this point, even if no loop optimizations were done.
2247 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2248 PhaseIdealLoop::verify(igvn);
2249 }
2251 if (range_check_cast_count() > 0) {
2252 // No more loop optimizations. Remove all range check dependent CastIINodes.
2253 C->remove_range_check_casts(igvn);
2254 igvn.optimize();
2255 }
2257 {
2258 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2259 PhaseMacroExpand mex(igvn);
2260 if (mex.expand_macro_nodes()) {
2261 assert(failing(), "must bail out w/ explicit message");
2262 return;
2263 }
2264 }
2266 } // (End scope of igvn; run destructor if necessary for asserts.)
2268 dump_inlining();
2269 // A method with only infinite loops has no edges entering loops from root
2270 {
2271 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2272 if (final_graph_reshaping()) {
2273 assert(failing(), "must bail out w/ explicit message");
2274 return;
2275 }
2276 }
2278 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2279 }
2282 //------------------------------Code_Gen---------------------------------------
2283 // Given a graph, generate code for it
2284 void Compile::Code_Gen() {
2285 if (failing()) {
2286 return;
2287 }
2289 // Perform instruction selection. You might think we could reclaim Matcher
2290 // memory PDQ, but actually the Matcher is used in generating spill code.
2291 // Internals of the Matcher (including some VectorSets) must remain live
2292 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2293 // set a bit in reclaimed memory.
2295 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2296 // nodes. Mapping is only valid at the root of each matched subtree.
2297 NOT_PRODUCT( verify_graph_edges(); )
2299 Matcher matcher;
2300 _matcher = &matcher;
2301 {
2302 TracePhase t2("matcher", &_t_matcher, true);
2303 matcher.match();
2304 }
2305 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2306 // nodes. Mapping is only valid at the root of each matched subtree.
2307 NOT_PRODUCT( verify_graph_edges(); )
2309 // If you have too many nodes, or if matching has failed, bail out
2310 check_node_count(0, "out of nodes matching instructions");
2311 if (failing()) {
2312 return;
2313 }
2315 // Build a proper-looking CFG
2316 PhaseCFG cfg(node_arena(), root(), matcher);
2317 _cfg = &cfg;
2318 {
2319 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2320 bool success = cfg.do_global_code_motion();
2321 if (!success) {
2322 return;
2323 }
2325 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2326 NOT_PRODUCT( verify_graph_edges(); )
2327 debug_only( cfg.verify(); )
2328 }
2330 PhaseChaitin regalloc(unique(), cfg, matcher);
2331 _regalloc = ®alloc;
2332 {
2333 TracePhase t2("regalloc", &_t_registerAllocation, true);
2334 // Perform register allocation. After Chaitin, use-def chains are
2335 // no longer accurate (at spill code) and so must be ignored.
2336 // Node->LRG->reg mappings are still accurate.
2337 _regalloc->Register_Allocate();
2339 // Bail out if the allocator builds too many nodes
2340 if (failing()) {
2341 return;
2342 }
2343 }
2345 // Prior to register allocation we kept empty basic blocks in case the
2346 // the allocator needed a place to spill. After register allocation we
2347 // are not adding any new instructions. If any basic block is empty, we
2348 // can now safely remove it.
2349 {
2350 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2351 cfg.remove_empty_blocks();
2352 if (do_freq_based_layout()) {
2353 PhaseBlockLayout layout(cfg);
2354 } else {
2355 cfg.set_loop_alignment();
2356 }
2357 cfg.fixup_flow();
2358 }
2360 // Apply peephole optimizations
2361 if( OptoPeephole ) {
2362 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2363 PhasePeephole peep( _regalloc, cfg);
2364 peep.do_transform();
2365 }
2367 // Do late expand if CPU requires this.
2368 if (Matcher::require_postalloc_expand) {
2369 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2370 cfg.postalloc_expand(_regalloc);
2371 }
2373 // Convert Nodes to instruction bits in a buffer
2374 {
2375 // %%%% workspace merge brought two timers together for one job
2376 TracePhase t2a("output", &_t_output, true);
2377 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2378 Output();
2379 }
2381 print_method(PHASE_FINAL_CODE);
2383 // He's dead, Jim.
2384 _cfg = (PhaseCFG*)0xdeadbeef;
2385 _regalloc = (PhaseChaitin*)0xdeadbeef;
2386 }
2389 //------------------------------dump_asm---------------------------------------
2390 // Dump formatted assembly
2391 #ifndef PRODUCT
2392 void Compile::dump_asm(int *pcs, uint pc_limit) {
2393 bool cut_short = false;
2394 tty->print_cr("#");
2395 tty->print("# "); _tf->dump(); tty->cr();
2396 tty->print_cr("#");
2398 // For all blocks
2399 int pc = 0x0; // Program counter
2400 char starts_bundle = ' ';
2401 _regalloc->dump_frame();
2403 Node *n = NULL;
2404 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2405 if (VMThread::should_terminate()) {
2406 cut_short = true;
2407 break;
2408 }
2409 Block* block = _cfg->get_block(i);
2410 if (block->is_connector() && !Verbose) {
2411 continue;
2412 }
2413 n = block->head();
2414 if (pcs && n->_idx < pc_limit) {
2415 tty->print("%3.3x ", pcs[n->_idx]);
2416 } else {
2417 tty->print(" ");
2418 }
2419 block->dump_head(_cfg);
2420 if (block->is_connector()) {
2421 tty->print_cr(" # Empty connector block");
2422 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2423 tty->print_cr(" # Block is sole successor of call");
2424 }
2426 // For all instructions
2427 Node *delay = NULL;
2428 for (uint j = 0; j < block->number_of_nodes(); j++) {
2429 if (VMThread::should_terminate()) {
2430 cut_short = true;
2431 break;
2432 }
2433 n = block->get_node(j);
2434 if (valid_bundle_info(n)) {
2435 Bundle* bundle = node_bundling(n);
2436 if (bundle->used_in_unconditional_delay()) {
2437 delay = n;
2438 continue;
2439 }
2440 if (bundle->starts_bundle()) {
2441 starts_bundle = '+';
2442 }
2443 }
2445 if (WizardMode) {
2446 n->dump();
2447 }
2449 if( !n->is_Region() && // Dont print in the Assembly
2450 !n->is_Phi() && // a few noisely useless nodes
2451 !n->is_Proj() &&
2452 !n->is_MachTemp() &&
2453 !n->is_SafePointScalarObject() &&
2454 !n->is_Catch() && // Would be nice to print exception table targets
2455 !n->is_MergeMem() && // Not very interesting
2456 !n->is_top() && // Debug info table constants
2457 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2458 ) {
2459 if (pcs && n->_idx < pc_limit)
2460 tty->print("%3.3x", pcs[n->_idx]);
2461 else
2462 tty->print(" ");
2463 tty->print(" %c ", starts_bundle);
2464 starts_bundle = ' ';
2465 tty->print("\t");
2466 n->format(_regalloc, tty);
2467 tty->cr();
2468 }
2470 // If we have an instruction with a delay slot, and have seen a delay,
2471 // then back up and print it
2472 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2473 assert(delay != NULL, "no unconditional delay instruction");
2474 if (WizardMode) delay->dump();
2476 if (node_bundling(delay)->starts_bundle())
2477 starts_bundle = '+';
2478 if (pcs && n->_idx < pc_limit)
2479 tty->print("%3.3x", pcs[n->_idx]);
2480 else
2481 tty->print(" ");
2482 tty->print(" %c ", starts_bundle);
2483 starts_bundle = ' ';
2484 tty->print("\t");
2485 delay->format(_regalloc, tty);
2486 tty->cr();
2487 delay = NULL;
2488 }
2490 // Dump the exception table as well
2491 if( n->is_Catch() && (Verbose || WizardMode) ) {
2492 // Print the exception table for this offset
2493 _handler_table.print_subtable_for(pc);
2494 }
2495 }
2497 if (pcs && n->_idx < pc_limit)
2498 tty->print_cr("%3.3x", pcs[n->_idx]);
2499 else
2500 tty->cr();
2502 assert(cut_short || delay == NULL, "no unconditional delay branch");
2504 } // End of per-block dump
2505 tty->cr();
2507 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2508 }
2509 #endif
2511 //------------------------------Final_Reshape_Counts---------------------------
2512 // This class defines counters to help identify when a method
2513 // may/must be executed using hardware with only 24-bit precision.
