Thu, 03 Jan 2013 15:09:55 -0800
8005522: use fast-string instructions on x86 for zeroing
Summary: use 'rep stosb' instead of 'rep stosq' when fast-string operations are available.
Reviewed-by: twisti, roland
1 /*
2 * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "code/exceptionHandlerTable.hpp"
30 #include "code/nmethod.hpp"
31 #include "compiler/compileLog.hpp"
32 #include "compiler/disassembler.hpp"
33 #include "compiler/oopMap.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/block.hpp"
36 #include "opto/c2compiler.hpp"
37 #include "opto/callGenerator.hpp"
38 #include "opto/callnode.hpp"
39 #include "opto/cfgnode.hpp"
40 #include "opto/chaitin.hpp"
41 #include "opto/compile.hpp"
42 #include "opto/connode.hpp"
43 #include "opto/divnode.hpp"
44 #include "opto/escape.hpp"
45 #include "opto/idealGraphPrinter.hpp"
46 #include "opto/loopnode.hpp"
47 #include "opto/machnode.hpp"
48 #include "opto/macro.hpp"
49 #include "opto/matcher.hpp"
50 #include "opto/memnode.hpp"
51 #include "opto/mulnode.hpp"
52 #include "opto/node.hpp"
53 #include "opto/opcodes.hpp"
54 #include "opto/output.hpp"
55 #include "opto/parse.hpp"
56 #include "opto/phaseX.hpp"
57 #include "opto/rootnode.hpp"
58 #include "opto/runtime.hpp"
59 #include "opto/stringopts.hpp"
60 #include "opto/type.hpp"
61 #include "opto/vectornode.hpp"
62 #include "runtime/arguments.hpp"
63 #include "runtime/signature.hpp"
64 #include "runtime/stubRoutines.hpp"
65 #include "runtime/timer.hpp"
66 #include "utilities/copy.hpp"
67 #ifdef TARGET_ARCH_MODEL_x86_32
68 # include "adfiles/ad_x86_32.hpp"
69 #endif
70 #ifdef TARGET_ARCH_MODEL_x86_64
71 # include "adfiles/ad_x86_64.hpp"
72 #endif
73 #ifdef TARGET_ARCH_MODEL_sparc
74 # include "adfiles/ad_sparc.hpp"
75 #endif
76 #ifdef TARGET_ARCH_MODEL_zero
77 # include "adfiles/ad_zero.hpp"
78 #endif
79 #ifdef TARGET_ARCH_MODEL_arm
80 # include "adfiles/ad_arm.hpp"
81 #endif
82 #ifdef TARGET_ARCH_MODEL_ppc
83 # include "adfiles/ad_ppc.hpp"
84 #endif
87 // -------------------- Compile::mach_constant_base_node -----------------------
88 // Constant table base node singleton.
89 MachConstantBaseNode* Compile::mach_constant_base_node() {
90 if (_mach_constant_base_node == NULL) {
91 _mach_constant_base_node = new (C) MachConstantBaseNode();
92 _mach_constant_base_node->add_req(C->root());
93 }
94 return _mach_constant_base_node;
95 }
98 /// Support for intrinsics.
100 // Return the index at which m must be inserted (or already exists).
101 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
102 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
103 #ifdef ASSERT
104 for (int i = 1; i < _intrinsics->length(); i++) {
105 CallGenerator* cg1 = _intrinsics->at(i-1);
106 CallGenerator* cg2 = _intrinsics->at(i);
107 assert(cg1->method() != cg2->method()
108 ? cg1->method() < cg2->method()
109 : cg1->is_virtual() < cg2->is_virtual(),
110 "compiler intrinsics list must stay sorted");
111 }
112 #endif
113 // Binary search sorted list, in decreasing intervals [lo, hi].
114 int lo = 0, hi = _intrinsics->length()-1;
115 while (lo <= hi) {
116 int mid = (uint)(hi + lo) / 2;
117 ciMethod* mid_m = _intrinsics->at(mid)->method();
118 if (m < mid_m) {
119 hi = mid-1;
120 } else if (m > mid_m) {
121 lo = mid+1;
122 } else {
123 // look at minor sort key
124 bool mid_virt = _intrinsics->at(mid)->is_virtual();
125 if (is_virtual < mid_virt) {
126 hi = mid-1;
127 } else if (is_virtual > mid_virt) {
128 lo = mid+1;
129 } else {
130 return mid; // exact match
131 }
132 }
133 }
134 return lo; // inexact match
135 }
137 void Compile::register_intrinsic(CallGenerator* cg) {
138 if (_intrinsics == NULL) {
139 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
140 }
141 // This code is stolen from ciObjectFactory::insert.
142 // Really, GrowableArray should have methods for
143 // insert_at, remove_at, and binary_search.
144 int len = _intrinsics->length();
145 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
146 if (index == len) {
147 _intrinsics->append(cg);
148 } else {
149 #ifdef ASSERT
150 CallGenerator* oldcg = _intrinsics->at(index);
151 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
152 #endif
153 _intrinsics->append(_intrinsics->at(len-1));
154 int pos;
155 for (pos = len-2; pos >= index; pos--) {
156 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
157 }
158 _intrinsics->at_put(index, cg);
159 }
160 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
161 }
163 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
164 assert(m->is_loaded(), "don't try this on unloaded methods");
165 if (_intrinsics != NULL) {
166 int index = intrinsic_insertion_index(m, is_virtual);
167 if (index < _intrinsics->length()
168 && _intrinsics->at(index)->method() == m
169 && _intrinsics->at(index)->is_virtual() == is_virtual) {
170 return _intrinsics->at(index);
171 }
172 }
173 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
174 if (m->intrinsic_id() != vmIntrinsics::_none &&
175 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
176 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
177 if (cg != NULL) {
178 // Save it for next time:
179 register_intrinsic(cg);
180 return cg;
181 } else {
182 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
183 }
184 }
185 return NULL;
186 }
188 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
189 // in library_call.cpp.
192 #ifndef PRODUCT
193 // statistics gathering...
195 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
196 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
198 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
199 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
200 int oflags = _intrinsic_hist_flags[id];
201 assert(flags != 0, "what happened?");
202 if (is_virtual) {
203 flags |= _intrinsic_virtual;
204 }
205 bool changed = (flags != oflags);
206 if ((flags & _intrinsic_worked) != 0) {
207 juint count = (_intrinsic_hist_count[id] += 1);
208 if (count == 1) {
209 changed = true; // first time
210 }
211 // increment the overall count also:
212 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
213 }
214 if (changed) {
215 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
216 // Something changed about the intrinsic's virtuality.
217 if ((flags & _intrinsic_virtual) != 0) {
218 // This is the first use of this intrinsic as a virtual call.
219 if (oflags != 0) {
220 // We already saw it as a non-virtual, so note both cases.
221 flags |= _intrinsic_both;
222 }
223 } else if ((oflags & _intrinsic_both) == 0) {
224 // This is the first use of this intrinsic as a non-virtual
225 flags |= _intrinsic_both;
226 }
227 }
228 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
229 }
230 // update the overall flags also:
231 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
232 return changed;
233 }
235 static char* format_flags(int flags, char* buf) {
236 buf[0] = 0;
237 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
238 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
239 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
240 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
241 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
242 if (buf[0] == 0) strcat(buf, ",");
243 assert(buf[0] == ',', "must be");
244 return &buf[1];
245 }
247 void Compile::print_intrinsic_statistics() {
248 char flagsbuf[100];
249 ttyLocker ttyl;
250 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
251 tty->print_cr("Compiler intrinsic usage:");
252 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
253 if (total == 0) total = 1; // avoid div0 in case of no successes
254 #define PRINT_STAT_LINE(name, c, f) \
255 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
256 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
257 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
258 int flags = _intrinsic_hist_flags[id];
259 juint count = _intrinsic_hist_count[id];
260 if ((flags | count) != 0) {
261 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
262 }
263 }
264 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
265 if (xtty != NULL) xtty->tail("statistics");
266 }
268 void Compile::print_statistics() {
269 { ttyLocker ttyl;
270 if (xtty != NULL) xtty->head("statistics type='opto'");
271 Parse::print_statistics();
272 PhaseCCP::print_statistics();
273 PhaseRegAlloc::print_statistics();
274 Scheduling::print_statistics();
275 PhasePeephole::print_statistics();
276 PhaseIdealLoop::print_statistics();
277 if (xtty != NULL) xtty->tail("statistics");
278 }
279 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
280 // put this under its own <statistics> element.
281 print_intrinsic_statistics();
282 }
283 }
284 #endif //PRODUCT
286 // Support for bundling info
287 Bundle* Compile::node_bundling(const Node *n) {
288 assert(valid_bundle_info(n), "oob");
289 return &_node_bundling_base[n->_idx];
290 }
292 bool Compile::valid_bundle_info(const Node *n) {
293 return (_node_bundling_limit > n->_idx);
294 }
297 void Compile::gvn_replace_by(Node* n, Node* nn) {
298 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
299 Node* use = n->last_out(i);
300 bool is_in_table = initial_gvn()->hash_delete(use);
301 uint uses_found = 0;
302 for (uint j = 0; j < use->len(); j++) {
303 if (use->in(j) == n) {
304 if (j < use->req())
305 use->set_req(j, nn);
306 else
307 use->set_prec(j, nn);
308 uses_found++;
309 }
310 }
311 if (is_in_table) {
312 // reinsert into table
313 initial_gvn()->hash_find_insert(use);
314 }
315 record_for_igvn(use);
316 i -= uses_found; // we deleted 1 or more copies of this edge
317 }
318 }
321 static inline bool not_a_node(const Node* n) {
322 if (n == NULL) return true;
323 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
324 if (*(address*)n == badAddress) return true; // kill by Node::destruct
325 return false;
326 }
328 // Identify all nodes that are reachable from below, useful.
329 // Use breadth-first pass that records state in a Unique_Node_List,
330 // recursive traversal is slower.
331 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
332 int estimated_worklist_size = unique();
333 useful.map( estimated_worklist_size, NULL ); // preallocate space
335 // Initialize worklist
336 if (root() != NULL) { useful.push(root()); }
337 // If 'top' is cached, declare it useful to preserve cached node
338 if( cached_top_node() ) { useful.push(cached_top_node()); }
340 // Push all useful nodes onto the list, breadthfirst
341 for( uint next = 0; next < useful.size(); ++next ) {
342 assert( next < unique(), "Unique useful nodes < total nodes");
343 Node *n = useful.at(next);
344 uint max = n->len();
345 for( uint i = 0; i < max; ++i ) {
346 Node *m = n->in(i);
347 if (not_a_node(m)) continue;
348 useful.push(m);
349 }
350 }
351 }
353 // Update dead_node_list with any missing dead nodes using useful
354 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
355 void Compile::update_dead_node_list(Unique_Node_List &useful) {
356 uint max_idx = unique();
357 VectorSet& useful_node_set = useful.member_set();
359 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
360 // If node with index node_idx is not in useful set,
361 // mark it as dead in dead node list.
362 if (! useful_node_set.test(node_idx) ) {
363 record_dead_node(node_idx);
364 }
365 }
366 }
368 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
369 int shift = 0;
370 for (int i = 0; i < inlines->length(); i++) {
371 CallGenerator* cg = inlines->at(i);
372 CallNode* call = cg->call_node();
373 if (shift > 0) {
374 inlines->at_put(i-shift, cg);
375 }
376 if (!useful.member(call)) {
377 shift++;
378 }
379 }
380 inlines->trunc_to(inlines->length()-shift);
381 }
383 // Disconnect all useless nodes by disconnecting those at the boundary.
384 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
385 uint next = 0;
386 while (next < useful.size()) {
387 Node *n = useful.at(next++);
388 // Use raw traversal of out edges since this code removes out edges
389 int max = n->outcnt();
390 for (int j = 0; j < max; ++j) {
391 Node* child = n->raw_out(j);
392 if (! useful.member(child)) {
393 assert(!child->is_top() || child != top(),
394 "If top is cached in Compile object it is in useful list");
395 // Only need to remove this out-edge to the useless node
396 n->raw_del_out(j);
397 --j;
398 --max;
399 }
400 }
401 if (n->outcnt() == 1 && n->has_special_unique_user()) {
402 record_for_igvn(n->unique_out());
403 }
404 }
405 // Remove useless macro and predicate opaq nodes
406 for (int i = C->macro_count()-1; i >= 0; i--) {
407 Node* n = C->macro_node(i);
408 if (!useful.member(n)) {
409 remove_macro_node(n);
410 }
411 }
412 // clean up the late inline lists
413 remove_useless_late_inlines(&_string_late_inlines, useful);
414 remove_useless_late_inlines(&_late_inlines, useful);
415 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
416 }
418 //------------------------------frame_size_in_words-----------------------------
419 // frame_slots in units of words
420 int Compile::frame_size_in_words() const {
421 // shift is 0 in LP32 and 1 in LP64
422 const int shift = (LogBytesPerWord - LogBytesPerInt);
423 int words = _frame_slots >> shift;
424 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
425 return words;
426 }
428 // ============================================================================
429 //------------------------------CompileWrapper---------------------------------
430 class CompileWrapper : public StackObj {
431 Compile *const _compile;
432 public:
433 CompileWrapper(Compile* compile);
435 ~CompileWrapper();
436 };
438 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
439 // the Compile* pointer is stored in the current ciEnv:
440 ciEnv* env = compile->env();
441 assert(env == ciEnv::current(), "must already be a ciEnv active");
442 assert(env->compiler_data() == NULL, "compile already active?");
443 env->set_compiler_data(compile);
444 assert(compile == Compile::current(), "sanity");
446 compile->set_type_dict(NULL);
447 compile->set_type_hwm(NULL);
448 compile->set_type_last_size(0);
449 compile->set_last_tf(NULL, NULL);
450 compile->set_indexSet_arena(NULL);
451 compile->set_indexSet_free_block_list(NULL);
452 compile->init_type_arena();
453 Type::Initialize(compile);
454 _compile->set_scratch_buffer_blob(NULL);
455 _compile->begin_method();
456 }
457 CompileWrapper::~CompileWrapper() {
458 _compile->end_method();
459 if (_compile->scratch_buffer_blob() != NULL)
460 BufferBlob::free(_compile->scratch_buffer_blob());
461 _compile->env()->set_compiler_data(NULL);
462 }
465 //----------------------------print_compile_messages---------------------------
466 void Compile::print_compile_messages() {
467 #ifndef PRODUCT
468 // Check if recompiling
469 if (_subsume_loads == false && PrintOpto) {
470 // Recompiling without allowing machine instructions to subsume loads
471 tty->print_cr("*********************************************************");
472 tty->print_cr("** Bailout: Recompile without subsuming loads **");
473 tty->print_cr("*********************************************************");
474 }
475 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
476 // Recompiling without escape analysis
477 tty->print_cr("*********************************************************");
478 tty->print_cr("** Bailout: Recompile without escape analysis **");
479 tty->print_cr("*********************************************************");
480 }
481 if (env()->break_at_compile()) {
482 // Open the debugger when compiling this method.
