Mon, 25 May 2020 14:24:27 +0800
8244407: JVM crashes after transformation in C2 IdealLoopTree::split_fall_in
Reviewed-by: thartmann, kvn, andrew
Contributed-by: zhouyong44@huawei.com
1 /*
2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "compiler/disassembler.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "jfr/jfrEvents.hpp"
36 #include "opto/addnode.hpp"
37 #include "opto/block.hpp"
38 #include "opto/c2compiler.hpp"
39 #include "opto/callGenerator.hpp"
40 #include "opto/callnode.hpp"
41 #include "opto/cfgnode.hpp"
42 #include "opto/chaitin.hpp"
43 #include "opto/compile.hpp"
44 #include "opto/connode.hpp"
45 #include "opto/divnode.hpp"
46 #include "opto/escape.hpp"
47 #include "opto/idealGraphPrinter.hpp"
48 #include "opto/loopnode.hpp"
49 #include "opto/machnode.hpp"
50 #include "opto/macro.hpp"
51 #include "opto/matcher.hpp"
52 #include "opto/mathexactnode.hpp"
53 #include "opto/memnode.hpp"
54 #include "opto/mulnode.hpp"
55 #include "opto/node.hpp"
56 #include "opto/opcodes.hpp"
57 #include "opto/output.hpp"
58 #include "opto/parse.hpp"
59 #include "opto/phaseX.hpp"
60 #include "opto/rootnode.hpp"
61 #include "opto/runtime.hpp"
62 #include "opto/stringopts.hpp"
63 #include "opto/type.hpp"
64 #include "opto/vectornode.hpp"
65 #include "runtime/arguments.hpp"
66 #include "runtime/signature.hpp"
67 #include "runtime/stubRoutines.hpp"
68 #include "runtime/timer.hpp"
69 #include "utilities/copy.hpp"
70 #if defined AD_MD_HPP
71 # include AD_MD_HPP
72 #elif defined TARGET_ARCH_MODEL_x86_32
73 # include "adfiles/ad_x86_32.hpp"
74 #elif defined TARGET_ARCH_MODEL_x86_64
75 # include "adfiles/ad_x86_64.hpp"
76 #elif defined TARGET_ARCH_MODEL_sparc
77 # include "adfiles/ad_sparc.hpp"
78 #elif defined TARGET_ARCH_MODEL_zero
79 # include "adfiles/ad_zero.hpp"
80 #elif defined TARGET_ARCH_MODEL_ppc_64
81 # include "adfiles/ad_ppc_64.hpp"
82 #endif
84 // -------------------- Compile::mach_constant_base_node -----------------------
85 // Constant table base node singleton.
86 MachConstantBaseNode* Compile::mach_constant_base_node() {
87 if (_mach_constant_base_node == NULL) {
88 _mach_constant_base_node = new (C) MachConstantBaseNode();
89 _mach_constant_base_node->add_req(C->root());
90 }
91 return _mach_constant_base_node;
92 }
95 /// Support for intrinsics.
97 // Return the index at which m must be inserted (or already exists).
98 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
99 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
100 #ifdef ASSERT
101 for (int i = 1; i < _intrinsics->length(); i++) {
102 CallGenerator* cg1 = _intrinsics->at(i-1);
103 CallGenerator* cg2 = _intrinsics->at(i);
104 assert(cg1->method() != cg2->method()
105 ? cg1->method() < cg2->method()
106 : cg1->is_virtual() < cg2->is_virtual(),
107 "compiler intrinsics list must stay sorted");
108 }
109 #endif
110 // Binary search sorted list, in decreasing intervals [lo, hi].
111 int lo = 0, hi = _intrinsics->length()-1;
112 while (lo <= hi) {
113 int mid = (uint)(hi + lo) / 2;
114 ciMethod* mid_m = _intrinsics->at(mid)->method();
115 if (m < mid_m) {
116 hi = mid-1;
117 } else if (m > mid_m) {
118 lo = mid+1;
119 } else {
120 // look at minor sort key
121 bool mid_virt = _intrinsics->at(mid)->is_virtual();
122 if (is_virtual < mid_virt) {
123 hi = mid-1;
124 } else if (is_virtual > mid_virt) {
125 lo = mid+1;
126 } else {
127 return mid; // exact match
128 }
129 }
130 }
131 return lo; // inexact match
132 }
134 void Compile::register_intrinsic(CallGenerator* cg) {
135 if (_intrinsics == NULL) {
136 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
137 }
138 // This code is stolen from ciObjectFactory::insert.
139 // Really, GrowableArray should have methods for
140 // insert_at, remove_at, and binary_search.
141 int len = _intrinsics->length();
142 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
143 if (index == len) {
144 _intrinsics->append(cg);
145 } else {
146 #ifdef ASSERT
147 CallGenerator* oldcg = _intrinsics->at(index);
148 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
149 #endif
150 _intrinsics->append(_intrinsics->at(len-1));
151 int pos;
152 for (pos = len-2; pos >= index; pos--) {
153 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
154 }
155 _intrinsics->at_put(index, cg);
156 }
157 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
158 }
160 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
161 assert(m->is_loaded(), "don't try this on unloaded methods");
162 if (_intrinsics != NULL) {
163 int index = intrinsic_insertion_index(m, is_virtual);
164 if (index < _intrinsics->length()
165 && _intrinsics->at(index)->method() == m
166 && _intrinsics->at(index)->is_virtual() == is_virtual) {
167 return _intrinsics->at(index);
168 }
169 }
170 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
171 if (m->intrinsic_id() != vmIntrinsics::_none &&
172 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
173 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
174 if (cg != NULL) {
175 // Save it for next time:
176 register_intrinsic(cg);
177 return cg;
178 } else {
179 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
180 }
181 }
182 return NULL;
183 }
185 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
186 // in library_call.cpp.
189 #ifndef PRODUCT
190 // statistics gathering...
192 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
193 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
195 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
196 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
197 int oflags = _intrinsic_hist_flags[id];
198 assert(flags != 0, "what happened?");
199 if (is_virtual) {
200 flags |= _intrinsic_virtual;
201 }
202 bool changed = (flags != oflags);
203 if ((flags & _intrinsic_worked) != 0) {
204 juint count = (_intrinsic_hist_count[id] += 1);
205 if (count == 1) {
206 changed = true; // first time
207 }
208 // increment the overall count also:
209 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
210 }
211 if (changed) {
212 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
213 // Something changed about the intrinsic's virtuality.
214 if ((flags & _intrinsic_virtual) != 0) {
215 // This is the first use of this intrinsic as a virtual call.
216 if (oflags != 0) {
217 // We already saw it as a non-virtual, so note both cases.
218 flags |= _intrinsic_both;
219 }
220 } else if ((oflags & _intrinsic_both) == 0) {
221 // This is the first use of this intrinsic as a non-virtual
222 flags |= _intrinsic_both;
223 }
224 }
225 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
226 }
227 // update the overall flags also:
228 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
229 return changed;
230 }
232 static char* format_flags(int flags, char* buf) {
233 buf[0] = 0;
234 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
235 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
236 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
237 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
238 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
239 if (buf[0] == 0) strcat(buf, ",");
240 assert(buf[0] == ',', "must be");
241 return &buf[1];
242 }
244 void Compile::print_intrinsic_statistics() {
245 char flagsbuf[100];
246 ttyLocker ttyl;
247 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
248 tty->print_cr("Compiler intrinsic usage:");
249 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
250 if (total == 0) total = 1; // avoid div0 in case of no successes
251 #define PRINT_STAT_LINE(name, c, f) \
252 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
253 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
254 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
255 int flags = _intrinsic_hist_flags[id];
256 juint count = _intrinsic_hist_count[id];
257 if ((flags | count) != 0) {
258 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
259 }
260 }
261 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
262 if (xtty != NULL) xtty->tail("statistics");
263 }
265 void Compile::print_statistics() {
266 { ttyLocker ttyl;
267 if (xtty != NULL) xtty->head("statistics type='opto'");
268 Parse::print_statistics();
269 PhaseCCP::print_statistics();
270 PhaseRegAlloc::print_statistics();
271 Scheduling::print_statistics();
272 PhasePeephole::print_statistics();
273 PhaseIdealLoop::print_statistics();
274 if (xtty != NULL) xtty->tail("statistics");
275 }
276 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
277 // put this under its own <statistics> element.
278 print_intrinsic_statistics();
279 }
280 }
281 #endif //PRODUCT
283 // Support for bundling info
284 Bundle* Compile::node_bundling(const Node *n) {
285 assert(valid_bundle_info(n), "oob");
286 return &_node_bundling_base[n->_idx];
287 }
289 bool Compile::valid_bundle_info(const Node *n) {
290 return (_node_bundling_limit > n->_idx);
291 }
294 void Compile::gvn_replace_by(Node* n, Node* nn) {
295 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
296 Node* use = n->last_out(i);
297 bool is_in_table = initial_gvn()->hash_delete(use);
298 uint uses_found = 0;
299 for (uint j = 0; j < use->len(); j++) {
300 if (use->in(j) == n) {
301 if (j < use->req())
302 use->set_req(j, nn);
303 else
304 use->set_prec(j, nn);
305 uses_found++;
306 }
307 }
308 if (is_in_table) {
309 // reinsert into table
310 initial_gvn()->hash_find_insert(use);
311 }
312 record_for_igvn(use);
313 i -= uses_found; // we deleted 1 or more copies of this edge
314 }
315 }
318 static inline bool not_a_node(const Node* n) {
319 if (n == NULL) return true;
320 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
321 if (*(address*)n == badAddress) return true; // kill by Node::destruct
322 return false;
323 }
325 // Identify all nodes that are reachable from below, useful.
326 // Use breadth-first pass that records state in a Unique_Node_List,
327 // recursive traversal is slower.
328 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
329 int estimated_worklist_size = live_nodes();
330 useful.map( estimated_worklist_size, NULL ); // preallocate space
332 // Initialize worklist
333 if (root() != NULL) { useful.push(root()); }
334 // If 'top' is cached, declare it useful to preserve cached node
335 if( cached_top_node() ) { useful.push(cached_top_node()); }
337 // Push all useful nodes onto the list, breadthfirst
338 for( uint next = 0; next < useful.size(); ++next ) {
339 assert( next < unique(), "Unique useful nodes < total nodes");
340 Node *n = useful.at(next);
341 uint max = n->len();
342 for( uint i = 0; i < max; ++i ) {
343 Node *m = n->in(i);
344 if (not_a_node(m)) continue;
345 useful.push(m);
346 }
347 }
348 }
350 // Update dead_node_list with any missing dead nodes using useful
351 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
352 void Compile::update_dead_node_list(Unique_Node_List &useful) {
353 uint max_idx = unique();
354 VectorSet& useful_node_set = useful.member_set();
356 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
357 // If node with index node_idx is not in useful set,
358 // mark it as dead in dead node list.