2514 struct Final_Reshape_Counts : public StackObj {
2515 int _call_count; // count non-inlined 'common' calls
2516 int _float_count; // count float ops requiring 24-bit precision
2517 int _double_count; // count double ops requiring more precision
2518 int _java_call_count; // count non-inlined 'java' calls
2519 int _inner_loop_count; // count loops which need alignment
2520 VectorSet _visited; // Visitation flags
2521 Node_List _tests; // Set of IfNodes & PCTableNodes
2523 Final_Reshape_Counts() :
2524 _call_count(0), _float_count(0), _double_count(0),
2525 _java_call_count(0), _inner_loop_count(0),
2526 _visited( Thread::current()->resource_area() ) { }
2528 void inc_call_count () { _call_count ++; }
2529 void inc_float_count () { _float_count ++; }
2530 void inc_double_count() { _double_count++; }
2531 void inc_java_call_count() { _java_call_count++; }
2532 void inc_inner_loop_count() { _inner_loop_count++; }
2534 int get_call_count () const { return _call_count ; }
2535 int get_float_count () const { return _float_count ; }
2536 int get_double_count() const { return _double_count; }
2537 int get_java_call_count() const { return _java_call_count; }
2538 int get_inner_loop_count() const { return _inner_loop_count; }
2539 };
2541 #ifdef ASSERT
2542 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2543 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2544 // Make sure the offset goes inside the instance layout.
2545 return k->contains_field_offset(tp->offset());
2546 // Note that OffsetBot and OffsetTop are very negative.
2547 }
2548 #endif
2550 // Eliminate trivially redundant StoreCMs and accumulate their
2551 // precedence edges.
2552 void Compile::eliminate_redundant_card_marks(Node* n) {
2553 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2554 if (n->in(MemNode::Address)->outcnt() > 1) {
2555 // There are multiple users of the same address so it might be
2556 // possible to eliminate some of the StoreCMs
2557 Node* mem = n->in(MemNode::Memory);
2558 Node* adr = n->in(MemNode::Address);
2559 Node* val = n->in(MemNode::ValueIn);
2560 Node* prev = n;
2561 bool done = false;
2562 // Walk the chain of StoreCMs eliminating ones that match. As
2563 // long as it's a chain of single users then the optimization is
2564 // safe. Eliminating partially redundant StoreCMs would require
2565 // cloning copies down the other paths.
2566 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2567 if (adr == mem->in(MemNode::Address) &&
2568 val == mem->in(MemNode::ValueIn)) {
2569 // redundant StoreCM
2570 if (mem->req() > MemNode::OopStore) {
2571 // Hasn't been processed by this code yet.
2572 n->add_prec(mem->in(MemNode::OopStore));
2573 } else {
2574 // Already converted to precedence edge
2575 for (uint i = mem->req(); i < mem->len(); i++) {
2576 // Accumulate any precedence edges
2577 if (mem->in(i) != NULL) {
2578 n->add_prec(mem->in(i));
2579 }
2580 }
2581 // Everything above this point has been processed.
2582 done = true;
2583 }
2584 // Eliminate the previous StoreCM
2585 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2586 assert(mem->outcnt() == 0, "should be dead");
2587 mem->disconnect_inputs(NULL, this);
2588 } else {
2589 prev = mem;
2590 }
2591 mem = prev->in(MemNode::Memory);
2592 }
2593 }
2594 }
2596 //------------------------------final_graph_reshaping_impl----------------------
2597 // Implement items 1-5 from final_graph_reshaping below.
2598 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2600 if ( n->outcnt() == 0 ) return; // dead node
2601 uint nop = n->Opcode();
2603 // Check for 2-input instruction with "last use" on right input.
2604 // Swap to left input. Implements item (2).
2605 if( n->req() == 3 && // two-input instruction
2606 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2607 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2608 n->in(2)->outcnt() == 1 &&// right use IS a last use
2609 !n->in(2)->is_Con() ) { // right use is not a constant
2610 // Check for commutative opcode
2611 switch( nop ) {
2612 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2613 case Op_MaxI: case Op_MinI:
2614 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2615 case Op_AndL: case Op_XorL: case Op_OrL:
2616 case Op_AndI: case Op_XorI: case Op_OrI: {
2617 // Move "last use" input to left by swapping inputs
2618 n->swap_edges(1, 2);
2619 break;
2620 }
2621 default:
2622 break;
2623 }
2624 }
2626 #ifdef ASSERT
2627 if( n->is_Mem() ) {
2628 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2629 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2630 // oop will be recorded in oop map if load crosses safepoint
2631 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2632 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2633 "raw memory operations should have control edge");
2634 }
2635 #endif
2636 // Count FPU ops and common calls, implements item (3)
2637 switch( nop ) {
2638 // Count all float operations that may use FPU
2639 case Op_AddF:
2640 case Op_SubF:
2641 case Op_MulF:
2642 case Op_DivF:
2643 case Op_NegF:
2644 case Op_ModF:
2645 case Op_ConvI2F:
2646 case Op_ConF:
2647 case Op_CmpF:
2648 case Op_CmpF3:
2649 // case Op_ConvL2F: // longs are split into 32-bit halves
2650 frc.inc_float_count();
2651 break;
2653 case Op_ConvF2D:
2654 case Op_ConvD2F:
2655 frc.inc_float_count();
2656 frc.inc_double_count();
2657 break;
2659 // Count all double operations that may use FPU
2660 case Op_AddD:
2661 case Op_SubD:
2662 case Op_MulD:
2663 case Op_DivD:
2664 case Op_NegD:
2665 case Op_ModD:
2666 case Op_ConvI2D:
2667 case Op_ConvD2I:
2668 // case Op_ConvL2D: // handled by leaf call
2669 // case Op_ConvD2L: // handled by leaf call
2670 case Op_ConD:
2671 case Op_CmpD:
2672 case Op_CmpD3:
2673 frc.inc_double_count();
2674 break;
2675 case Op_Opaque1: // Remove Opaque Nodes before matching
2676 case Op_Opaque2: // Remove Opaque Nodes before matching
2677 case Op_Opaque3:
2678 n->subsume_by(n->in(1), this);
2679 break;
2680 case Op_CallStaticJava:
2681 case Op_CallJava:
2682 case Op_CallDynamicJava:
2683 frc.inc_java_call_count(); // Count java call site;
2684 case Op_CallRuntime:
2685 case Op_CallLeaf:
2686 case Op_CallLeafNoFP: {
2687 assert( n->is_Call(), "" );
2688 CallNode *call = n->as_Call();
2689 // Count call sites where the FP mode bit would have to be flipped.
2690 // Do not count uncommon runtime calls:
2691 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2692 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2693 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2694 frc.inc_call_count(); // Count the call site
2695 } else { // See if uncommon argument is shared
2696 Node *n = call->in(TypeFunc::Parms);
2697 int nop = n->Opcode();
2698 // Clone shared simple arguments to uncommon calls, item (1).
