Fri, 15 Nov 2013 11:05:32 -0800
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
Summary: Add a field to C2 LoadNode and StoreNode classes which indicates whether the load/store should do an acquire/release on platforms which support it.
Reviewed-by: kvn
1 /*
2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "code/exceptionHandlerTable.hpp"
30 #include "code/nmethod.hpp"
31 #include "compiler/compileLog.hpp"
32 #include "compiler/disassembler.hpp"
33 #include "compiler/oopMap.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/block.hpp"
36 #include "opto/c2compiler.hpp"
37 #include "opto/callGenerator.hpp"
38 #include "opto/callnode.hpp"
39 #include "opto/cfgnode.hpp"
40 #include "opto/chaitin.hpp"
41 #include "opto/compile.hpp"
42 #include "opto/connode.hpp"
43 #include "opto/divnode.hpp"
44 #include "opto/escape.hpp"
45 #include "opto/idealGraphPrinter.hpp"
46 #include "opto/loopnode.hpp"
47 #include "opto/machnode.hpp"
48 #include "opto/macro.hpp"
49 #include "opto/matcher.hpp"
50 #include "opto/mathexactnode.hpp"
51 #include "opto/memnode.hpp"
52 #include "opto/mulnode.hpp"
53 #include "opto/node.hpp"
54 #include "opto/opcodes.hpp"
55 #include "opto/output.hpp"
56 #include "opto/parse.hpp"
57 #include "opto/phaseX.hpp"
58 #include "opto/rootnode.hpp"
59 #include "opto/runtime.hpp"
60 #include "opto/stringopts.hpp"
61 #include "opto/type.hpp"
62 #include "opto/vectornode.hpp"
63 #include "runtime/arguments.hpp"
64 #include "runtime/signature.hpp"
65 #include "runtime/stubRoutines.hpp"
66 #include "runtime/timer.hpp"
67 #include "trace/tracing.hpp"
68 #include "utilities/copy.hpp"
69 #ifdef TARGET_ARCH_MODEL_x86_32
70 # include "adfiles/ad_x86_32.hpp"
71 #endif
72 #ifdef TARGET_ARCH_MODEL_x86_64
73 # include "adfiles/ad_x86_64.hpp"
74 #endif
75 #ifdef TARGET_ARCH_MODEL_sparc
76 # include "adfiles/ad_sparc.hpp"
77 #endif
78 #ifdef TARGET_ARCH_MODEL_zero
79 # include "adfiles/ad_zero.hpp"
80 #endif
81 #ifdef TARGET_ARCH_MODEL_arm
82 # include "adfiles/ad_arm.hpp"
83 #endif
84 #ifdef TARGET_ARCH_MODEL_ppc_32
85 # include "adfiles/ad_ppc_32.hpp"
86 #endif
87 #ifdef TARGET_ARCH_MODEL_ppc_64
88 # include "adfiles/ad_ppc_64.hpp"
89 #endif
92 // -------------------- Compile::mach_constant_base_node -----------------------
93 // Constant table base node singleton.
94 MachConstantBaseNode* Compile::mach_constant_base_node() {
95 if (_mach_constant_base_node == NULL) {
96 _mach_constant_base_node = new (C) MachConstantBaseNode();
97 _mach_constant_base_node->add_req(C->root());
98 }
99 return _mach_constant_base_node;
100 }
103 /// Support for intrinsics.
105 // Return the index at which m must be inserted (or already exists).
106 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
107 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
108 #ifdef ASSERT
109 for (int i = 1; i < _intrinsics->length(); i++) {
110 CallGenerator* cg1 = _intrinsics->at(i-1);
111 CallGenerator* cg2 = _intrinsics->at(i);
112 assert(cg1->method() != cg2->method()
113 ? cg1->method() < cg2->method()
114 : cg1->is_virtual() < cg2->is_virtual(),
115 "compiler intrinsics list must stay sorted");
116 }
117 #endif
118 // Binary search sorted list, in decreasing intervals [lo, hi].
119 int lo = 0, hi = _intrinsics->length()-1;
120 while (lo <= hi) {
121 int mid = (uint)(hi + lo) / 2;
122 ciMethod* mid_m = _intrinsics->at(mid)->method();
123 if (m < mid_m) {
124 hi = mid-1;
125 } else if (m > mid_m) {
126 lo = mid+1;
127 } else {
128 // look at minor sort key
129 bool mid_virt = _intrinsics->at(mid)->is_virtual();
130 if (is_virtual < mid_virt) {
131 hi = mid-1;
132 } else if (is_virtual > mid_virt) {
133 lo = mid+1;
134 } else {
135 return mid; // exact match
136 }
137 }
138 }
139 return lo; // inexact match
140 }
142 void Compile::register_intrinsic(CallGenerator* cg) {
143 if (_intrinsics == NULL) {
144 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
145 }
146 // This code is stolen from ciObjectFactory::insert.
147 // Really, GrowableArray should have methods for
148 // insert_at, remove_at, and binary_search.
149 int len = _intrinsics->length();
150 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
151 if (index == len) {
152 _intrinsics->append(cg);
153 } else {
154 #ifdef ASSERT
155 CallGenerator* oldcg = _intrinsics->at(index);
156 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
157 #endif
158 _intrinsics->append(_intrinsics->at(len-1));
159 int pos;
160 for (pos = len-2; pos >= index; pos--) {
161 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
162 }
163 _intrinsics->at_put(index, cg);
164 }
165 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
166 }
168 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
169 assert(m->is_loaded(), "don't try this on unloaded methods");
170 if (_intrinsics != NULL) {
171 int index = intrinsic_insertion_index(m, is_virtual);
172 if (index < _intrinsics->length()
173 && _intrinsics->at(index)->method() == m
174 && _intrinsics->at(index)->is_virtual() == is_virtual) {
175 return _intrinsics->at(index);
176 }
177 }
178 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
179 if (m->intrinsic_id() != vmIntrinsics::_none &&
180 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
181 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
182 if (cg != NULL) {
183 // Save it for next time:
184 register_intrinsic(cg);
185 return cg;
186 } else {
187 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
188 }
189 }
190 return NULL;
191 }
193 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
194 // in library_call.cpp.
197 #ifndef PRODUCT
198 // statistics gathering...
200 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
201 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
203 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
204 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
205 int oflags = _intrinsic_hist_flags[id];
206 assert(flags != 0, "what happened?");
207 if (is_virtual) {
208 flags |= _intrinsic_virtual;
209 }
210 bool changed = (flags != oflags);
211 if ((flags & _intrinsic_worked) != 0) {
212 juint count = (_intrinsic_hist_count[id] += 1);
213 if (count == 1) {
214 changed = true; // first time
215 }
216 // increment the overall count also:
217 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
218 }
219 if (changed) {
220 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
221 // Something changed about the intrinsic's virtuality.
222 if ((flags & _intrinsic_virtual) != 0) {
223 // This is the first use of this intrinsic as a virtual call.
224 if (oflags != 0) {
225 // We already saw it as a non-virtual, so note both cases.
226 flags |= _intrinsic_both;
227 }
228 } else if ((oflags & _intrinsic_both) == 0) {
229 // This is the first use of this intrinsic as a non-virtual
230 flags |= _intrinsic_both;
231 }
232 }
233 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
234 }
235 // update the overall flags also:
236 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
237 return changed;
238 }
240 static char* format_flags(int flags, char* buf) {
241 buf[0] = 0;
242 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
243 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
244 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
245 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
246 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
247 if (buf[0] == 0) strcat(buf, ",");
248 assert(buf[0] == ',', "must be");
249 return &buf[1];
250 }
252 void Compile::print_intrinsic_statistics() {
253 char flagsbuf[100];
254 ttyLocker ttyl;
255 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
256 tty->print_cr("Compiler intrinsic usage:");
257 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
258 if (total == 0) total = 1; // avoid div0 in case of no successes
259 #define PRINT_STAT_LINE(name, c, f) \
260 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
261 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
262 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
263 int flags = _intrinsic_hist_flags[id];
264 juint count = _intrinsic_hist_count[id];
265 if ((flags | count) != 0) {
266 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
267 }
268 }
269 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
270 if (xtty != NULL) xtty->tail("statistics");
271 }
273 void Compile::print_statistics() {
274 { ttyLocker ttyl;
275 if (xtty != NULL) xtty->head("statistics type='opto'");
276 Parse::print_statistics();
277 PhaseCCP::print_statistics();
278 PhaseRegAlloc::print_statistics();
279 Scheduling::print_statistics();
280 PhasePeephole::print_statistics();
281 PhaseIdealLoop::print_statistics();
282 if (xtty != NULL) xtty->tail("statistics");
283 }
284 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
285 // put this under its own <statistics> element.
286 print_intrinsic_statistics();
287 }
288 }
289 #endif //PRODUCT
291 // Support for bundling info
292 Bundle* Compile::node_bundling(const Node *n) {
293 assert(valid_bundle_info(n), "oob");
294 return &_node_bundling_base[n->_idx];
295 }
297 bool Compile::valid_bundle_info(const Node *n) {
298 return (_node_bundling_limit > n->_idx);
299 }
302 void Compile::gvn_replace_by(Node* n, Node* nn) {
303 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
304 Node* use = n->last_out(i);
305 bool is_in_table = initial_gvn()->hash_delete(use);
306 uint uses_found = 0;
307 for (uint j = 0; j < use->len(); j++) {
308 if (use->in(j) == n) {
309 if (j < use->req())
310 use->set_req(j, nn);
311 else
312 use->set_prec(j, nn);
313 uses_found++;
314 }
315 }
316 if (is_in_table) {
317 // reinsert into table
318 initial_gvn()->hash_find_insert(use);
319 }
320 record_for_igvn(use);
321 i -= uses_found; // we deleted 1 or more copies of this edge
322 }
323 }
326 static inline bool not_a_node(const Node* n) {
327 if (n == NULL) return true;
328 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
329 if (*(address*)n == badAddress) return true; // kill by Node::destruct
330 return false;
331 }
333 // Identify all nodes that are reachable from below, useful.
334 // Use breadth-first pass that records state in a Unique_Node_List,
335 // recursive traversal is slower.
336 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
337 int estimated_worklist_size = unique();
338 useful.map( estimated_worklist_size, NULL ); // preallocate space
340 // Initialize worklist
341 if (root() != NULL) { useful.push(root()); }
342 // If 'top' is cached, declare it useful to preserve cached node
343 if( cached_top_node() ) { useful.push(cached_top_node()); }
345 // Push all useful nodes onto the list, breadthfirst
346 for( uint next = 0; next < useful.size(); ++next ) {
347 assert( next < unique(), "Unique useful nodes < total nodes");
348 Node *n = useful.at(next);
349 uint max = n->len();
350 for( uint i = 0; i < max; ++i ) {
351 Node *m = n->in(i);
352 if (not_a_node(m)) continue;
353 useful.push(m);
354 }
355 }
356 }
358 // Update dead_node_list with any missing dead nodes using useful
359 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
360 void Compile::update_dead_node_list(Unique_Node_List &useful) {
361 uint max_idx = unique();
362 VectorSet& useful_node_set = useful.member_set();
364 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
365 // If node with index node_idx is not in useful set,
366 // mark it as dead in dead node list.
367 if (! useful_node_set.test(node_idx) ) {
368 record_dead_node(node_idx);
369 }
370 }
371 }
373 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
374 int shift = 0;
375 for (int i = 0; i < inlines->length(); i++) {
376 CallGenerator* cg = inlines->at(i);
377 CallNode* call = cg->call_node();
378 if (shift > 0) {
379 inlines->at_put(i-shift, cg);
380 }
381 if (!useful.member(call)) {
382 shift++;
383 }
384 }
385 inlines->trunc_to(inlines->length()-shift);
386 }
388 // Disconnect all useless nodes by disconnecting those at the boundary.
389 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
390 uint next = 0;
391 while (next < useful.size()) {
392 Node *n = useful.at(next++);
393 // Use raw traversal of out edges since this code removes out edges
394 int max = n->outcnt();
395 for (int j = 0; j < max; ++j) {
396 Node* child = n->raw_out(j);
397 if (! useful.member(child)) {
398 assert(!child->is_top() || child != top(),
399 "If top is cached in Compile object it is in useful list");
400 // Only need to remove this out-edge to the useless node
401 n->raw_del_out(j);
402 --j;
403 --max;
404 }
405 }
406 if (n->outcnt() == 1 && n->has_special_unique_user()) {
407 record_for_igvn(n->unique_out());
408 }
409 }
410 // Remove useless macro and predicate opaq nodes
411 for (int i = C->macro_count()-1; i >= 0; i--) {
412 Node* n = C->macro_node(i);
413 if (!useful.member(n)) {
414 remove_macro_node(n);
415 }
416 }
417 // Remove useless expensive node
418 for (int i = C->expensive_count()-1; i >= 0; i--) {
419 Node* n = C->expensive_node(i);
420 if (!useful.member(n)) {
421 remove_expensive_node(n);
422 }
423 }
424 // clean up the late inline lists
425 remove_useless_late_inlines(&_string_late_inlines, useful);
426 remove_useless_late_inlines(&_boxing_late_inlines, useful);
427 remove_useless_late_inlines(&_late_inlines, useful);
428 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
429 }
431 //------------------------------frame_size_in_words-----------------------------
432 // frame_slots in units of words
433 int Compile::frame_size_in_words() const {
434 // shift is 0 in LP32 and 1 in LP64
435 const int shift = (LogBytesPerWord - LogBytesPerInt);
436 int words = _frame_slots >> shift;
437 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
438 return words;
439 }
441 // ============================================================================
442 //------------------------------CompileWrapper---------------------------------
443 class CompileWrapper : public StackObj {
444 Compile *const _compile;
445 public:
446 CompileWrapper(Compile* compile);
448 ~CompileWrapper();
449 };
451 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
452 // the Compile* pointer is stored in the current ciEnv:
453 ciEnv* env = compile->env();
454 assert(env == ciEnv::current(), "must already be a ciEnv active");
455 assert(env->compiler_data() == NULL, "compile already active?");
456 env->set_compiler_data(compile);
457 assert(compile == Compile::current(), "sanity");
459 compile->set_type_dict(NULL);
460 compile->set_type_hwm(NULL);
461 compile->set_type_last_size(0);
462 compile->set_last_tf(NULL, NULL);
463 compile->set_indexSet_arena(NULL);
464 compile->set_indexSet_free_block_list(NULL);
465 compile->init_type_arena();
466 Type::Initialize(compile);
467 _compile->set_scratch_buffer_blob(NULL);
468 _compile->begin_method();
469 }
470 CompileWrapper::~CompileWrapper() {
471 _compile->end_method();
472 if (_compile->scratch_buffer_blob() != NULL)
473 BufferBlob::free(_compile->scratch_buffer_blob());
474 _compile->env()->set_compiler_data(NULL);
475 }
478 //----------------------------print_compile_messages---------------------------
479 void Compile::print_compile_messages() {
480 #ifndef PRODUCT
481 // Check if recompiling
482 if (_subsume_loads == false && PrintOpto) {
483 // Recompiling without allowing machine instructions to subsume loads
484 tty->print_cr("*********************************************************");
485 tty->print_cr("** Bailout: Recompile without subsuming loads **");
486 tty->print_cr("*********************************************************");
487 }
488 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
489 // Recompiling without escape analysis
490 tty->print_cr("*********************************************************");
491 tty->print_cr("** Bailout: Recompile without escape analysis **");
492 tty->print_cr("*********************************************************");
493 }
494 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
495 // Recompiling without boxing elimination
496 tty->print_cr("*********************************************************");
497 tty->print_cr("** Bailout: Recompile without boxing elimination **");
498 tty->print_cr("*********************************************************");
499 }
500 if (env()->break_at_compile()) {
501 // Open the debugger when compiling this method.