483 tty->print("### Breaking when compiling: ");
484 method()->print_short_name();
485 tty->cr();
486 BREAKPOINT;
487 }
489 if( PrintOpto ) {
490 if (is_osr_compilation()) {
491 tty->print("[OSR]%3d", _compile_id);
492 } else {
493 tty->print("%3d", _compile_id);
494 }
495 }
496 #endif
497 }
500 //-----------------------init_scratch_buffer_blob------------------------------
501 // Construct a temporary BufferBlob and cache it for this compile.
502 void Compile::init_scratch_buffer_blob(int const_size) {
503 // If there is already a scratch buffer blob allocated and the
504 // constant section is big enough, use it. Otherwise free the
505 // current and allocate a new one.
506 BufferBlob* blob = scratch_buffer_blob();
507 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
508 // Use the current blob.
509 } else {
510 if (blob != NULL) {
511 BufferBlob::free(blob);
512 }
514 ResourceMark rm;
515 _scratch_const_size = const_size;
516 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
517 blob = BufferBlob::create("Compile::scratch_buffer", size);
518 // Record the buffer blob for next time.
519 set_scratch_buffer_blob(blob);
520 // Have we run out of code space?
521 if (scratch_buffer_blob() == NULL) {
522 // Let CompilerBroker disable further compilations.
523 record_failure("Not enough space for scratch buffer in CodeCache");
524 return;
525 }
526 }
528 // Initialize the relocation buffers
529 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
530 set_scratch_locs_memory(locs_buf);
531 }
534 //-----------------------scratch_emit_size-------------------------------------
535 // Helper function that computes size by emitting code
536 uint Compile::scratch_emit_size(const Node* n) {
537 // Start scratch_emit_size section.
538 set_in_scratch_emit_size(true);
540 // Emit into a trash buffer and count bytes emitted.
541 // This is a pretty expensive way to compute a size,
542 // but it works well enough if seldom used.
543 // All common fixed-size instructions are given a size
544 // method by the AD file.
545 // Note that the scratch buffer blob and locs memory are
546 // allocated at the beginning of the compile task, and
547 // may be shared by several calls to scratch_emit_size.
548 // The allocation of the scratch buffer blob is particularly
549 // expensive, since it has to grab the code cache lock.
550 BufferBlob* blob = this->scratch_buffer_blob();
551 assert(blob != NULL, "Initialize BufferBlob at start");
552 assert(blob->size() > MAX_inst_size, "sanity");
553 relocInfo* locs_buf = scratch_locs_memory();
554 address blob_begin = blob->content_begin();
555 address blob_end = (address)locs_buf;
556 assert(blob->content_contains(blob_end), "sanity");
557 CodeBuffer buf(blob_begin, blob_end - blob_begin);
558 buf.initialize_consts_size(_scratch_const_size);
559 buf.initialize_stubs_size(MAX_stubs_size);
560 assert(locs_buf != NULL, "sanity");
561 int lsize = MAX_locs_size / 3;
562 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
563 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
564 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
566 // Do the emission.
568 Label fakeL; // Fake label for branch instructions.
569 Label* saveL = NULL;
570 uint save_bnum = 0;
571 bool is_branch = n->is_MachBranch();
572 if (is_branch) {
573 MacroAssembler masm(&buf);
574 masm.bind(fakeL);
575 n->as_MachBranch()->save_label(&saveL, &save_bnum);
576 n->as_MachBranch()->label_set(&fakeL, 0);
577 }
578 n->emit(buf, this->regalloc());
579 if (is_branch) // Restore label.
580 n->as_MachBranch()->label_set(saveL, save_bnum);
582 // End scratch_emit_size section.
583 set_in_scratch_emit_size(false);
585 return buf.insts_size();
586 }
589 // ============================================================================
590 //------------------------------Compile standard-------------------------------
591 debug_only( int Compile::_debug_idx = 100000; )
593 // Compile a method. entry_bci is -1 for normal compilations and indicates
594 // the continuation bci for on stack replacement.
597 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis )
598 : Phase(Compiler),
599 _env(ci_env),
600 _log(ci_env->log()),
601 _compile_id(ci_env->compile_id()),
602 _save_argument_registers(false),
603 _stub_name(NULL),
604 _stub_function(NULL),
605 _stub_entry_point(NULL),
606 _method(target),
607 _entry_bci(osr_bci),
608 _initial_gvn(NULL),
609 _for_igvn(NULL),
610 _warm_calls(NULL),
611 _subsume_loads(subsume_loads),
612 _do_escape_analysis(do_escape_analysis),
613 _failure_reason(NULL),
614 _code_buffer("Compile::Fill_buffer"),
615 _orig_pc_slot(0),
616 _orig_pc_slot_offset_in_bytes(0),
617 _has_method_handle_invokes(false),
618 _mach_constant_base_node(NULL),
619 _node_bundling_limit(0),
620 _node_bundling_base(NULL),
621 _java_calls(0),
622 _inner_loops(0),
623 _scratch_const_size(-1),
624 _in_scratch_emit_size(false),
625 _dead_node_list(comp_arena()),
626 _dead_node_count(0),
627 #ifndef PRODUCT
628 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
629 _printer(IdealGraphPrinter::printer()),
630 #endif
631 _congraph(NULL),
632 _late_inlines(comp_arena(), 2, 0, NULL),
633 _string_late_inlines(comp_arena(), 2, 0, NULL),
634 _late_inlines_pos(0),
635 _number_of_mh_late_inlines(0),
636 _inlining_progress(false),
637 _inlining_incrementally(false),
638 _print_inlining_list(NULL),
639 _print_inlining(0) {
640 C = this;
642 CompileWrapper cw(this);
643 #ifndef PRODUCT
644 if (TimeCompiler2) {
645 tty->print(" ");
646 target->holder()->name()->print();
647 tty->print(".");
648 target->print_short_name();
649 tty->print(" ");
650 }
651 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
652 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
653 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
654 if (!print_opto_assembly) {
655 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
656 if (print_assembly && !Disassembler::can_decode()) {
657 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
658 print_opto_assembly = true;
659 }
660 }
661 set_print_assembly(print_opto_assembly);
662 set_parsed_irreducible_loop(false);
663 #endif
665 if (ProfileTraps) {
666 // Make sure the method being compiled gets its own MDO,
667 // so we can at least track the decompile_count().
668 method()->ensure_method_data();
669 }
671 Init(::AliasLevel);
674 print_compile_messages();
676 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
677 _ilt = InlineTree::build_inline_tree_root();
678 else
679 _ilt = NULL;
681 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
682 assert(num_alias_types() >= AliasIdxRaw, "");
684 #define MINIMUM_NODE_HASH 1023
685 // Node list that Iterative GVN will start with
686 Unique_Node_List for_igvn(comp_arena());
687 set_for_igvn(&for_igvn);
689 // GVN that will be run immediately on new nodes
690 uint estimated_size = method()->code_size()*4+64;
691 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
692 PhaseGVN gvn(node_arena(), estimated_size);
693 set_initial_gvn(&gvn);
695 if (PrintInlining) {
696 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
697 }
698 { // Scope for timing the parser
699 TracePhase t3("parse", &_t_parser, true);
701 // Put top into the hash table ASAP.
702 initial_gvn()->transform_no_reclaim(top());
704 // Set up tf(), start(), and find a CallGenerator.
705 CallGenerator* cg = NULL;
706 if (is_osr_compilation()) {
707 const TypeTuple *domain = StartOSRNode::osr_domain();
708 const TypeTuple *range = TypeTuple::make_range(method()->signature());
709 init_tf(TypeFunc::make(domain, range));
710 StartNode* s = new (this) StartOSRNode(root(), domain);
711 initial_gvn()->set_type_bottom(s);
712 init_start(s);
713 cg = CallGenerator::for_osr(method(), entry_bci());
714 } else {
715 // Normal case.
716 init_tf(TypeFunc::make(method()));
717 StartNode* s = new (this) StartNode(root(), tf()->domain());
718 initial_gvn()->set_type_bottom(s);
719 init_start(s);
720 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
721 // With java.lang.ref.reference.get() we must go through the
722 // intrinsic when G1 is enabled - even when get() is the root
723 // method of the compile - so that, if necessary, the value in
724 // the referent field of the reference object gets recorded by
725 // the pre-barrier code.
726 // Specifically, if G1 is enabled, the value in the referent
727 // field is recorded by the G1 SATB pre barrier. This will
728 // result in the referent being marked live and the reference
729 // object removed from the list of discovered references during
730 // reference processing.
731 cg = find_intrinsic(method(), false);
732 }
733 if (cg == NULL) {
734 float past_uses = method()->interpreter_invocation_count();
735 float expected_uses = past_uses;
736 cg = CallGenerator::for_inline(method(), expected_uses);
737 }
738 }
739 if (failing()) return;
740 if (cg == NULL) {
741 record_method_not_compilable_all_tiers("cannot parse method");
742 return;
743 }
744 JVMState* jvms = build_start_state(start(), tf());
745 if ((jvms = cg->generate(jvms)) == NULL) {
746 record_method_not_compilable("method parse failed");
747 return;
748 }
749 GraphKit kit(jvms);
751 if (!kit.stopped()) {
752 // Accept return values, and transfer control we know not where.
753 // This is done by a special, unique ReturnNode bound to root.
754 return_values(kit.jvms());
755 }
757 if (kit.has_exceptions()) {
758 // Any exceptions that escape from this call must be rethrown
759 // to whatever caller is dynamically above us on the stack.
760 // This is done by a special, unique RethrowNode bound to root.
761 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
762 }
764 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
766 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
767 inline_string_calls(true);
768 }
770 if (failing()) return;
772 print_method("Before RemoveUseless", 3);
774 // Remove clutter produced by parsing.
775 if (!failing()) {
776 ResourceMark rm;
777 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
778 }
779 }
781 // Note: Large methods are capped off in do_one_bytecode().
782 if (failing()) return;
784 // After parsing, node notes are no longer automagic.
785 // They must be propagated by register_new_node_with_optimizer(),
786 // clone(), or the like.
787 set_default_node_notes(NULL);
789 for (;;) {
790 int successes = Inline_Warm();
791 if (failing()) return;
792 if (successes == 0) break;
793 }
795 // Drain the list.
796 Finish_Warm();
797 #ifndef PRODUCT
798 if (_printer) {
799 _printer->print_inlining(this);
800 }
801 #endif
803 if (failing()) return;
804 NOT_PRODUCT( verify_graph_edges(); )
806 // Now optimize
807 Optimize();
808 if (failing()) return;
809 NOT_PRODUCT( verify_graph_edges(); )
811 #ifndef PRODUCT
812 if (PrintIdeal) {
813 ttyLocker ttyl; // keep the following output all in one block
814 // This output goes directly to the tty, not the compiler log.
815 // To enable tools to match it up with the compilation activity,
816 // be sure to tag this tty output with the compile ID.
817 if (xtty != NULL) {
818 xtty->head("ideal compile_id='%d'%s", compile_id(),
819 is_osr_compilation() ? " compile_kind='osr'" :
820 "");
821 }
822 root()->dump(9999);
823 if (xtty != NULL) {
824 xtty->tail("ideal");
825 }
826 }
827 #endif
829 // Now that we know the size of all the monitors we can add a fixed slot
830 // for the original deopt pc.
832 _orig_pc_slot = fixed_slots();
833 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
834 set_fixed_slots(next_slot);
836 // Now generate code
837 Code_Gen();
838 if (failing()) return;
840 // Check if we want to skip execution of all compiled code.