359 if (! useful_node_set.test(node_idx) ) {
360 record_dead_node(node_idx);
361 }
362 }
363 }
365 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
366 int shift = 0;
367 for (int i = 0; i < inlines->length(); i++) {
368 CallGenerator* cg = inlines->at(i);
369 CallNode* call = cg->call_node();
370 if (shift > 0) {
371 inlines->at_put(i-shift, cg);
372 }
373 if (!useful.member(call)) {
374 shift++;
375 }
376 }
377 inlines->trunc_to(inlines->length()-shift);
378 }
380 // Disconnect all useless nodes by disconnecting those at the boundary.
381 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
382 uint next = 0;
383 while (next < useful.size()) {
384 Node *n = useful.at(next++);
385 if (n->is_SafePoint()) {
386 // We're done with a parsing phase. Replaced nodes are not valid
387 // beyond that point.
388 n->as_SafePoint()->delete_replaced_nodes();
389 }
390 // Use raw traversal of out edges since this code removes out edges
391 int max = n->outcnt();
392 for (int j = 0; j < max; ++j) {
393 Node* child = n->raw_out(j);
394 if (! useful.member(child)) {
395 assert(!child->is_top() || child != top(),
396 "If top is cached in Compile object it is in useful list");
397 // Only need to remove this out-edge to the useless node
398 n->raw_del_out(j);
399 --j;
400 --max;
401 }
402 }
403 if (n->outcnt() == 1 && n->has_special_unique_user()) {
404 record_for_igvn(n->unique_out());
405 }
406 }
407 // Remove useless macro and predicate opaq nodes
408 for (int i = C->macro_count()-1; i >= 0; i--) {
409 Node* n = C->macro_node(i);
410 if (!useful.member(n)) {
411 remove_macro_node(n);
412 }
413 }
414 // Remove useless CastII nodes with range check dependency
415 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
416 Node* cast = range_check_cast_node(i);
417 if (!useful.member(cast)) {
418 remove_range_check_cast(cast);
419 }
420 }
421 // Remove useless expensive node
422 for (int i = C->expensive_count()-1; i >= 0; i--) {
423 Node* n = C->expensive_node(i);
424 if (!useful.member(n)) {
425 remove_expensive_node(n);
426 }
427 }
428 // clean up the late inline lists
429 remove_useless_late_inlines(&_string_late_inlines, useful);
430 remove_useless_late_inlines(&_boxing_late_inlines, useful);
431 remove_useless_late_inlines(&_late_inlines, useful);
432 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
433 }
435 //------------------------------frame_size_in_words-----------------------------
436 // frame_slots in units of words
437 int Compile::frame_size_in_words() const {
438 // shift is 0 in LP32 and 1 in LP64
439 const int shift = (LogBytesPerWord - LogBytesPerInt);
440 int words = _frame_slots >> shift;
441 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
442 return words;
443 }
445 // To bang the stack of this compiled method we use the stack size
446 // that the interpreter would need in case of a deoptimization. This
447 // removes the need to bang the stack in the deoptimization blob which
448 // in turn simplifies stack overflow handling.
449 int Compile::bang_size_in_bytes() const {
450 return MAX2(_interpreter_frame_size, frame_size_in_bytes());
451 }
453 // ============================================================================
454 //------------------------------CompileWrapper---------------------------------
455 class CompileWrapper : public StackObj {
456 Compile *const _compile;
457 public:
458 CompileWrapper(Compile* compile);
460 ~CompileWrapper();
461 };
463 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
464 // the Compile* pointer is stored in the current ciEnv:
465 ciEnv* env = compile->env();
466 assert(env == ciEnv::current(), "must already be a ciEnv active");
467 assert(env->compiler_data() == NULL, "compile already active?");
468 env->set_compiler_data(compile);
469 assert(compile == Compile::current(), "sanity");
471 compile->set_type_dict(NULL);
472 compile->set_type_hwm(NULL);
473 compile->set_type_last_size(0);
474 compile->set_last_tf(NULL, NULL);
475 compile->set_indexSet_arena(NULL);
476 compile->set_indexSet_free_block_list(NULL);
477 compile->init_type_arena();
478 Type::Initialize(compile);
479 _compile->set_scratch_buffer_blob(NULL);
480 _compile->begin_method();
481 }
482 CompileWrapper::~CompileWrapper() {
483 _compile->end_method();
484 if (_compile->scratch_buffer_blob() != NULL)
485 BufferBlob::free(_compile->scratch_buffer_blob());
486 _compile->env()->set_compiler_data(NULL);
487 }
490 //----------------------------print_compile_messages---------------------------
491 void Compile::print_compile_messages() {
492 #ifndef PRODUCT
493 // Check if recompiling
494 if (_subsume_loads == false && PrintOpto) {
495 // Recompiling without allowing machine instructions to subsume loads
496 tty->print_cr("*********************************************************");
497 tty->print_cr("** Bailout: Recompile without subsuming loads **");
498 tty->print_cr("*********************************************************");
499 }
500 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
501 // Recompiling without escape analysis
502 tty->print_cr("*********************************************************");
503 tty->print_cr("** Bailout: Recompile without escape analysis **");
504 tty->print_cr("*********************************************************");
505 }
506 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
507 // Recompiling without boxing elimination
508 tty->print_cr("*********************************************************");
509 tty->print_cr("** Bailout: Recompile without boxing elimination **");
510 tty->print_cr("*********************************************************");
511 }
512 if (env()->break_at_compile()) {
513 // Open the debugger when compiling this method.
514 tty->print("### Breaking when compiling: ");
515 method()->print_short_name();
516 tty->cr();
517 BREAKPOINT;
518 }
520 if( PrintOpto ) {
521 if (is_osr_compilation()) {
522 tty->print("[OSR]%3d", _compile_id);
523 } else {
524 tty->print("%3d", _compile_id);
525 }
526 }
527 #endif
528 }
531 //-----------------------init_scratch_buffer_blob------------------------------
532 // Construct a temporary BufferBlob and cache it for this compile.
533 void Compile::init_scratch_buffer_blob(int const_size) {
534 // If there is already a scratch buffer blob allocated and the
535 // constant section is big enough, use it. Otherwise free the
536 // current and allocate a new one.
537 BufferBlob* blob = scratch_buffer_blob();
538 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
539 // Use the current blob.
540 } else {
541 if (blob != NULL) {
542 BufferBlob::free(blob);
543 }
545 ResourceMark rm;
546 _scratch_const_size = const_size;
547 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
548 blob = BufferBlob::create("Compile::scratch_buffer", size);
549 // Record the buffer blob for next time.
550 set_scratch_buffer_blob(blob);
551 // Have we run out of code space?
552 if (scratch_buffer_blob() == NULL) {
553 // Let CompilerBroker disable further compilations.
554 record_failure("Not enough space for scratch buffer in CodeCache");
555 return;
556 }
557 }
559 // Initialize the relocation buffers
560 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
561 set_scratch_locs_memory(locs_buf);
562 }
565 //-----------------------scratch_emit_size-------------------------------------
566 // Helper function that computes size by emitting code
567 uint Compile::scratch_emit_size(const Node* n) {
568 // Start scratch_emit_size section.
569 set_in_scratch_emit_size(true);
571 // Emit into a trash buffer and count bytes emitted.
572 // This is a pretty expensive way to compute a size,
573 // but it works well enough if seldom used.
574 // All common fixed-size instructions are given a size
575 // method by the AD file.
576 // Note that the scratch buffer blob and locs memory are
577 // allocated at the beginning of the compile task, and
578 // may be shared by several calls to scratch_emit_size.
579 // The allocation of the scratch buffer blob is particularly
580 // expensive, since it has to grab the code cache lock.
581 BufferBlob* blob = this->scratch_buffer_blob();
582 assert(blob != NULL, "Initialize BufferBlob at start");
583 assert(blob->size() > MAX_inst_size, "sanity");
584 relocInfo* locs_buf = scratch_locs_memory();
585 address blob_begin = blob->content_begin();
586 address blob_end = (address)locs_buf;
587 assert(blob->content_contains(blob_end), "sanity");
588 CodeBuffer buf(blob_begin, blob_end - blob_begin);
589 buf.initialize_consts_size(_scratch_const_size);
590 buf.initialize_stubs_size(MAX_stubs_size);
591 assert(locs_buf != NULL, "sanity");
592 int lsize = MAX_locs_size / 3;
593 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
594 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
595 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
597 // Do the emission.
599 Label fakeL; // Fake label for branch instructions.
600 Label* saveL = NULL;
601 uint save_bnum = 0;
602 bool is_branch = n->is_MachBranch();
603 if (is_branch) {
604 MacroAssembler masm(&buf);
605 masm.bind(fakeL);
606 n->as_MachBranch()->save_label(&saveL, &save_bnum);
607 n->as_MachBranch()->label_set(&fakeL, 0);
608 }
609 n->emit(buf, this->regalloc());
611 // Emitting into the scratch buffer should not fail
612 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason()));
614 if (is_branch) // Restore label.
615 n->as_MachBranch()->label_set(saveL, save_bnum);
617 // End scratch_emit_size section.
618 set_in_scratch_emit_size(false);
620 return buf.insts_size();
621 }
624 // ============================================================================
625 //------------------------------Compile standard-------------------------------
626 debug_only( int Compile::_debug_idx = 100000; )
628 // Compile a method. entry_bci is -1 for normal compilations and indicates
629 // the continuation bci for on stack replacement.