502 tty->print("### Breaking when compiling: ");
503 method()->print_short_name();
504 tty->cr();
505 BREAKPOINT;
506 }
508 if( PrintOpto ) {
509 if (is_osr_compilation()) {
510 tty->print("[OSR]%3d", _compile_id);
511 } else {
512 tty->print("%3d", _compile_id);
513 }
514 }
515 #endif
516 }
519 //-----------------------init_scratch_buffer_blob------------------------------
520 // Construct a temporary BufferBlob and cache it for this compile.
521 void Compile::init_scratch_buffer_blob(int const_size) {
522 // If there is already a scratch buffer blob allocated and the
523 // constant section is big enough, use it. Otherwise free the
524 // current and allocate a new one.
525 BufferBlob* blob = scratch_buffer_blob();
526 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
527 // Use the current blob.
528 } else {
529 if (blob != NULL) {
530 BufferBlob::free(blob);
531 }
533 ResourceMark rm;
534 _scratch_const_size = const_size;
535 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
536 blob = BufferBlob::create("Compile::scratch_buffer", size);
537 // Record the buffer blob for next time.
538 set_scratch_buffer_blob(blob);
539 // Have we run out of code space?
540 if (scratch_buffer_blob() == NULL) {
541 // Let CompilerBroker disable further compilations.
542 record_failure("Not enough space for scratch buffer in CodeCache");
543 return;
544 }
545 }
547 // Initialize the relocation buffers
548 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
549 set_scratch_locs_memory(locs_buf);
550 }
553 //-----------------------scratch_emit_size-------------------------------------
554 // Helper function that computes size by emitting code
555 uint Compile::scratch_emit_size(const Node* n) {
556 // Start scratch_emit_size section.
557 set_in_scratch_emit_size(true);
559 // Emit into a trash buffer and count bytes emitted.
560 // This is a pretty expensive way to compute a size,
561 // but it works well enough if seldom used.
562 // All common fixed-size instructions are given a size
563 // method by the AD file.
564 // Note that the scratch buffer blob and locs memory are
565 // allocated at the beginning of the compile task, and
566 // may be shared by several calls to scratch_emit_size.
567 // The allocation of the scratch buffer blob is particularly
568 // expensive, since it has to grab the code cache lock.
569 BufferBlob* blob = this->scratch_buffer_blob();
570 assert(blob != NULL, "Initialize BufferBlob at start");
571 assert(blob->size() > MAX_inst_size, "sanity");
572 relocInfo* locs_buf = scratch_locs_memory();
573 address blob_begin = blob->content_begin();
574 address blob_end = (address)locs_buf;
575 assert(blob->content_contains(blob_end), "sanity");
576 CodeBuffer buf(blob_begin, blob_end - blob_begin);
577 buf.initialize_consts_size(_scratch_const_size);
578 buf.initialize_stubs_size(MAX_stubs_size);
579 assert(locs_buf != NULL, "sanity");
580 int lsize = MAX_locs_size / 3;
581 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
582 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
583 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
585 // Do the emission.
587 Label fakeL; // Fake label for branch instructions.
588 Label* saveL = NULL;
589 uint save_bnum = 0;
590 bool is_branch = n->is_MachBranch();
591 if (is_branch) {
592 MacroAssembler masm(&buf);
593 masm.bind(fakeL);
594 n->as_MachBranch()->save_label(&saveL, &save_bnum);
595 n->as_MachBranch()->label_set(&fakeL, 0);
596 }
597 n->emit(buf, this->regalloc());
598 if (is_branch) // Restore label.
599 n->as_MachBranch()->label_set(saveL, save_bnum);
601 // End scratch_emit_size section.
602 set_in_scratch_emit_size(false);
604 return buf.insts_size();
605 }
608 // ============================================================================
609 //------------------------------Compile standard-------------------------------
610 debug_only( int Compile::_debug_idx = 100000; )
612 // Compile a method. entry_bci is -1 for normal compilations and indicates
613 // the continuation bci for on stack replacement.
616 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
617 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
618 : Phase(Compiler),
619 _env(ci_env),
620 _log(ci_env->log()),
621 _compile_id(ci_env->compile_id()),
622 _save_argument_registers(false),
623 _stub_name(NULL),
624 _stub_function(NULL),
625 _stub_entry_point(NULL),
626 _method(target),
627 _entry_bci(osr_bci),
628 _initial_gvn(NULL),
629 _for_igvn(NULL),
630 _warm_calls(NULL),
631 _subsume_loads(subsume_loads),
632 _do_escape_analysis(do_escape_analysis),
633 _eliminate_boxing(eliminate_boxing),
634 _failure_reason(NULL),
635 _code_buffer("Compile::Fill_buffer"),
636 _orig_pc_slot(0),
637 _orig_pc_slot_offset_in_bytes(0),
638 _has_method_handle_invokes(false),
639 _mach_constant_base_node(NULL),
640 _node_bundling_limit(0),
641 _node_bundling_base(NULL),
642 _java_calls(0),
643 _inner_loops(0),
644 _scratch_const_size(-1),
645 _in_scratch_emit_size(false),
646 _dead_node_list(comp_arena()),
647 _dead_node_count(0),
648 #ifndef PRODUCT
649 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
650 _printer(IdealGraphPrinter::printer()),
651 #endif
652 _congraph(NULL),
653 _late_inlines(comp_arena(), 2, 0, NULL),
654 _string_late_inlines(comp_arena(), 2, 0, NULL),
655 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
656 _late_inlines_pos(0),
657 _number_of_mh_late_inlines(0),
658 _inlining_progress(false),
659 _inlining_incrementally(false),
660 _print_inlining_list(NULL),
661 _print_inlining_idx(0),
662 _preserve_jvm_state(0) {
663 C = this;
665 CompileWrapper cw(this);
666 #ifndef PRODUCT
667 if (TimeCompiler2) {
668 tty->print(" ");
669 target->holder()->name()->print();
670 tty->print(".");
671 target->print_short_name();
672 tty->print(" ");
673 }
674 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
675 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
676 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
677 if (!print_opto_assembly) {
678 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
679 if (print_assembly && !Disassembler::can_decode()) {
680 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
681 print_opto_assembly = true;
682 }
683 }
684 set_print_assembly(print_opto_assembly);
685 set_parsed_irreducible_loop(false);
686 #endif
687 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
688 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
690 if (ProfileTraps) {
691 // Make sure the method being compiled gets its own MDO,
692 // so we can at least track the decompile_count().
693 method()->ensure_method_data();
694 }
696 Init(::AliasLevel);
699 print_compile_messages();
701 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
702 _ilt = InlineTree::build_inline_tree_root();
703 else
704 _ilt = NULL;
706 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
707 assert(num_alias_types() >= AliasIdxRaw, "");
709 #define MINIMUM_NODE_HASH 1023
710 // Node list that Iterative GVN will start with
711 Unique_Node_List for_igvn(comp_arena());
712 set_for_igvn(&for_igvn);
714 // GVN that will be run immediately on new nodes
715 uint estimated_size = method()->code_size()*4+64;
716 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
717 PhaseGVN gvn(node_arena(), estimated_size);
718 set_initial_gvn(&gvn);
720 if (print_inlining() || print_intrinsics()) {
721 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
722 }
723 { // Scope for timing the parser
724 TracePhase t3("parse", &_t_parser, true);
726 // Put top into the hash table ASAP.
727 initial_gvn()->transform_no_reclaim(top());
729 // Set up tf(), start(), and find a CallGenerator.
730 CallGenerator* cg = NULL;
731 if (is_osr_compilation()) {
732 const TypeTuple *domain = StartOSRNode::osr_domain();
733 const TypeTuple *range = TypeTuple::make_range(method()->signature());
734 init_tf(TypeFunc::make(domain, range));
735 StartNode* s = new (this) StartOSRNode(root(), domain);
736 initial_gvn()->set_type_bottom(s);
737 init_start(s);
738 cg = CallGenerator::for_osr(method(), entry_bci());
739 } else {
740 // Normal case.
741 init_tf(TypeFunc::make(method()));
742 StartNode* s = new (this) StartNode(root(), tf()->domain());
743 initial_gvn()->set_type_bottom(s);
744 init_start(s);
745 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
746 // With java.lang.ref.reference.get() we must go through the
747 // intrinsic when G1 is enabled - even when get() is the root
748 // method of the compile - so that, if necessary, the value in
749 // the referent field of the reference object gets recorded by
750 // the pre-barrier code.
751 // Specifically, if G1 is enabled, the value in the referent
752 // field is recorded by the G1 SATB pre barrier. This will
753 // result in the referent being marked live and the reference
754 // object removed from the list of discovered references during
755 // reference processing.
756 cg = find_intrinsic(method(), false);
757 }
758 if (cg == NULL) {
759 float past_uses = method()->interpreter_invocation_count();
760 float expected_uses = past_uses;
761 cg = CallGenerator::for_inline(method(), expected_uses);
762 }
763 }
764 if (failing()) return;
765 if (cg == NULL) {
766 record_method_not_compilable_all_tiers("cannot parse method");
767 return;
768 }
769 JVMState* jvms = build_start_state(start(), tf());
770 if ((jvms = cg->generate(jvms, NULL)) == NULL) {
771 record_method_not_compilable("method parse failed");
772 return;
773 }
774 GraphKit kit(jvms);
776 if (!kit.stopped()) {
777 // Accept return values, and transfer control we know not where.
778 // This is done by a special, unique ReturnNode bound to root.
779 return_values(kit.jvms());
780 }
782 if (kit.has_exceptions()) {
783 // Any exceptions that escape from this call must be rethrown
784 // to whatever caller is dynamically above us on the stack.
785 // This is done by a special, unique RethrowNode bound to root.
786 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
787 }
789 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
791 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
792 inline_string_calls(true);
793 }
795 if (failing()) return;
797 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
799 // Remove clutter produced by parsing.
800 if (!failing()) {
801 ResourceMark rm;
802 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
803 }
804 }
806 // Note: Large methods are capped off in do_one_bytecode().
807 if (failing()) return;
809 // After parsing, node notes are no longer automagic.
810 // They must be propagated by register_new_node_with_optimizer(),
811 // clone(), or the like.
812 set_default_node_notes(NULL);
814 for (;;) {
815 int successes = Inline_Warm();
816 if (failing()) return;
817 if (successes == 0) break;
818 }
820 // Drain the list.
821 Finish_Warm();
822 #ifndef PRODUCT
823 if (_printer) {
824 _printer->print_inlining(this);
825 }
826 #endif
828 if (failing()) return;
829 NOT_PRODUCT( verify_graph_edges(); )
831 // Now optimize
832 Optimize();
833 if (failing()) return;
834 NOT_PRODUCT( verify_graph_edges(); )
836 #ifndef PRODUCT
837 if (PrintIdeal) {
838 ttyLocker ttyl; // keep the following output all in one block
839 // This output goes directly to the tty, not the compiler log.
840 // To enable tools to match it up with the compilation activity,
841 // be sure to tag this tty output with the compile ID.
842 if (xtty != NULL) {
843 xtty->head("ideal compile_id='%d'%s", compile_id(),
844 is_osr_compilation() ? " compile_kind='osr'" :
845 "");
846 }
847 root()->dump(9999);
848 if (xtty != NULL) {
849 xtty->tail("ideal");
850 }
851 }
852 #endif
854 // Now that we know the size of all the monitors we can add a fixed slot
855 // for the original deopt pc.
857 _orig_pc_slot = fixed_slots();
858 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
859 set_fixed_slots(next_slot);
861 // Now generate code
862 Code_Gen();
863 if (failing()) return;
865 // Check if we want to skip execution of all compiled code.
866 {
867 #ifndef PRODUCT
868 if (OptoNoExecute) {
869 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
870 return;
871 }
872 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
873 #endif
875 if (is_osr_compilation()) {
876 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
877 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
878 } else {
879 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
880 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
881 }
883 env()->register_method(_method, _entry_bci,
884 &_code_offsets,
885 _orig_pc_slot_offset_in_bytes,
886 code_buffer(),
887 frame_size_in_words(), _oop_map_set,
888 &_handler_table, &_inc_table,
889 compiler,
890 env()->comp_level(),
891 has_unsafe_access(),
892 SharedRuntime::is_wide_vector(max_vector_size())
893 );
895 if (log() != NULL) // Print code cache state into compiler log
896 log()->code_cache_state();
897 }
898 }
900 //------------------------------Compile----------------------------------------
901 // Compile a runtime stub
902 Compile::Compile( ciEnv* ci_env,
903 TypeFunc_generator generator,
904 address stub_function,
905 const char *stub_name,
906 int is_fancy_jump,
907 bool pass_tls,
908 bool save_arg_registers,
909 bool return_pc )
910 : Phase(Compiler),
911 _env(ci_env),
912 _log(ci_env->log()),
913 _compile_id(0),
914 _save_argument_registers(save_arg_registers),
915 _method(NULL),
916 _stub_name(stub_name),
917 _stub_function(stub_function),
918 _stub_entry_point(NULL),
919 _entry_bci(InvocationEntryBci),
920 _initial_gvn(NULL),
921 _for_igvn(NULL),
922 _warm_calls(NULL),
923 _orig_pc_slot(0),
924 _orig_pc_slot_offset_in_bytes(0),
925 _subsume_loads(true),
926 _do_escape_analysis(false),
927 _eliminate_boxing(false),
928 _failure_reason(NULL),
929 _code_buffer("Compile::Fill_buffer"),
930 _has_method_handle_invokes(false),
931 _mach_constant_base_node(NULL),
932 _node_bundling_limit(0),
933 _node_bundling_base(NULL),
934 _java_calls(0),
935 _inner_loops(0),
936 #ifndef PRODUCT
937 _trace_opto_output(TraceOptoOutput),
938 _printer(NULL),
939 #endif
940 _dead_node_list(comp_arena()),
941 _dead_node_count(0),
942 _congraph(NULL),
943 _number_of_mh_late_inlines(0),
944 _inlining_progress(false),
945 _inlining_incrementally(false),
946 _print_inlining_list(NULL),
947 _print_inlining_idx(0),
948 _preserve_jvm_state(0) {
949 C = this;
951 #ifndef PRODUCT
952 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
953 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
954 set_print_assembly(PrintFrameConverterAssembly);
955 set_parsed_irreducible_loop(false);
956 #endif
957 CompileWrapper cw(this);
958 Init(/*AliasLevel=*/ 0);
959 init_tf((*generator)());
961 {
962 // The following is a dummy for the sake of GraphKit::gen_stub
963 Unique_Node_List for_igvn(comp_arena());
964 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
965 PhaseGVN gvn(Thread::current()->resource_area(),255);
966 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
967 gvn.transform_no_reclaim(top());
969 GraphKit kit;
970 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
971 }
973 NOT_PRODUCT( verify_graph_edges(); )
974 Code_Gen();
975 if (failing()) return;
978 // Entry point will be accessed using compile->stub_entry_point();
979 if (code_buffer() == NULL) {
980 Matcher::soft_match_failure();
981 } else {
982 if (PrintAssembly && (WizardMode || Verbose))
983 tty->print_cr("### Stub::%s", stub_name);
985 if (!failing()) {
986 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
988 // Make the NMethod
989 // For now we mark the frame as never safe for profile stackwalking
990 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
991 code_buffer(),
992 CodeOffsets::frame_never_safe,
993 // _code_offsets.value(CodeOffsets::Frame_Complete),
994 frame_size_in_words(),
995 _oop_map_set,
996 save_arg_registers);
997 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
999 _stub_entry_point = rs->entry_point();
1000 }
1001 }
1002 }
1004 //------------------------------Init-------------------------------------------
1005 // Prepare for a single compilation
1006 void Compile::Init(int aliaslevel) {
1007 _unique = 0;
1008 _regalloc = NULL;
1010 _tf = NULL; // filled in later
1011 _top = NULL; // cached later
1012 _matcher = NULL; // filled in later
1013 _cfg = NULL; // filled in later
1015 set_24_bit_selection_and_mode(Use24BitFP, false);
1017 _node_note_array = NULL;
1018 _default_node_notes = NULL;
1020 _immutable_memory = NULL; // filled in at first inquiry
1022 // Globally visible Nodes
1023 // First set TOP to NULL to give safe behavior during creation of RootNode
1024 set_cached_top_node(NULL);
1025 set_root(new (this) RootNode());
1026 // Now that you have a Root to point to, create the real TOP
1027 set_cached_top_node( new (this) ConNode(Type::TOP) );
1028 set_recent_alloc(NULL, NULL);
1030 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1031 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1032 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1033 env()->set_dependencies(new Dependencies(env()));
1035 _fixed_slots = 0;
1036 set_has_split_ifs(false);
1037 set_has_loops(has_method() && method()->has_loops()); // first approximation
1038 set_has_stringbuilder(false);
1039 set_has_boxed_value(false);
1040 _trap_can_recompile = false; // no traps emitted yet
1041 _major_progress = true; // start out assuming good things will happen
1042 set_has_unsafe_access(false);
1043 set_max_vector_size(0);
1044 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1045 set_decompile_count(0);
1047 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1048 set_num_loop_opts(LoopOptsCount);
1049 set_do_inlining(Inline);
1050 set_max_inline_size(MaxInlineSize);
1051 set_freq_inline_size(FreqInlineSize);
1052 set_do_scheduling(OptoScheduling);
1053 set_do_count_invocations(false);
1054 set_do_method_data_update(false);
1056 if (debug_info()->recording_non_safepoints()) {
1057 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1058 (comp_arena(), 8, 0, NULL));
1059 set_default_node_notes(Node_Notes::make(this));
1060 }
1062 // // -- Initialize types before each compile --
1063 // // Update cached type information
1064 // if( _method && _method->constants() )
1065 // Type::update_loaded_types(_method, _method->constants());
1067 // Init alias_type map.