2699 if( n->outcnt() > 1 &&
2700 !n->is_Proj() &&
2701 nop != Op_CreateEx &&
2702 nop != Op_CheckCastPP &&
2703 nop != Op_DecodeN &&
2704 nop != Op_DecodeNKlass &&
2705 !n->is_Mem() ) {
2706 Node *x = n->clone();
2707 call->set_req( TypeFunc::Parms, x );
2708 }
2709 }
2710 break;
2711 }
2713 case Op_StoreD:
2714 case Op_LoadD:
2715 case Op_LoadD_unaligned:
2716 frc.inc_double_count();
2717 goto handle_mem;
2718 case Op_StoreF:
2719 case Op_LoadF:
2720 frc.inc_float_count();
2721 goto handle_mem;
2723 case Op_StoreCM:
2724 {
2725 // Convert OopStore dependence into precedence edge
2726 Node* prec = n->in(MemNode::OopStore);
2727 n->del_req(MemNode::OopStore);
2728 n->add_prec(prec);
2729 eliminate_redundant_card_marks(n);
2730 }
2732 // fall through
2734 case Op_StoreB:
2735 case Op_StoreC:
2736 case Op_StorePConditional:
2737 case Op_StoreI:
2738 case Op_StoreL:
2739 case Op_StoreIConditional:
2740 case Op_StoreLConditional:
2741 case Op_CompareAndSwapI:
2742 case Op_CompareAndSwapL:
2743 case Op_CompareAndSwapP:
2744 case Op_CompareAndSwapN:
2745 case Op_GetAndAddI:
2746 case Op_GetAndAddL:
2747 case Op_GetAndSetI:
2748 case Op_GetAndSetL:
2749 case Op_GetAndSetP:
2750 case Op_GetAndSetN:
2751 case Op_StoreP:
2752 case Op_StoreN:
2753 case Op_StoreNKlass:
2754 case Op_LoadB:
2755 case Op_LoadUB:
2756 case Op_LoadUS:
2757 case Op_LoadI:
2758 case Op_LoadKlass:
2759 case Op_LoadNKlass:
2760 case Op_LoadL:
2761 case Op_LoadL_unaligned:
2762 case Op_LoadPLocked:
2763 case Op_LoadP:
2764 case Op_LoadN:
2765 case Op_LoadRange:
2766 case Op_LoadS: {
2767 handle_mem:
2768 #ifdef ASSERT
2769 if( VerifyOptoOopOffsets ) {
2770 assert( n->is_Mem(), "" );
2771 MemNode *mem = (MemNode*)n;
2772 // Check to see if address types have grounded out somehow.
2773 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2774 assert( !tp || oop_offset_is_sane(tp), "" );
2775 }
2776 #endif
2777 break;
2778 }
2780 case Op_AddP: { // Assert sane base pointers
2781 Node *addp = n->in(AddPNode::Address);
2782 assert( !addp->is_AddP() ||
2783 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2784 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2785 "Base pointers must match" );
2786 #ifdef _LP64
2787 if ((UseCompressedOops || UseCompressedClassPointers) &&
2788 addp->Opcode() == Op_ConP &&
2789 addp == n->in(AddPNode::Base) &&
2790 n->in(AddPNode::Offset)->is_Con()) {
2791 // Use addressing with narrow klass to load with offset on x86.
2792 // On sparc loading 32-bits constant and decoding it have less
2793 // instructions (4) then load 64-bits constant (7).
2794 // Do this transformation here since IGVN will convert ConN back to ConP.
2795 const Type* t = addp->bottom_type();
2796 if (t->isa_oopptr() || t->isa_klassptr()) {
2797 Node* nn = NULL;
2799 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2801 // Look for existing ConN node of the same exact type.
2802 Node* r = root();
2803 uint cnt = r->outcnt();
2804 for (uint i = 0; i < cnt; i++) {
2805 Node* m = r->raw_out(i);
2806 if (m!= NULL && m->Opcode() == op &&
2807 m->bottom_type()->make_ptr() == t) {
2808 nn = m;
2809 break;
2810 }
2811 }
2812 if (nn != NULL) {
2813 // Decode a narrow oop to match address
2814 // [R12 + narrow_oop_reg<<3 + offset]
2815 if (t->isa_oopptr()) {
2816 nn = new (this) DecodeNNode(nn, t);
2817 } else {
2818 nn = new (this) DecodeNKlassNode(nn, t);
2819 }
2820 n->set_req(AddPNode::Base, nn);
2821 n->set_req(AddPNode::Address, nn);
2822 if (addp->outcnt() == 0) {
2823 addp->disconnect_inputs(NULL, this);
2824 }
2825 }
2826 }
2827 }
2828 #endif
2829 break;
2830 }
2832 #ifdef _LP64
2833 case Op_CastPP:
2834 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2835 Node* in1 = n->in(1);
2836 const Type* t = n->bottom_type();
2837 Node* new_in1 = in1->clone();
2838 new_in1->as_DecodeN()->set_type(t);
2840 if (!Matcher::narrow_oop_use_complex_address()) {
2841 //
2842 // x86, ARM and friends can handle 2 adds in addressing mode
2843 // and Matcher can fold a DecodeN node into address by using
2844 // a narrow oop directly and do implicit NULL check in address:
2845 //
2846 // [R12 + narrow_oop_reg<<3 + offset]
2847 // NullCheck narrow_oop_reg
2848 //
2849 // On other platforms (Sparc) we have to keep new DecodeN node and
2850 // use it to do implicit NULL check in address:
2851 //
2852 // decode_not_null narrow_oop_reg, base_reg
2853 // [base_reg + offset]
2854 // NullCheck base_reg
2855 //
2856 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2857 // to keep the information to which NULL check the new DecodeN node
2858 // corresponds to use it as value in implicit_null_check().
2859 //
2860 new_in1->set_req(0, n->in(0));
2861 }
2863 n->subsume_by(new_in1, this);
2864 if (in1->outcnt() == 0) {
2865 in1->disconnect_inputs(NULL, this);
2866 }
2867 }
2868 break;
2870 case Op_CmpP:
2871 // Do this transformation here to preserve CmpPNode::sub() and
2872 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2873 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2874 Node* in1 = n->in(1);
2875 Node* in2 = n->in(2);
2876 if (!in1->is_DecodeNarrowPtr()) {
2877 in2 = in1;
2878 in1 = n->in(2);
2879 }
2880 assert(in1->is_DecodeNarrowPtr(), "sanity");
2882 Node* new_in2 = NULL;
2883 if (in2->is_DecodeNarrowPtr()) {
2884 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2885 new_in2 = in2->in(1);
2886 } else if (in2->Opcode() == Op_ConP) {
2887 const Type* t = in2->bottom_type();
2888 if (t == TypePtr::NULL_PTR) {
2889 assert(in1->is_DecodeN(), "compare klass to null?");
2890 // Don't convert CmpP null check into CmpN if compressed
2891 // oops implicit null check is not generated.
2892 // This will allow to generate normal oop implicit null check.
2893 if (Matcher::gen_narrow_oop_implicit_null_checks())
2894 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2895 //
2896 // This transformation together with CastPP transformation above
2897 // will generated code for implicit NULL checks for compressed oops.
2898 //
2899 // The original code after Optimize()
2900 //
2901 // LoadN memory, narrow_oop_reg
2902 // decode narrow_oop_reg, base_reg
2903 // CmpP base_reg, NULL
2904 // CastPP base_reg // NotNull
2905 // Load [base_reg + offset], val_reg
2906 //
2907 // after these transformations will be
2908 //
2909 // LoadN memory, narrow_oop_reg
2910 // CmpN narrow_oop_reg, NULL
2911 // decode_not_null narrow_oop_reg, base_reg
2912 // Load [base_reg + offset], val_reg
2913 //
2914 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2915 // since narrow oops can be used in debug info now (see the code in
2916 // final_graph_reshaping_walk()).