632 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
633 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
634 : Phase(Compiler),
635 _env(ci_env),
636 _log(ci_env->log()),
637 _compile_id(ci_env->compile_id()),
638 _save_argument_registers(false),
639 _stub_name(NULL),
640 _stub_function(NULL),
641 _stub_entry_point(NULL),
642 _method(target),
643 _entry_bci(osr_bci),
644 _initial_gvn(NULL),
645 _for_igvn(NULL),
646 _warm_calls(NULL),
647 _subsume_loads(subsume_loads),
648 _do_escape_analysis(do_escape_analysis),
649 _eliminate_boxing(eliminate_boxing),
650 _failure_reason(NULL),
651 _code_buffer("Compile::Fill_buffer"),
652 _orig_pc_slot(0),
653 _orig_pc_slot_offset_in_bytes(0),
654 _has_method_handle_invokes(false),
655 _mach_constant_base_node(NULL),
656 _node_bundling_limit(0),
657 _node_bundling_base(NULL),
658 _java_calls(0),
659 _inner_loops(0),
660 _scratch_const_size(-1),
661 _in_scratch_emit_size(false),
662 _dead_node_list(comp_arena()),
663 _dead_node_count(0),
664 #ifndef PRODUCT
665 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
666 _in_dump_cnt(0),
667 _printer(IdealGraphPrinter::printer()),
668 #endif
669 _congraph(NULL),
670 _comp_arena(mtCompiler),
671 _node_arena(mtCompiler),
672 _old_arena(mtCompiler),
673 _Compile_types(mtCompiler),
674 _replay_inline_data(NULL),
675 _late_inlines(comp_arena(), 2, 0, NULL),
676 _string_late_inlines(comp_arena(), 2, 0, NULL),
677 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
678 _late_inlines_pos(0),
679 _number_of_mh_late_inlines(0),
680 _inlining_progress(false),
681 _inlining_incrementally(false),
682 _print_inlining_list(NULL),
683 _print_inlining_idx(0),
684 _interpreter_frame_size(0),
685 _max_node_limit(MaxNodeLimit) {
686 C = this;
688 CompileWrapper cw(this);
689 #ifndef PRODUCT
690 if (TimeCompiler2) {
691 tty->print(" ");
692 target->holder()->name()->print();
693 tty->print(".");
694 target->print_short_name();
695 tty->print(" ");
696 }
697 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
698 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
699 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
700 if (!print_opto_assembly) {
701 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
702 if (print_assembly && !Disassembler::can_decode()) {
703 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
704 print_opto_assembly = true;
705 }
706 }
707 set_print_assembly(print_opto_assembly);
708 set_parsed_irreducible_loop(false);
710 if (method()->has_option("ReplayInline")) {
711 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
712 }
713 #endif
714 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
715 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
716 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
718 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
719 // Make sure the method being compiled gets its own MDO,
720 // so we can at least track the decompile_count().
721 // Need MDO to record RTM code generation state.
722 method()->ensure_method_data();
723 }
725 Init(::AliasLevel);
728 print_compile_messages();
730 _ilt = InlineTree::build_inline_tree_root();
732 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
733 assert(num_alias_types() >= AliasIdxRaw, "");
735 #define MINIMUM_NODE_HASH 1023
736 // Node list that Iterative GVN will start with
737 Unique_Node_List for_igvn(comp_arena());
738 set_for_igvn(&for_igvn);
740 // GVN that will be run immediately on new nodes
741 uint estimated_size = method()->code_size()*4+64;
742 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
743 PhaseGVN gvn(node_arena(), estimated_size);
744 set_initial_gvn(&gvn);
746 if (print_inlining() || print_intrinsics()) {
747 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
748 }
749 { // Scope for timing the parser
750 TracePhase t3("parse", &_t_parser, true);
752 // Put top into the hash table ASAP.
753 initial_gvn()->transform_no_reclaim(top());
755 // Set up tf(), start(), and find a CallGenerator.
756 CallGenerator* cg = NULL;
757 if (is_osr_compilation()) {
758 const TypeTuple *domain = StartOSRNode::osr_domain();
759 const TypeTuple *range = TypeTuple::make_range(method()->signature());
760 init_tf(TypeFunc::make(domain, range));
761 StartNode* s = new (this) StartOSRNode(root(), domain);
762 initial_gvn()->set_type_bottom(s);
763 init_start(s);
764 cg = CallGenerator::for_osr(method(), entry_bci());
765 } else {
766 // Normal case.
767 init_tf(TypeFunc::make(method()));
768 StartNode* s = new (this) StartNode(root(), tf()->domain());
769 initial_gvn()->set_type_bottom(s);
770 init_start(s);
771 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
772 // With java.lang.ref.reference.get() we must go through the
773 // intrinsic when G1 is enabled - even when get() is the root
774 // method of the compile - so that, if necessary, the value in
775 // the referent field of the reference object gets recorded by
776 // the pre-barrier code.
777 // Specifically, if G1 is enabled, the value in the referent
778 // field is recorded by the G1 SATB pre barrier. This will
779 // result in the referent being marked live and the reference
780 // object removed from the list of discovered references during
781 // reference processing.
782 cg = find_intrinsic(method(), false);
783 }
784 if (cg == NULL) {
785 float past_uses = method()->interpreter_invocation_count();
786 float expected_uses = past_uses;
787 cg = CallGenerator::for_inline(method(), expected_uses);
788 }
789 }
790 if (failing()) return;
791 if (cg == NULL) {
792 record_method_not_compilable_all_tiers("cannot parse method");
793 return;
794 }
795 JVMState* jvms = build_start_state(start(), tf());
796 if ((jvms = cg->generate(jvms)) == NULL) {
797 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
798 record_method_not_compilable("method parse failed");
799 }
800 return;
801 }
802 GraphKit kit(jvms);
804 if (!kit.stopped()) {
805 // Accept return values, and transfer control we know not where.
806 // This is done by a special, unique ReturnNode bound to root.
807 return_values(kit.jvms());
808 }
810 if (kit.has_exceptions()) {
811 // Any exceptions that escape from this call must be rethrown
812 // to whatever caller is dynamically above us on the stack.
813 // This is done by a special, unique RethrowNode bound to root.
814 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
815 }
817 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
819 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
820 inline_string_calls(true);
821 }
823 if (failing()) return;
825 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
827 // Remove clutter produced by parsing.
828 if (!failing()) {
829 ResourceMark rm;
830 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
831 }
832 }
834 // Note: Large methods are capped off in do_one_bytecode().
835 if (failing()) return;
837 // After parsing, node notes are no longer automagic.
838 // They must be propagated by register_new_node_with_optimizer(),
839 // clone(), or the like.
840 set_default_node_notes(NULL);
842 for (;;) {
843 int successes = Inline_Warm();
844 if (failing()) return;
845 if (successes == 0) break;
846 }
848 // Drain the list.
849 Finish_Warm();
850 #ifndef PRODUCT
851 if (_printer) {
852 _printer->print_inlining(this);
853 }
854 #endif
856 if (failing()) return;
857 NOT_PRODUCT( verify_graph_edges(); )
859 // Now optimize
860 Optimize();
861 if (failing()) return;
862 NOT_PRODUCT( verify_graph_edges(); )
864 #ifndef PRODUCT
865 if (PrintIdeal) {
866 ttyLocker ttyl; // keep the following output all in one block
867 // This output goes directly to the tty, not the compiler log.
868 // To enable tools to match it up with the compilation activity,
869 // be sure to tag this tty output with the compile ID.
870 if (xtty != NULL) {
871 xtty->head("ideal compile_id='%d'%s", compile_id(),
872 is_osr_compilation() ? " compile_kind='osr'" :
873 "");
874 }
875 root()->dump(9999);
876 if (xtty != NULL) {
877 xtty->tail("ideal");
878 }
879 }
880 #endif
882 NOT_PRODUCT( verify_barriers(); )
884 // Dump compilation data to replay it.
885 if (method()->has_option("DumpReplay")) {
886 env()->dump_replay_data(_compile_id);
887 }
888 if (method()->has_option("DumpInline") && (ilt() != NULL)) {
889 env()->dump_inline_data(_compile_id);
890 }
892 // Now that we know the size of all the monitors we can add a fixed slot
893 // for the original deopt pc.
895 _orig_pc_slot = fixed_slots();
896 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
897 set_fixed_slots(next_slot);
899 // Compute when to use implicit null checks. Used by matching trap based
900 // nodes and NullCheck optimization.
901 set_allowed_deopt_reasons();
903 // Now generate code
904 Code_Gen();
905 if (failing()) return;
907 // Check if we want to skip execution of all compiled code.