1068 if (!_do_escape_analysis && aliaslevel == 3)
1069 aliaslevel = 2; // No unique types without escape analysis
1070 _AliasLevel = aliaslevel;
1071 const int grow_ats = 16;
1072 _max_alias_types = grow_ats;
1073 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1074 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1075 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1076 {
1077 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1078 }
1079 // Initialize the first few types.
1080 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1081 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1082 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1083 _num_alias_types = AliasIdxRaw+1;
1084 // Zero out the alias type cache.
1085 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1086 // A NULL adr_type hits in the cache right away. Preload the right answer.
1087 probe_alias_cache(NULL)->_index = AliasIdxTop;
1089 _intrinsics = NULL;
1090 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1091 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1092 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1093 register_library_intrinsics();
1094 }
1096 //---------------------------init_start----------------------------------------
1097 // Install the StartNode on this compile object.
1098 void Compile::init_start(StartNode* s) {
1099 if (failing())
1100 return; // already failing
1101 assert(s == start(), "");
1102 }
1104 StartNode* Compile::start() const {
1105 assert(!failing(), "");
1106 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1107 Node* start = root()->fast_out(i);
1108 if( start->is_Start() )
1109 return start->as_Start();
1110 }
1111 ShouldNotReachHere();
1112 return NULL;
1113 }
1115 //-------------------------------immutable_memory-------------------------------------
1116 // Access immutable memory
1117 Node* Compile::immutable_memory() {
1118 if (_immutable_memory != NULL) {
1119 return _immutable_memory;
1120 }
1121 StartNode* s = start();
1122 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1123 Node *p = s->fast_out(i);
1124 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1125 _immutable_memory = p;
1126 return _immutable_memory;
1127 }
1128 }
1129 ShouldNotReachHere();
1130 return NULL;
1131 }
1133 //----------------------set_cached_top_node------------------------------------
1134 // Install the cached top node, and make sure Node::is_top works correctly.
1135 void Compile::set_cached_top_node(Node* tn) {
1136 if (tn != NULL) verify_top(tn);
1137 Node* old_top = _top;
1138 _top = tn;
1139 // Calling Node::setup_is_top allows the nodes the chance to adjust
1140 // their _out arrays.
1141 if (_top != NULL) _top->setup_is_top();
1142 if (old_top != NULL) old_top->setup_is_top();
1143 assert(_top == NULL || top()->is_top(), "");
1144 }
1146 #ifdef ASSERT
1147 uint Compile::count_live_nodes_by_graph_walk() {
1148 Unique_Node_List useful(comp_arena());
1149 // Get useful node list by walking the graph.
1150 identify_useful_nodes(useful);
1151 return useful.size();
1152 }
1154 void Compile::print_missing_nodes() {
1156 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1157 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1158 return;
1159 }
1161 // This is an expensive function. It is executed only when the user
1162 // specifies VerifyIdealNodeCount option or otherwise knows the
1163 // additional work that needs to be done to identify reachable nodes
1164 // by walking the flow graph and find the missing ones using
1165 // _dead_node_list.
1167 Unique_Node_List useful(comp_arena());
1168 // Get useful node list by walking the graph.
1169 identify_useful_nodes(useful);
1171 uint l_nodes = C->live_nodes();
1172 uint l_nodes_by_walk = useful.size();
1174 if (l_nodes != l_nodes_by_walk) {
1175 if (_log != NULL) {
1176 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1177 _log->stamp();
1178 _log->end_head();
1179 }
1180 VectorSet& useful_member_set = useful.member_set();
1181 int last_idx = l_nodes_by_walk;
1182 for (int i = 0; i < last_idx; i++) {
1183 if (useful_member_set.test(i)) {
1184 if (_dead_node_list.test(i)) {
1185 if (_log != NULL) {
1186 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1187 }
1188 if (PrintIdealNodeCount) {
1189 // Print the log message to tty
1190 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1191 useful.at(i)->dump();
1192 }
1193 }
1194 }
1195 else if (! _dead_node_list.test(i)) {
1196 if (_log != NULL) {
1197 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1198 }
1199 if (PrintIdealNodeCount) {
1200 // Print the log message to tty
1201 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1202 }
1203 }
1204 }
1205 if (_log != NULL) {
1206 _log->tail("mismatched_nodes");
1207 }
1208 }
1209 }
1210 #endif
1212 #ifndef PRODUCT
1213 void Compile::verify_top(Node* tn) const {
1214 if (tn != NULL) {
1215 assert(tn->is_Con(), "top node must be a constant");
1216 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1217 assert(tn->in(0) != NULL, "must have live top node");
1218 }
1219 }
1220 #endif
1223 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1225 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1226 guarantee(arr != NULL, "");
1227 int num_blocks = arr->length();
1228 if (grow_by < num_blocks) grow_by = num_blocks;
1229 int num_notes = grow_by * _node_notes_block_size;
1230 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1231 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1232 while (num_notes > 0) {
1233 arr->append(notes);
1234 notes += _node_notes_block_size;
1235 num_notes -= _node_notes_block_size;
1236 }
1237 assert(num_notes == 0, "exact multiple, please");
1238 }
1240 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1241 if (source == NULL || dest == NULL) return false;
1243 if (dest->is_Con())
1244 return false; // Do not push debug info onto constants.
1246 #ifdef ASSERT
1247 // Leave a bread crumb trail pointing to the original node:
1248 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1249 dest->set_debug_orig(source);
1250 }
1251 #endif
1253 if (node_note_array() == NULL)
1254 return false; // Not collecting any notes now.
1256 // This is a copy onto a pre-existing node, which may already have notes.
1257 // If both nodes have notes, do not overwrite any pre-existing notes.
1258 Node_Notes* source_notes = node_notes_at(source->_idx);
1259 if (source_notes == NULL || source_notes->is_clear()) return false;
1260 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1261 if (dest_notes == NULL || dest_notes->is_clear()) {
1262 return set_node_notes_at(dest->_idx, source_notes);
1263 }
1265 Node_Notes merged_notes = (*source_notes);
1266 // The order of operations here ensures that dest notes will win...
1267 merged_notes.update_from(dest_notes);
1268 return set_node_notes_at(dest->_idx, &merged_notes);
1269 }
1272 //--------------------------allow_range_check_smearing-------------------------
1273 // Gating condition for coalescing similar range checks.
1274 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1275 // single covering check that is at least as strong as any of them.
1276 // If the optimization succeeds, the simplified (strengthened) range check
1277 // will always succeed. If it fails, we will deopt, and then give up
1278 // on the optimization.
1279 bool Compile::allow_range_check_smearing() const {
1280 // If this method has already thrown a range-check,
1281 // assume it was because we already tried range smearing
1282 // and it failed.
1283 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1284 return !already_trapped;
1285 }
1288 //------------------------------flatten_alias_type-----------------------------
1289 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1290 int offset = tj->offset();
1291 TypePtr::PTR ptr = tj->ptr();
1293 // Known instance (scalarizable allocation) alias only with itself.
1294 bool is_known_inst = tj->isa_oopptr() != NULL &&
1295 tj->is_oopptr()->is_known_instance();
1297 // Process weird unsafe references.
1298 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1299 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1300 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1301 tj = TypeOopPtr::BOTTOM;
1302 ptr = tj->ptr();
1303 offset = tj->offset();
1304 }
1306 // Array pointers need some flattening
1307 const TypeAryPtr *ta = tj->isa_aryptr();
1308 if (ta && ta->is_stable()) {
1309 // Erase stability property for alias analysis.
1310 tj = ta = ta->cast_to_stable(false);
1311 }
1312 if( ta && is_known_inst ) {
1313 if ( offset != Type::OffsetBot &&
1314 offset > arrayOopDesc::length_offset_in_bytes() ) {
1315 offset = Type::OffsetBot; // Flatten constant access into array body only
1316 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1317 }
1318 } else if( ta && _AliasLevel >= 2 ) {
1319 // For arrays indexed by constant indices, we flatten the alias
1320 // space to include all of the array body. Only the header, klass
1321 // and array length can be accessed un-aliased.
1322 if( offset != Type::OffsetBot ) {
1323 if( ta->const_oop() ) { // MethodData* or Method*
1324 offset = Type::OffsetBot; // Flatten constant access into array body
1325 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1326 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1327 // range is OK as-is.
1328 tj = ta = TypeAryPtr::RANGE;
1329 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1330 tj = TypeInstPtr::KLASS; // all klass loads look alike
1331 ta = TypeAryPtr::RANGE; // generic ignored junk
1332 ptr = TypePtr::BotPTR;
1333 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1334 tj = TypeInstPtr::MARK;
1335 ta = TypeAryPtr::RANGE; // generic ignored junk
1336 ptr = TypePtr::BotPTR;
1337 } else { // Random constant offset into array body
1338 offset = Type::OffsetBot; // Flatten constant access into array body
1339 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1340 }
1341 }
1342 // Arrays of fixed size alias with arrays of unknown size.
1343 if (ta->size() != TypeInt::POS) {
1344 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1345 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1346 }
1347 // Arrays of known objects become arrays of unknown objects.
1348 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1349 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1350 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1351 }
1352 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1353 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1354 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1355 }
1356 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1357 // cannot be distinguished by bytecode alone.
1358 if (ta->elem() == TypeInt::BOOL) {
1359 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1360 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1361 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1362 }
1363 // During the 2nd round of IterGVN, NotNull castings are removed.
1364 // Make sure the Bottom and NotNull variants alias the same.
1365 // Also, make sure exact and non-exact variants alias the same.
1366 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1367 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1368 }
1369 }
1371 // Oop pointers need some flattening
1372 const TypeInstPtr *to = tj->isa_instptr();
1373 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1374 ciInstanceKlass *k = to->klass()->as_instance_klass();
1375 if( ptr == TypePtr::Constant ) {
1376 if (to->klass() != ciEnv::current()->Class_klass() ||
1377 offset < k->size_helper() * wordSize) {
1378 // No constant oop pointers (such as Strings); they alias with
1379 // unknown strings.
1380 assert(!is_known_inst, "not scalarizable allocation");
1381 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1382 }
1383 } else if( is_known_inst ) {
1384 tj = to; // Keep NotNull and klass_is_exact for instance type
1385 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1386 // During the 2nd round of IterGVN, NotNull castings are removed.
1387 // Make sure the Bottom and NotNull variants alias the same.
1388 // Also, make sure exact and non-exact variants alias the same.
1389 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1390 }
1391 if (to->speculative() != NULL) {
1392 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1393 }
1394 // Canonicalize the holder of this field
1395 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1396 // First handle header references such as a LoadKlassNode, even if the
1397 // object's klass is unloaded at compile time (4965979).
1398 if (!is_known_inst) { // Do it only for non-instance types
1399 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1400 }
1401 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1402 // Static fields are in the space above the normal instance
1403 // fields in the java.lang.Class instance.
1404 if (to->klass() != ciEnv::current()->Class_klass()) {
1405 to = NULL;
1406 tj = TypeOopPtr::BOTTOM;
1407 offset = tj->offset();
1408 }
1409 } else {
1410 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1411 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1412 if( is_known_inst ) {
1413 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1414 } else {
1415 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1416 }
1417 }
1418 }
1419 }
1421 // Klass pointers to object array klasses need some flattening
1422 const TypeKlassPtr *tk = tj->isa_klassptr();
1423 if( tk ) {
1424 // If we are referencing a field within a Klass, we need
1425 // to assume the worst case of an Object. Both exact and
1426 // inexact types must flatten to the same alias class so
1427 // use NotNull as the PTR.
1428 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1430 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1431 TypeKlassPtr::OBJECT->klass(),
1432 offset);
1433 }
1435 ciKlass* klass = tk->klass();
1436 if( klass->is_obj_array_klass() ) {
1437 ciKlass* k = TypeAryPtr::OOPS->klass();
1438 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1439 k = TypeInstPtr::BOTTOM->klass();
1440 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1441 }
1443 // Check for precise loads from the primary supertype array and force them
1444 // to the supertype cache alias index. Check for generic array loads from
1445 // the primary supertype array and also force them to the supertype cache
1446 // alias index. Since the same load can reach both, we need to merge
1447 // these 2 disparate memories into the same alias class. Since the
1448 // primary supertype array is read-only, there's no chance of confusion
1449 // where we bypass an array load and an array store.
1450 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1451 if (offset == Type::OffsetBot ||
1452 (offset >= primary_supers_offset &&
1453 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1454 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1455 offset = in_bytes(Klass::secondary_super_cache_offset());
1456 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1457 }
1458 }
1460 // Flatten all Raw pointers together.