2917 //
2918 // At the end the code will be matched to
2919 // on x86:
2920 //
2921 // Load_narrow_oop memory, narrow_oop_reg
2922 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2923 // NullCheck narrow_oop_reg
2924 //
2925 // and on sparc:
2926 //
2927 // Load_narrow_oop memory, narrow_oop_reg
2928 // decode_not_null narrow_oop_reg, base_reg
2929 // Load [base_reg + offset], val_reg
2930 // NullCheck base_reg
2931 //
2932 } else if (t->isa_oopptr()) {
2933 new_in2 = ConNode::make(this, t->make_narrowoop());
2934 } else if (t->isa_klassptr()) {
2935 new_in2 = ConNode::make(this, t->make_narrowklass());
2936 }
2937 }
2938 if (new_in2 != NULL) {
2939 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2940 n->subsume_by(cmpN, this);
2941 if (in1->outcnt() == 0) {
2942 in1->disconnect_inputs(NULL, this);
2943 }
2944 if (in2->outcnt() == 0) {
2945 in2->disconnect_inputs(NULL, this);
2946 }
2947 }
2948 }
2949 break;
2951 case Op_DecodeN:
2952 case Op_DecodeNKlass:
2953 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2954 // DecodeN could be pinned when it can't be fold into
2955 // an address expression, see the code for Op_CastPP above.
2956 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2957 break;
2959 case Op_EncodeP:
2960 case Op_EncodePKlass: {
2961 Node* in1 = n->in(1);
2962 if (in1->is_DecodeNarrowPtr()) {
2963 n->subsume_by(in1->in(1), this);
2964 } else if (in1->Opcode() == Op_ConP) {
2965 const Type* t = in1->bottom_type();
2966 if (t == TypePtr::NULL_PTR) {
2967 assert(t->isa_oopptr(), "null klass?");
2968 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2969 } else if (t->isa_oopptr()) {
2970 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2971 } else if (t->isa_klassptr()) {
2972 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2973 }
2974 }
2975 if (in1->outcnt() == 0) {
2976 in1->disconnect_inputs(NULL, this);
2977 }
2978 break;
2979 }
2981 case Op_Proj: {
2982 if (OptimizeStringConcat) {
2983 ProjNode* p = n->as_Proj();
2984 if (p->_is_io_use) {
2985 // Separate projections were used for the exception path which
2986 // are normally removed by a late inline. If it wasn't inlined
2987 // then they will hang around and should just be replaced with
2988 // the original one.
2989 Node* proj = NULL;
2990 // Replace with just one
2991 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2992 Node *use = i.get();
2993 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2994 proj = use;
2995 break;
2996 }
2997 }
2998 assert(proj != NULL, "must be found");
2999 p->subsume_by(proj, this);
3000 }
3001 }
3002 break;
3003 }
3005 case Op_Phi:
3006 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3007 // The EncodeP optimization may create Phi with the same edges
3008 // for all paths. It is not handled well by Register Allocator.
3009 Node* unique_in = n->in(1);
3010 assert(unique_in != NULL, "");
3011 uint cnt = n->req();
3012 for (uint i = 2; i < cnt; i++) {
3013 Node* m = n->in(i);
3014 assert(m != NULL, "");
3015 if (unique_in != m)
3016 unique_in = NULL;
3017 }
3018 if (unique_in != NULL) {
3019 n->subsume_by(unique_in, this);
3020 }
3021 }
3022 break;
3024 #endif
3026 #ifdef ASSERT
3027 case Op_CastII:
3028 // Verify that all range check dependent CastII nodes were removed.
3029 if (n->isa_CastII()->has_range_check()) {
3030 n->dump(3);
3031 assert(false, "Range check dependent CastII node was not removed");
3032 }
3033 break;
3034 #endif
3036 case Op_ModI:
3037 if (UseDivMod) {
3038 // Check if a%b and a/b both exist
3039 Node* d = n->find_similar(Op_DivI);
3040 if (d) {
3041 // Replace them with a fused divmod if supported
3042 if (Matcher::has_match_rule(Op_DivModI)) {
3043 DivModINode* divmod = DivModINode::make(this, n);
3044 d->subsume_by(divmod->div_proj(), this);
3045 n->subsume_by(divmod->mod_proj(), this);
3046 } else {
3047 // replace a%b with a-((a/b)*b)
3048 Node* mult = new (this) MulINode(d, d->in(2));
3049 Node* sub = new (this) SubINode(d->in(1), mult);
3050 n->subsume_by(sub, this);
3051 }
3052 }
3053 }
3054 break;
3056 case Op_ModL:
3057 if (UseDivMod) {
3058 // Check if a%b and a/b both exist
3059 Node* d = n->find_similar(Op_DivL);
3060 if (d) {
3061 // Replace them with a fused divmod if supported
3062 if (Matcher::has_match_rule(Op_DivModL)) {
3063 DivModLNode* divmod = DivModLNode::make(this, n);
3064 d->subsume_by(divmod->div_proj(), this);
3065 n->subsume_by(divmod->mod_proj(), this);
3066 } else {
3067 // replace a%b with a-((a/b)*b)
3068 Node* mult = new (this) MulLNode(d, d->in(2));
3069 Node* sub = new (this) SubLNode(d->in(1), mult);
3070 n->subsume_by(sub, this);
3071 }
3072 }
3073 }
3074 break;
3076 case Op_LoadVector:
3077 case Op_StoreVector:
3078 break;
3080 case Op_PackB:
3081 case Op_PackS:
3082 case Op_PackI:
3083 case Op_PackF:
3084 case Op_PackL:
3085 case Op_PackD:
3086 if (n->req()-1 > 2) {
3087 // Replace many operand PackNodes with a binary tree for matching
3088 PackNode* p = (PackNode*) n;
3089 Node* btp = p->binary_tree_pack(this, 1, n->req());
3090 n->subsume_by(btp, this);
3091 }
3092 break;
3093 case Op_Loop:
3094 case Op_CountedLoop:
3095 if (n->as_Loop()->is_inner_loop()) {
3096 frc.inc_inner_loop_count();
3097 }
3098 break;
3099 case Op_LShiftI:
3100 case Op_RShiftI:
3101 case Op_URShiftI:
3102 case Op_LShiftL:
3103 case Op_RShiftL:
3104 case Op_URShiftL:
3105 if (Matcher::need_masked_shift_count) {
3106 // The cpu's shift instructions don't restrict the count to the
3107 // lower 5/6 bits. We need to do the masking ourselves.
3108 Node* in2 = n->in(2);
3109 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3110 const TypeInt* t = in2->find_int_type();
3111 if (t != NULL && t->is_con()) {
3112 juint shift = t->get_con();
3113 if (shift > mask) { // Unsigned cmp
3114 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
3115 }
3116 } else {
3117 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3118 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3119 n->set_req(2, shift);
3120 }
3121 }
3122 if (in2->outcnt() == 0) { // Remove dead node
3123 in2->disconnect_inputs(NULL, this);
3124 }
3125 }
3126 break;
3127 case Op_MemBarStoreStore:
3128 case Op_MemBarRelease:
3129 // Break the link with AllocateNode: it is no longer useful and
3130 // confuses register allocation.
3131 if (n->req() > MemBarNode::Precedent) {
3132 n->set_req(MemBarNode::Precedent, top());
3133 }
3134 break;
3135 default:
3136 assert( !n->is_Call(), "" );
3137 assert( !n->is_Mem(), "" );
3138 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3139 break;
3140 }
3142 // Collect CFG split points
3143 if (n->is_MultiBranch())
3144 frc._tests.push(n);
3145 }
3147 //------------------------------final_graph_reshaping_walk---------------------
3148 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3149 // requires that the walk visits a node's inputs before visiting the node.
3150 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3151 ResourceArea *area = Thread::current()->resource_area();
3152 Unique_Node_List sfpt(area);
3154 frc._visited.set(root->_idx); // first, mark node as visited
3155 uint cnt = root->req();
3156 Node *n = root;
3157 uint i = 0;
3158 while (true) {
3159 if (i < cnt) {
3160 // Place all non-visited non-null inputs onto stack
3161 Node* m = n->in(i);
3162 ++i;
3163 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3164 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3165 // compute worst case interpreter size in case of a deoptimization
3166 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3168 sfpt.push(m);
3169 }
3170 cnt = m->req();
3171 nstack.push(n, i); // put on stack parent and next input's index
3172 n = m;
3173 i = 0;
3174 }
3175 } else {
3176 // Now do post-visit work
3177 final_graph_reshaping_impl( n, frc );
3178 if (nstack.is_empty())
3179 break; // finished
3180 n = nstack.node(); // Get node from stack
3181 cnt = n->req();
3182 i = nstack.index();
3183 nstack.pop(); // Shift to the next node on stack
3184 }
3185 }
3187 // Skip next transformation if compressed oops are not used.