908 {
909 #ifndef PRODUCT
910 if (OptoNoExecute) {
911 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
912 return;
913 }
914 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
915 #endif
917 if (is_osr_compilation()) {
918 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
919 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
920 } else {
921 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
922 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
923 }
925 env()->register_method(_method, _entry_bci,
926 &_code_offsets,
927 _orig_pc_slot_offset_in_bytes,
928 code_buffer(),
929 frame_size_in_words(), _oop_map_set,
930 &_handler_table, &_inc_table,
931 compiler,
932 env()->comp_level(),
933 has_unsafe_access(),
934 SharedRuntime::is_wide_vector(max_vector_size()),
935 rtm_state()
936 );
938 if (log() != NULL) // Print code cache state into compiler log
939 log()->code_cache_state();
940 }
941 }
943 //------------------------------Compile----------------------------------------
944 // Compile a runtime stub
945 Compile::Compile( ciEnv* ci_env,
946 TypeFunc_generator generator,
947 address stub_function,
948 const char *stub_name,
949 int is_fancy_jump,
950 bool pass_tls,
951 bool save_arg_registers,
952 bool return_pc )
953 : Phase(Compiler),
954 _env(ci_env),
955 _log(ci_env->log()),
956 _compile_id(0),
957 _save_argument_registers(save_arg_registers),
958 _method(NULL),
959 _stub_name(stub_name),
960 _stub_function(stub_function),
961 _stub_entry_point(NULL),
962 _entry_bci(InvocationEntryBci),
963 _initial_gvn(NULL),
964 _for_igvn(NULL),
965 _warm_calls(NULL),
966 _orig_pc_slot(0),
967 _orig_pc_slot_offset_in_bytes(0),
968 _subsume_loads(true),
969 _do_escape_analysis(false),
970 _eliminate_boxing(false),
971 _failure_reason(NULL),
972 _code_buffer("Compile::Fill_buffer"),
973 _has_method_handle_invokes(false),
974 _mach_constant_base_node(NULL),
975 _node_bundling_limit(0),
976 _node_bundling_base(NULL),
977 _java_calls(0),
978 _inner_loops(0),
979 #ifndef PRODUCT
980 _trace_opto_output(TraceOptoOutput),
981 _in_dump_cnt(0),
982 _printer(NULL),
983 #endif
984 _comp_arena(mtCompiler),
985 _node_arena(mtCompiler),
986 _old_arena(mtCompiler),
987 _Compile_types(mtCompiler),
988 _dead_node_list(comp_arena()),
989 _dead_node_count(0),
990 _congraph(NULL),
991 _replay_inline_data(NULL),
992 _number_of_mh_late_inlines(0),
993 _inlining_progress(false),
994 _inlining_incrementally(false),
995 _print_inlining_list(NULL),
996 _print_inlining_idx(0),
997 _allowed_reasons(0),
998 _interpreter_frame_size(0),
999 _max_node_limit(MaxNodeLimit) {
1000 C = this;
1002 #ifndef PRODUCT
1003 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
1004 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
1005 set_print_assembly(PrintFrameConverterAssembly);
1006 set_parsed_irreducible_loop(false);
1007 #endif
1008 set_has_irreducible_loop(false); // no loops
1010 CompileWrapper cw(this);
1011 Init(/*AliasLevel=*/ 0);
1012 init_tf((*generator)());
1014 {
1015 // The following is a dummy for the sake of GraphKit::gen_stub
1016 Unique_Node_List for_igvn(comp_arena());
1017 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1018 PhaseGVN gvn(Thread::current()->resource_area(),255);
1019 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1020 gvn.transform_no_reclaim(top());
1022 GraphKit kit;
1023 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1024 }
1026 NOT_PRODUCT( verify_graph_edges(); )
1027 Code_Gen();
1028 if (failing()) return;
1031 // Entry point will be accessed using compile->stub_entry_point();
1032 if (code_buffer() == NULL) {
1033 Matcher::soft_match_failure();
1034 } else {
1035 if (PrintAssembly && (WizardMode || Verbose))
1036 tty->print_cr("### Stub::%s", stub_name);
1038 if (!failing()) {
1039 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1041 // Make the NMethod
1042 // For now we mark the frame as never safe for profile stackwalking
1043 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1044 code_buffer(),
1045 CodeOffsets::frame_never_safe,
1046 // _code_offsets.value(CodeOffsets::Frame_Complete),
1047 frame_size_in_words(),
1048 _oop_map_set,
1049 save_arg_registers);
1050 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1052 _stub_entry_point = rs->entry_point();
1053 }
1054 }
1055 }
1057 //------------------------------Init-------------------------------------------
1058 // Prepare for a single compilation
1059 void Compile::Init(int aliaslevel) {
1060 _unique = 0;
1061 _regalloc = NULL;
1063 _tf = NULL; // filled in later
1064 _top = NULL; // cached later
1065 _matcher = NULL; // filled in later
1066 _cfg = NULL; // filled in later
1068 set_24_bit_selection_and_mode(Use24BitFP, false);
1070 _node_note_array = NULL;
1071 _default_node_notes = NULL;
1073 _immutable_memory = NULL; // filled in at first inquiry
1075 // Globally visible Nodes
1076 // First set TOP to NULL to give safe behavior during creation of RootNode
1077 set_cached_top_node(NULL);
1078 set_root(new (this) RootNode());
1079 // Now that you have a Root to point to, create the real TOP
1080 set_cached_top_node( new (this) ConNode(Type::TOP) );
1081 set_recent_alloc(NULL, NULL);
1083 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1084 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1085 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1086 env()->set_dependencies(new Dependencies(env()));
1088 _fixed_slots = 0;
1089 set_has_split_ifs(false);
1090 set_has_loops(has_method() && method()->has_loops()); // first approximation
1091 set_has_stringbuilder(false);
1092 set_has_boxed_value(false);
1093 _trap_can_recompile = false; // no traps emitted yet
1094 _major_progress = true; // start out assuming good things will happen
1095 set_has_unsafe_access(false);
1096 set_max_vector_size(0);
1097 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1098 set_decompile_count(0);
1100 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1101 set_num_loop_opts(LoopOptsCount);
1102 set_do_inlining(Inline);
1103 set_max_inline_size(MaxInlineSize);
1104 set_freq_inline_size(FreqInlineSize);
1105 set_do_scheduling(OptoScheduling);
1106 set_do_count_invocations(false);
1107 set_do_method_data_update(false);
1108 set_rtm_state(NoRTM); // No RTM lock eliding by default
1109 method_has_option_value("MaxNodeLimit", _max_node_limit);
1110 #if INCLUDE_RTM_OPT
1111 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1112 int rtm_state = method()->method_data()->rtm_state();
1113 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1114 // Don't generate RTM lock eliding code.
1115 set_rtm_state(NoRTM);
1116 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1117 // Generate RTM lock eliding code without abort ratio calculation code.
1118 set_rtm_state(UseRTM);
1119 } else if (UseRTMDeopt) {
1120 // Generate RTM lock eliding code and include abort ratio calculation
1121 // code if UseRTMDeopt is on.
1122 set_rtm_state(ProfileRTM);
1123 }
1124 }
1125 #endif
1126 if (debug_info()->recording_non_safepoints()) {
1127 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1128 (comp_arena(), 8, 0, NULL));
1129 set_default_node_notes(Node_Notes::make(this));
1130 }
1132 // // -- Initialize types before each compile --
1133 // // Update cached type information
1134 // if( _method && _method->constants() )
1135 // Type::update_loaded_types(_method, _method->constants());
1137 // Init alias_type map.
1138 if (!_do_escape_analysis && aliaslevel == 3)
1139 aliaslevel = 2; // No unique types without escape analysis
1140 _AliasLevel = aliaslevel;
1141 const int grow_ats = 16;
1142 _max_alias_types = grow_ats;
1143 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1144 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1145 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1146 {
1147 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1148 }
1149 // Initialize the first few types.
1150 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1151 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1152 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1153 _num_alias_types = AliasIdxRaw+1;
1154 // Zero out the alias type cache.
1155 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1156 // A NULL adr_type hits in the cache right away. Preload the right answer.
1157 probe_alias_cache(NULL)->_index = AliasIdxTop;
1159 _intrinsics = NULL;
1160 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1161 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1162 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1163 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1164 register_library_intrinsics();
1165 }
1167 //---------------------------init_start----------------------------------------
1168 // Install the StartNode on this compile object.
1169 void Compile::init_start(StartNode* s) {
1170 if (failing())
1171 return; // already failing
1172 assert(s == start(), "");
1173 }
1175 StartNode* Compile::start() const {
1176 assert(!failing(), "");
1177 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1178 Node* start = root()->fast_out(i);
1179 if( start->is_Start() )
1180 return start->as_Start();
1181 }
1182 fatal("Did not find Start node!");
1183 return NULL;
1184 }
1186 //-------------------------------immutable_memory-------------------------------------
1187 // Access immutable memory
1188 Node* Compile::immutable_memory() {
1189 if (_immutable_memory != NULL) {
1190 return _immutable_memory;
1191 }
1192 StartNode* s = start();
1193 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1194 Node *p = s->fast_out(i);
1195 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1196 _immutable_memory = p;
1197 return _immutable_memory;
1198 }
1199 }
1200 ShouldNotReachHere();
1201 return NULL;
1202 }
1204 //----------------------set_cached_top_node------------------------------------
1205 // Install the cached top node, and make sure Node::is_top works correctly.
1206 void Compile::set_cached_top_node(Node* tn) {
1207 if (tn != NULL) verify_top(tn);
1208 Node* old_top = _top;
1209 _top = tn;
1210 // Calling Node::setup_is_top allows the nodes the chance to adjust
1211 // their _out arrays.
1212 if (_top != NULL) _top->setup_is_top();
1213 if (old_top != NULL) old_top->setup_is_top();
1214 assert(_top == NULL || top()->is_top(), "");
1215 }
1217 #ifdef ASSERT
1218 uint Compile::count_live_nodes_by_graph_walk() {
1219 Unique_Node_List useful(comp_arena());
1220 // Get useful node list by walking the graph.
1221 identify_useful_nodes(useful);
1222 return useful.size();
1223 }
1225 void Compile::print_missing_nodes() {
1227 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1228 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1229 return;
1230 }
1232 // This is an expensive function. It is executed only when the user
1233 // specifies VerifyIdealNodeCount option or otherwise knows the
1234 // additional work that needs to be done to identify reachable nodes
1235 // by walking the flow graph and find the missing ones using
1236 // _dead_node_list.
1238 Unique_Node_List useful(comp_arena());
1239 // Get useful node list by walking the graph.
1240 identify_useful_nodes(useful);
1242 uint l_nodes = C->live_nodes();
1243 uint l_nodes_by_walk = useful.size();
1245 if (l_nodes != l_nodes_by_walk) {
1246 if (_log != NULL) {
1247 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1248 _log->stamp();
1249 _log->end_head();
1250 }
1251 VectorSet& useful_member_set = useful.member_set();
1252 int last_idx = l_nodes_by_walk;
1253 for (int i = 0; i < last_idx; i++) {
1254 if (useful_member_set.test(i)) {
1255 if (_dead_node_list.test(i)) {
1256 if (_log != NULL) {
1257 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1258 }
1259 if (PrintIdealNodeCount) {
1260 // Print the log message to tty
1261 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1262 useful.at(i)->dump();
1263 }
1264 }
1265 }
1266 else if (! _dead_node_list.test(i)) {
1267 if (_log != NULL) {
1268 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1269 }
1270 if (PrintIdealNodeCount) {
1271 // Print the log message to tty
1272 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1273 }
1274 }
1275 }
1276 if (_log != NULL) {
1277 _log->tail("mismatched_nodes");
1278 }
1279 }
1280 }
1281 #endif
1283 #ifndef PRODUCT
1284 void Compile::verify_top(Node* tn) const {
1285 if (tn != NULL) {
1286 assert(tn->is_Con(), "top node must be a constant");
1287 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1288 assert(tn->in(0) != NULL, "must have live top node");
1289 }
1290 }
1291 #endif
1294 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1296 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1297 guarantee(arr != NULL, "");
1298 int num_blocks = arr->length();
1299 if (grow_by < num_blocks) grow_by = num_blocks;
1300 int num_notes = grow_by * _node_notes_block_size;
1301 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1302 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1303 while (num_notes > 0) {
1304 arr->append(notes);
1305 notes += _node_notes_block_size;
1306 num_notes -= _node_notes_block_size;
1307 }
1308 assert(num_notes == 0, "exact multiple, please");
1309 }
1311 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1312 if (source == NULL || dest == NULL) return false;
1314 if (dest->is_Con())
1315 return false; // Do not push debug info onto constants.