1461 if (tj->base() == Type::RawPtr)
1462 tj = TypeRawPtr::BOTTOM;
1464 if (tj->base() == Type::AnyPtr)
1465 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1467 // Flatten all to bottom for now
1468 switch( _AliasLevel ) {
1469 case 0:
1470 tj = TypePtr::BOTTOM;
1471 break;
1472 case 1: // Flatten to: oop, static, field or array
1473 switch (tj->base()) {
1474 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1475 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1476 case Type::AryPtr: // do not distinguish arrays at all
1477 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1478 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1479 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1480 default: ShouldNotReachHere();
1481 }
1482 break;
1483 case 2: // No collapsing at level 2; keep all splits
1484 case 3: // No collapsing at level 3; keep all splits
1485 break;
1486 default:
1487 Unimplemented();
1488 }
1490 offset = tj->offset();
1491 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1493 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1494 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1495 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1496 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1497 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1498 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1499 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1500 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1501 assert( tj->ptr() != TypePtr::TopPTR &&
1502 tj->ptr() != TypePtr::AnyNull &&
1503 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1504 // assert( tj->ptr() != TypePtr::Constant ||
1505 // tj->base() == Type::RawPtr ||
1506 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1508 return tj;
1509 }
1511 void Compile::AliasType::Init(int i, const TypePtr* at) {
1512 _index = i;
1513 _adr_type = at;
1514 _field = NULL;
1515 _element = NULL;
1516 _is_rewritable = true; // default
1517 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1518 if (atoop != NULL && atoop->is_known_instance()) {
1519 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1520 _general_index = Compile::current()->get_alias_index(gt);
1521 } else {
1522 _general_index = 0;
1523 }
1524 }
1526 //---------------------------------print_on------------------------------------
1527 #ifndef PRODUCT
1528 void Compile::AliasType::print_on(outputStream* st) {
1529 if (index() < 10)
1530 st->print("@ <%d> ", index());
1531 else st->print("@ <%d>", index());
1532 st->print(is_rewritable() ? " " : " RO");
1533 int offset = adr_type()->offset();
1534 if (offset == Type::OffsetBot)
1535 st->print(" +any");
1536 else st->print(" +%-3d", offset);
1537 st->print(" in ");
1538 adr_type()->dump_on(st);
1539 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1540 if (field() != NULL && tjp) {
1541 if (tjp->klass() != field()->holder() ||
1542 tjp->offset() != field()->offset_in_bytes()) {
1543 st->print(" != ");
1544 field()->print();
1545 st->print(" ***");
1546 }
1547 }
1548 }
1550 void print_alias_types() {
1551 Compile* C = Compile::current();
1552 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1553 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1554 C->alias_type(idx)->print_on(tty);
1555 tty->cr();
1556 }
1557 }
1558 #endif
1561 //----------------------------probe_alias_cache--------------------------------
1562 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1563 intptr_t key = (intptr_t) adr_type;
1564 key ^= key >> logAliasCacheSize;
1565 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1566 }
1569 //-----------------------------grow_alias_types--------------------------------
1570 void Compile::grow_alias_types() {
1571 const int old_ats = _max_alias_types; // how many before?
1572 const int new_ats = old_ats; // how many more?
1573 const int grow_ats = old_ats+new_ats; // how many now?
1574 _max_alias_types = grow_ats;
1575 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1576 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1577 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1578 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1579 }
1582 //--------------------------------find_alias_type------------------------------
1583 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1584 if (_AliasLevel == 0)
1585 return alias_type(AliasIdxBot);
1587 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1588 if (ace->_adr_type == adr_type) {
1589 return alias_type(ace->_index);
1590 }
1592 // Handle special cases.
1593 if (adr_type == NULL) return alias_type(AliasIdxTop);
1594 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1596 // Do it the slow way.
1597 const TypePtr* flat = flatten_alias_type(adr_type);
1599 #ifdef ASSERT
1600 assert(flat == flatten_alias_type(flat), "idempotent");
1601 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1602 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1603 const TypeOopPtr* foop = flat->is_oopptr();
1604 // Scalarizable allocations have exact klass always.
1605 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1606 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1607 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1608 }
1609 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1610 #endif
1612 int idx = AliasIdxTop;
1613 for (int i = 0; i < num_alias_types(); i++) {
1614 if (alias_type(i)->adr_type() == flat) {
1615 idx = i;
1616 break;
1617 }
1618 }
1620 if (idx == AliasIdxTop) {
1621 if (no_create) return NULL;
1622 // Grow the array if necessary.
1623 if (_num_alias_types == _max_alias_types) grow_alias_types();
1624 // Add a new alias type.
1625 idx = _num_alias_types++;
1626 _alias_types[idx]->Init(idx, flat);
1627 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1628 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1629 if (flat->isa_instptr()) {
1630 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1631 && flat->is_instptr()->klass() == env()->Class_klass())
1632 alias_type(idx)->set_rewritable(false);
1633 }
1634 if (flat->isa_aryptr()) {
1635 #ifdef ASSERT
1636 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1637 // (T_BYTE has the weakest alignment and size restrictions...)
1638 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1639 #endif
1640 if (flat->offset() == TypePtr::OffsetBot) {
1641 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1642 }
1643 }
1644 if (flat->isa_klassptr()) {
1645 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1646 alias_type(idx)->set_rewritable(false);
1647 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1648 alias_type(idx)->set_rewritable(false);
1649 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1650 alias_type(idx)->set_rewritable(false);
1651 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1652 alias_type(idx)->set_rewritable(false);
1653 }
1654 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1655 // but the base pointer type is not distinctive enough to identify
1656 // references into JavaThread.)
1658 // Check for final fields.
1659 const TypeInstPtr* tinst = flat->isa_instptr();
1660 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1661 ciField* field;
1662 if (tinst->const_oop() != NULL &&
1663 tinst->klass() == ciEnv::current()->Class_klass() &&
1664 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1665 // static field
1666 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1667 field = k->get_field_by_offset(tinst->offset(), true);
1668 } else {
1669 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1670 field = k->get_field_by_offset(tinst->offset(), false);
1671 }
1672 assert(field == NULL ||
1673 original_field == NULL ||
1674 (field->holder() == original_field->holder() &&
1675 field->offset() == original_field->offset() &&
1676 field->is_static() == original_field->is_static()), "wrong field?");
1677 // Set field() and is_rewritable() attributes.
1678 if (field != NULL) alias_type(idx)->set_field(field);
1679 }
1680 }
1682 // Fill the cache for next time.
1683 ace->_adr_type = adr_type;
1684 ace->_index = idx;
1685 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1687 // Might as well try to fill the cache for the flattened version, too.
1688 AliasCacheEntry* face = probe_alias_cache(flat);
1689 if (face->_adr_type == NULL) {
1690 face->_adr_type = flat;
1691 face->_index = idx;
1692 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1693 }
1695 return alias_type(idx);
1696 }
1699 Compile::AliasType* Compile::alias_type(ciField* field) {
1700 const TypeOopPtr* t;
1701 if (field->is_static())
1702 t = TypeInstPtr::make(field->holder()->java_mirror());
1703 else
1704 t = TypeOopPtr::make_from_klass_raw(field->holder());
1705 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1706 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1707 return atp;
1708 }
1711 //------------------------------have_alias_type--------------------------------
1712 bool Compile::have_alias_type(const TypePtr* adr_type) {
1713 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1714 if (ace->_adr_type == adr_type) {
1715 return true;
1716 }
1718 // Handle special cases.
1719 if (adr_type == NULL) return true;
1720 if (adr_type == TypePtr::BOTTOM) return true;
1722 return find_alias_type(adr_type, true, NULL) != NULL;
1723 }
1725 //-----------------------------must_alias--------------------------------------
1726 // True if all values of the given address type are in the given alias category.
1727 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1728 if (alias_idx == AliasIdxBot) return true; // the universal category
1729 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1730 if (alias_idx == AliasIdxTop) return false; // the empty category
1731 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1733 // the only remaining possible overlap is identity
1734 int adr_idx = get_alias_index(adr_type);
1735 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1736 assert(adr_idx == alias_idx ||
1737 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1738 && adr_type != TypeOopPtr::BOTTOM),
1739 "should not be testing for overlap with an unsafe pointer");
1740 return adr_idx == alias_idx;
1741 }
1743 //------------------------------can_alias--------------------------------------
1744 // True if any values of the given address type are in the given alias category.
1745 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1746 if (alias_idx == AliasIdxTop) return false; // the empty category
1747 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1748 if (alias_idx == AliasIdxBot) return true; // the universal category
1749 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1751 // the only remaining possible overlap is identity
1752 int adr_idx = get_alias_index(adr_type);
1753 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1754 return adr_idx == alias_idx;
1755 }
1759 //---------------------------pop_warm_call-------------------------------------
1760 WarmCallInfo* Compile::pop_warm_call() {
1761 WarmCallInfo* wci = _warm_calls;
1762 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1763 return wci;
1764 }
1766 //----------------------------Inline_Warm--------------------------------------
1767 int Compile::Inline_Warm() {
1768 // If there is room, try to inline some more warm call sites.
1769 // %%% Do a graph index compaction pass when we think we're out of space?
1770 if (!InlineWarmCalls) return 0;
1772 int calls_made_hot = 0;
1773 int room_to_grow = NodeCountInliningCutoff - unique();
1774 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1775 int amount_grown = 0;
1776 WarmCallInfo* call;
1777 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1778 int est_size = (int)call->size();
1779 if (est_size > (room_to_grow - amount_grown)) {
1780 // This one won't fit anyway. Get rid of it.
1781 call->make_cold();
1782 continue;
1783 }
1784 call->make_hot();
1785 calls_made_hot++;
1786 amount_grown += est_size;
1787 amount_to_grow -= est_size;
1788 }
1790 if (calls_made_hot > 0) set_major_progress();
1791 return calls_made_hot;
1792 }
1795 //----------------------------Finish_Warm--------------------------------------
1796 void Compile::Finish_Warm() {
1797 if (!InlineWarmCalls) return;
1798 if (failing()) return;
1799 if (warm_calls() == NULL) return;
1801 // Clean up loose ends, if we are out of space for inlining.
1802 WarmCallInfo* call;
1803 while ((call = pop_warm_call()) != NULL) {
1804 call->make_cold();
1805 }
1806 }
1808 //---------------------cleanup_loop_predicates-----------------------
1809 // Remove the opaque nodes that protect the predicates so that all unused
1810 // checks and uncommon_traps will be eliminated from the ideal graph
1811 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1812 if (predicate_count()==0) return;
1813 for (int i = predicate_count(); i > 0; i--) {
1814 Node * n = predicate_opaque1_node(i-1);
1815 assert(n->Opcode() == Op_Opaque1, "must be");
1816 igvn.replace_node(n, n->in(1));
1817 }
1818 assert(predicate_count()==0, "should be clean!");
1819 }
1821 // StringOpts and late inlining of string methods
1822 void Compile::inline_string_calls(bool parse_time) {
1823 {
1824 // remove useless nodes to make the usage analysis simpler
1825 ResourceMark rm;
1826 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1827 }
1829 {
1830 ResourceMark rm;
1831 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1832 PhaseStringOpts pso(initial_gvn(), for_igvn());
1833 print_method(PHASE_AFTER_STRINGOPTS, 3);
1834 }
1836 // now inline anything that we skipped the first time around
1837 if (!parse_time) {
1838 _late_inlines_pos = _late_inlines.length();
1839 }
1841 while (_string_late_inlines.length() > 0) {
1842 CallGenerator* cg = _string_late_inlines.pop();
1843 cg->do_late_inline();
1844 if (failing()) return;
1845 }
1846 _string_late_inlines.trunc_to(0);
1847 }
1849 // Late inlining of boxing methods
1850 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1851 if (_boxing_late_inlines.length() > 0) {
1852 assert(has_boxed_value(), "inconsistent");
1854 PhaseGVN* gvn = initial_gvn();
1855 set_inlining_incrementally(true);
1857 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1858 for_igvn()->clear();
1859 gvn->replace_with(&igvn);
1861 while (_boxing_late_inlines.length() > 0) {
1862 CallGenerator* cg = _boxing_late_inlines.pop();
1863 cg->do_late_inline();
1864 if (failing()) return;
1865 }
1866 _boxing_late_inlines.trunc_to(0);
1868 {
1869 ResourceMark rm;
1870 PhaseRemoveUseless pru(gvn, for_igvn());
1871 }
1873 igvn = PhaseIterGVN(gvn);
1874 igvn.optimize();
1876 set_inlining_progress(false);
1877 set_inlining_incrementally(false);
1878 }
1879 }
1881 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1882 assert(IncrementalInline, "incremental inlining should be on");
1883 PhaseGVN* gvn = initial_gvn();
1885 set_inlining_progress(false);
1886 for_igvn()->clear();
1887 gvn->replace_with(&igvn);
1889 int i = 0;
1891 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1892 CallGenerator* cg = _late_inlines.at(i);
1893 _late_inlines_pos = i+1;
1894 cg->do_late_inline();
1895 if (failing()) return;
1896 }
1897 int j = 0;
1898 for (; i < _late_inlines.length(); i++, j++) {
1899 _late_inlines.at_put(j, _late_inlines.at(i));
1900 }
1901 _late_inlines.trunc_to(j);
1903 {
1904 ResourceMark rm;
1905 PhaseRemoveUseless pru(gvn, for_igvn());
1906 }
1908 igvn = PhaseIterGVN(gvn);
1909 }
1911 // Perform incremental inlining until bound on number of live nodes is reached
1912 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1913 PhaseGVN* gvn = initial_gvn();
1915 set_inlining_incrementally(true);
1916 set_inlining_progress(true);
1917 uint low_live_nodes = 0;
1919 while(inlining_progress() && _late_inlines.length() > 0) {
1921 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1922 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1923 // PhaseIdealLoop is expensive so we only try it once we are
1924 // out of loop and we only try it again if the previous helped
1925 // got the number of nodes down significantly
1926 PhaseIdealLoop ideal_loop( igvn, false, true );
1927 if (failing()) return;
1928 low_live_nodes = live_nodes();
1929 _major_progress = true;
1930 }
1932 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1933 break;
1934 }
1935 }
1937 inline_incrementally_one(igvn);
1939 if (failing()) return;
1941 igvn.optimize();
1943 if (failing()) return;
1944 }
1946 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1948 if (_string_late_inlines.length() > 0) {
1949 assert(has_stringbuilder(), "inconsistent");
1950 for_igvn()->clear();
1951 initial_gvn()->replace_with(&igvn);
1953 inline_string_calls(false);
1955 if (failing()) return;
1957 {
1958 ResourceMark rm;
1959 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1960 }
1962 igvn = PhaseIterGVN(gvn);
1964 igvn.optimize();
1965 }
1967 set_inlining_incrementally(false);
1968 }
1971 //------------------------------Optimize---------------------------------------
1972 // Given a graph, optimize it.
1973 void Compile::Optimize() {
1974 TracePhase t1("optimizer", &_t_optimizer, true);
1976 #ifndef PRODUCT
1977 if (env()->break_at_compile()) {
1978 BREAKPOINT;
1979 }
1981 #endif
1983 ResourceMark rm;
1984 int loop_opts_cnt;
1986 NOT_PRODUCT( verify_graph_edges(); )
1988 print_method(PHASE_AFTER_PARSING);
1990 {
1991 // Iterative Global Value Numbering, including ideal transforms
1992 // Initialize IterGVN with types and values from parse-time GVN
1993 PhaseIterGVN igvn(initial_gvn());
1994 {
1995 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1996 igvn.optimize();
1997 }
1999 print_method(PHASE_ITER_GVN1, 2);
2001 if (failing()) return;
2003 {
2004 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2005 inline_incrementally(igvn);
2006 }
2008 print_method(PHASE_INCREMENTAL_INLINE, 2);
2010 if (failing()) return;
2012 if (eliminate_boxing()) {
2013 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2014 // Inline valueOf() methods now.
2015 inline_boxing_calls(igvn);
2017 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2019 if (failing()) return;
2020 }
2022 // Remove the speculative part of types and clean up the graph from
2023 // the extra CastPP nodes whose only purpose is to carry them. Do
2024 // that early so that optimizations are not disrupted by the extra
2025 // CastPP nodes.