3188 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3189 (!UseCompressedOops && !UseCompressedClassPointers))
3190 return;
3192 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3193 // It could be done for an uncommon traps or any safepoints/calls
3194 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3195 while (sfpt.size() > 0) {
3196 n = sfpt.pop();
3197 JVMState *jvms = n->as_SafePoint()->jvms();
3198 assert(jvms != NULL, "sanity");
3199 int start = jvms->debug_start();
3200 int end = n->req();
3201 bool is_uncommon = (n->is_CallStaticJava() &&
3202 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3203 for (int j = start; j < end; j++) {
3204 Node* in = n->in(j);
3205 if (in->is_DecodeNarrowPtr()) {
3206 bool safe_to_skip = true;
3207 if (!is_uncommon ) {
3208 // Is it safe to skip?
3209 for (uint i = 0; i < in->outcnt(); i++) {
3210 Node* u = in->raw_out(i);
3211 if (!u->is_SafePoint() ||
3212 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3213 safe_to_skip = false;
3214 }
3215 }
3216 }
3217 if (safe_to_skip) {
3218 n->set_req(j, in->in(1));
3219 }
3220 if (in->outcnt() == 0) {
3221 in->disconnect_inputs(NULL, this);
3222 }
3223 }
3224 }
3225 }
3226 }
3228 //------------------------------final_graph_reshaping--------------------------
3229 // Final Graph Reshaping.
3230 //
3231 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3232 // and not commoned up and forced early. Must come after regular
3233 // optimizations to avoid GVN undoing the cloning. Clone constant
3234 // inputs to Loop Phis; these will be split by the allocator anyways.
3235 // Remove Opaque nodes.
3236 // (2) Move last-uses by commutative operations to the left input to encourage
3237 // Intel update-in-place two-address operations and better register usage
3238 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3239 // calls canonicalizing them back.
3240 // (3) Count the number of double-precision FP ops, single-precision FP ops
3241 // and call sites. On Intel, we can get correct rounding either by
3242 // forcing singles to memory (requires extra stores and loads after each
3243 // FP bytecode) or we can set a rounding mode bit (requires setting and
3244 // clearing the mode bit around call sites). The mode bit is only used
3245 // if the relative frequency of single FP ops to calls is low enough.
3246 // This is a key transform for SPEC mpeg_audio.
3247 // (4) Detect infinite loops; blobs of code reachable from above but not
3248 // below. Several of the Code_Gen algorithms fail on such code shapes,
3249 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3250 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3251 // Detection is by looking for IfNodes where only 1 projection is
3252 // reachable from below or CatchNodes missing some targets.
3253 // (5) Assert for insane oop offsets in debug mode.
3255 bool Compile::final_graph_reshaping() {
3256 // an infinite loop may have been eliminated by the optimizer,
3257 // in which case the graph will be empty.
3258 if (root()->req() == 1) {
3259 record_method_not_compilable("trivial infinite loop");
3260 return true;
3261 }
3263 // Expensive nodes have their control input set to prevent the GVN
3264 // from freely commoning them. There's no GVN beyond this point so
3265 // no need to keep the control input. We want the expensive nodes to
3266 // be freely moved to the least frequent code path by gcm.
3267 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3268 for (int i = 0; i < expensive_count(); i++) {
3269 _expensive_nodes->at(i)->set_req(0, NULL);
3270 }
3272 Final_Reshape_Counts frc;
3274 // Visit everybody reachable!
3275 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3276 Node_Stack nstack(live_nodes() >> 1);
3277 final_graph_reshaping_walk(nstack, root(), frc);
3279 // Check for unreachable (from below) code (i.e., infinite loops).
3280 for( uint i = 0; i < frc._tests.size(); i++ ) {
3281 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3282 // Get number of CFG targets.
3283 // Note that PCTables include exception targets after calls.
3284 uint required_outcnt = n->required_outcnt();
3285 if (n->outcnt() != required_outcnt) {
3286 // Check for a few special cases. Rethrow Nodes never take the
3287 // 'fall-thru' path, so expected kids is 1 less.
3288 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3289 if (n->in(0)->in(0)->is_Call()) {
3290 CallNode *call = n->in(0)->in(0)->as_Call();
3291 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3292 required_outcnt--; // Rethrow always has 1 less kid
3293 } else if (call->req() > TypeFunc::Parms &&
3294 call->is_CallDynamicJava()) {
3295 // Check for null receiver. In such case, the optimizer has
3296 // detected that the virtual call will always result in a null
3297 // pointer exception. The fall-through projection of this CatchNode
3298 // will not be populated.
3299 Node *arg0 = call->in(TypeFunc::Parms);
3300 if (arg0->is_Type() &&
3301 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3302 required_outcnt--;
3303 }
3304 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3305 call->req() > TypeFunc::Parms+1 &&
3306 call->is_CallStaticJava()) {
3307 // Check for negative array length. In such case, the optimizer has
3308 // detected that the allocation attempt will always result in an
3309 // exception. There is no fall-through projection of this CatchNode .
3310 Node *arg1 = call->in(TypeFunc::Parms+1);
3311 if (arg1->is_Type() &&
3312 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3313 required_outcnt--;
3314 }
3315 }
3316 }
3317 }
3318 // Recheck with a better notion of 'required_outcnt'
3319 if (n->outcnt() != required_outcnt) {
3320 record_method_not_compilable("malformed control flow");
3321 return true; // Not all targets reachable!
3322 }
3323 }
3324 // Check that I actually visited all kids. Unreached kids
3325 // must be infinite loops.
3326 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3327 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3328 record_method_not_compilable("infinite loop");
3329 return true; // Found unvisited kid; must be unreach
3330 }
3331 }
3333 // If original bytecodes contained a mixture of floats and doubles
3334 // check if the optimizer has made it homogenous, item (3).
3335 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3336 frc.get_float_count() > 32 &&
3337 frc.get_double_count() == 0 &&
3338 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3339 set_24_bit_selection_and_mode( false, true );
3340 }
3342 set_java_calls(frc.get_java_call_count());
3343 set_inner_loops(frc.get_inner_loop_count());
3345 // No infinite loops, no reason to bail out.
3346 return false;
3347 }
3349 //-----------------------------too_many_traps----------------------------------
3350 // Report if there are too many traps at the current method and bci.
3351 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3352 bool Compile::too_many_traps(ciMethod* method,
3353 int bci,
3354 Deoptimization::DeoptReason reason) {
3355 ciMethodData* md = method->method_data();
3356 if (md->is_empty()) {
3357 // Assume the trap has not occurred, or that it occurred only
3358 // because of a transient condition during start-up in the interpreter.
3359 return false;
3360 }
3361 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3362 if (md->has_trap_at(bci, m, reason) != 0) {
3363 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3364 // Also, if there are multiple reasons, or if there is no per-BCI record,
3365 // assume the worst.
3366 if (log())
3367 log()->elem("observe trap='%s' count='%d'",
3368 Deoptimization::trap_reason_name(reason),
3369 md->trap_count(reason));
3370 return true;
3371 } else {
3372 // Ignore method/bci and see if there have been too many globally.
3373 return too_many_traps(reason, md);
3374 }
3375 }
3377 // Less-accurate variant which does not require a method and bci.
3378 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3379 ciMethodData* logmd) {
3380 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3381 // Too many traps globally.
3382 // Note that we use cumulative trap_count, not just md->trap_count.