841 {
842 #ifndef PRODUCT
843 if (OptoNoExecute) {
844 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
845 return;
846 }
847 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
848 #endif
850 if (is_osr_compilation()) {
851 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
852 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
853 } else {
854 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
855 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
856 }
858 env()->register_method(_method, _entry_bci,
859 &_code_offsets,
860 _orig_pc_slot_offset_in_bytes,
861 code_buffer(),
862 frame_size_in_words(), _oop_map_set,
863 &_handler_table, &_inc_table,
864 compiler,
865 env()->comp_level(),
866 has_unsafe_access(),
867 SharedRuntime::is_wide_vector(max_vector_size())
868 );
870 if (log() != NULL) // Print code cache state into compiler log
871 log()->code_cache_state();
872 }
873 }
875 //------------------------------Compile----------------------------------------
876 // Compile a runtime stub
877 Compile::Compile( ciEnv* ci_env,
878 TypeFunc_generator generator,
879 address stub_function,
880 const char *stub_name,
881 int is_fancy_jump,
882 bool pass_tls,
883 bool save_arg_registers,
884 bool return_pc )
885 : Phase(Compiler),
886 _env(ci_env),
887 _log(ci_env->log()),
888 _compile_id(-1),
889 _save_argument_registers(save_arg_registers),
890 _method(NULL),
891 _stub_name(stub_name),
892 _stub_function(stub_function),
893 _stub_entry_point(NULL),
894 _entry_bci(InvocationEntryBci),
895 _initial_gvn(NULL),
896 _for_igvn(NULL),
897 _warm_calls(NULL),
898 _orig_pc_slot(0),
899 _orig_pc_slot_offset_in_bytes(0),
900 _subsume_loads(true),
901 _do_escape_analysis(false),
902 _failure_reason(NULL),
903 _code_buffer("Compile::Fill_buffer"),
904 _has_method_handle_invokes(false),
905 _mach_constant_base_node(NULL),
906 _node_bundling_limit(0),
907 _node_bundling_base(NULL),
908 _java_calls(0),
909 _inner_loops(0),
910 #ifndef PRODUCT
911 _trace_opto_output(TraceOptoOutput),
912 _printer(NULL),
913 #endif
914 _dead_node_list(comp_arena()),
915 _dead_node_count(0),
916 _congraph(NULL),
917 _number_of_mh_late_inlines(0),
918 _inlining_progress(false),
919 _inlining_incrementally(false),
920 _print_inlining_list(NULL),
921 _print_inlining(0) {
922 C = this;
924 #ifndef PRODUCT
925 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
926 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
927 set_print_assembly(PrintFrameConverterAssembly);
928 set_parsed_irreducible_loop(false);
929 #endif
930 CompileWrapper cw(this);
931 Init(/*AliasLevel=*/ 0);
932 init_tf((*generator)());
934 {
935 // The following is a dummy for the sake of GraphKit::gen_stub
936 Unique_Node_List for_igvn(comp_arena());
937 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
938 PhaseGVN gvn(Thread::current()->resource_area(),255);
939 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
940 gvn.transform_no_reclaim(top());
942 GraphKit kit;
943 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
944 }
946 NOT_PRODUCT( verify_graph_edges(); )
947 Code_Gen();
948 if (failing()) return;
951 // Entry point will be accessed using compile->stub_entry_point();
952 if (code_buffer() == NULL) {
953 Matcher::soft_match_failure();
954 } else {
955 if (PrintAssembly && (WizardMode || Verbose))
956 tty->print_cr("### Stub::%s", stub_name);
958 if (!failing()) {
959 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
961 // Make the NMethod
962 // For now we mark the frame as never safe for profile stackwalking
963 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
964 code_buffer(),
965 CodeOffsets::frame_never_safe,
966 // _code_offsets.value(CodeOffsets::Frame_Complete),
967 frame_size_in_words(),
968 _oop_map_set,
969 save_arg_registers);
970 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
972 _stub_entry_point = rs->entry_point();
973 }
974 }
975 }
977 //------------------------------Init-------------------------------------------
978 // Prepare for a single compilation
979 void Compile::Init(int aliaslevel) {
980 _unique = 0;
981 _regalloc = NULL;
983 _tf = NULL; // filled in later
984 _top = NULL; // cached later
985 _matcher = NULL; // filled in later
986 _cfg = NULL; // filled in later
988 set_24_bit_selection_and_mode(Use24BitFP, false);
990 _node_note_array = NULL;
991 _default_node_notes = NULL;
993 _immutable_memory = NULL; // filled in at first inquiry
995 // Globally visible Nodes
996 // First set TOP to NULL to give safe behavior during creation of RootNode
997 set_cached_top_node(NULL);
998 set_root(new (this) RootNode());
999 // Now that you have a Root to point to, create the real TOP
1000 set_cached_top_node( new (this) ConNode(Type::TOP) );
1001 set_recent_alloc(NULL, NULL);
1003 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1004 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1005 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1006 env()->set_dependencies(new Dependencies(env()));
1008 _fixed_slots = 0;
1009 set_has_split_ifs(false);
1010 set_has_loops(has_method() && method()->has_loops()); // first approximation
1011 set_has_stringbuilder(false);
1012 _trap_can_recompile = false; // no traps emitted yet
1013 _major_progress = true; // start out assuming good things will happen
1014 set_has_unsafe_access(false);
1015 set_max_vector_size(0);
1016 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1017 set_decompile_count(0);
1019 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1020 set_num_loop_opts(LoopOptsCount);
1021 set_do_inlining(Inline);
1022 set_max_inline_size(MaxInlineSize);
1023 set_freq_inline_size(FreqInlineSize);
1024 set_do_scheduling(OptoScheduling);
1025 set_do_count_invocations(false);
1026 set_do_method_data_update(false);
1028 if (debug_info()->recording_non_safepoints()) {
1029 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1030 (comp_arena(), 8, 0, NULL));
1031 set_default_node_notes(Node_Notes::make(this));
1032 }
1034 // // -- Initialize types before each compile --
1035 // // Update cached type information
1036 // if( _method && _method->constants() )
1037 // Type::update_loaded_types(_method, _method->constants());
1039 // Init alias_type map.
1040 if (!_do_escape_analysis && aliaslevel == 3)
1041 aliaslevel = 2; // No unique types without escape analysis
1042 _AliasLevel = aliaslevel;
1043 const int grow_ats = 16;
1044 _max_alias_types = grow_ats;
1045 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1046 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1047 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1048 {
1049 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1050 }
1051 // Initialize the first few types.
1052 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1053 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1054 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1055 _num_alias_types = AliasIdxRaw+1;
1056 // Zero out the alias type cache.
1057 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1058 // A NULL adr_type hits in the cache right away. Preload the right answer.
1059 probe_alias_cache(NULL)->_index = AliasIdxTop;
1061 _intrinsics = NULL;
1062 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1063 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1064 register_library_intrinsics();
1065 }
1067 //---------------------------init_start----------------------------------------
1068 // Install the StartNode on this compile object.
1069 void Compile::init_start(StartNode* s) {
1070 if (failing())
1071 return; // already failing
1072 assert(s == start(), "");
1073 }
1075 StartNode* Compile::start() const {
1076 assert(!failing(), "");
1077 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1078 Node* start = root()->fast_out(i);
1079 if( start->is_Start() )
1080 return start->as_Start();
1081 }
1082 ShouldNotReachHere();
1083 return NULL;
1084 }
1086 //-------------------------------immutable_memory-------------------------------------
1087 // Access immutable memory
1088 Node* Compile::immutable_memory() {
1089 if (_immutable_memory != NULL) {
1090 return _immutable_memory;
1091 }
1092 StartNode* s = start();
1093 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1094 Node *p = s->fast_out(i);
1095 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1096 _immutable_memory = p;
1097 return _immutable_memory;
1098 }
1099 }
1100 ShouldNotReachHere();
1101 return NULL;
1102 }
1104 //----------------------set_cached_top_node------------------------------------
1105 // Install the cached top node, and make sure Node::is_top works correctly.
1106 void Compile::set_cached_top_node(Node* tn) {
1107 if (tn != NULL) verify_top(tn);
1108 Node* old_top = _top;
1109 _top = tn;
1110 // Calling Node::setup_is_top allows the nodes the chance to adjust
1111 // their _out arrays.
1112 if (_top != NULL) _top->setup_is_top();
1113 if (old_top != NULL) old_top->setup_is_top();
1114 assert(_top == NULL || top()->is_top(), "");
1115 }
1117 #ifdef ASSERT
1118 uint Compile::count_live_nodes_by_graph_walk() {
1119 Unique_Node_List useful(comp_arena());
1120 // Get useful node list by walking the graph.
1121 identify_useful_nodes(useful);
1122 return useful.size();
1123 }
1125 void Compile::print_missing_nodes() {
1127 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1128 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1129 return;
1130 }
1132 // This is an expensive function. It is executed only when the user
1133 // specifies VerifyIdealNodeCount option or otherwise knows the
1134 // additional work that needs to be done to identify reachable nodes
1135 // by walking the flow graph and find the missing ones using
1136 // _dead_node_list.
1138 Unique_Node_List useful(comp_arena());
1139 // Get useful node list by walking the graph.
1140 identify_useful_nodes(useful);
1142 uint l_nodes = C->live_nodes();
1143 uint l_nodes_by_walk = useful.size();
1145 if (l_nodes != l_nodes_by_walk) {
1146 if (_log != NULL) {
1147 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1148 _log->stamp();
1149 _log->end_head();
1150 }
1151 VectorSet& useful_member_set = useful.member_set();
1152 int last_idx = l_nodes_by_walk;
1153 for (int i = 0; i < last_idx; i++) {
1154 if (useful_member_set.test(i)) {
1155 if (_dead_node_list.test(i)) {
1156 if (_log != NULL) {
1157 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1158 }
1159 if (PrintIdealNodeCount) {
1160 // Print the log message to tty
1161 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1162 useful.at(i)->dump();
1163 }
1164 }
1165 }
1166 else if (! _dead_node_list.test(i)) {
1167 if (_log != NULL) {
1168 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1169 }
1170 if (PrintIdealNodeCount) {
1171 // Print the log message to tty
1172 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1173 }
1174 }
1175 }
1176 if (_log != NULL) {
1177 _log->tail("mismatched_nodes");
1178 }
1179 }
1180 }
1181 #endif
1183 #ifndef PRODUCT
1184 void Compile::verify_top(Node* tn) const {
1185 if (tn != NULL) {
1186 assert(tn->is_Con(), "top node must be a constant");
1187 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1188 assert(tn->in(0) != NULL, "must have live top node");
1189 }
1190 }
1191 #endif
1194 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1196 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1197 guarantee(arr != NULL, "");
1198 int num_blocks = arr->length();
1199 if (grow_by < num_blocks) grow_by = num_blocks;
1200 int num_notes = grow_by * _node_notes_block_size;
1201 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1202 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1203 while (num_notes > 0) {
1204 arr->append(notes);
1205 notes += _node_notes_block_size;
1206 num_notes -= _node_notes_block_size;
1207 }
1208 assert(num_notes == 0, "exact multiple, please");
1209 }
1211 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1212 if (source == NULL || dest == NULL) return false;
1214 if (dest->is_Con())
1215 return false; // Do not push debug info onto constants.
1217 #ifdef ASSERT
1218 // Leave a bread crumb trail pointing to the original node:
1219 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1220 dest->set_debug_orig(source);
1221 }
1222 #endif
1224 if (node_note_array() == NULL)
1225 return false; // Not collecting any notes now.
1227 // This is a copy onto a pre-existing node, which may already have notes.
1228 // If both nodes have notes, do not overwrite any pre-existing notes.
1229 Node_Notes* source_notes = node_notes_at(source->_idx);
1230 if (source_notes == NULL || source_notes->is_clear()) return false;
1231 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1232 if (dest_notes == NULL || dest_notes->is_clear()) {
1233 return set_node_notes_at(dest->_idx, source_notes);
1234 }
1236 Node_Notes merged_notes = (*source_notes);
1237 // The order of operations here ensures that dest notes will win...
1238 merged_notes.update_from(dest_notes);
1239 return set_node_notes_at(dest->_idx, &merged_notes);
1240 }
1243 //--------------------------allow_range_check_smearing-------------------------
1244 // Gating condition for coalescing similar range checks.
1245 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1246 // single covering check that is at least as strong as any of them.
1247 // If the optimization succeeds, the simplified (strengthened) range check
1248 // will always succeed. If it fails, we will deopt, and then give up
1249 // on the optimization.
1250 bool Compile::allow_range_check_smearing() const {
1251 // If this method has already thrown a range-check,
1252 // assume it was because we already tried range smearing
1253 // and it failed.
1254 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1255 return !already_trapped;
1256 }
1259 //------------------------------flatten_alias_type-----------------------------
1260 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1261 int offset = tj->offset();
1262 TypePtr::PTR ptr = tj->ptr();
1264 // Known instance (scalarizable allocation) alias only with itself.
1265 bool is_known_inst = tj->isa_oopptr() != NULL &&
1266 tj->is_oopptr()->is_known_instance();
1268 // Process weird unsafe references.
1269 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1270 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1271 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1272 tj = TypeOopPtr::BOTTOM;
1273 ptr = tj->ptr();
1274 offset = tj->offset();
1275 }
1277 // Array pointers need some flattening
1278 const TypeAryPtr *ta = tj->isa_aryptr();
1279 if( ta && is_known_inst ) {
1280 if ( offset != Type::OffsetBot &&
1281 offset > arrayOopDesc::length_offset_in_bytes() ) {
1282 offset = Type::OffsetBot; // Flatten constant access into array body only
1283 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1284 }
1285 } else if( ta && _AliasLevel >= 2 ) {
1286 // For arrays indexed by constant indices, we flatten the alias
1287 // space to include all of the array body. Only the header, klass
1288 // and array length can be accessed un-aliased.
1289 if( offset != Type::OffsetBot ) {
1290 if( ta->const_oop() ) { // MethodData* or Method*
1291 offset = Type::OffsetBot; // Flatten constant access into array body
1292 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1293 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1294 // range is OK as-is.
1295 tj = ta = TypeAryPtr::RANGE;
1296 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1297 tj = TypeInstPtr::KLASS; // all klass loads look alike
1298 ta = TypeAryPtr::RANGE; // generic ignored junk
1299 ptr = TypePtr::BotPTR;
1300 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1301 tj = TypeInstPtr::MARK;
1302 ta = TypeAryPtr::RANGE; // generic ignored junk
1303 ptr = TypePtr::BotPTR;
1304 } else { // Random constant offset into array body
1305 offset = Type::OffsetBot; // Flatten constant access into array body
1306 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1307 }
1308 }
1309 // Arrays of fixed size alias with arrays of unknown size.
1310 if (ta->size() != TypeInt::POS) {
1311 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1312 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1313 }
1314 // Arrays of known objects become arrays of unknown objects.
1315 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1316 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1317 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1318 }
1319 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1320 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1321 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1322 }
1323 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1324 // cannot be distinguished by bytecode alone.
1325 if (ta->elem() == TypeInt::BOOL) {
1326 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1327 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1328 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1329 }
1330 // During the 2nd round of IterGVN, NotNull castings are removed.
1331 // Make sure the Bottom and NotNull variants alias the same.
1332 // Also, make sure exact and non-exact variants alias the same.
1333 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1334 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1335 }
1336 }
1338 // Oop pointers need some flattening
1339 const TypeInstPtr *to = tj->isa_instptr();
1340 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1341 ciInstanceKlass *k = to->klass()->as_instance_klass();
1342 if( ptr == TypePtr::Constant ) {
1343 if (to->klass() != ciEnv::current()->Class_klass() ||
1344 offset < k->size_helper() * wordSize) {
1345 // No constant oop pointers (such as Strings); they alias with
1346 // unknown strings.
1347 assert(!is_known_inst, "not scalarizable allocation");
1348 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1349 }
1350 } else if( is_known_inst ) {
1351 tj = to; // Keep NotNull and klass_is_exact for instance type
1352 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1353 // During the 2nd round of IterGVN, NotNull castings are removed.
1354 // Make sure the Bottom and NotNull variants alias the same.