2026 remove_speculative_types(igvn);
2028 // No more new expensive nodes will be added to the list from here
2029 // so keep only the actual candidates for optimizations.
2030 cleanup_expensive_nodes(igvn);
2032 // Perform escape analysis
2033 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2034 if (has_loops()) {
2035 // Cleanup graph (remove dead nodes).
2036 TracePhase t2("idealLoop", &_t_idealLoop, true);
2037 PhaseIdealLoop ideal_loop( igvn, false, true );
2038 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2039 if (failing()) return;
2040 }
2041 ConnectionGraph::do_analysis(this, &igvn);
2043 if (failing()) return;
2045 // Optimize out fields loads from scalar replaceable allocations.
2046 igvn.optimize();
2047 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2049 if (failing()) return;
2051 if (congraph() != NULL && macro_count() > 0) {
2052 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2053 PhaseMacroExpand mexp(igvn);
2054 mexp.eliminate_macro_nodes();
2055 igvn.set_delay_transform(false);
2057 igvn.optimize();
2058 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2060 if (failing()) return;
2061 }
2062 }
2064 // Loop transforms on the ideal graph. Range Check Elimination,
2065 // peeling, unrolling, etc.
2067 // Set loop opts counter
2068 loop_opts_cnt = num_loop_opts();
2069 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2070 {
2071 TracePhase t2("idealLoop", &_t_idealLoop, true);
2072 PhaseIdealLoop ideal_loop( igvn, true );
2073 loop_opts_cnt--;
2074 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2075 if (failing()) return;
2076 }
2077 // Loop opts pass if partial peeling occurred in previous pass
2078 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2079 TracePhase t3("idealLoop", &_t_idealLoop, true);
2080 PhaseIdealLoop ideal_loop( igvn, false );
2081 loop_opts_cnt--;
2082 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2083 if (failing()) return;
2084 }
2085 // Loop opts pass for loop-unrolling before CCP
2086 if(major_progress() && (loop_opts_cnt > 0)) {
2087 TracePhase t4("idealLoop", &_t_idealLoop, true);
2088 PhaseIdealLoop ideal_loop( igvn, false );
2089 loop_opts_cnt--;
2090 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2091 }
2092 if (!failing()) {
2093 // Verify that last round of loop opts produced a valid graph
2094 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2095 PhaseIdealLoop::verify(igvn);
2096 }
2097 }
2098 if (failing()) return;
2100 // Conditional Constant Propagation;
2101 PhaseCCP ccp( &igvn );
2102 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2103 {
2104 TracePhase t2("ccp", &_t_ccp, true);
2105 ccp.do_transform();
2106 }
2107 print_method(PHASE_CPP1, 2);
2109 assert( true, "Break here to ccp.dump_old2new_map()");
2111 // Iterative Global Value Numbering, including ideal transforms
2112 {
2113 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2114 igvn = ccp;
2115 igvn.optimize();
2116 }
2118 print_method(PHASE_ITER_GVN2, 2);
2120 if (failing()) return;
2122 // Loop transforms on the ideal graph. Range Check Elimination,
2123 // peeling, unrolling, etc.
2124 if(loop_opts_cnt > 0) {
2125 debug_only( int cnt = 0; );
2126 while(major_progress() && (loop_opts_cnt > 0)) {
2127 TracePhase t2("idealLoop", &_t_idealLoop, true);
2128 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2129 PhaseIdealLoop ideal_loop( igvn, true);
2130 loop_opts_cnt--;
2131 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2132 if (failing()) return;
2133 }
2134 }
2136 {
2137 // Verify that all previous optimizations produced a valid graph
2138 // at least to this point, even if no loop optimizations were done.
2139 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2140 PhaseIdealLoop::verify(igvn);
2141 }
2143 {
2144 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2145 PhaseMacroExpand mex(igvn);
2146 if (mex.expand_macro_nodes()) {
2147 assert(failing(), "must bail out w/ explicit message");
2148 return;
2149 }
2150 }
2152 } // (End scope of igvn; run destructor if necessary for asserts.)
2154 dump_inlining();
2155 // A method with only infinite loops has no edges entering loops from root
2156 {
2157 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2158 if (final_graph_reshaping()) {
2159 assert(failing(), "must bail out w/ explicit message");
2160 return;
2161 }
2162 }
2164 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2165 }
2168 //------------------------------Code_Gen---------------------------------------
2169 // Given a graph, generate code for it
2170 void Compile::Code_Gen() {
2171 if (failing()) {
2172 return;
2173 }
2175 // Perform instruction selection. You might think we could reclaim Matcher
2176 // memory PDQ, but actually the Matcher is used in generating spill code.
2177 // Internals of the Matcher (including some VectorSets) must remain live
2178 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2179 // set a bit in reclaimed memory.
2181 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2182 // nodes. Mapping is only valid at the root of each matched subtree.
2183 NOT_PRODUCT( verify_graph_edges(); )
2185 Matcher matcher;
2186 _matcher = &matcher;
2187 {
2188 TracePhase t2("matcher", &_t_matcher, true);
2189 matcher.match();
2190 }
2191 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2192 // nodes. Mapping is only valid at the root of each matched subtree.
2193 NOT_PRODUCT( verify_graph_edges(); )
2195 // If you have too many nodes, or if matching has failed, bail out
2196 check_node_count(0, "out of nodes matching instructions");
2197 if (failing()) {
2198 return;
2199 }
2201 // Build a proper-looking CFG
2202 PhaseCFG cfg(node_arena(), root(), matcher);
2203 _cfg = &cfg;
2204 {
2205 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2206 bool success = cfg.do_global_code_motion();
2207 if (!success) {
2208 return;
2209 }
2211 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2212 NOT_PRODUCT( verify_graph_edges(); )
2213 debug_only( cfg.verify(); )
2214 }
2216 PhaseChaitin regalloc(unique(), cfg, matcher);
2217 _regalloc = ®alloc;
2218 {
2219 TracePhase t2("regalloc", &_t_registerAllocation, true);
2220 // Perform register allocation. After Chaitin, use-def chains are
2221 // no longer accurate (at spill code) and so must be ignored.
2222 // Node->LRG->reg mappings are still accurate.
2223 _regalloc->Register_Allocate();
2225 // Bail out if the allocator builds too many nodes
2226 if (failing()) {
2227 return;
2228 }
2229 }
2231 // Prior to register allocation we kept empty basic blocks in case the
2232 // the allocator needed a place to spill. After register allocation we
2233 // are not adding any new instructions. If any basic block is empty, we
2234 // can now safely remove it.
2235 {
2236 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2237 cfg.remove_empty_blocks();
2238 if (do_freq_based_layout()) {
2239 PhaseBlockLayout layout(cfg);
2240 } else {
2241 cfg.set_loop_alignment();
2242 }
2243 cfg.fixup_flow();
2244 }
2246 // Apply peephole optimizations
2247 if( OptoPeephole ) {
2248 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2249 PhasePeephole peep( _regalloc, cfg);
2250 peep.do_transform();
2251 }
2253 // Convert Nodes to instruction bits in a buffer
2254 {
2255 // %%%% workspace merge brought two timers together for one job
2256 TracePhase t2a("output", &_t_output, true);
2257 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2258 Output();
2259 }
2261 print_method(PHASE_FINAL_CODE);
2263 // He's dead, Jim.
2264 _cfg = (PhaseCFG*)0xdeadbeef;
2265 _regalloc = (PhaseChaitin*)0xdeadbeef;
2266 }
2269 //------------------------------dump_asm---------------------------------------
2270 // Dump formatted assembly
2271 #ifndef PRODUCT
2272 void Compile::dump_asm(int *pcs, uint pc_limit) {
2273 bool cut_short = false;
2274 tty->print_cr("#");
2275 tty->print("# "); _tf->dump(); tty->cr();
2276 tty->print_cr("#");
2278 // For all blocks
2279 int pc = 0x0; // Program counter
2280 char starts_bundle = ' ';
2281 _regalloc->dump_frame();
2283 Node *n = NULL;
2284 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2285 if (VMThread::should_terminate()) {
2286 cut_short = true;
2287 break;
2288 }
2289 Block* block = _cfg->get_block(i);
2290 if (block->is_connector() && !Verbose) {
2291 continue;
2292 }
2293 n = block->head();
2294 if (pcs && n->_idx < pc_limit) {
2295 tty->print("%3.3x ", pcs[n->_idx]);
2296 } else {
2297 tty->print(" ");
2298 }
2299 block->dump_head(_cfg);
2300 if (block->is_connector()) {
2301 tty->print_cr(" # Empty connector block");
2302 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2303 tty->print_cr(" # Block is sole successor of call");
2304 }
2306 // For all instructions
2307 Node *delay = NULL;
2308 for (uint j = 0; j < block->number_of_nodes(); j++) {
2309 if (VMThread::should_terminate()) {
2310 cut_short = true;
2311 break;
2312 }
2313 n = block->get_node(j);
2314 if (valid_bundle_info(n)) {
2315 Bundle* bundle = node_bundling(n);
2316 if (bundle->used_in_unconditional_delay()) {
2317 delay = n;
2318 continue;
2319 }
2320 if (bundle->starts_bundle()) {
2321 starts_bundle = '+';
2322 }
2323 }
2325 if (WizardMode) {
2326 n->dump();
2327 }
2329 if( !n->is_Region() && // Dont print in the Assembly
2330 !n->is_Phi() && // a few noisely useless nodes
2331 !n->is_Proj() &&
2332 !n->is_MachTemp() &&
2333 !n->is_SafePointScalarObject() &&
2334 !n->is_Catch() && // Would be nice to print exception table targets
2335 !n->is_MergeMem() && // Not very interesting
2336 !n->is_top() && // Debug info table constants
2337 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2338 ) {
2339 if (pcs && n->_idx < pc_limit)
2340 tty->print("%3.3x", pcs[n->_idx]);
2341 else
2342 tty->print(" ");
2343 tty->print(" %c ", starts_bundle);
2344 starts_bundle = ' ';
2345 tty->print("\t");
2346 n->format(_regalloc, tty);
2347 tty->cr();
2348 }
2350 // If we have an instruction with a delay slot, and have seen a delay,
2351 // then back up and print it
2352 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2353 assert(delay != NULL, "no unconditional delay instruction");
2354 if (WizardMode) delay->dump();
2356 if (node_bundling(delay)->starts_bundle())
2357 starts_bundle = '+';
2358 if (pcs && n->_idx < pc_limit)
2359 tty->print("%3.3x", pcs[n->_idx]);
2360 else
2361 tty->print(" ");
2362 tty->print(" %c ", starts_bundle);
2363 starts_bundle = ' ';
2364 tty->print("\t");
2365 delay->format(_regalloc, tty);
2366 tty->print_cr("");
2367 delay = NULL;
2368 }
2370 // Dump the exception table as well
2371 if( n->is_Catch() && (Verbose || WizardMode) ) {
2372 // Print the exception table for this offset
2373 _handler_table.print_subtable_for(pc);
2374 }
2375 }
2377 if (pcs && n->_idx < pc_limit)
2378 tty->print_cr("%3.3x", pcs[n->_idx]);
2379 else
2380 tty->print_cr("");
2382 assert(cut_short || delay == NULL, "no unconditional delay branch");
2384 } // End of per-block dump
2385 tty->print_cr("");
2387 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2388 }
2389 #endif
2391 //------------------------------Final_Reshape_Counts---------------------------
2392 // This class defines counters to help identify when a method
2393 // may/must be executed using hardware with only 24-bit precision.
2394 struct Final_Reshape_Counts : public StackObj {
2395 int _call_count; // count non-inlined 'common' calls
2396 int _float_count; // count float ops requiring 24-bit precision
2397 int _double_count; // count double ops requiring more precision
2398 int _java_call_count; // count non-inlined 'java' calls
2399 int _inner_loop_count; // count loops which need alignment
2400 VectorSet _visited; // Visitation flags
2401 Node_List _tests; // Set of IfNodes & PCTableNodes
2403 Final_Reshape_Counts() :
2404 _call_count(0), _float_count(0), _double_count(0),
2405 _java_call_count(0), _inner_loop_count(0),
2406 _visited( Thread::current()->resource_area() ) { }
2408 void inc_call_count () { _call_count ++; }
2409 void inc_float_count () { _float_count ++; }
2410 void inc_double_count() { _double_count++; }
2411 void inc_java_call_count() { _java_call_count++; }
2412 void inc_inner_loop_count() { _inner_loop_count++; }
2414 int get_call_count () const { return _call_count ; }
2415 int get_float_count () const { return _float_count ; }
2416 int get_double_count() const { return _double_count; }
2417 int get_java_call_count() const { return _java_call_count; }
2418 int get_inner_loop_count() const { return _inner_loop_count; }
2419 };
2421 #ifdef ASSERT
2422 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2423 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2424 // Make sure the offset goes inside the instance layout.
2425 return k->contains_field_offset(tp->offset());
2426 // Note that OffsetBot and OffsetTop are very negative.
2427 }
2428 #endif
2430 // Eliminate trivially redundant StoreCMs and accumulate their
2431 // precedence edges.
2432 void Compile::eliminate_redundant_card_marks(Node* n) {
2433 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2434 if (n->in(MemNode::Address)->outcnt() > 1) {
2435 // There are multiple users of the same address so it might be
2436 // possible to eliminate some of the StoreCMs
2437 Node* mem = n->in(MemNode::Memory);
2438 Node* adr = n->in(MemNode::Address);
2439 Node* val = n->in(MemNode::ValueIn);
2440 Node* prev = n;
2441 bool done = false;
2442 // Walk the chain of StoreCMs eliminating ones that match. As
2443 // long as it's a chain of single users then the optimization is
2444 // safe. Eliminating partially redundant StoreCMs would require
2445 // cloning copies down the other paths.
2446 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2447 if (adr == mem->in(MemNode::Address) &&
2448 val == mem->in(MemNode::ValueIn)) {
2449 // redundant StoreCM
2450 if (mem->req() > MemNode::OopStore) {
2451 // Hasn't been processed by this code yet.
2452 n->add_prec(mem->in(MemNode::OopStore));
2453 } else {
2454 // Already converted to precedence edge
2455 for (uint i = mem->req(); i < mem->len(); i++) {
2456 // Accumulate any precedence edges
2457 if (mem->in(i) != NULL) {
2458 n->add_prec(mem->in(i));
2459 }
2460 }
2461 // Everything above this point has been processed.
2462 done = true;
2463 }
2464 // Eliminate the previous StoreCM
2465 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2466 assert(mem->outcnt() == 0, "should be dead");
2467 mem->disconnect_inputs(NULL, this);
2468 } else {
2469 prev = mem;
2470 }
2471 mem = prev->in(MemNode::Memory);
2472 }
2473 }
2474 }
2476 //------------------------------final_graph_reshaping_impl----------------------
2477 // Implement items 1-5 from final_graph_reshaping below.
2478 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2480 if ( n->outcnt() == 0 ) return; // dead node
2481 uint nop = n->Opcode();
2483 // Check for 2-input instruction with "last use" on right input.
2484 // Swap to left input. Implements item (2).