3383 if (log()) {
3384 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3385 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3386 Deoptimization::trap_reason_name(reason),
3387 mcount, trap_count(reason));
3388 }
3389 return true;
3390 } else {
3391 // The coast is clear.
3392 return false;
3393 }
3394 }
3396 //--------------------------too_many_recompiles--------------------------------
3397 // Report if there are too many recompiles at the current method and bci.
3398 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3399 // Is not eager to return true, since this will cause the compiler to use
3400 // Action_none for a trap point, to avoid too many recompilations.
3401 bool Compile::too_many_recompiles(ciMethod* method,
3402 int bci,
3403 Deoptimization::DeoptReason reason) {
3404 ciMethodData* md = method->method_data();
3405 if (md->is_empty()) {
3406 // Assume the trap has not occurred, or that it occurred only
3407 // because of a transient condition during start-up in the interpreter.
3408 return false;
3409 }
3410 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3411 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3412 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3413 Deoptimization::DeoptReason per_bc_reason
3414 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3415 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3416 if ((per_bc_reason == Deoptimization::Reason_none
3417 || md->has_trap_at(bci, m, reason) != 0)
3418 // The trap frequency measure we care about is the recompile count:
3419 && md->trap_recompiled_at(bci, m)
3420 && md->overflow_recompile_count() >= bc_cutoff) {
3421 // Do not emit a trap here if it has already caused recompilations.
3422 // Also, if there are multiple reasons, or if there is no per-BCI record,
3423 // assume the worst.
3424 if (log())
3425 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3426 Deoptimization::trap_reason_name(reason),
3427 md->trap_count(reason),
3428 md->overflow_recompile_count());
3429 return true;
3430 } else if (trap_count(reason) != 0
3431 && decompile_count() >= m_cutoff) {
3432 // Too many recompiles globally, and we have seen this sort of trap.
3433 // Use cumulative decompile_count, not just md->decompile_count.
3434 if (log())
3435 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3436 Deoptimization::trap_reason_name(reason),
3437 md->trap_count(reason), trap_count(reason),
3438 md->decompile_count(), decompile_count());
3439 return true;
3440 } else {
3441 // The coast is clear.
3442 return false;
3443 }
3444 }
3446 // Compute when not to trap. Used by matching trap based nodes and
3447 // NullCheck optimization.
3448 void Compile::set_allowed_deopt_reasons() {
3449 _allowed_reasons = 0;
3450 if (is_method_compilation()) {
3451 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3452 assert(rs < BitsPerInt, "recode bit map");
3453 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3454 _allowed_reasons |= nth_bit(rs);
3455 }
3456 }
3457 }
3458 }
3460 #ifndef PRODUCT
3461 //------------------------------verify_graph_edges---------------------------
3462 // Walk the Graph and verify that there is a one-to-one correspondence
3463 // between Use-Def edges and Def-Use edges in the graph.
3464 void Compile::verify_graph_edges(bool no_dead_code) {
3465 if (VerifyGraphEdges) {
3466 ResourceArea *area = Thread::current()->resource_area();
3467 Unique_Node_List visited(area);
3468 // Call recursive graph walk to check edges
3469 _root->verify_edges(visited);
3470 if (no_dead_code) {
3471 // Now make sure that no visited node is used by an unvisited node.
3472 bool dead_nodes = 0;
3473 Unique_Node_List checked(area);
3474 while (visited.size() > 0) {
3475 Node* n = visited.pop();
3476 checked.push(n);
3477 for (uint i = 0; i < n->outcnt(); i++) {
3478 Node* use = n->raw_out(i);
3479 if (checked.member(use)) continue; // already checked
3480 if (visited.member(use)) continue; // already in the graph
3481 if (use->is_Con()) continue; // a dead ConNode is OK
3482 // At this point, we have found a dead node which is DU-reachable.
3483 if (dead_nodes++ == 0)
3484 tty->print_cr("*** Dead nodes reachable via DU edges:");
3485 use->dump(2);
3486 tty->print_cr("---");
3487 checked.push(use); // No repeats; pretend it is now checked.
3488 }
3489 }
3490 assert(dead_nodes == 0, "using nodes must be reachable from root");
3491 }
3492 }
3493 }
3495 // Verify GC barriers consistency
3496 // Currently supported:
3497 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3498 void Compile::verify_barriers() {
3499 if (UseG1GC) {
3500 // Verify G1 pre-barriers
3501 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3503 ResourceArea *area = Thread::current()->resource_area();
3504 Unique_Node_List visited(area);
3505 Node_List worklist(area);
3506 // We're going to walk control flow backwards starting from the Root
3507 worklist.push(_root);
3508 while (worklist.size() > 0) {
3509 Node* x = worklist.pop();
3510 if (x == NULL || x == top()) continue;
3511 if (visited.member(x)) {
3512 continue;
3513 } else {
3514 visited.push(x);
3515 }
3517 if (x->is_Region()) {
3518 for (uint i = 1; i < x->req(); i++) {
3519 worklist.push(x->in(i));
3520 }
3521 } else {
3522 worklist.push(x->in(0));
3523 // We are looking for the pattern:
3524 // /->ThreadLocal
3525 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3526 // \->ConI(0)
3527 // We want to verify that the If and the LoadB have the same control
3528 // See GraphKit::g1_write_barrier_pre()
3529 if (x->is_If()) {
3530 IfNode *iff = x->as_If();
3531 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3532 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3533 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3534 && cmp->in(1)->is_Load()) {
3535 LoadNode* load = cmp->in(1)->as_Load();
3536 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3537 && load->in(2)->in(3)->is_Con()
3538 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3540 Node* if_ctrl = iff->in(0);
3541 Node* load_ctrl = load->in(0);
3543 if (if_ctrl != load_ctrl) {
3544 // Skip possible CProj->NeverBranch in infinite loops
3545 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3546 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3547 if_ctrl = if_ctrl->in(0)->in(0);
3548 }
3549 }
3550 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3551 }
3552 }
3553 }
3554 }
3555 }
3556 }
3557 }
3558 }
3560 #endif
3562 // The Compile object keeps track of failure reasons separately from the ciEnv.
3563 // This is required because there is not quite a 1-1 relation between the
3564 // ciEnv and its compilation task and the Compile object. Note that one
3565 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3566 // to backtrack and retry without subsuming loads. Other than this backtracking
3567 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3568 // by the logic in C2Compiler.
3569 void Compile::record_failure(const char* reason) {
3570 if (log() != NULL) {
3571 log()->elem("failure reason='%s' phase='compile'", reason);
3572 }
3573 if (_failure_reason == NULL) {
3574 // Record the first failure reason.
3575 _failure_reason = reason;
3576 }
3578 EventCompilerFailure event;
3579 if (event.should_commit()) {
3580 event.set_compileID(Compile::compile_id());
3581 event.set_failure(reason);
3582 event.commit();
3583 }
3585 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3586 C->print_method(PHASE_FAILURE);
3587 }
3588 _root = NULL; // flush the graph, too
3589 }
3591 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3592 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3593 _phase_name(name), _dolog(dolog)
3594 {
3595 if (dolog) {
3596 C = Compile::current();
3597 _log = C->log();
3598 } else {
3599 C = NULL;
3600 _log = NULL;
3601 }
3602 if (_log != NULL) {
3603 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3604 _log->stamp();
3605 _log->end_head();
3606 }
3607 }
3609 Compile::TracePhase::~TracePhase() {
3611 C = Compile::current();
3612 if (_dolog) {
3613 _log = C->log();
3614 } else {
3615 _log = NULL;
3616 }
3618 #ifdef ASSERT
3619 if (PrintIdealNodeCount) {
3620 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3621 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3622 }
3624 if (VerifyIdealNodeCount) {
3625 Compile::current()->print_missing_nodes();
3626 }
3627 #endif
3629 if (_log != NULL) {
3630 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3631 }
3632 }
3634 //=============================================================================
3635 // Two Constant's are equal when the type and the value are equal.