1317 #ifdef ASSERT
1318 // Leave a bread crumb trail pointing to the original node:
1319 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1320 dest->set_debug_orig(source);
1321 }
1322 #endif
1324 if (node_note_array() == NULL)
1325 return false; // Not collecting any notes now.
1327 // This is a copy onto a pre-existing node, which may already have notes.
1328 // If both nodes have notes, do not overwrite any pre-existing notes.
1329 Node_Notes* source_notes = node_notes_at(source->_idx);
1330 if (source_notes == NULL || source_notes->is_clear()) return false;
1331 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1332 if (dest_notes == NULL || dest_notes->is_clear()) {
1333 return set_node_notes_at(dest->_idx, source_notes);
1334 }
1336 Node_Notes merged_notes = (*source_notes);
1337 // The order of operations here ensures that dest notes will win...
1338 merged_notes.update_from(dest_notes);
1339 return set_node_notes_at(dest->_idx, &merged_notes);
1340 }
1343 //--------------------------allow_range_check_smearing-------------------------
1344 // Gating condition for coalescing similar range checks.
1345 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1346 // single covering check that is at least as strong as any of them.
1347 // If the optimization succeeds, the simplified (strengthened) range check
1348 // will always succeed. If it fails, we will deopt, and then give up
1349 // on the optimization.
1350 bool Compile::allow_range_check_smearing() const {
1351 // If this method has already thrown a range-check,
1352 // assume it was because we already tried range smearing
1353 // and it failed.
1354 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1355 return !already_trapped;
1356 }
1359 //------------------------------flatten_alias_type-----------------------------
1360 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1361 int offset = tj->offset();
1362 TypePtr::PTR ptr = tj->ptr();
1364 // Known instance (scalarizable allocation) alias only with itself.
1365 bool is_known_inst = tj->isa_oopptr() != NULL &&
1366 tj->is_oopptr()->is_known_instance();
1368 // Process weird unsafe references.
1369 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1370 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1371 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1372 tj = TypeOopPtr::BOTTOM;
1373 ptr = tj->ptr();
1374 offset = tj->offset();
1375 }
1377 // Array pointers need some flattening
1378 const TypeAryPtr *ta = tj->isa_aryptr();
1379 if (ta && ta->is_stable()) {
1380 // Erase stability property for alias analysis.
1381 tj = ta = ta->cast_to_stable(false);
1382 }
1383 if( ta && is_known_inst ) {
1384 if ( offset != Type::OffsetBot &&
1385 offset > arrayOopDesc::length_offset_in_bytes() ) {
1386 offset = Type::OffsetBot; // Flatten constant access into array body only
1387 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1388 }
1389 } else if( ta && _AliasLevel >= 2 ) {
1390 // For arrays indexed by constant indices, we flatten the alias
1391 // space to include all of the array body. Only the header, klass
1392 // and array length can be accessed un-aliased.
1393 if( offset != Type::OffsetBot ) {
1394 if( ta->const_oop() ) { // MethodData* or Method*
1395 offset = Type::OffsetBot; // Flatten constant access into array body
1396 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1397 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1398 // range is OK as-is.
1399 tj = ta = TypeAryPtr::RANGE;
1400 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1401 tj = TypeInstPtr::KLASS; // all klass loads look alike
1402 ta = TypeAryPtr::RANGE; // generic ignored junk
1403 ptr = TypePtr::BotPTR;
1404 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1405 tj = TypeInstPtr::MARK;
1406 ta = TypeAryPtr::RANGE; // generic ignored junk
1407 ptr = TypePtr::BotPTR;
1408 } else { // Random constant offset into array body
1409 offset = Type::OffsetBot; // Flatten constant access into array body
1410 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1411 }
1412 }
1413 // Arrays of fixed size alias with arrays of unknown size.
1414 if (ta->size() != TypeInt::POS) {
1415 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1416 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1417 }
1418 // Arrays of known objects become arrays of unknown objects.
1419 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1420 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1421 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1422 }
1423 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1424 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1425 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1426 }
1427 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1428 // cannot be distinguished by bytecode alone.
1429 if (ta->elem() == TypeInt::BOOL) {
1430 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1431 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1432 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1433 }
1434 // During the 2nd round of IterGVN, NotNull castings are removed.
1435 // Make sure the Bottom and NotNull variants alias the same.
1436 // Also, make sure exact and non-exact variants alias the same.
1437 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1438 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1439 }
1440 }
1442 // Oop pointers need some flattening
1443 const TypeInstPtr *to = tj->isa_instptr();
1444 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1445 ciInstanceKlass *k = to->klass()->as_instance_klass();
1446 if( ptr == TypePtr::Constant ) {
1447 if (to->klass() != ciEnv::current()->Class_klass() ||
1448 offset < k->size_helper() * wordSize) {
1449 // No constant oop pointers (such as Strings); they alias with
1450 // unknown strings.
1451 assert(!is_known_inst, "not scalarizable allocation");
1452 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1453 }
1454 } else if( is_known_inst ) {
1455 tj = to; // Keep NotNull and klass_is_exact for instance type
1456 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1457 // During the 2nd round of IterGVN, NotNull castings are removed.
1458 // Make sure the Bottom and NotNull variants alias the same.
1459 // Also, make sure exact and non-exact variants alias the same.
1460 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1461 }
1462 if (to->speculative() != NULL) {
1463 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1464 }
1465 // Canonicalize the holder of this field
1466 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1467 // First handle header references such as a LoadKlassNode, even if the
1468 // object's klass is unloaded at compile time (4965979).
1469 if (!is_known_inst) { // Do it only for non-instance types
1470 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1471 }
1472 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1473 // Static fields are in the space above the normal instance
1474 // fields in the java.lang.Class instance.
1475 if (to->klass() != ciEnv::current()->Class_klass()) {
1476 to = NULL;
1477 tj = TypeOopPtr::BOTTOM;
1478 offset = tj->offset();
1479 }
1480 } else {
1481 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1482 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1483 if( is_known_inst ) {
1484 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1485 } else {
1486 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1487 }
1488 }
1489 }
1490 }
1492 // Klass pointers to object array klasses need some flattening
1493 const TypeKlassPtr *tk = tj->isa_klassptr();
1494 if( tk ) {
1495 // If we are referencing a field within a Klass, we need
1496 // to assume the worst case of an Object. Both exact and
1497 // inexact types must flatten to the same alias class so
1498 // use NotNull as the PTR.
1499 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1501 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1502 TypeKlassPtr::OBJECT->klass(),
1503 offset);
1504 }
1506 ciKlass* klass = tk->klass();
1507 if( klass->is_obj_array_klass() ) {
1508 ciKlass* k = TypeAryPtr::OOPS->klass();
1509 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1510 k = TypeInstPtr::BOTTOM->klass();
1511 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1512 }
1514 // Check for precise loads from the primary supertype array and force them
1515 // to the supertype cache alias index. Check for generic array loads from
1516 // the primary supertype array and also force them to the supertype cache
1517 // alias index. Since the same load can reach both, we need to merge
1518 // these 2 disparate memories into the same alias class. Since the
1519 // primary supertype array is read-only, there's no chance of confusion
1520 // where we bypass an array load and an array store.
1521 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1522 if (offset == Type::OffsetBot ||
1523 (offset >= primary_supers_offset &&
1524 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1525 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1526 offset = in_bytes(Klass::secondary_super_cache_offset());
1527 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1528 }
1529 }
1531 // Flatten all Raw pointers together.
1532 if (tj->base() == Type::RawPtr)
1533 tj = TypeRawPtr::BOTTOM;
1535 if (tj->base() == Type::AnyPtr)
1536 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1538 // Flatten all to bottom for now
1539 switch( _AliasLevel ) {
1540 case 0:
1541 tj = TypePtr::BOTTOM;
1542 break;
1543 case 1: // Flatten to: oop, static, field or array
1544 switch (tj->base()) {
1545 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1546 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1547 case Type::AryPtr: // do not distinguish arrays at all
1548 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1549 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1550 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1551 default: ShouldNotReachHere();
1552 }
1553 break;
1554 case 2: // No collapsing at level 2; keep all splits
1555 case 3: // No collapsing at level 3; keep all splits
1556 break;
1557 default:
1558 Unimplemented();
1559 }
1561 offset = tj->offset();
1562 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1564 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1565 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1566 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1567 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1568 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1569 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1570 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1571 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1572 assert( tj->ptr() != TypePtr::TopPTR &&
1573 tj->ptr() != TypePtr::AnyNull &&
1574 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1575 // assert( tj->ptr() != TypePtr::Constant ||
1576 // tj->base() == Type::RawPtr ||
1577 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1579 return tj;
1580 }
1582 void Compile::AliasType::Init(int i, const TypePtr* at) {
1583 _index = i;
1584 _adr_type = at;
1585 _field = NULL;
1586 _element = NULL;
1587 _is_rewritable = true; // default
1588 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1589 if (atoop != NULL && atoop->is_known_instance()) {
1590 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1591 _general_index = Compile::current()->get_alias_index(gt);
1592 } else {
1593 _general_index = 0;
1594 }
1595 }
1597 BasicType Compile::AliasType::basic_type() const {
1598 if (element() != NULL) {
1599 const Type* element = adr_type()->is_aryptr()->elem();
1600 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1601 } if (field() != NULL) {
1602 return field()->layout_type();
1603 } else {
1604 return T_ILLEGAL; // unknown
1605 }
1606 }
1608 //---------------------------------print_on------------------------------------
1609 #ifndef PRODUCT
1610 void Compile::AliasType::print_on(outputStream* st) {
1611 if (index() < 10)
1612 st->print("@ <%d> ", index());
1613 else st->print("@ <%d>", index());
1614 st->print(is_rewritable() ? " " : " RO");
1615 int offset = adr_type()->offset();
1616 if (offset == Type::OffsetBot)
1617 st->print(" +any");
1618 else st->print(" +%-3d", offset);
1619 st->print(" in ");
1620 adr_type()->dump_on(st);
1621 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1622 if (field() != NULL && tjp) {
1623 if (tjp->klass() != field()->holder() ||
1624 tjp->offset() != field()->offset_in_bytes()) {
1625 st->print(" != ");
1626 field()->print();
1627 st->print(" ***");
1628 }
1629 }
1630 }
1632 void print_alias_types() {
1633 Compile* C = Compile::current();
1634 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1635 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1636 C->alias_type(idx)->print_on(tty);
1637 tty->cr();
1638 }
1639 }
1640 #endif
1643 //----------------------------probe_alias_cache--------------------------------
1644 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1645 intptr_t key = (intptr_t) adr_type;
1646 key ^= key >> logAliasCacheSize;
1647 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1648 }
1651 //-----------------------------grow_alias_types--------------------------------
1652 void Compile::grow_alias_types() {
1653 const int old_ats = _max_alias_types; // how many before?