2485 if( n->req() == 3 && // two-input instruction
2486 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2487 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2488 n->in(2)->outcnt() == 1 &&// right use IS a last use
2489 !n->in(2)->is_Con() ) { // right use is not a constant
2490 // Check for commutative opcode
2491 switch( nop ) {
2492 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2493 case Op_MaxI: case Op_MinI:
2494 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2495 case Op_AndL: case Op_XorL: case Op_OrL:
2496 case Op_AndI: case Op_XorI: case Op_OrI: {
2497 // Move "last use" input to left by swapping inputs
2498 n->swap_edges(1, 2);
2499 break;
2500 }
2501 default:
2502 break;
2503 }
2504 }
2506 #ifdef ASSERT
2507 if( n->is_Mem() ) {
2508 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2509 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2510 // oop will be recorded in oop map if load crosses safepoint
2511 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2512 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2513 "raw memory operations should have control edge");
2514 }
2515 #endif
2516 // Count FPU ops and common calls, implements item (3)
2517 switch( nop ) {
2518 // Count all float operations that may use FPU
2519 case Op_AddF:
2520 case Op_SubF:
2521 case Op_MulF:
2522 case Op_DivF:
2523 case Op_NegF:
2524 case Op_ModF:
2525 case Op_ConvI2F:
2526 case Op_ConF:
2527 case Op_CmpF:
2528 case Op_CmpF3:
2529 // case Op_ConvL2F: // longs are split into 32-bit halves
2530 frc.inc_float_count();
2531 break;
2533 case Op_ConvF2D:
2534 case Op_ConvD2F:
2535 frc.inc_float_count();
2536 frc.inc_double_count();
2537 break;
2539 // Count all double operations that may use FPU
2540 case Op_AddD:
2541 case Op_SubD:
2542 case Op_MulD:
2543 case Op_DivD:
2544 case Op_NegD:
2545 case Op_ModD:
2546 case Op_ConvI2D:
2547 case Op_ConvD2I:
2548 // case Op_ConvL2D: // handled by leaf call
2549 // case Op_ConvD2L: // handled by leaf call
2550 case Op_ConD:
2551 case Op_CmpD:
2552 case Op_CmpD3:
2553 frc.inc_double_count();
2554 break;
2555 case Op_Opaque1: // Remove Opaque Nodes before matching
2556 case Op_Opaque2: // Remove Opaque Nodes before matching
2557 n->subsume_by(n->in(1), this);
2558 break;
2559 case Op_CallStaticJava:
2560 case Op_CallJava:
2561 case Op_CallDynamicJava:
2562 frc.inc_java_call_count(); // Count java call site;
2563 case Op_CallRuntime:
2564 case Op_CallLeaf:
2565 case Op_CallLeafNoFP: {
2566 assert( n->is_Call(), "" );
2567 CallNode *call = n->as_Call();
2568 // Count call sites where the FP mode bit would have to be flipped.
2569 // Do not count uncommon runtime calls:
2570 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2571 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2572 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2573 frc.inc_call_count(); // Count the call site
2574 } else { // See if uncommon argument is shared
2575 Node *n = call->in(TypeFunc::Parms);
2576 int nop = n->Opcode();
2577 // Clone shared simple arguments to uncommon calls, item (1).
2578 if( n->outcnt() > 1 &&
2579 !n->is_Proj() &&
2580 nop != Op_CreateEx &&
2581 nop != Op_CheckCastPP &&
2582 nop != Op_DecodeN &&
2583 nop != Op_DecodeNKlass &&
2584 !n->is_Mem() ) {
2585 Node *x = n->clone();
2586 call->set_req( TypeFunc::Parms, x );
2587 }
2588 }
2589 break;
2590 }
2592 case Op_StoreD:
2593 case Op_LoadD:
2594 case Op_LoadD_unaligned:
2595 frc.inc_double_count();
2596 goto handle_mem;
2597 case Op_StoreF:
2598 case Op_LoadF:
2599 frc.inc_float_count();
2600 goto handle_mem;
2602 case Op_StoreCM:
2603 {
2604 // Convert OopStore dependence into precedence edge
2605 Node* prec = n->in(MemNode::OopStore);
2606 n->del_req(MemNode::OopStore);
2607 n->add_prec(prec);
2608 eliminate_redundant_card_marks(n);
2609 }
2611 // fall through
2613 case Op_StoreB:
2614 case Op_StoreC:
2615 case Op_StorePConditional:
2616 case Op_StoreI:
2617 case Op_StoreL:
2618 case Op_StoreIConditional:
2619 case Op_StoreLConditional:
2620 case Op_CompareAndSwapI:
2621 case Op_CompareAndSwapL:
2622 case Op_CompareAndSwapP:
2623 case Op_CompareAndSwapN:
2624 case Op_GetAndAddI:
2625 case Op_GetAndAddL:
2626 case Op_GetAndSetI:
2627 case Op_GetAndSetL:
2628 case Op_GetAndSetP:
2629 case Op_GetAndSetN:
2630 case Op_StoreP:
2631 case Op_StoreN:
2632 case Op_StoreNKlass:
2633 case Op_LoadB:
2634 case Op_LoadUB:
2635 case Op_LoadUS:
2636 case Op_LoadI:
2637 case Op_LoadKlass:
2638 case Op_LoadNKlass:
2639 case Op_LoadL:
2640 case Op_LoadL_unaligned:
2641 case Op_LoadPLocked:
2642 case Op_LoadP:
2643 case Op_LoadN:
2644 case Op_LoadRange:
2645 case Op_LoadS: {
2646 handle_mem:
2647 #ifdef ASSERT
2648 if( VerifyOptoOopOffsets ) {
2649 assert( n->is_Mem(), "" );
2650 MemNode *mem = (MemNode*)n;
2651 // Check to see if address types have grounded out somehow.
2652 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2653 assert( !tp || oop_offset_is_sane(tp), "" );
2654 }
2655 #endif
2656 break;
2657 }
2659 case Op_AddP: { // Assert sane base pointers
2660 Node *addp = n->in(AddPNode::Address);
2661 assert( !addp->is_AddP() ||
2662 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2663 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2664 "Base pointers must match" );
2665 #ifdef _LP64
2666 if ((UseCompressedOops || UseCompressedClassPointers) &&
2667 addp->Opcode() == Op_ConP &&
2668 addp == n->in(AddPNode::Base) &&
2669 n->in(AddPNode::Offset)->is_Con()) {
2670 // Use addressing with narrow klass to load with offset on x86.
2671 // On sparc loading 32-bits constant and decoding it have less
2672 // instructions (4) then load 64-bits constant (7).
2673 // Do this transformation here since IGVN will convert ConN back to ConP.
2674 const Type* t = addp->bottom_type();
2675 if (t->isa_oopptr() || t->isa_klassptr()) {
2676 Node* nn = NULL;
2678 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2680 // Look for existing ConN node of the same exact type.
2681 Node* r = root();
2682 uint cnt = r->outcnt();
2683 for (uint i = 0; i < cnt; i++) {
2684 Node* m = r->raw_out(i);
2685 if (m!= NULL && m->Opcode() == op &&
2686 m->bottom_type()->make_ptr() == t) {
2687 nn = m;
2688 break;
2689 }
2690 }
2691 if (nn != NULL) {
2692 // Decode a narrow oop to match address
2693 // [R12 + narrow_oop_reg<<3 + offset]
2694 if (t->isa_oopptr()) {
2695 nn = new (this) DecodeNNode(nn, t);
2696 } else {
2697 nn = new (this) DecodeNKlassNode(nn, t);
2698 }
2699 n->set_req(AddPNode::Base, nn);
2700 n->set_req(AddPNode::Address, nn);
2701 if (addp->outcnt() == 0) {
2702 addp->disconnect_inputs(NULL, this);
2703 }
2704 }
2705 }
2706 }
2707 #endif
2708 break;
2709 }
2711 #ifdef _LP64
2712 case Op_CastPP:
2713 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2714 Node* in1 = n->in(1);
2715 const Type* t = n->bottom_type();
2716 Node* new_in1 = in1->clone();
2717 new_in1->as_DecodeN()->set_type(t);
2719 if (!Matcher::narrow_oop_use_complex_address()) {
2720 //
2721 // x86, ARM and friends can handle 2 adds in addressing mode
2722 // and Matcher can fold a DecodeN node into address by using
2723 // a narrow oop directly and do implicit NULL check in address:
2724 //
2725 // [R12 + narrow_oop_reg<<3 + offset]
2726 // NullCheck narrow_oop_reg
2727 //
2728 // On other platforms (Sparc) we have to keep new DecodeN node and
2729 // use it to do implicit NULL check in address:
2730 //
2731 // decode_not_null narrow_oop_reg, base_reg
2732 // [base_reg + offset]
2733 // NullCheck base_reg
2734 //
2735 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2736 // to keep the information to which NULL check the new DecodeN node
2737 // corresponds to use it as value in implicit_null_check().
2738 //
2739 new_in1->set_req(0, n->in(0));
2740 }
2742 n->subsume_by(new_in1, this);
2743 if (in1->outcnt() == 0) {
2744 in1->disconnect_inputs(NULL, this);
2745 }
2746 }
2747 break;
2749 case Op_CmpP:
2750 // Do this transformation here to preserve CmpPNode::sub() and
2751 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2752 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2753 Node* in1 = n->in(1);
2754 Node* in2 = n->in(2);
2755 if (!in1->is_DecodeNarrowPtr()) {
2756 in2 = in1;
2757 in1 = n->in(2);
2758 }
2759 assert(in1->is_DecodeNarrowPtr(), "sanity");
2761 Node* new_in2 = NULL;
2762 if (in2->is_DecodeNarrowPtr()) {
2763 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2764 new_in2 = in2->in(1);
2765 } else if (in2->Opcode() == Op_ConP) {
2766 const Type* t = in2->bottom_type();
2767 if (t == TypePtr::NULL_PTR) {
2768 assert(in1->is_DecodeN(), "compare klass to null?");
2769 // Don't convert CmpP null check into CmpN if compressed
2770 // oops implicit null check is not generated.
2771 // This will allow to generate normal oop implicit null check.
2772 if (Matcher::gen_narrow_oop_implicit_null_checks())
2773 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2774 //
2775 // This transformation together with CastPP transformation above
2776 // will generated code for implicit NULL checks for compressed oops.
2777 //
2778 // The original code after Optimize()
2779 //
2780 // LoadN memory, narrow_oop_reg
2781 // decode narrow_oop_reg, base_reg
2782 // CmpP base_reg, NULL
2783 // CastPP base_reg // NotNull
2784 // Load [base_reg + offset], val_reg
2785 //
2786 // after these transformations will be
2787 //
2788 // LoadN memory, narrow_oop_reg
2789 // CmpN narrow_oop_reg, NULL
2790 // decode_not_null narrow_oop_reg, base_reg
2791 // Load [base_reg + offset], val_reg
2792 //
2793 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2794 // since narrow oops can be used in debug info now (see the code in
2795 // final_graph_reshaping_walk()).
2796 //
2797 // At the end the code will be matched to
2798 // on x86:
2799 //
2800 // Load_narrow_oop memory, narrow_oop_reg
2801 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2802 // NullCheck narrow_oop_reg
2803 //
2804 // and on sparc:
2805 //
2806 // Load_narrow_oop memory, narrow_oop_reg
2807 // decode_not_null narrow_oop_reg, base_reg
2808 // Load [base_reg + offset], val_reg
2809 // NullCheck base_reg
2810 //
2811 } else if (t->isa_oopptr()) {
2812 new_in2 = ConNode::make(this, t->make_narrowoop());
2813 } else if (t->isa_klassptr()) {
2814 new_in2 = ConNode::make(this, t->make_narrowklass());
2815 }
2816 }
2817 if (new_in2 != NULL) {
2818 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2819 n->subsume_by(cmpN, this);
2820 if (in1->outcnt() == 0) {
2821 in1->disconnect_inputs(NULL, this);
2822 }
2823 if (in2->outcnt() == 0) {
2824 in2->disconnect_inputs(NULL, this);
2825 }
2826 }
2827 }
2828 break;
2830 case Op_DecodeN:
2831 case Op_DecodeNKlass:
2832 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2833 // DecodeN could be pinned when it can't be fold into
2834 // an address expression, see the code for Op_CastPP above.
2835 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2836 break;
2838 case Op_EncodeP:
2839 case Op_EncodePKlass: {
2840 Node* in1 = n->in(1);
2841 if (in1->is_DecodeNarrowPtr()) {
2842 n->subsume_by(in1->in(1), this);
2843 } else if (in1->Opcode() == Op_ConP) {
2844 const Type* t = in1->bottom_type();
2845 if (t == TypePtr::NULL_PTR) {
2846 assert(t->isa_oopptr(), "null klass?");
2847 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2848 } else if (t->isa_oopptr()) {
2849 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2850 } else if (t->isa_klassptr()) {
2851 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2852 }
2853 }
2854 if (in1->outcnt() == 0) {
2855 in1->disconnect_inputs(NULL, this);
2856 }
2857 break;
2858 }
2860 case Op_Proj: {
2861 if (OptimizeStringConcat) {
2862 ProjNode* p = n->as_Proj();
2863 if (p->_is_io_use) {
2864 // Separate projections were used for the exception path which
2865 // are normally removed by a late inline. If it wasn't inlined
2866 // then they will hang around and should just be replaced with
2867 // the original one.
2868 Node* proj = NULL;
2869 // Replace with just one
2870 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2871 Node *use = i.get();
2872 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2873 proj = use;
2874 break;
2875 }
2876 }
2877 assert(proj != NULL, "must be found");
2878 p->subsume_by(proj, this);
2879 }
2880 }
2881 break;
2882 }
2884 case Op_Phi:
2885 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
2886 // The EncodeP optimization may create Phi with the same edges
2887 // for all paths. It is not handled well by Register Allocator.
2888 Node* unique_in = n->in(1);
2889 assert(unique_in != NULL, "");
2890 uint cnt = n->req();
2891 for (uint i = 2; i < cnt; i++) {
2892 Node* m = n->in(i);
2893 assert(m != NULL, "");
2894 if (unique_in != m)
2895 unique_in = NULL;
2896 }
2897 if (unique_in != NULL) {
2898 n->subsume_by(unique_in, this);
2899 }
2900 }
2901 break;
2903 #endif
2905 case Op_ModI:
2906 if (UseDivMod) {
2907 // Check if a%b and a/b both exist
2908 Node* d = n->find_similar(Op_DivI);
2909 if (d) {
2910 // Replace them with a fused divmod if supported
2911 if (Matcher::has_match_rule(Op_DivModI)) {
2912 DivModINode* divmod = DivModINode::make(this, n);
2913 d->subsume_by(divmod->div_proj(), this);
2914 n->subsume_by(divmod->mod_proj(), this);
2915 } else {
2916 // replace a%b with a-((a/b)*b)
2917 Node* mult = new (this) MulINode(d, d->in(2));
2918 Node* sub = new (this) SubINode(d->in(1), mult);
2919 n->subsume_by(sub, this);
2920 }
2921 }
2922 }
2923 break;
2925 case Op_ModL:
2926 if (UseDivMod) {
2927 // Check if a%b and a/b both exist
2928 Node* d = n->find_similar(Op_DivL);
2929 if (d) {
2930 // Replace them with a fused divmod if supported
2931 if (Matcher::has_match_rule(Op_DivModL)) {
2932 DivModLNode* divmod = DivModLNode::make(this, n);
2933 d->subsume_by(divmod->div_proj(), this);
2934 n->subsume_by(divmod->mod_proj(), this);
2935 } else {
2936 // replace a%b with a-((a/b)*b)
2937 Node* mult = new (this) MulLNode(d, d->in(2));
2938 Node* sub = new (this) SubLNode(d->in(1), mult);
2939 n->subsume_by(sub, this);
2940 }
2941 }
2942 }
2943 break;
2945 case Op_LoadVector:
2946 case Op_StoreVector:
2947 break;
2949 case Op_PackB:
2950 case Op_PackS:
2951 case Op_PackI:
2952 case Op_PackF:
2953 case Op_PackL:
2954 case Op_PackD:
2955 if (n->req()-1 > 2) {
2956 // Replace many operand PackNodes with a binary tree for matching
2957 PackNode* p = (PackNode*) n;
2958 Node* btp = p->binary_tree_pack(this, 1, n->req());
2959 n->subsume_by(btp, this);
2960 }
2961 break;
2962 case Op_Loop:
2963 case Op_CountedLoop:
2964 if (n->as_Loop()->is_inner_loop()) {
2965 frc.inc_inner_loop_count();
2966 }
2967 break;
2968 case Op_LShiftI:
2969 case Op_RShiftI:
2970 case Op_URShiftI:
2971 case Op_LShiftL:
2972 case Op_RShiftL:
2973 case Op_URShiftL:
2974 if (Matcher::need_masked_shift_count) {
2975 // The cpu's shift instructions don't restrict the count to the
2976 // lower 5/6 bits. We need to do the masking ourselves.