3636 bool Compile::Constant::operator==(const Constant& other) {
3637 if (type() != other.type() ) return false;
3638 if (can_be_reused() != other.can_be_reused()) return false;
3639 // For floating point values we compare the bit pattern.
3640 switch (type()) {
3641 case T_FLOAT: return (_v._value.i == other._v._value.i);
3642 case T_LONG:
3643 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3644 case T_OBJECT:
3645 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3646 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3647 case T_METADATA: return (_v._metadata == other._v._metadata);
3648 default: ShouldNotReachHere();
3649 }
3650 return false;
3651 }
3653 static int type_to_size_in_bytes(BasicType t) {
3654 switch (t) {
3655 case T_LONG: return sizeof(jlong );
3656 case T_FLOAT: return sizeof(jfloat );
3657 case T_DOUBLE: return sizeof(jdouble);
3658 case T_METADATA: return sizeof(Metadata*);
3659 // We use T_VOID as marker for jump-table entries (labels) which
3660 // need an internal word relocation.
3661 case T_VOID:
3662 case T_ADDRESS:
3663 case T_OBJECT: return sizeof(jobject);
3664 }
3666 ShouldNotReachHere();
3667 return -1;
3668 }
3670 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3671 // sort descending
3672 if (a->freq() > b->freq()) return -1;
3673 if (a->freq() < b->freq()) return 1;
3674 return 0;
3675 }
3677 void Compile::ConstantTable::calculate_offsets_and_size() {
3678 // First, sort the array by frequencies.
3679 _constants.sort(qsort_comparator);
3681 #ifdef ASSERT
3682 // Make sure all jump-table entries were sorted to the end of the
3683 // array (they have a negative frequency).
3684 bool found_void = false;
3685 for (int i = 0; i < _constants.length(); i++) {
3686 Constant con = _constants.at(i);
3687 if (con.type() == T_VOID)
3688 found_void = true; // jump-tables
3689 else
3690 assert(!found_void, "wrong sorting");
3691 }
3692 #endif
3694 int offset = 0;
3695 for (int i = 0; i < _constants.length(); i++) {
3696 Constant* con = _constants.adr_at(i);
3698 // Align offset for type.
3699 int typesize = type_to_size_in_bytes(con->type());
3700 offset = align_size_up(offset, typesize);
3701 con->set_offset(offset); // set constant's offset
3703 if (con->type() == T_VOID) {
3704 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3705 offset = offset + typesize * n->outcnt(); // expand jump-table
3706 } else {
3707 offset = offset + typesize;
3708 }
3709 }
3711 // Align size up to the next section start (which is insts; see
3712 // CodeBuffer::align_at_start).
3713 assert(_size == -1, "already set?");
3714 _size = align_size_up(offset, CodeEntryAlignment);
3715 }
3717 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3718 MacroAssembler _masm(&cb);
3719 for (int i = 0; i < _constants.length(); i++) {
3720 Constant con = _constants.at(i);
3721 address constant_addr = NULL;
3722 switch (con.type()) {
3723 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3724 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3725 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3726 case T_OBJECT: {
3727 jobject obj = con.get_jobject();
3728 int oop_index = _masm.oop_recorder()->find_index(obj);
3729 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3730 break;
3731 }
3732 case T_ADDRESS: {
3733 address addr = (address) con.get_jobject();
3734 constant_addr = _masm.address_constant(addr);
3735 break;
3736 }
3737 // We use T_VOID as marker for jump-table entries (labels) which
3738 // need an internal word relocation.
3739 case T_VOID: {
3740 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3741 // Fill the jump-table with a dummy word. The real value is
3742 // filled in later in fill_jump_table.
3743 address dummy = (address) n;
3744 constant_addr = _masm.address_constant(dummy);
3745 // Expand jump-table
3746 for (uint i = 1; i < n->outcnt(); i++) {
3747 address temp_addr = _masm.address_constant(dummy + i);
3748 assert(temp_addr, "consts section too small");
3749 }
3750 break;
3751 }
3752 case T_METADATA: {
3753 Metadata* obj = con.get_metadata();
3754 int metadata_index = _masm.oop_recorder()->find_index(obj);
3755 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3756 break;
3757 }
3758 default: ShouldNotReachHere();
3759 }
3760 assert(constant_addr, "consts section too small");
3761 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3762 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())));
3763 }
3764 }
3766 int Compile::ConstantTable::find_offset(Constant& con) const {
3767 int idx = _constants.find(con);
3768 assert(idx != -1, "constant must be in constant table");
3769 int offset = _constants.at(idx).offset();
3770 assert(offset != -1, "constant table not emitted yet?");
3771 return offset;
3772 }
3774 void Compile::ConstantTable::add(Constant& con) {
3775 if (con.can_be_reused()) {
3776 int idx = _constants.find(con);
3777 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3778 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3779 return;
3780 }
3781 }
3782 (void) _constants.append(con);
3783 }
3785 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3786 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3787 Constant con(type, value, b->_freq);
3788 add(con);
3789 return con;
3790 }
3792 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3793 Constant con(metadata);
3794 add(con);
3795 return con;
3796 }
3798 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3799 jvalue value;
3800 BasicType type = oper->type()->basic_type();
3801 switch (type) {
3802 case T_LONG: value.j = oper->constantL(); break;
3803 case T_FLOAT: value.f = oper->constantF(); break;
3804 case T_DOUBLE: value.d = oper->constantD(); break;
3805 case T_OBJECT:
3806 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3807 case T_METADATA: return add((Metadata*)oper->constant()); break;
3808 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3809 }
3810 return add(n, type, value);
3811 }
3813 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3814 jvalue value;
3815 // We can use the node pointer here to identify the right jump-table
3816 // as this method is called from Compile::Fill_buffer right before
3817 // the MachNodes are emitted and the jump-table is filled (means the
3818 // MachNode pointers do not change anymore).
3819 value.l = (jobject) n;
3820 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3821 add(con);
3822 return con;
3823 }
3825 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3826 // If called from Compile::scratch_emit_size do nothing.
3827 if (Compile::current()->in_scratch_emit_size()) return;
3829 assert(labels.is_nonempty(), "must be");
3830 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3832 // Since MachConstantNode::constant_offset() also contains
3833 // table_base_offset() we need to subtract the table_base_offset()
3834 // to get the plain offset into the constant table.
3835 int offset = n->constant_offset() - table_base_offset();
3837 MacroAssembler _masm(&cb);
3838 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3840 for (uint i = 0; i < n->outcnt(); i++) {
3841 address* constant_addr = &jump_table_base[i];
3842 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)));
3843 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3844 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3845 }
3846 }
3848 void Compile::dump_inlining() {
3849 if (print_inlining() || print_intrinsics()) {
3850 // Print inlining message for candidates that we couldn't inline
3851 // for lack of space or non constant receiver
3852 for (int i = 0; i < _late_inlines.length(); i++) {
3853 CallGenerator* cg = _late_inlines.at(i);
3854 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3855 }
3856 Unique_Node_List useful;
3857 useful.push(root());
3858 for (uint next = 0; next < useful.size(); ++next) {
3859 Node* n = useful.at(next);
3860 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3861 CallNode* call = n->as_Call();
3862 CallGenerator* cg = call->generator();
3863 cg->print_inlining_late("receiver not constant");
3864 }
3865 uint max = n->len();
3866 for ( uint i = 0; i < max; ++i ) {
3867 Node *m = n->in(i);
3868 if ( m == NULL ) continue;
3869 useful.push(m);
3870 }
3871 }
3872 for (int i = 0; i < _print_inlining_list->length(); i++) {
3873 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
3874 }
3875 }
3876 }
3878 // Dump inlining replay data to the stream.
3879 // Don't change thread state and acquire any locks.