1355 // Also, make sure exact and non-exact variants alias the same.
1356 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1357 }
1358 // Canonicalize the holder of this field
1359 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1360 // First handle header references such as a LoadKlassNode, even if the
1361 // object's klass is unloaded at compile time (4965979).
1362 if (!is_known_inst) { // Do it only for non-instance types
1363 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1364 }
1365 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1366 // Static fields are in the space above the normal instance
1367 // fields in the java.lang.Class instance.
1368 if (to->klass() != ciEnv::current()->Class_klass()) {
1369 to = NULL;
1370 tj = TypeOopPtr::BOTTOM;
1371 offset = tj->offset();
1372 }
1373 } else {
1374 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1375 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1376 if( is_known_inst ) {
1377 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1378 } else {
1379 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1380 }
1381 }
1382 }
1383 }
1385 // Klass pointers to object array klasses need some flattening
1386 const TypeKlassPtr *tk = tj->isa_klassptr();
1387 if( tk ) {
1388 // If we are referencing a field within a Klass, we need
1389 // to assume the worst case of an Object. Both exact and
1390 // inexact types must flatten to the same alias class so
1391 // use NotNull as the PTR.
1392 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1394 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1395 TypeKlassPtr::OBJECT->klass(),
1396 offset);
1397 }
1399 ciKlass* klass = tk->klass();
1400 if( klass->is_obj_array_klass() ) {
1401 ciKlass* k = TypeAryPtr::OOPS->klass();
1402 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1403 k = TypeInstPtr::BOTTOM->klass();
1404 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1405 }
1407 // Check for precise loads from the primary supertype array and force them
1408 // to the supertype cache alias index. Check for generic array loads from
1409 // the primary supertype array and also force them to the supertype cache
1410 // alias index. Since the same load can reach both, we need to merge
1411 // these 2 disparate memories into the same alias class. Since the
1412 // primary supertype array is read-only, there's no chance of confusion
1413 // where we bypass an array load and an array store.
1414 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1415 if (offset == Type::OffsetBot ||
1416 (offset >= primary_supers_offset &&
1417 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1418 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1419 offset = in_bytes(Klass::secondary_super_cache_offset());
1420 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1421 }
1422 }
1424 // Flatten all Raw pointers together.
1425 if (tj->base() == Type::RawPtr)
1426 tj = TypeRawPtr::BOTTOM;
1428 if (tj->base() == Type::AnyPtr)
1429 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1431 // Flatten all to bottom for now
1432 switch( _AliasLevel ) {
1433 case 0:
1434 tj = TypePtr::BOTTOM;
1435 break;
1436 case 1: // Flatten to: oop, static, field or array
1437 switch (tj->base()) {
1438 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1439 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1440 case Type::AryPtr: // do not distinguish arrays at all
1441 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1442 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1443 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1444 default: ShouldNotReachHere();
1445 }
1446 break;
1447 case 2: // No collapsing at level 2; keep all splits
1448 case 3: // No collapsing at level 3; keep all splits
1449 break;
1450 default:
1451 Unimplemented();
1452 }
1454 offset = tj->offset();
1455 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1457 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1458 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1459 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1460 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1461 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1462 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1463 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1464 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1465 assert( tj->ptr() != TypePtr::TopPTR &&
1466 tj->ptr() != TypePtr::AnyNull &&
1467 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1468 // assert( tj->ptr() != TypePtr::Constant ||
1469 // tj->base() == Type::RawPtr ||
1470 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1472 return tj;
1473 }
1475 void Compile::AliasType::Init(int i, const TypePtr* at) {
1476 _index = i;
1477 _adr_type = at;
1478 _field = NULL;
1479 _is_rewritable = true; // default
1480 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1481 if (atoop != NULL && atoop->is_known_instance()) {
1482 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1483 _general_index = Compile::current()->get_alias_index(gt);
1484 } else {
1485 _general_index = 0;
1486 }
1487 }
1489 //---------------------------------print_on------------------------------------
1490 #ifndef PRODUCT
1491 void Compile::AliasType::print_on(outputStream* st) {
1492 if (index() < 10)
1493 st->print("@ <%d> ", index());
1494 else st->print("@ <%d>", index());
1495 st->print(is_rewritable() ? " " : " RO");
1496 int offset = adr_type()->offset();
1497 if (offset == Type::OffsetBot)
1498 st->print(" +any");
1499 else st->print(" +%-3d", offset);
1500 st->print(" in ");
1501 adr_type()->dump_on(st);
1502 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1503 if (field() != NULL && tjp) {
1504 if (tjp->klass() != field()->holder() ||
1505 tjp->offset() != field()->offset_in_bytes()) {
1506 st->print(" != ");
1507 field()->print();
1508 st->print(" ***");
1509 }
1510 }
1511 }
1513 void print_alias_types() {
1514 Compile* C = Compile::current();
1515 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1516 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1517 C->alias_type(idx)->print_on(tty);
1518 tty->cr();
1519 }
1520 }
1521 #endif
1524 //----------------------------probe_alias_cache--------------------------------
1525 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1526 intptr_t key = (intptr_t) adr_type;
1527 key ^= key >> logAliasCacheSize;
1528 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1529 }
1532 //-----------------------------grow_alias_types--------------------------------
1533 void Compile::grow_alias_types() {
1534 const int old_ats = _max_alias_types; // how many before?
1535 const int new_ats = old_ats; // how many more?
1536 const int grow_ats = old_ats+new_ats; // how many now?
1537 _max_alias_types = grow_ats;
1538 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1539 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1540 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1541 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1542 }
1545 //--------------------------------find_alias_type------------------------------
1546 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1547 if (_AliasLevel == 0)
1548 return alias_type(AliasIdxBot);
1550 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1551 if (ace->_adr_type == adr_type) {
1552 return alias_type(ace->_index);
1553 }
1555 // Handle special cases.
1556 if (adr_type == NULL) return alias_type(AliasIdxTop);
1557 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1559 // Do it the slow way.
1560 const TypePtr* flat = flatten_alias_type(adr_type);
1562 #ifdef ASSERT
1563 assert(flat == flatten_alias_type(flat), "idempotent");
1564 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1565 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1566 const TypeOopPtr* foop = flat->is_oopptr();
1567 // Scalarizable allocations have exact klass always.
1568 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1569 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1570 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1571 }
1572 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1573 #endif
1575 int idx = AliasIdxTop;
1576 for (int i = 0; i < num_alias_types(); i++) {
1577 if (alias_type(i)->adr_type() == flat) {
1578 idx = i;
1579 break;
1580 }
1581 }
1583 if (idx == AliasIdxTop) {
1584 if (no_create) return NULL;
1585 // Grow the array if necessary.
1586 if (_num_alias_types == _max_alias_types) grow_alias_types();
1587 // Add a new alias type.
1588 idx = _num_alias_types++;
1589 _alias_types[idx]->Init(idx, flat);
1590 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1591 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1592 if (flat->isa_instptr()) {
1593 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1594 && flat->is_instptr()->klass() == env()->Class_klass())
1595 alias_type(idx)->set_rewritable(false);
1596 }
1597 if (flat->isa_klassptr()) {
1598 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1599 alias_type(idx)->set_rewritable(false);
1600 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1601 alias_type(idx)->set_rewritable(false);
1602 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1603 alias_type(idx)->set_rewritable(false);
1604 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1605 alias_type(idx)->set_rewritable(false);
1606 }
1607 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1608 // but the base pointer type is not distinctive enough to identify
1609 // references into JavaThread.)
1611 // Check for final fields.
1612 const TypeInstPtr* tinst = flat->isa_instptr();
1613 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1614 ciField* field;
1615 if (tinst->const_oop() != NULL &&
1616 tinst->klass() == ciEnv::current()->Class_klass() &&
1617 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1618 // static field
1619 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1620 field = k->get_field_by_offset(tinst->offset(), true);
1621 } else {
1622 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1623 field = k->get_field_by_offset(tinst->offset(), false);
1624 }
1625 assert(field == NULL ||
1626 original_field == NULL ||
1627 (field->holder() == original_field->holder() &&
1628 field->offset() == original_field->offset() &&
1629 field->is_static() == original_field->is_static()), "wrong field?");
1630 // Set field() and is_rewritable() attributes.
1631 if (field != NULL) alias_type(idx)->set_field(field);
1632 }
1633 }
1635 // Fill the cache for next time.
1636 ace->_adr_type = adr_type;
1637 ace->_index = idx;
1638 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1640 // Might as well try to fill the cache for the flattened version, too.
1641 AliasCacheEntry* face = probe_alias_cache(flat);
1642 if (face->_adr_type == NULL) {
1643 face->_adr_type = flat;
1644 face->_index = idx;
1645 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1646 }
1648 return alias_type(idx);
1649 }
1652 Compile::AliasType* Compile::alias_type(ciField* field) {
1653 const TypeOopPtr* t;
1654 if (field->is_static())
1655 t = TypeInstPtr::make(field->holder()->java_mirror());
1656 else
1657 t = TypeOopPtr::make_from_klass_raw(field->holder());
1658 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1659 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1660 return atp;
1661 }
1664 //------------------------------have_alias_type--------------------------------
1665 bool Compile::have_alias_type(const TypePtr* adr_type) {
1666 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1667 if (ace->_adr_type == adr_type) {
1668 return true;
1669 }
1671 // Handle special cases.
1672 if (adr_type == NULL) return true;
1673 if (adr_type == TypePtr::BOTTOM) return true;
1675 return find_alias_type(adr_type, true, NULL) != NULL;
1676 }
1678 //-----------------------------must_alias--------------------------------------
1679 // True if all values of the given address type are in the given alias category.
1680 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1681 if (alias_idx == AliasIdxBot) return true; // the universal category
1682 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1683 if (alias_idx == AliasIdxTop) return false; // the empty category
1684 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1686 // the only remaining possible overlap is identity
1687 int adr_idx = get_alias_index(adr_type);
1688 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1689 assert(adr_idx == alias_idx ||
1690 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1691 && adr_type != TypeOopPtr::BOTTOM),
1692 "should not be testing for overlap with an unsafe pointer");
1693 return adr_idx == alias_idx;
1694 }
1696 //------------------------------can_alias--------------------------------------
1697 // True if any values of the given address type are in the given alias category.
1698 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1699 if (alias_idx == AliasIdxTop) return false; // the empty category
1700 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1701 if (alias_idx == AliasIdxBot) return true; // the universal category
1702 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1704 // the only remaining possible overlap is identity
1705 int adr_idx = get_alias_index(adr_type);
1706 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1707 return adr_idx == alias_idx;
1708 }
1712 //---------------------------pop_warm_call-------------------------------------
1713 WarmCallInfo* Compile::pop_warm_call() {
1714 WarmCallInfo* wci = _warm_calls;
1715 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1716 return wci;
1717 }
1719 //----------------------------Inline_Warm--------------------------------------
1720 int Compile::Inline_Warm() {
1721 // If there is room, try to inline some more warm call sites.
1722 // %%% Do a graph index compaction pass when we think we're out of space?
1723 if (!InlineWarmCalls) return 0;
1725 int calls_made_hot = 0;
1726 int room_to_grow = NodeCountInliningCutoff - unique();
1727 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1728 int amount_grown = 0;
1729 WarmCallInfo* call;
1730 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1731 int est_size = (int)call->size();
1732 if (est_size > (room_to_grow - amount_grown)) {
1733 // This one won't fit anyway. Get rid of it.
1734 call->make_cold();
1735 continue;
1736 }
1737 call->make_hot();
1738 calls_made_hot++;
1739 amount_grown += est_size;
1740 amount_to_grow -= est_size;
1741 }
1743 if (calls_made_hot > 0) set_major_progress();
1744 return calls_made_hot;
1745 }
1748 //----------------------------Finish_Warm--------------------------------------
1749 void Compile::Finish_Warm() {
1750 if (!InlineWarmCalls) return;
1751 if (failing()) return;
1752 if (warm_calls() == NULL) return;
1754 // Clean up loose ends, if we are out of space for inlining.
1755 WarmCallInfo* call;
1756 while ((call = pop_warm_call()) != NULL) {
1757 call->make_cold();
1758 }
1759 }
1761 //---------------------cleanup_loop_predicates-----------------------
1762 // Remove the opaque nodes that protect the predicates so that all unused
1763 // checks and uncommon_traps will be eliminated from the ideal graph
1764 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1765 if (predicate_count()==0) return;
1766 for (int i = predicate_count(); i > 0; i--) {
1767 Node * n = predicate_opaque1_node(i-1);
1768 assert(n->Opcode() == Op_Opaque1, "must be");
1769 igvn.replace_node(n, n->in(1));
1770 }
1771 assert(predicate_count()==0, "should be clean!");
1772 }
1774 // StringOpts and late inlining of string methods
1775 void Compile::inline_string_calls(bool parse_time) {
1776 {
1777 // remove useless nodes to make the usage analysis simpler
1778 ResourceMark rm;
1779 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1780 }
1782 {
1783 ResourceMark rm;
1784 print_method("Before StringOpts", 3);
1785 PhaseStringOpts pso(initial_gvn(), for_igvn());
1786 print_method("After StringOpts", 3);
1787 }
1789 // now inline anything that we skipped the first time around
1790 if (!parse_time) {
1791 _late_inlines_pos = _late_inlines.length();
1792 }
1794 while (_string_late_inlines.length() > 0) {
1795 CallGenerator* cg = _string_late_inlines.pop();
1796 cg->do_late_inline();
1797 if (failing()) return;
1798 }
1799 _string_late_inlines.trunc_to(0);
1800 }
1802 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1803 assert(IncrementalInline, "incremental inlining should be on");
1804 PhaseGVN* gvn = initial_gvn();
1806 set_inlining_progress(false);
1807 for_igvn()->clear();
1808 gvn->replace_with(&igvn);
1810 int i = 0;
1812 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1813 CallGenerator* cg = _late_inlines.at(i);
1814 _late_inlines_pos = i+1;
1815 cg->do_late_inline();
1816 if (failing()) return;
1817 }
1818 int j = 0;
1819 for (; i < _late_inlines.length(); i++, j++) {
1820 _late_inlines.at_put(j, _late_inlines.at(i));
1821 }
1822 _late_inlines.trunc_to(j);
1824 {
1825 ResourceMark rm;
1826 PhaseRemoveUseless pru(C->initial_gvn(), C->for_igvn());
1827 }
1829 igvn = PhaseIterGVN(gvn);
1830 }
1832 // Perform incremental inlining until bound on number of live nodes is reached
1833 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1834 PhaseGVN* gvn = initial_gvn();
1836 set_inlining_incrementally(true);
1837 set_inlining_progress(true);
1838 uint low_live_nodes = 0;
1840 while(inlining_progress() && _late_inlines.length() > 0) {
1842 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1843 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1844 // PhaseIdealLoop is expensive so we only try it once we are
1845 // out of loop and we only try it again if the previous helped
1846 // got the number of nodes down significantly
1847 PhaseIdealLoop ideal_loop( igvn, false, true );
1848 if (failing()) return;
1849 low_live_nodes = live_nodes();
1850 _major_progress = true;
1851 }
1853 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1854 break;
1855 }
1856 }
1858 inline_incrementally_one(igvn);
1860 if (failing()) return;
1862 igvn.optimize();
1864 if (failing()) return;
1865 }
1867 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1869 if (_string_late_inlines.length() > 0) {
1870 assert(has_stringbuilder(), "inconsistent");
1871 for_igvn()->clear();
1872 initial_gvn()->replace_with(&igvn);
1874 inline_string_calls(false);
1876 if (failing()) return;
1878 {
1879 ResourceMark rm;
1880 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1881 }
1883 igvn = PhaseIterGVN(gvn);
1885 igvn.optimize();
1886 }
1888 set_inlining_incrementally(false);
1889 }
1892 //------------------------------Optimize---------------------------------------
1893 // Given a graph, optimize it.