2977 Node* in2 = n->in(2);
2978 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2979 const TypeInt* t = in2->find_int_type();
2980 if (t != NULL && t->is_con()) {
2981 juint shift = t->get_con();
2982 if (shift > mask) { // Unsigned cmp
2983 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
2984 }
2985 } else {
2986 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2987 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
2988 n->set_req(2, shift);
2989 }
2990 }
2991 if (in2->outcnt() == 0) { // Remove dead node
2992 in2->disconnect_inputs(NULL, this);
2993 }
2994 }
2995 break;
2996 case Op_MemBarStoreStore:
2997 case Op_MemBarRelease:
2998 // Break the link with AllocateNode: it is no longer useful and
2999 // confuses register allocation.
3000 if (n->req() > MemBarNode::Precedent) {
3001 n->set_req(MemBarNode::Precedent, top());
3002 }
3003 break;
3004 // Must set a control edge on all nodes that produce a FlagsProj
3005 // so they can't escape the block that consumes the flags.
3006 // Must also set the non throwing branch as the control
3007 // for all nodes that depends on the result. Unless the node
3008 // already have a control that isn't the control of the
3009 // flag producer
3010 case Op_FlagsProj:
3011 {
3012 MathExactNode* math = (MathExactNode*) n->in(0);
3013 Node* ctrl = math->control_node();
3014 Node* non_throwing = math->non_throwing_branch();
3015 math->set_req(0, ctrl);
3017 Node* result = math->result_node();
3018 if (result != NULL) {
3019 for (DUIterator_Fast jmax, j = result->fast_outs(jmax); j < jmax; j++) {
3020 Node* out = result->fast_out(j);
3021 // Phi nodes shouldn't be moved. They would only match below if they
3022 // had the same control as the MathExactNode. The only time that
3023 // would happen is if the Phi is also an input to the MathExact
3024 if (!out->is_Phi()) {
3025 if (out->in(0) == NULL) {
3026 out->set_req(0, non_throwing);
3027 } else if (out->in(0) == ctrl) {
3028 out->set_req(0, non_throwing);
3029 }
3030 }
3031 }
3032 }
3033 }
3034 break;
3035 default:
3036 assert( !n->is_Call(), "" );
3037 assert( !n->is_Mem(), "" );
3038 break;
3039 }
3041 // Collect CFG split points
3042 if (n->is_MultiBranch())
3043 frc._tests.push(n);
3044 }
3046 //------------------------------final_graph_reshaping_walk---------------------
3047 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3048 // requires that the walk visits a node's inputs before visiting the node.
3049 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3050 ResourceArea *area = Thread::current()->resource_area();
3051 Unique_Node_List sfpt(area);
3053 frc._visited.set(root->_idx); // first, mark node as visited
3054 uint cnt = root->req();
3055 Node *n = root;
3056 uint i = 0;
3057 while (true) {
3058 if (i < cnt) {
3059 // Place all non-visited non-null inputs onto stack
3060 Node* m = n->in(i);
3061 ++i;
3062 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3063 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
3064 sfpt.push(m);
3065 cnt = m->req();
3066 nstack.push(n, i); // put on stack parent and next input's index
3067 n = m;
3068 i = 0;
3069 }
3070 } else {
3071 // Now do post-visit work
3072 final_graph_reshaping_impl( n, frc );
3073 if (nstack.is_empty())
3074 break; // finished
3075 n = nstack.node(); // Get node from stack
3076 cnt = n->req();
3077 i = nstack.index();
3078 nstack.pop(); // Shift to the next node on stack
3079 }
3080 }
3082 // Skip next transformation if compressed oops are not used.
3083 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3084 (!UseCompressedOops && !UseCompressedClassPointers))
3085 return;
3087 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3088 // It could be done for an uncommon traps or any safepoints/calls
3089 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3090 while (sfpt.size() > 0) {
3091 n = sfpt.pop();
3092 JVMState *jvms = n->as_SafePoint()->jvms();
3093 assert(jvms != NULL, "sanity");
3094 int start = jvms->debug_start();
3095 int end = n->req();
3096 bool is_uncommon = (n->is_CallStaticJava() &&
3097 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3098 for (int j = start; j < end; j++) {
3099 Node* in = n->in(j);
3100 if (in->is_DecodeNarrowPtr()) {
3101 bool safe_to_skip = true;
3102 if (!is_uncommon ) {
3103 // Is it safe to skip?
3104 for (uint i = 0; i < in->outcnt(); i++) {
3105 Node* u = in->raw_out(i);
3106 if (!u->is_SafePoint() ||
3107 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3108 safe_to_skip = false;
3109 }
3110 }
3111 }
3112 if (safe_to_skip) {
3113 n->set_req(j, in->in(1));
3114 }
3115 if (in->outcnt() == 0) {
3116 in->disconnect_inputs(NULL, this);
3117 }
3118 }
3119 }
3120 }
3121 }
3123 //------------------------------final_graph_reshaping--------------------------
3124 // Final Graph Reshaping.
3125 //
3126 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3127 // and not commoned up and forced early. Must come after regular
3128 // optimizations to avoid GVN undoing the cloning. Clone constant
3129 // inputs to Loop Phis; these will be split by the allocator anyways.
3130 // Remove Opaque nodes.
3131 // (2) Move last-uses by commutative operations to the left input to encourage
3132 // Intel update-in-place two-address operations and better register usage
3133 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3134 // calls canonicalizing them back.
3135 // (3) Count the number of double-precision FP ops, single-precision FP ops
3136 // and call sites. On Intel, we can get correct rounding either by
3137 // forcing singles to memory (requires extra stores and loads after each
3138 // FP bytecode) or we can set a rounding mode bit (requires setting and
3139 // clearing the mode bit around call sites). The mode bit is only used
3140 // if the relative frequency of single FP ops to calls is low enough.
3141 // This is a key transform for SPEC mpeg_audio.
3142 // (4) Detect infinite loops; blobs of code reachable from above but not
3143 // below. Several of the Code_Gen algorithms fail on such code shapes,
3144 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3145 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3146 // Detection is by looking for IfNodes where only 1 projection is
3147 // reachable from below or CatchNodes missing some targets.
3148 // (5) Assert for insane oop offsets in debug mode.
3150 bool Compile::final_graph_reshaping() {
3151 // an infinite loop may have been eliminated by the optimizer,
3152 // in which case the graph will be empty.
3153 if (root()->req() == 1) {
3154 record_method_not_compilable("trivial infinite loop");
3155 return true;
3156 }
3158 // Expensive nodes have their control input set to prevent the GVN
3159 // from freely commoning them. There's no GVN beyond this point so
3160 // no need to keep the control input. We want the expensive nodes to
3161 // be freely moved to the least frequent code path by gcm.
3162 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3163 for (int i = 0; i < expensive_count(); i++) {
3164 _expensive_nodes->at(i)->set_req(0, NULL);
3165 }
3167 Final_Reshape_Counts frc;
3169 // Visit everybody reachable!
3170 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3171 Node_Stack nstack(unique() >> 1);
3172 final_graph_reshaping_walk(nstack, root(), frc);
3174 // Check for unreachable (from below) code (i.e., infinite loops).
3175 for( uint i = 0; i < frc._tests.size(); i++ ) {
3176 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3177 // Get number of CFG targets.
3178 // Note that PCTables include exception targets after calls.
3179 uint required_outcnt = n->required_outcnt();
3180 if (n->outcnt() != required_outcnt) {
3181 // Check for a few special cases. Rethrow Nodes never take the
3182 // 'fall-thru' path, so expected kids is 1 less.
3183 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3184 if (n->in(0)->in(0)->is_Call()) {
3185 CallNode *call = n->in(0)->in(0)->as_Call();
3186 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3187 required_outcnt--; // Rethrow always has 1 less kid
3188 } else if (call->req() > TypeFunc::Parms &&
3189 call->is_CallDynamicJava()) {
3190 // Check for null receiver. In such case, the optimizer has
3191 // detected that the virtual call will always result in a null
3192 // pointer exception. The fall-through projection of this CatchNode
3193 // will not be populated.
3194 Node *arg0 = call->in(TypeFunc::Parms);
3195 if (arg0->is_Type() &&
3196 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3197 required_outcnt--;
3198 }
3199 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3200 call->req() > TypeFunc::Parms+1 &&
3201 call->is_CallStaticJava()) {
3202 // Check for negative array length. In such case, the optimizer has
3203 // detected that the allocation attempt will always result in an
3204 // exception. There is no fall-through projection of this CatchNode .
3205 Node *arg1 = call->in(TypeFunc::Parms+1);
3206 if (arg1->is_Type() &&
3207 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3208 required_outcnt--;
3209 }
3210 }
3211 }
3212 }
3213 // Recheck with a better notion of 'required_outcnt'
3214 if (n->outcnt() != required_outcnt) {
3215 record_method_not_compilable("malformed control flow");
3216 return true; // Not all targets reachable!
3217 }
3218 }
3219 // Check that I actually visited all kids. Unreached kids
3220 // must be infinite loops.
3221 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3222 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3223 record_method_not_compilable("infinite loop");
3224 return true; // Found unvisited kid; must be unreach
3225 }
3226 }
3228 // If original bytecodes contained a mixture of floats and doubles
3229 // check if the optimizer has made it homogenous, item (3).
3230 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3231 frc.get_float_count() > 32 &&
3232 frc.get_double_count() == 0 &&
3233 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3234 set_24_bit_selection_and_mode( false, true );
3235 }
3237 set_java_calls(frc.get_java_call_count());
3238 set_inner_loops(frc.get_inner_loop_count());
3240 // No infinite loops, no reason to bail out.
3241 return false;
3242 }
3244 //-----------------------------too_many_traps----------------------------------
3245 // Report if there are too many traps at the current method and bci.
3246 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3247 bool Compile::too_many_traps(ciMethod* method,
3248 int bci,
3249 Deoptimization::DeoptReason reason) {
3250 ciMethodData* md = method->method_data();
3251 if (md->is_empty()) {
3252 // Assume the trap has not occurred, or that it occurred only
3253 // because of a transient condition during start-up in the interpreter.
3254 return false;
3255 }
3256 if (md->has_trap_at(bci, reason) != 0) {
3257 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3258 // Also, if there are multiple reasons, or if there is no per-BCI record,
3259 // assume the worst.
3260 if (log())
3261 log()->elem("observe trap='%s' count='%d'",
3262 Deoptimization::trap_reason_name(reason),
3263 md->trap_count(reason));
3264 return true;
3265 } else {
3266 // Ignore method/bci and see if there have been too many globally.
3267 return too_many_traps(reason, md);
3268 }
3269 }
3271 // Less-accurate variant which does not require a method and bci.
3272 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3273 ciMethodData* logmd) {
3274 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
3275 // Too many traps globally.
3276 // Note that we use cumulative trap_count, not just md->trap_count.
3277 if (log()) {
3278 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3279 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3280 Deoptimization::trap_reason_name(reason),
3281 mcount, trap_count(reason));
3282 }
3283 return true;
3284 } else {
3285 // The coast is clear.
3286 return false;
3287 }
3288 }
3290 //--------------------------too_many_recompiles--------------------------------
3291 // Report if there are too many recompiles at the current method and bci.
3292 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3293 // Is not eager to return true, since this will cause the compiler to use
3294 // Action_none for a trap point, to avoid too many recompilations.
3295 bool Compile::too_many_recompiles(ciMethod* method,
3296 int bci,
3297 Deoptimization::DeoptReason reason) {
3298 ciMethodData* md = method->method_data();
3299 if (md->is_empty()) {
3300 // Assume the trap has not occurred, or that it occurred only
3301 // because of a transient condition during start-up in the interpreter.
3302 return false;
3303 }
3304 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3305 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3306 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3307 Deoptimization::DeoptReason per_bc_reason
3308 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3309 if ((per_bc_reason == Deoptimization::Reason_none
3310 || md->has_trap_at(bci, reason) != 0)
3311 // The trap frequency measure we care about is the recompile count:
3312 && md->trap_recompiled_at(bci)
3313 && md->overflow_recompile_count() >= bc_cutoff) {
3314 // Do not emit a trap here if it has already caused recompilations.
3315 // Also, if there are multiple reasons, or if there is no per-BCI record,
3316 // assume the worst.
3317 if (log())
3318 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3319 Deoptimization::trap_reason_name(reason),
3320 md->trap_count(reason),
3321 md->overflow_recompile_count());
3322 return true;
3323 } else if (trap_count(reason) != 0
3324 && decompile_count() >= m_cutoff) {
3325 // Too many recompiles globally, and we have seen this sort of trap.
3326 // Use cumulative decompile_count, not just md->decompile_count.
3327 if (log())
3328 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3329 Deoptimization::trap_reason_name(reason),
3330 md->trap_count(reason), trap_count(reason),
3331 md->decompile_count(), decompile_count());
3332 return true;
3333 } else {
3334 // The coast is clear.
3335 return false;
3336 }
3337 }
3340 #ifndef PRODUCT
3341 //------------------------------verify_graph_edges---------------------------
3342 // Walk the Graph and verify that there is a one-to-one correspondence
3343 // between Use-Def edges and Def-Use edges in the graph.
3344 void Compile::verify_graph_edges(bool no_dead_code) {
3345 if (VerifyGraphEdges) {
3346 ResourceArea *area = Thread::current()->resource_area();
3347 Unique_Node_List visited(area);
3348 // Call recursive graph walk to check edges
3349 _root->verify_edges(visited);
3350 if (no_dead_code) {
3351 // Now make sure that no visited node is used by an unvisited node.
3352 bool dead_nodes = 0;
3353 Unique_Node_List checked(area);
3354 while (visited.size() > 0) {
3355 Node* n = visited.pop();
3356 checked.push(n);
3357 for (uint i = 0; i < n->outcnt(); i++) {
3358 Node* use = n->raw_out(i);
3359 if (checked.member(use)) continue; // already checked
3360 if (visited.member(use)) continue; // already in the graph
3361 if (use->is_Con()) continue; // a dead ConNode is OK
3362 // At this point, we have found a dead node which is DU-reachable.
3363 if (dead_nodes++ == 0)
3364 tty->print_cr("*** Dead nodes reachable via DU edges:");
3365 use->dump(2);
3366 tty->print_cr("---");
3367 checked.push(use); // No repeats; pretend it is now checked.
3368 }
3369 }
3370 assert(dead_nodes == 0, "using nodes must be reachable from root");
3371 }
3372 }
3373 }
3374 #endif
3376 // The Compile object keeps track of failure reasons separately from the ciEnv.