3880 void Compile::dump_inline_data(outputStream* out) {
3881 InlineTree* inl_tree = ilt();
3882 if (inl_tree != NULL) {
3883 out->print(" inline %d", inl_tree->count());
3884 inl_tree->dump_replay_data(out);
3885 }
3886 }
3888 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3889 if (n1->Opcode() < n2->Opcode()) return -1;
3890 else if (n1->Opcode() > n2->Opcode()) return 1;
3892 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()));
3893 for (uint i = 1; i < n1->req(); i++) {
3894 if (n1->in(i) < n2->in(i)) return -1;
3895 else if (n1->in(i) > n2->in(i)) return 1;
3896 }
3898 return 0;
3899 }
3901 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3902 Node* n1 = *n1p;
3903 Node* n2 = *n2p;
3905 return cmp_expensive_nodes(n1, n2);
3906 }
3908 void Compile::sort_expensive_nodes() {
3909 if (!expensive_nodes_sorted()) {
3910 _expensive_nodes->sort(cmp_expensive_nodes);
3911 }
3912 }
3914 bool Compile::expensive_nodes_sorted() const {
3915 for (int i = 1; i < _expensive_nodes->length(); i++) {
3916 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3917 return false;
3918 }
3919 }
3920 return true;
3921 }
3923 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3924 if (_expensive_nodes->length() == 0) {
3925 return false;
3926 }
3928 assert(OptimizeExpensiveOps, "optimization off?");
3930 // Take this opportunity to remove dead nodes from the list
3931 int j = 0;
3932 for (int i = 0; i < _expensive_nodes->length(); i++) {
3933 Node* n = _expensive_nodes->at(i);
3934 if (!n->is_unreachable(igvn)) {
3935 assert(n->is_expensive(), "should be expensive");
3936 _expensive_nodes->at_put(j, n);
3937 j++;
3938 }
3939 }
3940 _expensive_nodes->trunc_to(j);
3942 // Then sort the list so that similar nodes are next to each other
3943 // and check for at least two nodes of identical kind with same data
3944 // inputs.
3945 sort_expensive_nodes();
3947 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3948 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3949 return true;
3950 }
3951 }
3953 return false;
3954 }
3956 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3957 if (_expensive_nodes->length() == 0) {
3958 return;
3959 }
3961 assert(OptimizeExpensiveOps, "optimization off?");
3963 // Sort to bring similar nodes next to each other and clear the
3964 // control input of nodes for which there's only a single copy.
3965 sort_expensive_nodes();
3967 int j = 0;
3968 int identical = 0;
3969 int i = 0;
3970 for (; i < _expensive_nodes->length()-1; i++) {
3971 assert(j <= i, "can't write beyond current index");
3972 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3973 identical++;
3974 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3975 continue;
3976 }
3977 if (identical > 0) {
3978 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3979 identical = 0;
3980 } else {
3981 Node* n = _expensive_nodes->at(i);
3982 igvn.hash_delete(n);
3983 n->set_req(0, NULL);
3984 igvn.hash_insert(n);
3985 }
3986 }
3987 if (identical > 0) {
3988 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3989 } else if (_expensive_nodes->length() >= 1) {
3990 Node* n = _expensive_nodes->at(i);
3991 igvn.hash_delete(n);
3992 n->set_req(0, NULL);
3993 igvn.hash_insert(n);
3994 }
3995 _expensive_nodes->trunc_to(j);
3996 }
3998 void Compile::add_expensive_node(Node * n) {
3999 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4000 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4001 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4002 if (OptimizeExpensiveOps) {
4003 _expensive_nodes->append(n);
4004 } else {
4005 // Clear control input and let IGVN optimize expensive nodes if
4006 // OptimizeExpensiveOps is off.
4007 n->set_req(0, NULL);
4008 }
4009 }
4011 /**
4012 * Remove the speculative part of types and clean up the graph
4013 */
4014 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4015 if (UseTypeSpeculation) {
4016 Unique_Node_List worklist;
4017 worklist.push(root());
4018 int modified = 0;
4019 // Go over all type nodes that carry a speculative type, drop the
4020 // speculative part of the type and enqueue the node for an igvn
4021 // which may optimize it out.
4022 for (uint next = 0; next < worklist.size(); ++next) {
4023 Node *n = worklist.at(next);
4024 if (n->is_Type()) {
4025 TypeNode* tn = n->as_Type();
4026 const Type* t = tn->type();
4027 const Type* t_no_spec = t->remove_speculative();
4028 if (t_no_spec != t) {
4029 bool in_hash = igvn.hash_delete(n);
4030 assert(in_hash, "node should be in igvn hash table");
4031 tn->set_type(t_no_spec);
4032 igvn.hash_insert(n);
4033 igvn._worklist.push(n); // give it a chance to go away
4034 modified++;
4035 }
4036 }
4037 uint max = n->len();
4038 for( uint i = 0; i < max; ++i ) {
4039 Node *m = n->in(i);
4040 if (not_a_node(m)) continue;
4041 worklist.push(m);
4042 }
4043 }
4044 // Drop the speculative part of all types in the igvn's type table
4045 igvn.remove_speculative_types();
4046 if (modified > 0) {
4047 igvn.optimize();
4048 }
4049 #ifdef ASSERT
4050 // Verify that after the IGVN is over no speculative type has resurfaced
4051 worklist.clear();
4052 worklist.push(root());
4053 for (uint next = 0; next < worklist.size(); ++next) {
4054 Node *n = worklist.at(next);
4055 const Type* t = igvn.type_or_null(n);
4056 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4057 if (n->is_Type()) {
4058 t = n->as_Type()->type();
4059 assert(t == t->remove_speculative(), "no more speculative types");
4060 }
4061 uint max = n->len();
4062 for( uint i = 0; i < max; ++i ) {
4063 Node *m = n->in(i);
4064 if (not_a_node(m)) continue;
4065 worklist.push(m);
4066 }
4067 }
4068 igvn.check_no_speculative_types();
4069 #endif
4070 }
4071 }
4073 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4074 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4075 if (ctrl != NULL) {
4076 // Express control dependency by a CastII node with a narrow type.
4077 value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */);
4078 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4079 // node from floating above the range check during loop optimizations. Otherwise, the
4080 // ConvI2L node may be eliminated independently of the range check, causing the data path
4081 // to become TOP while the control path is still there (although it's unreachable).
4082 value->set_req(0, ctrl);
4083 // Save CastII node to remove it after loop optimizations.
4084 phase->C->add_range_check_cast(value);
4085 value = phase->transform(value);
4086 }
4087 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4088 return phase->transform(new (phase->C) ConvI2LNode(value, ltype));
4089 }
4091 // Auxiliary method to support randomized stressing/fuzzing.
4092 //
4093 // This method can be called the arbitrary number of times, with current count
4094 // as the argument. The logic allows selecting a single candidate from the
4095 // running list of candidates as follows:
4096 // int count = 0;
4097 // Cand* selected = null;
4098 // while(cand = cand->next()) {
4099 // if (randomized_select(++count)) {
4100 // selected = cand;
4101 // }
4102 // }
4103 //
4104 // Including count equalizes the chances any candidate is "selected".
4105 // This is useful when we don't have the complete list of candidates to choose
4106 // from uniformly. In this case, we need to adjust the randomicity of the
4107 // selection, or else we will end up biasing the selection towards the latter
4108 // candidates.
4109 //
4110 // Quick back-envelope calculation shows that for the list of n candidates
4111 // the equal probability for the candidate to persist as "best" can be
4112 // achieved by replacing it with "next" k-th candidate with the probability
4113 // of 1/k. It can be easily shown that by the end of the run, the
4114 // probability for any candidate is converged to 1/n, thus giving the
4115 // uniform distribution among all the candidates.
4116 //
4117 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4118 #define RANDOMIZED_DOMAIN_POW 29
4119 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4120 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4121 bool Compile::randomized_select(int count) {
4122 assert(count > 0, "only positive");
4123 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4124 }