1894 void Compile::Optimize() {
1895 TracePhase t1("optimizer", &_t_optimizer, true);
1897 #ifndef PRODUCT
1898 if (env()->break_at_compile()) {
1899 BREAKPOINT;
1900 }
1902 #endif
1904 ResourceMark rm;
1905 int loop_opts_cnt;
1907 NOT_PRODUCT( verify_graph_edges(); )
1909 print_method("After Parsing");
1911 {
1912 // Iterative Global Value Numbering, including ideal transforms
1913 // Initialize IterGVN with types and values from parse-time GVN
1914 PhaseIterGVN igvn(initial_gvn());
1915 {
1916 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1917 igvn.optimize();
1918 }
1920 print_method("Iter GVN 1", 2);
1922 if (failing()) return;
1924 inline_incrementally(igvn);
1926 print_method("Incremental Inline", 2);
1928 if (failing()) return;
1930 // Perform escape analysis
1931 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
1932 if (has_loops()) {
1933 // Cleanup graph (remove dead nodes).
1934 TracePhase t2("idealLoop", &_t_idealLoop, true);
1935 PhaseIdealLoop ideal_loop( igvn, false, true );
1936 if (major_progress()) print_method("PhaseIdealLoop before EA", 2);
1937 if (failing()) return;
1938 }
1939 ConnectionGraph::do_analysis(this, &igvn);
1941 if (failing()) return;
1943 // Optimize out fields loads from scalar replaceable allocations.
1944 igvn.optimize();
1945 print_method("Iter GVN after EA", 2);
1947 if (failing()) return;
1949 if (congraph() != NULL && macro_count() > 0) {
1950 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
1951 PhaseMacroExpand mexp(igvn);
1952 mexp.eliminate_macro_nodes();
1953 igvn.set_delay_transform(false);
1955 igvn.optimize();
1956 print_method("Iter GVN after eliminating allocations and locks", 2);
1958 if (failing()) return;
1959 }
1960 }
1962 // Loop transforms on the ideal graph. Range Check Elimination,
1963 // peeling, unrolling, etc.
1965 // Set loop opts counter
1966 loop_opts_cnt = num_loop_opts();
1967 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
1968 {
1969 TracePhase t2("idealLoop", &_t_idealLoop, true);
1970 PhaseIdealLoop ideal_loop( igvn, true );
1971 loop_opts_cnt--;
1972 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
1973 if (failing()) return;
1974 }
1975 // Loop opts pass if partial peeling occurred in previous pass
1976 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
1977 TracePhase t3("idealLoop", &_t_idealLoop, true);
1978 PhaseIdealLoop ideal_loop( igvn, false );
1979 loop_opts_cnt--;
1980 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
1981 if (failing()) return;
1982 }
1983 // Loop opts pass for loop-unrolling before CCP
1984 if(major_progress() && (loop_opts_cnt > 0)) {
1985 TracePhase t4("idealLoop", &_t_idealLoop, true);
1986 PhaseIdealLoop ideal_loop( igvn, false );
1987 loop_opts_cnt--;
1988 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
1989 }
1990 if (!failing()) {
1991 // Verify that last round of loop opts produced a valid graph
1992 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1993 PhaseIdealLoop::verify(igvn);
1994 }
1995 }
1996 if (failing()) return;
1998 // Conditional Constant Propagation;
1999 PhaseCCP ccp( &igvn );
2000 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2001 {
2002 TracePhase t2("ccp", &_t_ccp, true);
2003 ccp.do_transform();
2004 }
2005 print_method("PhaseCPP 1", 2);
2007 assert( true, "Break here to ccp.dump_old2new_map()");
2009 // Iterative Global Value Numbering, including ideal transforms
2010 {
2011 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2012 igvn = ccp;
2013 igvn.optimize();
2014 }
2016 print_method("Iter GVN 2", 2);
2018 if (failing()) return;
2020 // Loop transforms on the ideal graph. Range Check Elimination,
2021 // peeling, unrolling, etc.
2022 if(loop_opts_cnt > 0) {
2023 debug_only( int cnt = 0; );
2024 while(major_progress() && (loop_opts_cnt > 0)) {
2025 TracePhase t2("idealLoop", &_t_idealLoop, true);
2026 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2027 PhaseIdealLoop ideal_loop( igvn, true);
2028 loop_opts_cnt--;
2029 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
2030 if (failing()) return;
2031 }
2032 }
2034 {
2035 // Verify that all previous optimizations produced a valid graph
2036 // at least to this point, even if no loop optimizations were done.
2037 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2038 PhaseIdealLoop::verify(igvn);
2039 }
2041 {
2042 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2043 PhaseMacroExpand mex(igvn);
2044 if (mex.expand_macro_nodes()) {
2045 assert(failing(), "must bail out w/ explicit message");
2046 return;
2047 }
2048 }
2050 } // (End scope of igvn; run destructor if necessary for asserts.)
2052 dump_inlining();
2053 // A method with only infinite loops has no edges entering loops from root
2054 {
2055 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2056 if (final_graph_reshaping()) {
2057 assert(failing(), "must bail out w/ explicit message");
2058 return;
2059 }
2060 }
2062 print_method("Optimize finished", 2);
2063 }
2066 //------------------------------Code_Gen---------------------------------------
2067 // Given a graph, generate code for it
2068 void Compile::Code_Gen() {
2069 if (failing()) return;
2071 // Perform instruction selection. You might think we could reclaim Matcher
2072 // memory PDQ, but actually the Matcher is used in generating spill code.
2073 // Internals of the Matcher (including some VectorSets) must remain live
2074 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2075 // set a bit in reclaimed memory.
2077 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2078 // nodes. Mapping is only valid at the root of each matched subtree.
2079 NOT_PRODUCT( verify_graph_edges(); )
2081 Node_List proj_list;
2082 Matcher m(proj_list);
2083 _matcher = &m;
2084 {
2085 TracePhase t2("matcher", &_t_matcher, true);
2086 m.match();
2087 }
2088 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2089 // nodes. Mapping is only valid at the root of each matched subtree.
2090 NOT_PRODUCT( verify_graph_edges(); )
2092 // If you have too many nodes, or if matching has failed, bail out
2093 check_node_count(0, "out of nodes matching instructions");
2094 if (failing()) return;
2096 // Build a proper-looking CFG
2097 PhaseCFG cfg(node_arena(), root(), m);
2098 _cfg = &cfg;
2099 {
2100 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2101 cfg.Dominators();
2102 if (failing()) return;
2104 NOT_PRODUCT( verify_graph_edges(); )
2106 cfg.Estimate_Block_Frequency();
2107 cfg.GlobalCodeMotion(m,unique(),proj_list);
2108 if (failing()) return;
2110 print_method("Global code motion", 2);
2112 NOT_PRODUCT( verify_graph_edges(); )
2114 debug_only( cfg.verify(); )
2115 }
2116 NOT_PRODUCT( verify_graph_edges(); )
2118 PhaseChaitin regalloc(unique(),cfg,m);
2119 _regalloc = ®alloc;
2120 {
2121 TracePhase t2("regalloc", &_t_registerAllocation, true);
2122 // Perform any platform dependent preallocation actions. This is used,
2123 // for example, to avoid taking an implicit null pointer exception
2124 // using the frame pointer on win95.
2125 _regalloc->pd_preallocate_hook();
2127 // Perform register allocation. After Chaitin, use-def chains are
2128 // no longer accurate (at spill code) and so must be ignored.
2129 // Node->LRG->reg mappings are still accurate.
2130 _regalloc->Register_Allocate();
2132 // Bail out if the allocator builds too many nodes
2133 if (failing()) return;
2134 }
2136 // Prior to register allocation we kept empty basic blocks in case the
2137 // the allocator needed a place to spill. After register allocation we
2138 // are not adding any new instructions. If any basic block is empty, we
2139 // can now safely remove it.
2140 {
2141 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2142 cfg.remove_empty();
2143 if (do_freq_based_layout()) {
2144 PhaseBlockLayout layout(cfg);
2145 } else {
2146 cfg.set_loop_alignment();
2147 }
2148 cfg.fixup_flow();
2149 }
2151 // Perform any platform dependent postallocation verifications.
2152 debug_only( _regalloc->pd_postallocate_verify_hook(); )
2154 // Apply peephole optimizations
2155 if( OptoPeephole ) {
2156 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2157 PhasePeephole peep( _regalloc, cfg);
2158 peep.do_transform();
2159 }
2161 // Convert Nodes to instruction bits in a buffer
2162 {
2163 // %%%% workspace merge brought two timers together for one job
2164 TracePhase t2a("output", &_t_output, true);
2165 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2166 Output();
2167 }
2169 print_method("Final Code");
2171 // He's dead, Jim.
2172 _cfg = (PhaseCFG*)0xdeadbeef;
2173 _regalloc = (PhaseChaitin*)0xdeadbeef;
2174 }
2177 //------------------------------dump_asm---------------------------------------
2178 // Dump formatted assembly
2179 #ifndef PRODUCT
2180 void Compile::dump_asm(int *pcs, uint pc_limit) {
2181 bool cut_short = false;
2182 tty->print_cr("#");
2183 tty->print("# "); _tf->dump(); tty->cr();
2184 tty->print_cr("#");
2186 // For all blocks
2187 int pc = 0x0; // Program counter
2188 char starts_bundle = ' ';
2189 _regalloc->dump_frame();
2191 Node *n = NULL;
2192 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
2193 if (VMThread::should_terminate()) { cut_short = true; break; }
2194 Block *b = _cfg->_blocks[i];
2195 if (b->is_connector() && !Verbose) continue;
2196 n = b->_nodes[0];
2197 if (pcs && n->_idx < pc_limit)
2198 tty->print("%3.3x ", pcs[n->_idx]);
2199 else
2200 tty->print(" ");
2201 b->dump_head( &_cfg->_bbs );
2202 if (b->is_connector()) {
2203 tty->print_cr(" # Empty connector block");
2204 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2205 tty->print_cr(" # Block is sole successor of call");
2206 }
2208 // For all instructions
2209 Node *delay = NULL;
2210 for( uint j = 0; j<b->_nodes.size(); j++ ) {
2211 if (VMThread::should_terminate()) { cut_short = true; break; }
2212 n = b->_nodes[j];
2213 if (valid_bundle_info(n)) {
2214 Bundle *bundle = node_bundling(n);
2215 if (bundle->used_in_unconditional_delay()) {
2216 delay = n;
2217 continue;
2218 }
2219 if (bundle->starts_bundle())
2220 starts_bundle = '+';
2221 }
2223 if (WizardMode) n->dump();
2225 if( !n->is_Region() && // Dont print in the Assembly
2226 !n->is_Phi() && // a few noisely useless nodes
2227 !n->is_Proj() &&
2228 !n->is_MachTemp() &&
2229 !n->is_SafePointScalarObject() &&
2230 !n->is_Catch() && // Would be nice to print exception table targets
2231 !n->is_MergeMem() && // Not very interesting
2232 !n->is_top() && // Debug info table constants
2233 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2234 ) {
2235 if (pcs && n->_idx < pc_limit)
2236 tty->print("%3.3x", pcs[n->_idx]);
2237 else
2238 tty->print(" ");
2239 tty->print(" %c ", starts_bundle);
2240 starts_bundle = ' ';
2241 tty->print("\t");
2242 n->format(_regalloc, tty);
2243 tty->cr();
2244 }
2246 // If we have an instruction with a delay slot, and have seen a delay,
2247 // then back up and print it
2248 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2249 assert(delay != NULL, "no unconditional delay instruction");
2250 if (WizardMode) delay->dump();
2252 if (node_bundling(delay)->starts_bundle())
2253 starts_bundle = '+';
2254 if (pcs && n->_idx < pc_limit)
2255 tty->print("%3.3x", pcs[n->_idx]);
2256 else
2257 tty->print(" ");
2258 tty->print(" %c ", starts_bundle);
2259 starts_bundle = ' ';
2260 tty->print("\t");
2261 delay->format(_regalloc, tty);
2262 tty->print_cr("");
2263 delay = NULL;
2264 }
2266 // Dump the exception table as well
2267 if( n->is_Catch() && (Verbose || WizardMode) ) {
2268 // Print the exception table for this offset
2269 _handler_table.print_subtable_for(pc);
2270 }
2271 }
2273 if (pcs && n->_idx < pc_limit)
2274 tty->print_cr("%3.3x", pcs[n->_idx]);
2275 else
2276 tty->print_cr("");
2278 assert(cut_short || delay == NULL, "no unconditional delay branch");
2280 } // End of per-block dump
2281 tty->print_cr("");
2283 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2284 }
2285 #endif
2287 //------------------------------Final_Reshape_Counts---------------------------
2288 // This class defines counters to help identify when a method
2289 // may/must be executed using hardware with only 24-bit precision.