3377 // This is required because there is not quite a 1-1 relation between the
3378 // ciEnv and its compilation task and the Compile object. Note that one
3379 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3380 // to backtrack and retry without subsuming loads. Other than this backtracking
3381 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3382 // by the logic in C2Compiler.
3383 void Compile::record_failure(const char* reason) {
3384 if (log() != NULL) {
3385 log()->elem("failure reason='%s' phase='compile'", reason);
3386 }
3387 if (_failure_reason == NULL) {
3388 // Record the first failure reason.
3389 _failure_reason = reason;
3390 }
3392 EventCompilerFailure event;
3393 if (event.should_commit()) {
3394 event.set_compileID(Compile::compile_id());
3395 event.set_failure(reason);
3396 event.commit();
3397 }
3399 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3400 C->print_method(PHASE_FAILURE);
3401 }
3402 _root = NULL; // flush the graph, too
3403 }
3405 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3406 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3407 _phase_name(name), _dolog(dolog)
3408 {
3409 if (dolog) {
3410 C = Compile::current();
3411 _log = C->log();
3412 } else {
3413 C = NULL;
3414 _log = NULL;
3415 }
3416 if (_log != NULL) {
3417 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3418 _log->stamp();
3419 _log->end_head();
3420 }
3421 }
3423 Compile::TracePhase::~TracePhase() {
3425 C = Compile::current();
3426 if (_dolog) {
3427 _log = C->log();
3428 } else {
3429 _log = NULL;
3430 }
3432 #ifdef ASSERT
3433 if (PrintIdealNodeCount) {
3434 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3435 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3436 }
3438 if (VerifyIdealNodeCount) {
3439 Compile::current()->print_missing_nodes();
3440 }
3441 #endif
3443 if (_log != NULL) {
3444 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3445 }
3446 }
3448 //=============================================================================
3449 // Two Constant's are equal when the type and the value are equal.
3450 bool Compile::Constant::operator==(const Constant& other) {
3451 if (type() != other.type() ) return false;
3452 if (can_be_reused() != other.can_be_reused()) return false;
3453 // For floating point values we compare the bit pattern.
3454 switch (type()) {
3455 case T_FLOAT: return (_v._value.i == other._v._value.i);
3456 case T_LONG:
3457 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3458 case T_OBJECT:
3459 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3460 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3461 case T_METADATA: return (_v._metadata == other._v._metadata);
3462 default: ShouldNotReachHere();
3463 }
3464 return false;
3465 }
3467 static int type_to_size_in_bytes(BasicType t) {
3468 switch (t) {
3469 case T_LONG: return sizeof(jlong );
3470 case T_FLOAT: return sizeof(jfloat );
3471 case T_DOUBLE: return sizeof(jdouble);
3472 case T_METADATA: return sizeof(Metadata*);
3473 // We use T_VOID as marker for jump-table entries (labels) which
3474 // need an internal word relocation.
3475 case T_VOID:
3476 case T_ADDRESS:
3477 case T_OBJECT: return sizeof(jobject);
3478 }
3480 ShouldNotReachHere();
3481 return -1;
3482 }
3484 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3485 // sort descending
3486 if (a->freq() > b->freq()) return -1;
3487 if (a->freq() < b->freq()) return 1;
3488 return 0;
3489 }
3491 void Compile::ConstantTable::calculate_offsets_and_size() {
3492 // First, sort the array by frequencies.
3493 _constants.sort(qsort_comparator);
3495 #ifdef ASSERT
3496 // Make sure all jump-table entries were sorted to the end of the
3497 // array (they have a negative frequency).
3498 bool found_void = false;
3499 for (int i = 0; i < _constants.length(); i++) {
3500 Constant con = _constants.at(i);
3501 if (con.type() == T_VOID)
3502 found_void = true; // jump-tables
3503 else
3504 assert(!found_void, "wrong sorting");
3505 }
3506 #endif
3508 int offset = 0;
3509 for (int i = 0; i < _constants.length(); i++) {
3510 Constant* con = _constants.adr_at(i);
3512 // Align offset for type.
3513 int typesize = type_to_size_in_bytes(con->type());
3514 offset = align_size_up(offset, typesize);
3515 con->set_offset(offset); // set constant's offset
3517 if (con->type() == T_VOID) {
3518 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3519 offset = offset + typesize * n->outcnt(); // expand jump-table
3520 } else {
3521 offset = offset + typesize;
3522 }
3523 }
3525 // Align size up to the next section start (which is insts; see
3526 // CodeBuffer::align_at_start).
3527 assert(_size == -1, "already set?");
3528 _size = align_size_up(offset, CodeEntryAlignment);
3529 }
3531 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3532 MacroAssembler _masm(&cb);
3533 for (int i = 0; i < _constants.length(); i++) {
3534 Constant con = _constants.at(i);
3535 address constant_addr;
3536 switch (con.type()) {
3537 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3538 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3539 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3540 case T_OBJECT: {
3541 jobject obj = con.get_jobject();
3542 int oop_index = _masm.oop_recorder()->find_index(obj);
3543 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3544 break;
3545 }
3546 case T_ADDRESS: {
3547 address addr = (address) con.get_jobject();
3548 constant_addr = _masm.address_constant(addr);
3549 break;
3550 }
3551 // We use T_VOID as marker for jump-table entries (labels) which
3552 // need an internal word relocation.
3553 case T_VOID: {
3554 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3555 // Fill the jump-table with a dummy word. The real value is
3556 // filled in later in fill_jump_table.
3557 address dummy = (address) n;
3558 constant_addr = _masm.address_constant(dummy);
3559 // Expand jump-table
3560 for (uint i = 1; i < n->outcnt(); i++) {
3561 address temp_addr = _masm.address_constant(dummy + i);
3562 assert(temp_addr, "consts section too small");
3563 }
3564 break;
3565 }
3566 case T_METADATA: {
3567 Metadata* obj = con.get_metadata();
3568 int metadata_index = _masm.oop_recorder()->find_index(obj);
3569 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3570 break;
3571 }
3572 default: ShouldNotReachHere();
3573 }
3574 assert(constant_addr, "consts section too small");
3575 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3576 }
3577 }
3579 int Compile::ConstantTable::find_offset(Constant& con) const {
3580 int idx = _constants.find(con);
3581 assert(idx != -1, "constant must be in constant table");
3582 int offset = _constants.at(idx).offset();
3583 assert(offset != -1, "constant table not emitted yet?");
3584 return offset;
3585 }
3587 void Compile::ConstantTable::add(Constant& con) {
3588 if (con.can_be_reused()) {
3589 int idx = _constants.find(con);
3590 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3591 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3592 return;
3593 }
3594 }
3595 (void) _constants.append(con);
3596 }
3598 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3599 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3600 Constant con(type, value, b->_freq);
3601 add(con);
3602 return con;
3603 }
3605 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3606 Constant con(metadata);
3607 add(con);
3608 return con;
3609 }
3611 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3612 jvalue value;
3613 BasicType type = oper->type()->basic_type();
3614 switch (type) {
3615 case T_LONG: value.j = oper->constantL(); break;
3616 case T_FLOAT: value.f = oper->constantF(); break;
3617 case T_DOUBLE: value.d = oper->constantD(); break;
3618 case T_OBJECT:
3619 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3620 case T_METADATA: return add((Metadata*)oper->constant()); break;
3621 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3622 }
3623 return add(n, type, value);
3624 }
3626 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3627 jvalue value;
3628 // We can use the node pointer here to identify the right jump-table
3629 // as this method is called from Compile::Fill_buffer right before
3630 // the MachNodes are emitted and the jump-table is filled (means the
3631 // MachNode pointers do not change anymore).
3632 value.l = (jobject) n;
3633 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3634 add(con);
3635 return con;
3636 }
3638 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3639 // If called from Compile::scratch_emit_size do nothing.
3640 if (Compile::current()->in_scratch_emit_size()) return;
3642 assert(labels.is_nonempty(), "must be");
3643 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3645 // Since MachConstantNode::constant_offset() also contains
3646 // table_base_offset() we need to subtract the table_base_offset()
3647 // to get the plain offset into the constant table.
3648 int offset = n->constant_offset() - table_base_offset();
3650 MacroAssembler _masm(&cb);
3651 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3653 for (uint i = 0; i < n->outcnt(); i++) {
3654 address* constant_addr = &jump_table_base[i];
3655 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i)));
3656 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3657 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3658 }
3659 }
3661 void Compile::dump_inlining() {
3662 if (print_inlining() || print_intrinsics()) {
3663 // Print inlining message for candidates that we couldn't inline
3664 // for lack of space or non constant receiver
3665 for (int i = 0; i < _late_inlines.length(); i++) {
3666 CallGenerator* cg = _late_inlines.at(i);
3667 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3668 }
3669 Unique_Node_List useful;
3670 useful.push(root());
3671 for (uint next = 0; next < useful.size(); ++next) {
3672 Node* n = useful.at(next);
3673 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3674 CallNode* call = n->as_Call();
3675 CallGenerator* cg = call->generator();
3676 cg->print_inlining_late("receiver not constant");
3677 }
3678 uint max = n->len();
3679 for ( uint i = 0; i < max; ++i ) {
3680 Node *m = n->in(i);
3681 if ( m == NULL ) continue;
3682 useful.push(m);
3683 }
3684 }
3685 for (int i = 0; i < _print_inlining_list->length(); i++) {
3686 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string());
3687 }
3688 }
3689 }
3691 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3692 if (n1->Opcode() < n2->Opcode()) return -1;
3693 else if (n1->Opcode() > n2->Opcode()) return 1;
3695 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()));
3696 for (uint i = 1; i < n1->req(); i++) {
3697 if (n1->in(i) < n2->in(i)) return -1;
3698 else if (n1->in(i) > n2->in(i)) return 1;
3699 }
3701 return 0;
3702 }
3704 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3705 Node* n1 = *n1p;
3706 Node* n2 = *n2p;
3708 return cmp_expensive_nodes(n1, n2);
3709 }
3711 void Compile::sort_expensive_nodes() {
3712 if (!expensive_nodes_sorted()) {
3713 _expensive_nodes->sort(cmp_expensive_nodes);
3714 }
3715 }
3717 bool Compile::expensive_nodes_sorted() const {
3718 for (int i = 1; i < _expensive_nodes->length(); i++) {
3719 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3720 return false;
3721 }
3722 }
3723 return true;
3724 }
3726 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3727 if (_expensive_nodes->length() == 0) {
3728 return false;
3729 }
3731 assert(OptimizeExpensiveOps, "optimization off?");
3733 // Take this opportunity to remove dead nodes from the list
3734 int j = 0;
3735 for (int i = 0; i < _expensive_nodes->length(); i++) {
3736 Node* n = _expensive_nodes->at(i);
3737 if (!n->is_unreachable(igvn)) {
3738 assert(n->is_expensive(), "should be expensive");
3739 _expensive_nodes->at_put(j, n);
3740 j++;
3741 }
3742 }
3743 _expensive_nodes->trunc_to(j);
3745 // Then sort the list so that similar nodes are next to each other
3746 // and check for at least two nodes of identical kind with same data
3747 // inputs.
3748 sort_expensive_nodes();
3750 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3751 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3752 return true;
3753 }
3754 }
3756 return false;
3757 }
3759 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3760 if (_expensive_nodes->length() == 0) {
3761 return;
3762 }
3764 assert(OptimizeExpensiveOps, "optimization off?");
3766 // Sort to bring similar nodes next to each other and clear the
3767 // control input of nodes for which there's only a single copy.
3768 sort_expensive_nodes();
3770 int j = 0;
3771 int identical = 0;
3772 int i = 0;
3773 for (; i < _expensive_nodes->length()-1; i++) {
3774 assert(j <= i, "can't write beyond current index");
3775 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3776 identical++;
3777 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3778 continue;
3779 }
3780 if (identical > 0) {
3781 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3782 identical = 0;
3783 } else {
3784 Node* n = _expensive_nodes->at(i);
3785 igvn.hash_delete(n);
3786 n->set_req(0, NULL);
3787 igvn.hash_insert(n);
3788 }
3789 }
3790 if (identical > 0) {
3791 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3792 } else if (_expensive_nodes->length() >= 1) {
3793 Node* n = _expensive_nodes->at(i);
3794 igvn.hash_delete(n);
3795 n->set_req(0, NULL);
3796 igvn.hash_insert(n);
3797 }
3798 _expensive_nodes->trunc_to(j);
3799 }
3801 void Compile::add_expensive_node(Node * n) {
3802 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3803 assert(n->is_expensive(), "expensive nodes with non-null control here only");
3804 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
3805 if (OptimizeExpensiveOps) {
3806 _expensive_nodes->append(n);
3807 } else {
3808 // Clear control input and let IGVN optimize expensive nodes if
3809 // OptimizeExpensiveOps is off.
3810 n->set_req(0, NULL);
3811 }
3812 }
3814 /**
3815 * Remove the speculative part of types and clean up the graph
3816 */
3817 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
3818 if (UseTypeSpeculation) {
3819 Unique_Node_List worklist;
3820 worklist.push(root());
3821 int modified = 0;
3822 // Go over all type nodes that carry a speculative type, drop the
3823 // speculative part of the type and enqueue the node for an igvn
3824 // which may optimize it out.
3825 for (uint next = 0; next < worklist.size(); ++next) {
3826 Node *n = worklist.at(next);
3827 if (n->is_Type() && n->as_Type()->type()->isa_oopptr() != NULL &&
3828 n->as_Type()->type()->is_oopptr()->speculative() != NULL) {
3829 TypeNode* tn = n->as_Type();
3830 const TypeOopPtr* t = tn->type()->is_oopptr();
3831 bool in_hash = igvn.hash_delete(n);
3832 assert(in_hash, "node should be in igvn hash table");
3833 tn->set_type(t->remove_speculative());
3834 igvn.hash_insert(n);
3835 igvn._worklist.push(n); // give it a chance to go away
3836 modified++;
3837 }
3838 uint max = n->len();
3839 for( uint i = 0; i < max; ++i ) {
3840 Node *m = n->in(i);
3841 if (not_a_node(m)) continue;
3842 worklist.push(m);
3843 }
3844 }
3845 // Drop the speculative part of all types in the igvn's type table
3846 igvn.remove_speculative_types();
3847 if (modified > 0) {
3848 igvn.optimize();
3849 }
3850 }
3851 }
3853 // Auxiliary method to support randomized stressing/fuzzing.
3854 //
3855 // This method can be called the arbitrary number of times, with current count
3856 // as the argument. The logic allows selecting a single candidate from the
3857 // running list of candidates as follows:
3858 // int count = 0;
3859 // Cand* selected = null;
3860 // while(cand = cand->next()) {
3861 // if (randomized_select(++count)) {
3862 // selected = cand;
3863 // }
3864 // }
3865 //
3866 // Including count equalizes the chances any candidate is "selected".
3867 // This is useful when we don't have the complete list of candidates to choose
3868 // from uniformly. In this case, we need to adjust the randomicity of the
3869 // selection, or else we will end up biasing the selection towards the latter
3870 // candidates.
3871 //
3872 // Quick back-envelope calculation shows that for the list of n candidates
3873 // the equal probability for the candidate to persist as "best" can be
3874 // achieved by replacing it with "next" k-th candidate with the probability
3875 // of 1/k. It can be easily shown that by the end of the run, the
3876 // probability for any candidate is converged to 1/n, thus giving the
3877 // uniform distribution among all the candidates.
3878 //
3879 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
3880 #define RANDOMIZED_DOMAIN_POW 29
3881 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
3882 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
3883 bool Compile::randomized_select(int count) {
3884 assert(count > 0, "only positive");
3885 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
3886 }