1654 const int new_ats = old_ats; // how many more?
1655 const int grow_ats = old_ats+new_ats; // how many now?
1656 _max_alias_types = grow_ats;
1657 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1658 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1659 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1660 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1661 }
1664 //--------------------------------find_alias_type------------------------------
1665 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1666 if (_AliasLevel == 0)
1667 return alias_type(AliasIdxBot);
1669 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1670 if (ace->_adr_type == adr_type) {
1671 return alias_type(ace->_index);
1672 }
1674 // Handle special cases.
1675 if (adr_type == NULL) return alias_type(AliasIdxTop);
1676 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1678 // Do it the slow way.
1679 const TypePtr* flat = flatten_alias_type(adr_type);
1681 #ifdef ASSERT
1682 {
1683 ResourceMark rm;
1684 assert(flat == flatten_alias_type(flat),
1685 err_msg("not idempotent: adr_type = %s; flat = %s => %s", Type::str(adr_type),
1686 Type::str(flat), Type::str(flatten_alias_type(flat))));
1687 assert(flat != TypePtr::BOTTOM,
1688 err_msg("cannot alias-analyze an untyped ptr: adr_type = %s", Type::str(adr_type)));
1689 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1690 const TypeOopPtr* foop = flat->is_oopptr();
1691 // Scalarizable allocations have exact klass always.
1692 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1693 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1694 assert(foop == flatten_alias_type(xoop),
1695 err_msg("exactness must not affect alias type: foop = %s; xoop = %s",
1696 Type::str(foop), Type::str(xoop)));
1697 }
1698 }
1699 #endif
1701 int idx = AliasIdxTop;
1702 for (int i = 0; i < num_alias_types(); i++) {
1703 if (alias_type(i)->adr_type() == flat) {
1704 idx = i;
1705 break;
1706 }
1707 }
1709 if (idx == AliasIdxTop) {
1710 if (no_create) return NULL;
1711 // Grow the array if necessary.
1712 if (_num_alias_types == _max_alias_types) grow_alias_types();
1713 // Add a new alias type.
1714 idx = _num_alias_types++;
1715 _alias_types[idx]->Init(idx, flat);
1716 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1717 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1718 if (flat->isa_instptr()) {
1719 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1720 && flat->is_instptr()->klass() == env()->Class_klass())
1721 alias_type(idx)->set_rewritable(false);
1722 }
1723 if (flat->isa_aryptr()) {
1724 #ifdef ASSERT
1725 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1726 // (T_BYTE has the weakest alignment and size restrictions...)
1727 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1728 #endif
1729 if (flat->offset() == TypePtr::OffsetBot) {
1730 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1731 }
1732 }
1733 if (flat->isa_klassptr()) {
1734 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1735 alias_type(idx)->set_rewritable(false);
1736 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1737 alias_type(idx)->set_rewritable(false);
1738 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1739 alias_type(idx)->set_rewritable(false);
1740 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1741 alias_type(idx)->set_rewritable(false);
1742 }
1743 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1744 // but the base pointer type is not distinctive enough to identify
1745 // references into JavaThread.)
1747 // Check for final fields.
1748 const TypeInstPtr* tinst = flat->isa_instptr();
1749 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1750 ciField* field;
1751 if (tinst->const_oop() != NULL &&
1752 tinst->klass() == ciEnv::current()->Class_klass() &&
1753 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1754 // static field
1755 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1756 field = k->get_field_by_offset(tinst->offset(), true);
1757 } else {
1758 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1759 field = k->get_field_by_offset(tinst->offset(), false);
1760 }
1761 assert(field == NULL ||
1762 original_field == NULL ||
1763 (field->holder() == original_field->holder() &&
1764 field->offset() == original_field->offset() &&
1765 field->is_static() == original_field->is_static()), "wrong field?");
1766 // Set field() and is_rewritable() attributes.
1767 if (field != NULL) alias_type(idx)->set_field(field);
1768 }
1769 }
1771 // Fill the cache for next time.
1772 ace->_adr_type = adr_type;
1773 ace->_index = idx;
1774 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1776 // Might as well try to fill the cache for the flattened version, too.
1777 AliasCacheEntry* face = probe_alias_cache(flat);
1778 if (face->_adr_type == NULL) {
1779 face->_adr_type = flat;
1780 face->_index = idx;
1781 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1782 }
1784 return alias_type(idx);
1785 }
1788 Compile::AliasType* Compile::alias_type(ciField* field) {
1789 const TypeOopPtr* t;
1790 if (field->is_static())
1791 t = TypeInstPtr::make(field->holder()->java_mirror());
1792 else
1793 t = TypeOopPtr::make_from_klass_raw(field->holder());
1794 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1795 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1796 return atp;
1797 }
1800 //------------------------------have_alias_type--------------------------------
1801 bool Compile::have_alias_type(const TypePtr* adr_type) {
1802 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1803 if (ace->_adr_type == adr_type) {
1804 return true;
1805 }
1807 // Handle special cases.
1808 if (adr_type == NULL) return true;
1809 if (adr_type == TypePtr::BOTTOM) return true;
1811 return find_alias_type(adr_type, true, NULL) != NULL;
1812 }
1814 //-----------------------------must_alias--------------------------------------
1815 // True if all values of the given address type are in the given alias category.
1816 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1817 if (alias_idx == AliasIdxBot) return true; // the universal category
1818 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1819 if (alias_idx == AliasIdxTop) return false; // the empty category
1820 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1822 // the only remaining possible overlap is identity
1823 int adr_idx = get_alias_index(adr_type);
1824 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1825 assert(adr_idx == alias_idx ||
1826 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1827 && adr_type != TypeOopPtr::BOTTOM),
1828 "should not be testing for overlap with an unsafe pointer");
1829 return adr_idx == alias_idx;
1830 }
1832 //------------------------------can_alias--------------------------------------
1833 // True if any values of the given address type are in the given alias category.
1834 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1835 if (alias_idx == AliasIdxTop) return false; // the empty category
1836 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1837 if (alias_idx == AliasIdxBot) return true; // the universal category
1838 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1840 // the only remaining possible overlap is identity
1841 int adr_idx = get_alias_index(adr_type);
1842 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1843 return adr_idx == alias_idx;
1844 }
1848 //---------------------------pop_warm_call-------------------------------------
1849 WarmCallInfo* Compile::pop_warm_call() {
1850 WarmCallInfo* wci = _warm_calls;
1851 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1852 return wci;
1853 }
1855 //----------------------------Inline_Warm--------------------------------------
1856 int Compile::Inline_Warm() {
1857 // If there is room, try to inline some more warm call sites.
1858 // %%% Do a graph index compaction pass when we think we're out of space?
1859 if (!InlineWarmCalls) return 0;
1861 int calls_made_hot = 0;
1862 int room_to_grow = NodeCountInliningCutoff - unique();
1863 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1864 int amount_grown = 0;
1865 WarmCallInfo* call;
1866 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1867 int est_size = (int)call->size();
1868 if (est_size > (room_to_grow - amount_grown)) {
1869 // This one won't fit anyway. Get rid of it.
1870 call->make_cold();
1871 continue;
1872 }
1873 call->make_hot();
1874 calls_made_hot++;
1875 amount_grown += est_size;
1876 amount_to_grow -= est_size;
1877 }
1879 if (calls_made_hot > 0) set_major_progress();
1880 return calls_made_hot;
1881 }
1884 //----------------------------Finish_Warm--------------------------------------
1885 void Compile::Finish_Warm() {
1886 if (!InlineWarmCalls) return;
1887 if (failing()) return;
1888 if (warm_calls() == NULL) return;
1890 // Clean up loose ends, if we are out of space for inlining.
1891 WarmCallInfo* call;
1892 while ((call = pop_warm_call()) != NULL) {
1893 call->make_cold();
1894 }
1895 }
1897 //---------------------cleanup_loop_predicates-----------------------
1898 // Remove the opaque nodes that protect the predicates so that all unused
1899 // checks and uncommon_traps will be eliminated from the ideal graph
1900 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1901 if (predicate_count()==0) return;
1902 for (int i = predicate_count(); i > 0; i--) {
1903 Node * n = predicate_opaque1_node(i-1);
1904 assert(n->Opcode() == Op_Opaque1, "must be");
1905 igvn.replace_node(n, n->in(1));
1906 }
1907 assert(predicate_count()==0, "should be clean!");
1908 }
1910 void Compile::add_range_check_cast(Node* n) {
1911 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1912 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1913 _range_check_casts->append(n);
1914 }
1916 // Remove all range check dependent CastIINodes.