2290 struct Final_Reshape_Counts : public StackObj {
2291 int _call_count; // count non-inlined 'common' calls
2292 int _float_count; // count float ops requiring 24-bit precision
2293 int _double_count; // count double ops requiring more precision
2294 int _java_call_count; // count non-inlined 'java' calls
2295 int _inner_loop_count; // count loops which need alignment
2296 VectorSet _visited; // Visitation flags
2297 Node_List _tests; // Set of IfNodes & PCTableNodes
2299 Final_Reshape_Counts() :
2300 _call_count(0), _float_count(0), _double_count(0),
2301 _java_call_count(0), _inner_loop_count(0),
2302 _visited( Thread::current()->resource_area() ) { }
2304 void inc_call_count () { _call_count ++; }
2305 void inc_float_count () { _float_count ++; }
2306 void inc_double_count() { _double_count++; }
2307 void inc_java_call_count() { _java_call_count++; }
2308 void inc_inner_loop_count() { _inner_loop_count++; }
2310 int get_call_count () const { return _call_count ; }
2311 int get_float_count () const { return _float_count ; }
2312 int get_double_count() const { return _double_count; }
2313 int get_java_call_count() const { return _java_call_count; }
2314 int get_inner_loop_count() const { return _inner_loop_count; }
2315 };
2317 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2318 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2319 // Make sure the offset goes inside the instance layout.
2320 return k->contains_field_offset(tp->offset());
2321 // Note that OffsetBot and OffsetTop are very negative.
2322 }
2324 // Eliminate trivially redundant StoreCMs and accumulate their
2325 // precedence edges.
2326 void Compile::eliminate_redundant_card_marks(Node* n) {
2327 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2328 if (n->in(MemNode::Address)->outcnt() > 1) {
2329 // There are multiple users of the same address so it might be
2330 // possible to eliminate some of the StoreCMs
2331 Node* mem = n->in(MemNode::Memory);
2332 Node* adr = n->in(MemNode::Address);
2333 Node* val = n->in(MemNode::ValueIn);
2334 Node* prev = n;
2335 bool done = false;
2336 // Walk the chain of StoreCMs eliminating ones that match. As
2337 // long as it's a chain of single users then the optimization is
2338 // safe. Eliminating partially redundant StoreCMs would require
2339 // cloning copies down the other paths.
2340 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2341 if (adr == mem->in(MemNode::Address) &&
2342 val == mem->in(MemNode::ValueIn)) {
2343 // redundant StoreCM
2344 if (mem->req() > MemNode::OopStore) {
2345 // Hasn't been processed by this code yet.
2346 n->add_prec(mem->in(MemNode::OopStore));
2347 } else {
2348 // Already converted to precedence edge
2349 for (uint i = mem->req(); i < mem->len(); i++) {
2350 // Accumulate any precedence edges
2351 if (mem->in(i) != NULL) {
2352 n->add_prec(mem->in(i));
2353 }
2354 }
2355 // Everything above this point has been processed.
2356 done = true;
2357 }
2358 // Eliminate the previous StoreCM
2359 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2360 assert(mem->outcnt() == 0, "should be dead");
2361 mem->disconnect_inputs(NULL, this);
2362 } else {
2363 prev = mem;
2364 }
2365 mem = prev->in(MemNode::Memory);
2366 }
2367 }
2368 }
2370 //------------------------------final_graph_reshaping_impl----------------------
2371 // Implement items 1-5 from final_graph_reshaping below.
2372 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2374 if ( n->outcnt() == 0 ) return; // dead node
2375 uint nop = n->Opcode();
2377 // Check for 2-input instruction with "last use" on right input.
2378 // Swap to left input. Implements item (2).
2379 if( n->req() == 3 && // two-input instruction
2380 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2381 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2382 n->in(2)->outcnt() == 1 &&// right use IS a last use
2383 !n->in(2)->is_Con() ) { // right use is not a constant
2384 // Check for commutative opcode
2385 switch( nop ) {
2386 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2387 case Op_MaxI: case Op_MinI:
2388 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2389 case Op_AndL: case Op_XorL: case Op_OrL:
2390 case Op_AndI: case Op_XorI: case Op_OrI: {
2391 // Move "last use" input to left by swapping inputs
2392 n->swap_edges(1, 2);
2393 break;
2394 }
2395 default:
2396 break;
2397 }
2398 }
2400 #ifdef ASSERT
2401 if( n->is_Mem() ) {
2402 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2403 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2404 // oop will be recorded in oop map if load crosses safepoint
2405 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2406 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2407 "raw memory operations should have control edge");
2408 }
2409 #endif
2410 // Count FPU ops and common calls, implements item (3)
2411 switch( nop ) {
2412 // Count all float operations that may use FPU
2413 case Op_AddF:
2414 case Op_SubF:
2415 case Op_MulF:
2416 case Op_DivF:
2417 case Op_NegF:
2418 case Op_ModF:
2419 case Op_ConvI2F:
2420 case Op_ConF:
2421 case Op_CmpF:
2422 case Op_CmpF3:
2423 // case Op_ConvL2F: // longs are split into 32-bit halves
2424 frc.inc_float_count();
2425 break;
2427 case Op_ConvF2D:
2428 case Op_ConvD2F:
2429 frc.inc_float_count();
2430 frc.inc_double_count();
2431 break;
2433 // Count all double operations that may use FPU
2434 case Op_AddD:
2435 case Op_SubD:
2436 case Op_MulD:
2437 case Op_DivD:
2438 case Op_NegD:
2439 case Op_ModD:
2440 case Op_ConvI2D:
2441 case Op_ConvD2I:
2442 // case Op_ConvL2D: // handled by leaf call
2443 // case Op_ConvD2L: // handled by leaf call
2444 case Op_ConD:
2445 case Op_CmpD:
2446 case Op_CmpD3:
2447 frc.inc_double_count();
2448 break;
2449 case Op_Opaque1: // Remove Opaque Nodes before matching
2450 case Op_Opaque2: // Remove Opaque Nodes before matching
2451 n->subsume_by(n->in(1), this);
2452 break;
2453 case Op_CallStaticJava:
2454 case Op_CallJava:
2455 case Op_CallDynamicJava:
2456 frc.inc_java_call_count(); // Count java call site;
2457 case Op_CallRuntime:
2458 case Op_CallLeaf:
2459 case Op_CallLeafNoFP: {
2460 assert( n->is_Call(), "" );
2461 CallNode *call = n->as_Call();
2462 // Count call sites where the FP mode bit would have to be flipped.
2463 // Do not count uncommon runtime calls:
2464 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2465 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2466 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2467 frc.inc_call_count(); // Count the call site
2468 } else { // See if uncommon argument is shared
2469 Node *n = call->in(TypeFunc::Parms);
2470 int nop = n->Opcode();
2471 // Clone shared simple arguments to uncommon calls, item (1).
2472 if( n->outcnt() > 1 &&
2473 !n->is_Proj() &&
2474 nop != Op_CreateEx &&
2475 nop != Op_CheckCastPP &&
2476 nop != Op_DecodeN &&
2477 nop != Op_DecodeNKlass &&
2478 !n->is_Mem() ) {
2479 Node *x = n->clone();
2480 call->set_req( TypeFunc::Parms, x );
2481 }
2482 }
2483 break;
2484 }
2486 case Op_StoreD:
2487 case Op_LoadD:
2488 case Op_LoadD_unaligned:
2489 frc.inc_double_count();
2490 goto handle_mem;
2491 case Op_StoreF:
2492 case Op_LoadF:
2493 frc.inc_float_count();
2494 goto handle_mem;
2496 case Op_StoreCM:
2497 {
2498 // Convert OopStore dependence into precedence edge
2499 Node* prec = n->in(MemNode::OopStore);
2500 n->del_req(MemNode::OopStore);
2501 n->add_prec(prec);
2502 eliminate_redundant_card_marks(n);
2503 }
2505 // fall through
2507 case Op_StoreB:
2508 case Op_StoreC:
2509 case Op_StorePConditional:
2510 case Op_StoreI:
2511 case Op_StoreL:
2512 case Op_StoreIConditional:
2513 case Op_StoreLConditional:
2514 case Op_CompareAndSwapI:
2515 case Op_CompareAndSwapL:
2516 case Op_CompareAndSwapP:
2517 case Op_CompareAndSwapN:
2518 case Op_GetAndAddI:
2519 case Op_GetAndAddL:
2520 case Op_GetAndSetI:
2521 case Op_GetAndSetL:
2522 case Op_GetAndSetP:
2523 case Op_GetAndSetN:
2524 case Op_StoreP:
2525 case Op_StoreN:
2526 case Op_StoreNKlass:
2527 case Op_LoadB:
2528 case Op_LoadUB:
2529 case Op_LoadUS:
2530 case Op_LoadI:
2531 case Op_LoadKlass:
2532 case Op_LoadNKlass:
2533 case Op_LoadL:
2534 case Op_LoadL_unaligned:
2535 case Op_LoadPLocked:
2536 case Op_LoadP:
2537 case Op_LoadN:
2538 case Op_LoadRange:
2539 case Op_LoadS: {
2540 handle_mem:
2541 #ifdef ASSERT
2542 if( VerifyOptoOopOffsets ) {
2543 assert( n->is_Mem(), "" );
2544 MemNode *mem = (MemNode*)n;
2545 // Check to see if address types have grounded out somehow.
2546 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2547 assert( !tp || oop_offset_is_sane(tp), "" );
2548 }
2549 #endif
2550 break;
2551 }
2553 case Op_AddP: { // Assert sane base pointers
2554 Node *addp = n->in(AddPNode::Address);
2555 assert( !addp->is_AddP() ||
2556 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2557 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2558 "Base pointers must match" );
2559 #ifdef _LP64
2560 if ((UseCompressedOops || UseCompressedKlassPointers) &&
2561 addp->Opcode() == Op_ConP &&
2562 addp == n->in(AddPNode::Base) &&
2563 n->in(AddPNode::Offset)->is_Con()) {
2564 // Use addressing with narrow klass to load with offset on x86.
2565 // On sparc loading 32-bits constant and decoding it have less
2566 // instructions (4) then load 64-bits constant (7).
2567 // Do this transformation here since IGVN will convert ConN back to ConP.
2568 const Type* t = addp->bottom_type();
2569 if (t->isa_oopptr() || t->isa_klassptr()) {
2570 Node* nn = NULL;
2572 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2574 // Look for existing ConN node of the same exact type.
2575 Node* r = root();
2576 uint cnt = r->outcnt();
2577 for (uint i = 0; i < cnt; i++) {
2578 Node* m = r->raw_out(i);
2579 if (m!= NULL && m->Opcode() == op &&
2580 m->bottom_type()->make_ptr() == t) {
2581 nn = m;
2582 break;
2583 }
2584 }
2585 if (nn != NULL) {
2586 // Decode a narrow oop to match address
2587 // [R12 + narrow_oop_reg<<3 + offset]
2588 if (t->isa_oopptr()) {
2589 nn = new (this) DecodeNNode(nn, t);
2590 } else {
2591 nn = new (this) DecodeNKlassNode(nn, t);
2592 }
2593 n->set_req(AddPNode::Base, nn);
2594 n->set_req(AddPNode::Address, nn);
2595 if (addp->outcnt() == 0) {
2596 addp->disconnect_inputs(NULL, this);
2597 }
2598 }
2599 }
2600 }
2601 #endif
2602 break;
2603 }
2605 #ifdef _LP64
2606 case Op_CastPP:
2607 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2608 Node* in1 = n->in(1);
2609 const Type* t = n->bottom_type();
2610 Node* new_in1 = in1->clone();
2611 new_in1->as_DecodeN()->set_type(t);
2613 if (!Matcher::narrow_oop_use_complex_address()) {
2614 //
2615 // x86, ARM and friends can handle 2 adds in addressing mode
2616 // and Matcher can fold a DecodeN node into address by using
2617 // a narrow oop directly and do implicit NULL check in address:
2618 //
2619 // [R12 + narrow_oop_reg<<3 + offset]
2620 // NullCheck narrow_oop_reg
2621 //
2622 // On other platforms (Sparc) we have to keep new DecodeN node and
2623 // use it to do implicit NULL check in address:
2624 //
2625 // decode_not_null narrow_oop_reg, base_reg
2626 // [base_reg + offset]
2627 // NullCheck base_reg
2628 //
2629 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2630 // to keep the information to which NULL check the new DecodeN node
2631 // corresponds to use it as value in implicit_null_check().
2632 //
2633 new_in1->set_req(0, n->in(0));
2634 }
2636 n->subsume_by(new_in1, this);
2637 if (in1->outcnt() == 0) {
2638 in1->disconnect_inputs(NULL, this);
2639 }
2640 }
2641 break;
2643 case Op_CmpP:
2644 // Do this transformation here to preserve CmpPNode::sub() and
2645 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2646 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2647 Node* in1 = n->in(1);
2648 Node* in2 = n->in(2);
2649 if (!in1->is_DecodeNarrowPtr()) {
2650 in2 = in1;
2651 in1 = n->in(2);
2652 }
2653 assert(in1->is_DecodeNarrowPtr(), "sanity");
2655 Node* new_in2 = NULL;
2656 if (in2->is_DecodeNarrowPtr()) {
2657 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2658 new_in2 = in2->in(1);
2659 } else if (in2->Opcode() == Op_ConP) {
2660 const Type* t = in2->bottom_type();
2661 if (t == TypePtr::NULL_PTR) {
2662 assert(in1->is_DecodeN(), "compare klass to null?");
2663 // Don't convert CmpP null check into CmpN if compressed
2664 // oops implicit null check is not generated.
2665 // This will allow to generate normal oop implicit null check.
2666 if (Matcher::gen_narrow_oop_implicit_null_checks())
2667 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2668 //
2669 // This transformation together with CastPP transformation above
2670 // will generated code for implicit NULL checks for compressed oops.
2671 //
2672 // The original code after Optimize()
2673 //
2674 // LoadN memory, narrow_oop_reg
2675 // decode narrow_oop_reg, base_reg
2676 // CmpP base_reg, NULL
2677 // CastPP base_reg // NotNull
2678 // Load [base_reg + offset], val_reg
2679 //
2680 // after these transformations will be
2681 //
2682 // LoadN memory, narrow_oop_reg
2683 // CmpN narrow_oop_reg, NULL
2684 // decode_not_null narrow_oop_reg, base_reg
2685 // Load [base_reg + offset], val_reg
2686 //
2687 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2688 // since narrow oops can be used in debug info now (see the code in
2689 // final_graph_reshaping_walk()).
2690 //
2691 // At the end the code will be matched to
2692 // on x86:
2693 //
2694 // Load_narrow_oop memory, narrow_oop_reg
2695 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2696 // NullCheck narrow_oop_reg
2697 //
2698 // and on sparc:
2699 //
2700 // Load_narrow_oop memory, narrow_oop_reg
2701 // decode_not_null narrow_oop_reg, base_reg
2702 // Load [base_reg + offset], val_reg
2703 // NullCheck base_reg
2704 //
2705 } else if (t->isa_oopptr()) {
2706 new_in2 = ConNode::make(this, t->make_narrowoop());
2707 } else if (t->isa_klassptr()) {
2708 new_in2 = ConNode::make(this, t->make_narrowklass());
2709 }
2710 }
2711 if (new_in2 != NULL) {
2712 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2713 n->subsume_by(cmpN, this);
2714 if (in1->outcnt() == 0) {
2715 in1->disconnect_inputs(NULL, this);
2716 }
2717 if (in2->outcnt() == 0) {
2718 in2->disconnect_inputs(NULL, this);
2719 }
2720 }
2721 }
2722 break;
2724 case Op_DecodeN:
2725 case Op_DecodeNKlass:
2726 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2727 // DecodeN could be pinned when it can't be fold into
2728 // an address expression, see the code for Op_CastPP above.
2729 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2730 break;
2732 case Op_EncodeP:
2733 case Op_EncodePKlass: {
2734 Node* in1 = n->in(1);
2735 if (in1->is_DecodeNarrowPtr()) {
2736 n->subsume_by(in1->in(1), this);
2737 } else if (in1->Opcode() == Op_ConP) {
2738 const Type* t = in1->bottom_type();
2739 if (t == TypePtr::NULL_PTR) {
2740 assert(t->isa_oopptr(), "null klass?");
2741 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2742 } else if (t->isa_oopptr()) {
2743 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2744 } else if (t->isa_klassptr()) {
2745 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2746 }
2747 }
2748 if (in1->outcnt() == 0) {
2749 in1->disconnect_inputs(NULL, this);
2750 }
2751 break;
2752 }
2754 case Op_Proj: {
2755 if (OptimizeStringConcat) {
2756 ProjNode* p = n->as_Proj();
2757 if (p->_is_io_use) {
2758 // Separate projections were used for the exception path which
2759 // are normally removed by a late inline. If it wasn't inlined
2760 // then they will hang around and should just be replaced with
2761 // the original one.
2762 Node* proj = NULL;
2763 // Replace with just one
2764 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2765 Node *use = i.get();
2766 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2767 proj = use;
2768 break;
2769 }
2770 }
2771 assert(proj != NULL, "must be found");
2772 p->subsume_by(proj, this);
2773 }
2774 }
2775 break;
2776 }
2778 case Op_Phi:
2779 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
2780 // The EncodeP optimization may create Phi with the same edges
2781 // for all paths. It is not handled well by Register Allocator.
2782 Node* unique_in = n->in(1);
2783 assert(unique_in != NULL, "");
2784 uint cnt = n->req();
2785 for (uint i = 2; i < cnt; i++) {
2786 Node* m = n->in(i);
2787 assert(m != NULL, "");
2788 if (unique_in != m)
2789 unique_in = NULL;
2790 }
2791 if (unique_in != NULL) {
2792 n->subsume_by(unique_in, this);
2793 }
2794 }
2795 break;
2797 #endif
2799 case Op_ModI:
2800 if (UseDivMod) {
2801 // Check if a%b and a/b both exist
2802 Node* d = n->find_similar(Op_DivI);
2803 if (d) {
2804 // Replace them with a fused divmod if supported
2805 if (Matcher::has_match_rule(Op_DivModI)) {
2806 DivModINode* divmod = DivModINode::make(this, n);
2807 d->subsume_by(divmod->div_proj(), this);
2808 n->subsume_by(divmod->mod_proj(), this);
2809 } else {
2810 // replace a%b with a-((a/b)*b)
2811 Node* mult = new (this) MulINode(d, d->in(2));
2812 Node* sub = new (this) SubINode(d->in(1), mult);
2813 n->subsume_by(sub, this);
2814 }
2815 }
2816 }
2817 break;
2819 case Op_ModL:
2820 if (UseDivMod) {
2821 // Check if a%b and a/b both exist
2822 Node* d = n->find_similar(Op_DivL);
2823 if (d) {
2824 // Replace them with a fused divmod if supported
2825 if (Matcher::has_match_rule(Op_DivModL)) {
2826 DivModLNode* divmod = DivModLNode::make(this, n);
2827 d->subsume_by(divmod->div_proj(), this);
2828 n->subsume_by(divmod->mod_proj(), this);
2829 } else {
2830 // replace a%b with a-((a/b)*b)
2831 Node* mult = new (this) MulLNode(d, d->in(2));
2832 Node* sub = new (this) SubLNode(d->in(1), mult);
2833 n->subsume_by(sub, this);
2834 }
2835 }
2836 }
2837 break;
2839 case Op_LoadVector:
2840 case Op_StoreVector:
2841 break;
2843 case Op_PackB:
2844 case Op_PackS:
2845 case Op_PackI:
2846 case Op_PackF:
2847 case Op_PackL:
2848 case Op_PackD:
2849 if (n->req()-1 > 2) {
2850 // Replace many operand PackNodes with a binary tree for matching
2851 PackNode* p = (PackNode*) n;
2852 Node* btp = p->binary_tree_pack(this, 1, n->req());
2853 n->subsume_by(btp, this);
2854 }
2855 break;
2856 case Op_Loop:
2857 case Op_CountedLoop:
2858 if (n->as_Loop()->is_inner_loop()) {
2859 frc.inc_inner_loop_count();
2860 }
2861 break;
2862 case Op_LShiftI:
2863 case Op_RShiftI:
2864 case Op_URShiftI:
2865 case Op_LShiftL:
2866 case Op_RShiftL:
2867 case Op_URShiftL:
2868 if (Matcher::need_masked_shift_count) {
2869 // The cpu's shift instructions don't restrict the count to the
2870 // lower 5/6 bits. We need to do the masking ourselves.
2871 Node* in2 = n->in(2);
2872 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2873 const TypeInt* t = in2->find_int_type();
2874 if (t != NULL && t->is_con()) {
2875 juint shift = t->get_con();
2876 if (shift > mask) { // Unsigned cmp
2877 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
2878 }
2879 } else {
2880 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2881 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
2882 n->set_req(2, shift);
2883 }
2884 }
2885 if (in2->outcnt() == 0) { // Remove dead node
2886 in2->disconnect_inputs(NULL, this);
2887 }
2888 }
2889 break;
2890 default:
2891 assert( !n->is_Call(), "" );
2892 assert( !n->is_Mem(), "" );
2893 break;
2894 }
2896 // Collect CFG split points
2897 if (n->is_MultiBranch())
2898 frc._tests.push(n);
2899 }
2901 //------------------------------final_graph_reshaping_walk---------------------
2902 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2903 // requires that the walk visits a node's inputs before visiting the node.
2904 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
2905 ResourceArea *area = Thread::current()->resource_area();
2906 Unique_Node_List sfpt(area);
2908 frc._visited.set(root->_idx); // first, mark node as visited
2909 uint cnt = root->req();
2910 Node *n = root;
2911 uint i = 0;
2912 while (true) {
2913 if (i < cnt) {
2914 // Place all non-visited non-null inputs onto stack
2915 Node* m = n->in(i);
2916 ++i;
2917 if (m != NULL && !frc._visited.test_set(m->_idx)) {
2918 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
2919 sfpt.push(m);
2920 cnt = m->req();
2921 nstack.push(n, i); // put on stack parent and next input's index
2922 n = m;
2923 i = 0;
2924 }
2925 } else {
2926 // Now do post-visit work
2927 final_graph_reshaping_impl( n, frc );
2928 if (nstack.is_empty())
2929 break; // finished
2930 n = nstack.node(); // Get node from stack
2931 cnt = n->req();
2932 i = nstack.index();
2933 nstack.pop(); // Shift to the next node on stack
2934 }
2935 }
2937 // Skip next transformation if compressed oops are not used.
2938 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
2939 (!UseCompressedOops && !UseCompressedKlassPointers))
2940 return;
2942 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
2943 // It could be done for an uncommon traps or any safepoints/calls
2944 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
2945 while (sfpt.size() > 0) {
2946 n = sfpt.pop();
2947 JVMState *jvms = n->as_SafePoint()->jvms();
2948 assert(jvms != NULL, "sanity");
2949 int start = jvms->debug_start();
2950 int end = n->req();
2951 bool is_uncommon = (n->is_CallStaticJava() &&
2952 n->as_CallStaticJava()->uncommon_trap_request() != 0);
2953 for (int j = start; j < end; j++) {
2954 Node* in = n->in(j);
2955 if (in->is_DecodeNarrowPtr()) {
2956 bool safe_to_skip = true;
2957 if (!is_uncommon ) {
2958 // Is it safe to skip?
2959 for (uint i = 0; i < in->outcnt(); i++) {
2960 Node* u = in->raw_out(i);
2961 if (!u->is_SafePoint() ||
2962 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
2963 safe_to_skip = false;
2964 }
2965 }
2966 }
2967 if (safe_to_skip) {
2968 n->set_req(j, in->in(1));
2969 }
2970 if (in->outcnt() == 0) {
2971 in->disconnect_inputs(NULL, this);
2972 }
2973 }
2974 }
2975 }
2976 }
2978 //------------------------------final_graph_reshaping--------------------------
2979 // Final Graph Reshaping.
2980 //
2981 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
2982 // and not commoned up and forced early. Must come after regular
2983 // optimizations to avoid GVN undoing the cloning. Clone constant
2984 // inputs to Loop Phis; these will be split by the allocator anyways.
2985 // Remove Opaque nodes.
2986 // (2) Move last-uses by commutative operations to the left input to encourage
2987 // Intel update-in-place two-address operations and better register usage
2988 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
2989 // calls canonicalizing them back.
2990 // (3) Count the number of double-precision FP ops, single-precision FP ops
2991 // and call sites. On Intel, we can get correct rounding either by
2992 // forcing singles to memory (requires extra stores and loads after each
2993 // FP bytecode) or we can set a rounding mode bit (requires setting and
2994 // clearing the mode bit around call sites). The mode bit is only used
2995 // if the relative frequency of single FP ops to calls is low enough.
2996 // This is a key transform for SPEC mpeg_audio.
2997 // (4) Detect infinite loops; blobs of code reachable from above but not
2998 // below. Several of the Code_Gen algorithms fail on such code shapes,
2999 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3000 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3001 // Detection is by looking for IfNodes where only 1 projection is
3002 // reachable from below or CatchNodes missing some targets.
3003 // (5) Assert for insane oop offsets in debug mode.
3005 bool Compile::final_graph_reshaping() {
3006 // an infinite loop may have been eliminated by the optimizer,
3007 // in which case the graph will be empty.
3008 if (root()->req() == 1) {
3009 record_method_not_compilable("trivial infinite loop");
3010 return true;
3011 }
3013 Final_Reshape_Counts frc;
3015 // Visit everybody reachable!
3016 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3017 Node_Stack nstack(unique() >> 1);
3018 final_graph_reshaping_walk(nstack, root(), frc);
3020 // Check for unreachable (from below) code (i.e., infinite loops).
3021 for( uint i = 0; i < frc._tests.size(); i++ ) {
3022 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3023 // Get number of CFG targets.
3024 // Note that PCTables include exception targets after calls.
3025 uint required_outcnt = n->required_outcnt();
3026 if (n->outcnt() != required_outcnt) {
3027 // Check for a few special cases. Rethrow Nodes never take the
3028 // 'fall-thru' path, so expected kids is 1 less.
3029 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3030 if (n->in(0)->in(0)->is_Call()) {
3031 CallNode *call = n->in(0)->in(0)->as_Call();
3032 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3033 required_outcnt--; // Rethrow always has 1 less kid
3034 } else if (call->req() > TypeFunc::Parms &&
3035 call->is_CallDynamicJava()) {
3036 // Check for null receiver. In such case, the optimizer has
3037 // detected that the virtual call will always result in a null
3038 // pointer exception. The fall-through projection of this CatchNode
3039 // will not be populated.
3040 Node *arg0 = call->in(TypeFunc::Parms);
3041 if (arg0->is_Type() &&
3042 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3043 required_outcnt--;
3044 }
3045 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3046 call->req() > TypeFunc::Parms+1 &&
3047 call->is_CallStaticJava()) {
3048 // Check for negative array length. In such case, the optimizer has
3049 // detected that the allocation attempt will always result in an
3050 // exception. There is no fall-through projection of this CatchNode .
3051 Node *arg1 = call->in(TypeFunc::Parms+1);
3052 if (arg1->is_Type() &&
3053 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3054 required_outcnt--;
3055 }
3056 }
3057 }
3058 }
3059 // Recheck with a better notion of 'required_outcnt'
3060 if (n->outcnt() != required_outcnt) {
3061 record_method_not_compilable("malformed control flow");
3062 return true; // Not all targets reachable!
3063 }
3064 }
3065 // Check that I actually visited all kids. Unreached kids
3066 // must be infinite loops.
3067 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3068 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3069 record_method_not_compilable("infinite loop");
3070 return true; // Found unvisited kid; must be unreach
3071 }
3072 }
3074 // If original bytecodes contained a mixture of floats and doubles
3075 // check if the optimizer has made it homogenous, item (3).
3076 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3077 frc.get_float_count() > 32 &&
3078 frc.get_double_count() == 0 &&
3079 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3080 set_24_bit_selection_and_mode( false, true );
3081 }
3083 set_java_calls(frc.get_java_call_count());
3084 set_inner_loops(frc.get_inner_loop_count());
3086 // No infinite loops, no reason to bail out.