1917 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1918 for (int i = range_check_cast_count(); i > 0; i--) {
1919 Node* cast = range_check_cast_node(i-1);
1920 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1921 igvn.replace_node(cast, cast->in(1));
1922 }
1923 assert(range_check_cast_count() == 0, "should be empty");
1924 }
1926 // StringOpts and late inlining of string methods
1927 void Compile::inline_string_calls(bool parse_time) {
1928 {
1929 // remove useless nodes to make the usage analysis simpler
1930 ResourceMark rm;
1931 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1932 }
1934 {
1935 ResourceMark rm;
1936 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1937 PhaseStringOpts pso(initial_gvn(), for_igvn());
1938 print_method(PHASE_AFTER_STRINGOPTS, 3);
1939 }
1941 // now inline anything that we skipped the first time around
1942 if (!parse_time) {
1943 _late_inlines_pos = _late_inlines.length();
1944 }
1946 while (_string_late_inlines.length() > 0) {
1947 CallGenerator* cg = _string_late_inlines.pop();
1948 cg->do_late_inline();
1949 if (failing()) return;
1950 }
1951 _string_late_inlines.trunc_to(0);
1952 }
1954 // Late inlining of boxing methods
1955 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1956 if (_boxing_late_inlines.length() > 0) {
1957 assert(has_boxed_value(), "inconsistent");
1959 PhaseGVN* gvn = initial_gvn();
1960 set_inlining_incrementally(true);
1962 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1963 for_igvn()->clear();
1964 gvn->replace_with(&igvn);
1966 _late_inlines_pos = _late_inlines.length();
1968 while (_boxing_late_inlines.length() > 0) {
1969 CallGenerator* cg = _boxing_late_inlines.pop();
1970 cg->do_late_inline();
1971 if (failing()) return;
1972 }
1973 _boxing_late_inlines.trunc_to(0);
1975 {
1976 ResourceMark rm;
1977 PhaseRemoveUseless pru(gvn, for_igvn());
1978 }
1980 igvn = PhaseIterGVN(gvn);
1981 igvn.optimize();
1983 set_inlining_progress(false);
1984 set_inlining_incrementally(false);
1985 }
1986 }
1988 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1989 assert(IncrementalInline, "incremental inlining should be on");
1990 PhaseGVN* gvn = initial_gvn();
1992 set_inlining_progress(false);
1993 for_igvn()->clear();
1994 gvn->replace_with(&igvn);
1996 int i = 0;
1998 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1999 CallGenerator* cg = _late_inlines.at(i);
2000 _late_inlines_pos = i+1;
2001 cg->do_late_inline();
2002 if (failing()) return;
2003 }
2004 int j = 0;
2005 for (; i < _late_inlines.length(); i++, j++) {
2006 _late_inlines.at_put(j, _late_inlines.at(i));
2007 }
2008 _late_inlines.trunc_to(j);
2010 {
2011 ResourceMark rm;
2012 PhaseRemoveUseless pru(gvn, for_igvn());
2013 }
2015 igvn = PhaseIterGVN(gvn);
2016 }
2018 // Perform incremental inlining until bound on number of live nodes is reached
2019 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2020 PhaseGVN* gvn = initial_gvn();
2022 set_inlining_incrementally(true);
2023 set_inlining_progress(true);
2024 uint low_live_nodes = 0;
2026 while(inlining_progress() && _late_inlines.length() > 0) {
2028 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2029 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2030 // PhaseIdealLoop is expensive so we only try it once we are
2031 // out of live nodes and we only try it again if the previous
2032 // helped got the number of nodes down significantly
2033 PhaseIdealLoop ideal_loop( igvn, false, true );
2034 if (failing()) return;
2035 low_live_nodes = live_nodes();
2036 _major_progress = true;
2037 }
2039 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2040 break;
2041 }
2042 }
2044 inline_incrementally_one(igvn);
2046 if (failing()) return;
2048 igvn.optimize();
2050 if (failing()) return;
2051 }
2053 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2055 if (_string_late_inlines.length() > 0) {
2056 assert(has_stringbuilder(), "inconsistent");
2057 for_igvn()->clear();
2058 initial_gvn()->replace_with(&igvn);
2060 inline_string_calls(false);
2062 if (failing()) return;
2064 {
2065 ResourceMark rm;
2066 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2067 }
2069 igvn = PhaseIterGVN(gvn);
2071 igvn.optimize();
2072 }
2074 set_inlining_incrementally(false);
2075 }
2078 //------------------------------Optimize---------------------------------------
2079 // Given a graph, optimize it.
2080 void Compile::Optimize() {
2081 TracePhase t1("optimizer", &_t_optimizer, true);
2083 #ifndef PRODUCT
2084 if (env()->break_at_compile()) {
2085 BREAKPOINT;
2086 }
2088 #endif
2090 ResourceMark rm;
2091 int loop_opts_cnt;
2093 NOT_PRODUCT( verify_graph_edges(); )
2095 print_method(PHASE_AFTER_PARSING);
2097 {
2098 // Iterative Global Value Numbering, including ideal transforms
2099 // Initialize IterGVN with types and values from parse-time GVN
2100 PhaseIterGVN igvn(initial_gvn());
2101 {
2102 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2103 igvn.optimize();
2104 }
2106 print_method(PHASE_ITER_GVN1, 2);
2108 if (failing()) return;
2110 {
2111 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2112 inline_incrementally(igvn);
2113 }
2115 print_method(PHASE_INCREMENTAL_INLINE, 2);
2117 if (failing()) return;
2119 if (eliminate_boxing()) {
2120 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2121 // Inline valueOf() methods now.
2122 inline_boxing_calls(igvn);
2124 if (AlwaysIncrementalInline) {
2125 inline_incrementally(igvn);
2126 }
2128 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2130 if (failing()) return;
2131 }
2133 // Remove the speculative part of types and clean up the graph from
2134 // the extra CastPP nodes whose only purpose is to carry them. Do
2135 // that early so that optimizations are not disrupted by the extra
2136 // CastPP nodes.
2137 remove_speculative_types(igvn);
2139 // No more new expensive nodes will be added to the list from here
2140 // so keep only the actual candidates for optimizations.
2141 cleanup_expensive_nodes(igvn);
2143 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2144 NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);)
2145 initial_gvn()->replace_with(&igvn);
2146 for_igvn()->clear();
2147 Unique_Node_List new_worklist(C->comp_arena());
2148 {
2149 ResourceMark rm;
2150 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2151 }
2152 set_for_igvn(&new_worklist);
2153 igvn = PhaseIterGVN(initial_gvn());
2154 igvn.optimize();
2155 }
2157 // Perform escape analysis
2158 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2159 if (has_loops()) {
2160 // Cleanup graph (remove dead nodes).
2161 TracePhase t2("idealLoop", &_t_idealLoop, true);
2162 PhaseIdealLoop ideal_loop( igvn, false, true );
2163 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2164 if (failing()) return;
2165 }
2166 ConnectionGraph::do_analysis(this, &igvn);
2168 if (failing()) return;
2170 // Optimize out fields loads from scalar replaceable allocations.
2171 igvn.optimize();
2172 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2174 if (failing()) return;
2176 if (congraph() != NULL && macro_count() > 0) {
2177 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2178 PhaseMacroExpand mexp(igvn);
2179 mexp.eliminate_macro_nodes();
2180 igvn.set_delay_transform(false);
2182 igvn.optimize();
2183 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2185 if (failing()) return;
2186 }
2187 }
2189 // Loop transforms on the ideal graph. Range Check Elimination,
2190 // peeling, unrolling, etc.
2192 // Set loop opts counter
2193 loop_opts_cnt = num_loop_opts();
2194 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2195 {
2196 TracePhase t2("idealLoop", &_t_idealLoop, true);
2197 PhaseIdealLoop ideal_loop( igvn, true );
2198 loop_opts_cnt--;
2199 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2200 if (failing()) return;
2201 }
2202 // Loop opts pass if partial peeling occurred in previous pass
2203 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2204 TracePhase t3("idealLoop", &_t_idealLoop, true);
2205 PhaseIdealLoop ideal_loop( igvn, false );
2206 loop_opts_cnt--;
2207 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2208 if (failing()) return;
2209 }
2210 // Loop opts pass for loop-unrolling before CCP
2211 if(major_progress() && (loop_opts_cnt > 0)) {
2212 TracePhase t4("idealLoop", &_t_idealLoop, true);
2213 PhaseIdealLoop ideal_loop( igvn, false );
2214 loop_opts_cnt--;
2215 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2216 }
2217 if (!failing()) {
2218 // Verify that last round of loop opts produced a valid graph
2219 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2220 PhaseIdealLoop::verify(igvn);
2221 }
2222 }
2223 if (failing()) return;
2225 // Conditional Constant Propagation;
2226 PhaseCCP ccp( &igvn );
2227 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2228 {
2229 TracePhase t2("ccp", &_t_ccp, true);
2230 ccp.do_transform();
2231 }
2232 print_method(PHASE_CPP1, 2);
2234 assert( true, "Break here to ccp.dump_old2new_map()");
2236 // Iterative Global Value Numbering, including ideal transforms
2237 {
2238 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2239 igvn = ccp;
2240 igvn.optimize();
2241 }
2243 print_method(PHASE_ITER_GVN2, 2);
2245 if (failing()) return;
2247 // Loop transforms on the ideal graph. Range Check Elimination,
2248 // peeling, unrolling, etc.
2249 if(loop_opts_cnt > 0) {
2250 debug_only( int cnt = 0; );
2251 while(major_progress() && (loop_opts_cnt > 0)) {
2252 TracePhase t2("idealLoop", &_t_idealLoop, true);
2253 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2254 PhaseIdealLoop ideal_loop( igvn, true);
2255 loop_opts_cnt--;
2256 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2257 if (failing()) return;
2258 }
2259 }
2261 {
2262 // Verify that all previous optimizations produced a valid graph
2263 // at least to this point, even if no loop optimizations were done.
2264 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2265 PhaseIdealLoop::verify(igvn);
2266 }
2268 if (range_check_cast_count() > 0) {
2269 // No more loop optimizations. Remove all range check dependent CastIINodes.
2270 C->remove_range_check_casts(igvn);
2271 igvn.optimize();
2272 }
2274 {
2275 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2276 PhaseMacroExpand mex(igvn);
2277 if (mex.expand_macro_nodes()) {
2278 assert(failing(), "must bail out w/ explicit message");
2279 return;
2280 }
2281 }
2283 } // (End scope of igvn; run destructor if necessary for asserts.)
2285 dump_inlining();
2286 // A method with only infinite loops has no edges entering loops from root
2287 {
2288 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2289 if (final_graph_reshaping()) {
2290 assert(failing(), "must bail out w/ explicit message");
2291 return;
2292 }
2293 }
2295 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2296 }
2299 //------------------------------Code_Gen---------------------------------------
2300 // Given a graph, generate code for it
2301 void Compile::Code_Gen() {
2302 if (failing()) {
2303 return;
2304 }
2306 // Perform instruction selection. You might think we could reclaim Matcher
2307 // memory PDQ, but actually the Matcher is used in generating spill code.
2308 // Internals of the Matcher (including some VectorSets) must remain live
2309 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2310 // set a bit in reclaimed memory.
2312 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2313 // nodes. Mapping is only valid at the root of each matched subtree.
2314 NOT_PRODUCT( verify_graph_edges(); )
2316 Matcher matcher;
2317 _matcher = &matcher;
2318 {
2319 TracePhase t2("matcher", &_t_matcher, true);
2320 matcher.match();
2321 }
2322 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2323 // nodes. Mapping is only valid at the root of each matched subtree.