3087 return false;
3088 }
3090 //-----------------------------too_many_traps----------------------------------
3091 // Report if there are too many traps at the current method and bci.
3092 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3093 bool Compile::too_many_traps(ciMethod* method,
3094 int bci,
3095 Deoptimization::DeoptReason reason) {
3096 ciMethodData* md = method->method_data();
3097 if (md->is_empty()) {
3098 // Assume the trap has not occurred, or that it occurred only
3099 // because of a transient condition during start-up in the interpreter.
3100 return false;
3101 }
3102 if (md->has_trap_at(bci, reason) != 0) {
3103 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3104 // Also, if there are multiple reasons, or if there is no per-BCI record,
3105 // assume the worst.
3106 if (log())
3107 log()->elem("observe trap='%s' count='%d'",
3108 Deoptimization::trap_reason_name(reason),
3109 md->trap_count(reason));
3110 return true;
3111 } else {
3112 // Ignore method/bci and see if there have been too many globally.
3113 return too_many_traps(reason, md);
3114 }
3115 }
3117 // Less-accurate variant which does not require a method and bci.
3118 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3119 ciMethodData* logmd) {
3120 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
3121 // Too many traps globally.
3122 // Note that we use cumulative trap_count, not just md->trap_count.
3123 if (log()) {
3124 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3125 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3126 Deoptimization::trap_reason_name(reason),
3127 mcount, trap_count(reason));
3128 }
3129 return true;
3130 } else {
3131 // The coast is clear.
3132 return false;
3133 }
3134 }
3136 //--------------------------too_many_recompiles--------------------------------
3137 // Report if there are too many recompiles at the current method and bci.
3138 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3139 // Is not eager to return true, since this will cause the compiler to use
3140 // Action_none for a trap point, to avoid too many recompilations.
3141 bool Compile::too_many_recompiles(ciMethod* method,
3142 int bci,
3143 Deoptimization::DeoptReason reason) {
3144 ciMethodData* md = method->method_data();
3145 if (md->is_empty()) {
3146 // Assume the trap has not occurred, or that it occurred only
3147 // because of a transient condition during start-up in the interpreter.
3148 return false;
3149 }
3150 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3151 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3152 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3153 Deoptimization::DeoptReason per_bc_reason
3154 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3155 if ((per_bc_reason == Deoptimization::Reason_none
3156 || md->has_trap_at(bci, reason) != 0)
3157 // The trap frequency measure we care about is the recompile count:
3158 && md->trap_recompiled_at(bci)
3159 && md->overflow_recompile_count() >= bc_cutoff) {
3160 // Do not emit a trap here if it has already caused recompilations.
3161 // Also, if there are multiple reasons, or if there is no per-BCI record,
3162 // assume the worst.
3163 if (log())
3164 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3165 Deoptimization::trap_reason_name(reason),
3166 md->trap_count(reason),
3167 md->overflow_recompile_count());
3168 return true;
3169 } else if (trap_count(reason) != 0
3170 && decompile_count() >= m_cutoff) {
3171 // Too many recompiles globally, and we have seen this sort of trap.
3172 // Use cumulative decompile_count, not just md->decompile_count.
3173 if (log())
3174 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3175 Deoptimization::trap_reason_name(reason),
3176 md->trap_count(reason), trap_count(reason),
3177 md->decompile_count(), decompile_count());
3178 return true;
3179 } else {
3180 // The coast is clear.
3181 return false;
3182 }
3183 }
3186 #ifndef PRODUCT
3187 //------------------------------verify_graph_edges---------------------------
3188 // Walk the Graph and verify that there is a one-to-one correspondence
3189 // between Use-Def edges and Def-Use edges in the graph.
3190 void Compile::verify_graph_edges(bool no_dead_code) {
3191 if (VerifyGraphEdges) {
3192 ResourceArea *area = Thread::current()->resource_area();
3193 Unique_Node_List visited(area);
3194 // Call recursive graph walk to check edges
3195 _root->verify_edges(visited);
3196 if (no_dead_code) {
3197 // Now make sure that no visited node is used by an unvisited node.
3198 bool dead_nodes = 0;
3199 Unique_Node_List checked(area);
3200 while (visited.size() > 0) {
3201 Node* n = visited.pop();
3202 checked.push(n);
3203 for (uint i = 0; i < n->outcnt(); i++) {
3204 Node* use = n->raw_out(i);
3205 if (checked.member(use)) continue; // already checked
3206 if (visited.member(use)) continue; // already in the graph
3207 if (use->is_Con()) continue; // a dead ConNode is OK
3208 // At this point, we have found a dead node which is DU-reachable.
3209 if (dead_nodes++ == 0)
3210 tty->print_cr("*** Dead nodes reachable via DU edges:");
3211 use->dump(2);
3212 tty->print_cr("---");
3213 checked.push(use); // No repeats; pretend it is now checked.
3214 }
3215 }
3216 assert(dead_nodes == 0, "using nodes must be reachable from root");
3217 }
3218 }
3219 }
3220 #endif
3222 // The Compile object keeps track of failure reasons separately from the ciEnv.
3223 // This is required because there is not quite a 1-1 relation between the
3224 // ciEnv and its compilation task and the Compile object. Note that one
3225 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3226 // to backtrack and retry without subsuming loads. Other than this backtracking
3227 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3228 // by the logic in C2Compiler.
3229 void Compile::record_failure(const char* reason) {
3230 if (log() != NULL) {
3231 log()->elem("failure reason='%s' phase='compile'", reason);
3232 }
3233 if (_failure_reason == NULL) {
3234 // Record the first failure reason.
3235 _failure_reason = reason;
3236 }
3237 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3238 C->print_method(_failure_reason);
3239 }
3240 _root = NULL; // flush the graph, too
3241 }
3243 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3244 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3245 _phase_name(name), _dolog(dolog)
3246 {
3247 if (dolog) {
3248 C = Compile::current();
3249 _log = C->log();
3250 } else {
3251 C = NULL;
3252 _log = NULL;
3253 }
3254 if (_log != NULL) {
3255 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3256 _log->stamp();
3257 _log->end_head();
3258 }
3259 }
3261 Compile::TracePhase::~TracePhase() {
3263 C = Compile::current();
3264 if (_dolog) {
3265 _log = C->log();
3266 } else {
3267 _log = NULL;
3268 }
3270 #ifdef ASSERT
3271 if (PrintIdealNodeCount) {
3272 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3273 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3274 }
3276 if (VerifyIdealNodeCount) {
3277 Compile::current()->print_missing_nodes();
3278 }
3279 #endif
3281 if (_log != NULL) {
3282 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3283 }
3284 }
3286 //=============================================================================
3287 // Two Constant's are equal when the type and the value are equal.
3288 bool Compile::Constant::operator==(const Constant& other) {
3289 if (type() != other.type() ) return false;
3290 if (can_be_reused() != other.can_be_reused()) return false;
3291 // For floating point values we compare the bit pattern.
3292 switch (type()) {
3293 case T_FLOAT: return (_v._value.i == other._v._value.i);
3294 case T_LONG:
3295 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3296 case T_OBJECT:
3297 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3298 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3299 case T_METADATA: return (_v._metadata == other._v._metadata);
3300 default: ShouldNotReachHere();
3301 }
3302 return false;
3303 }
3305 static int type_to_size_in_bytes(BasicType t) {
3306 switch (t) {
3307 case T_LONG: return sizeof(jlong );
3308 case T_FLOAT: return sizeof(jfloat );
3309 case T_DOUBLE: return sizeof(jdouble);
3310 case T_METADATA: return sizeof(Metadata*);
3311 // We use T_VOID as marker for jump-table entries (labels) which
3312 // need an internal word relocation.
3313 case T_VOID:
3314 case T_ADDRESS:
3315 case T_OBJECT: return sizeof(jobject);
3316 }
3318 ShouldNotReachHere();
3319 return -1;
3320 }
3322 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3323 // sort descending
3324 if (a->freq() > b->freq()) return -1;
3325 if (a->freq() < b->freq()) return 1;
3326 return 0;
3327 }
3329 void Compile::ConstantTable::calculate_offsets_and_size() {
3330 // First, sort the array by frequencies.
3331 _constants.sort(qsort_comparator);
3333 #ifdef ASSERT
3334 // Make sure all jump-table entries were sorted to the end of the
3335 // array (they have a negative frequency).
3336 bool found_void = false;
3337 for (int i = 0; i < _constants.length(); i++) {
3338 Constant con = _constants.at(i);
3339 if (con.type() == T_VOID)
3340 found_void = true; // jump-tables
3341 else
3342 assert(!found_void, "wrong sorting");
3343 }
3344 #endif
3346 int offset = 0;
3347 for (int i = 0; i < _constants.length(); i++) {
3348 Constant* con = _constants.adr_at(i);
3350 // Align offset for type.
3351 int typesize = type_to_size_in_bytes(con->type());
3352 offset = align_size_up(offset, typesize);
3353 con->set_offset(offset); // set constant's offset
3355 if (con->type() == T_VOID) {
3356 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3357 offset = offset + typesize * n->outcnt(); // expand jump-table
3358 } else {
3359 offset = offset + typesize;
3360 }
3361 }
3363 // Align size up to the next section start (which is insts; see
3364 // CodeBuffer::align_at_start).
3365 assert(_size == -1, "already set?");
3366 _size = align_size_up(offset, CodeEntryAlignment);
3367 }
3369 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3370 MacroAssembler _masm(&cb);
3371 for (int i = 0; i < _constants.length(); i++) {
3372 Constant con = _constants.at(i);
3373 address constant_addr;
3374 switch (con.type()) {
3375 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3376 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3377 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3378 case T_OBJECT: {
3379 jobject obj = con.get_jobject();
3380 int oop_index = _masm.oop_recorder()->find_index(obj);
3381 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3382 break;
3383 }
3384 case T_ADDRESS: {
3385 address addr = (address) con.get_jobject();
3386 constant_addr = _masm.address_constant(addr);
3387 break;
3388 }
3389 // We use T_VOID as marker for jump-table entries (labels) which
3390 // need an internal word relocation.
3391 case T_VOID: {
3392 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3393 // Fill the jump-table with a dummy word. The real value is
3394 // filled in later in fill_jump_table.
3395 address dummy = (address) n;
3396 constant_addr = _masm.address_constant(dummy);
3397 // Expand jump-table
3398 for (uint i = 1; i < n->outcnt(); i++) {
3399 address temp_addr = _masm.address_constant(dummy + i);
3400 assert(temp_addr, "consts section too small");
3401 }
3402 break;
3403 }
3404 case T_METADATA: {
3405 Metadata* obj = con.get_metadata();
3406 int metadata_index = _masm.oop_recorder()->find_index(obj);
3407 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3408 break;
3409 }
3410 default: ShouldNotReachHere();
3411 }
3412 assert(constant_addr, "consts section too small");
3413 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3414 }
3415 }
3417 int Compile::ConstantTable::find_offset(Constant& con) const {
3418 int idx = _constants.find(con);
3419 assert(idx != -1, "constant must be in constant table");
3420 int offset = _constants.at(idx).offset();
3421 assert(offset != -1, "constant table not emitted yet?");
3422 return offset;
3423 }
3425 void Compile::ConstantTable::add(Constant& con) {
3426 if (con.can_be_reused()) {
3427 int idx = _constants.find(con);
3428 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3429 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3430 return;
3431 }
3432 }
3433 (void) _constants.append(con);
3434 }
3436 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3437 Block* b = Compile::current()->cfg()->_bbs[n->_idx];
3438 Constant con(type, value, b->_freq);
3439 add(con);
3440 return con;
3441 }
3443 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3444 Constant con(metadata);
3445 add(con);
3446 return con;
3447 }
3449 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3450 jvalue value;
3451 BasicType type = oper->type()->basic_type();
3452 switch (type) {
3453 case T_LONG: value.j = oper->constantL(); break;
3454 case T_FLOAT: value.f = oper->constantF(); break;
3455 case T_DOUBLE: value.d = oper->constantD(); break;
3456 case T_OBJECT:
3457 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3458 case T_METADATA: return add((Metadata*)oper->constant()); break;
3459 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3460 }
3461 return add(n, type, value);
3462 }
3464 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3465 jvalue value;
3466 // We can use the node pointer here to identify the right jump-table
3467 // as this method is called from Compile::Fill_buffer right before
3468 // the MachNodes are emitted and the jump-table is filled (means the
3469 // MachNode pointers do not change anymore).
3470 value.l = (jobject) n;
3471 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3472 add(con);
3473 return con;
3474 }
3476 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3477 // If called from Compile::scratch_emit_size do nothing.
3478 if (Compile::current()->in_scratch_emit_size()) return;
3480 assert(labels.is_nonempty(), "must be");
3481 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3483 // Since MachConstantNode::constant_offset() also contains
3484 // table_base_offset() we need to subtract the table_base_offset()
3485 // to get the plain offset into the constant table.
3486 int offset = n->constant_offset() - table_base_offset();
3488 MacroAssembler _masm(&cb);
3489 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3491 for (uint i = 0; i < n->outcnt(); i++) {
3492 address* constant_addr = &jump_table_base[i];
3493 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)));
3494 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3495 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3496 }
3497 }
3499 void Compile::dump_inlining() {
3500 if (PrintInlining) {
3501 // Print inlining message for candidates that we couldn't inline
3502 // for lack of space or non constant receiver
3503 for (int i = 0; i < _late_inlines.length(); i++) {
3504 CallGenerator* cg = _late_inlines.at(i);
3505 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3506 }
3507 Unique_Node_List useful;
3508 useful.push(root());
3509 for (uint next = 0; next < useful.size(); ++next) {
3510 Node* n = useful.at(next);
3511 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3512 CallNode* call = n->as_Call();
3513 CallGenerator* cg = call->generator();
3514 cg->print_inlining_late("receiver not constant");
3515 }
3516 uint max = n->len();
3517 for ( uint i = 0; i < max; ++i ) {
3518 Node *m = n->in(i);
3519 if ( m == NULL ) continue;
3520 useful.push(m);
3521 }
3522 }
3523 for (int i = 0; i < _print_inlining_list->length(); i++) {
3524 tty->print(_print_inlining_list->at(i).ss()->as_string());
3525 }
3526 }
3527 }