2324 NOT_PRODUCT( verify_graph_edges(); )
2326 // If you have too many nodes, or if matching has failed, bail out
2327 check_node_count(0, "out of nodes matching instructions");
2328 if (failing()) {
2329 return;
2330 }
2332 // Build a proper-looking CFG
2333 PhaseCFG cfg(node_arena(), root(), matcher);
2334 _cfg = &cfg;
2335 {
2336 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2337 bool success = cfg.do_global_code_motion();
2338 if (!success) {
2339 return;
2340 }
2342 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2343 NOT_PRODUCT( verify_graph_edges(); )
2344 debug_only( cfg.verify(); )
2345 }
2347 PhaseChaitin regalloc(unique(), cfg, matcher);
2348 _regalloc = ®alloc;
2349 {
2350 TracePhase t2("regalloc", &_t_registerAllocation, true);
2351 // Perform register allocation. After Chaitin, use-def chains are
2352 // no longer accurate (at spill code) and so must be ignored.
2353 // Node->LRG->reg mappings are still accurate.
2354 _regalloc->Register_Allocate();
2356 // Bail out if the allocator builds too many nodes
2357 if (failing()) {
2358 return;
2359 }
2360 }
2362 // Prior to register allocation we kept empty basic blocks in case the
2363 // the allocator needed a place to spill. After register allocation we
2364 // are not adding any new instructions. If any basic block is empty, we
2365 // can now safely remove it.
2366 {
2367 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2368 cfg.remove_empty_blocks();
2369 if (do_freq_based_layout()) {
2370 PhaseBlockLayout layout(cfg);
2371 } else {
2372 cfg.set_loop_alignment();
2373 }
2374 cfg.fixup_flow();
2375 }
2377 // Apply peephole optimizations
2378 if( OptoPeephole ) {
2379 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2380 PhasePeephole peep( _regalloc, cfg);
2381 peep.do_transform();
2382 }
2384 // Do late expand if CPU requires this.
2385 if (Matcher::require_postalloc_expand) {
2386 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2387 cfg.postalloc_expand(_regalloc);
2388 }
2390 // Convert Nodes to instruction bits in a buffer
2391 {
2392 // %%%% workspace merge brought two timers together for one job
2393 TracePhase t2a("output", &_t_output, true);
2394 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2395 Output();
2396 }
2398 print_method(PHASE_FINAL_CODE);
2400 // He's dead, Jim.
2401 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2402 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2403 }
2406 //------------------------------dump_asm---------------------------------------
2407 // Dump formatted assembly
2408 #ifndef PRODUCT
2409 void Compile::dump_asm(int *pcs, uint pc_limit) {
2410 bool cut_short = false;
2411 tty->print_cr("#");
2412 tty->print("# "); _tf->dump(); tty->cr();
2413 tty->print_cr("#");
2415 // For all blocks
2416 int pc = 0x0; // Program counter
2417 char starts_bundle = ' ';
2418 _regalloc->dump_frame();
2420 Node *n = NULL;
2421 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2422 if (VMThread::should_terminate()) {
2423 cut_short = true;
2424 break;
2425 }
2426 Block* block = _cfg->get_block(i);
2427 if (block->is_connector() && !Verbose) {
2428 continue;
2429 }
2430 n = block->head();
2431 if (pcs && n->_idx < pc_limit) {
2432 tty->print("%3.3x ", pcs[n->_idx]);
2433 } else {
2434 tty->print(" ");
2435 }
2436 block->dump_head(_cfg);
2437 if (block->is_connector()) {
2438 tty->print_cr(" # Empty connector block");
2439 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2440 tty->print_cr(" # Block is sole successor of call");
2441 }
2443 // For all instructions
2444 Node *delay = NULL;
2445 for (uint j = 0; j < block->number_of_nodes(); j++) {
2446 if (VMThread::should_terminate()) {
2447 cut_short = true;
2448 break;
2449 }
2450 n = block->get_node(j);
2451 if (valid_bundle_info(n)) {
2452 Bundle* bundle = node_bundling(n);
2453 if (bundle->used_in_unconditional_delay()) {
2454 delay = n;
2455 continue;
2456 }
2457 if (bundle->starts_bundle()) {
2458 starts_bundle = '+';
2459 }
2460 }
2462 if (WizardMode) {
2463 n->dump();
2464 }
2466 if( !n->is_Region() && // Dont print in the Assembly
2467 !n->is_Phi() && // a few noisely useless nodes
2468 !n->is_Proj() &&
2469 !n->is_MachTemp() &&
2470 !n->is_SafePointScalarObject() &&
2471 !n->is_Catch() && // Would be nice to print exception table targets
2472 !n->is_MergeMem() && // Not very interesting
2473 !n->is_top() && // Debug info table constants
2474 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2475 ) {
2476 if (pcs && n->_idx < pc_limit)
2477 tty->print("%3.3x", pcs[n->_idx]);
2478 else
2479 tty->print(" ");
2480 tty->print(" %c ", starts_bundle);
2481 starts_bundle = ' ';
2482 tty->print("\t");
2483 n->format(_regalloc, tty);
2484 tty->cr();
2485 }
2487 // If we have an instruction with a delay slot, and have seen a delay,
2488 // then back up and print it
2489 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2490 assert(delay != NULL, "no unconditional delay instruction");
2491 if (WizardMode) delay->dump();
2493 if (node_bundling(delay)->starts_bundle())
2494 starts_bundle = '+';
2495 if (pcs && n->_idx < pc_limit)
2496 tty->print("%3.3x", pcs[n->_idx]);
2497 else
2498 tty->print(" ");
2499 tty->print(" %c ", starts_bundle);
2500 starts_bundle = ' ';
2501 tty->print("\t");
2502 delay->format(_regalloc, tty);
2503 tty->cr();
2504 delay = NULL;
2505 }
2507 // Dump the exception table as well
2508 if( n->is_Catch() && (Verbose || WizardMode) ) {
2509 // Print the exception table for this offset
2510 _handler_table.print_subtable_for(pc);
2511 }
2512 }
2514 if (pcs && n->_idx < pc_limit)
2515 tty->print_cr("%3.3x", pcs[n->_idx]);
2516 else
2517 tty->cr();
2519 assert(cut_short || delay == NULL, "no unconditional delay branch");
2521 } // End of per-block dump
2522 tty->cr();
2524 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2525 }
2526 #endif
2528 //------------------------------Final_Reshape_Counts---------------------------
2529 // This class defines counters to help identify when a method
2530 // may/must be executed using hardware with only 24-bit precision.
2531 struct Final_Reshape_Counts : public StackObj {
2532 int _call_count; // count non-inlined 'common' calls
2533 int _float_count; // count float ops requiring 24-bit precision
2534 int _double_count; // count double ops requiring more precision
2535 int _java_call_count; // count non-inlined 'java' calls
2536 int _inner_loop_count; // count loops which need alignment
2537 VectorSet _visited; // Visitation flags
2538 Node_List _tests; // Set of IfNodes & PCTableNodes
2540 Final_Reshape_Counts() :
2541 _call_count(0), _float_count(0), _double_count(0),
2542 _java_call_count(0), _inner_loop_count(0),
2543 _visited( Thread::current()->resource_area() ) { }
2545 void inc_call_count () { _call_count ++; }
2546 void inc_float_count () { _float_count ++; }
2547 void inc_double_count() { _double_count++; }
2548 void inc_java_call_count() { _java_call_count++; }
2549 void inc_inner_loop_count() { _inner_loop_count++; }
2551 int get_call_count () const { return _call_count ; }
2552 int get_float_count () const { return _float_count ; }
2553 int get_double_count() const { return _double_count; }
2554 int get_java_call_count() const { return _java_call_count; }
2555 int get_inner_loop_count() const { return _inner_loop_count; }
2556 };
2558 #ifdef ASSERT
2559 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2560 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2561 // Make sure the offset goes inside the instance layout.
2562 return k->contains_field_offset(tp->offset());
2563 // Note that OffsetBot and OffsetTop are very negative.
2564 }
2565 #endif
2567 // Eliminate trivially redundant StoreCMs and accumulate their
2568 // precedence edges.
2569 void Compile::eliminate_redundant_card_marks(Node* n) {
2570 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2571 if (n->in(MemNode::Address)->outcnt() > 1) {
2572 // There are multiple users of the same address so it might be
2573 // possible to eliminate some of the StoreCMs
2574 Node* mem = n->in(MemNode::Memory);
2575 Node* adr = n->in(MemNode::Address);
2576 Node* val = n->in(MemNode::ValueIn);
2577 Node* prev = n;
2578 bool done = false;
2579 // Walk the chain of StoreCMs eliminating ones that match. As
2580 // long as it's a chain of single users then the optimization is
2581 // safe. Eliminating partially redundant StoreCMs would require
2582 // cloning copies down the other paths.
2583 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2584 if (adr == mem->in(MemNode::Address) &&
2585 val == mem->in(MemNode::ValueIn)) {
2586 // redundant StoreCM
2587 if (mem->req() > MemNode::OopStore) {
2588 // Hasn't been processed by this code yet.
2589 n->add_prec(mem->in(MemNode::OopStore));
2590 } else {
2591 // Already converted to precedence edge
2592 for (uint i = mem->req(); i < mem->len(); i++) {
2593 // Accumulate any precedence edges
2594 if (mem->in(i) != NULL) {
2595 n->add_prec(mem->in(i));
2596 }
2597 }
2598 // Everything above this point has been processed.
2599 done = true;
2600 }
2601 // Eliminate the previous StoreCM
2602 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2603 assert(mem->outcnt() == 0, "should be dead");
2604 mem->disconnect_inputs(NULL, this);
2605 } else {
2606 prev = mem;
2607 }
2608 mem = prev->in(MemNode::Memory);
2609 }
2610 }
2611 }
2613 //------------------------------final_graph_reshaping_impl----------------------
2614 // Implement items 1-5 from final_graph_reshaping below.
2615 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2617 if ( n->outcnt() == 0 ) return; // dead node
2618 uint nop = n->Opcode();
2620 // Check for 2-input instruction with "last use" on right input.
2621 // Swap to left input. Implements item (2).
2622 if( n->req() == 3 && // two-input instruction
2623 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2624 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2625 n->in(2)->outcnt() == 1 &&// right use IS a last use
2626 !n->in(2)->is_Con() ) { // right use is not a constant
2627 // Check for commutative opcode
2628 switch( nop ) {
2629 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2630 case Op_MaxI: case Op_MinI:
2631 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2632 case Op_AndL: case Op_XorL: case Op_OrL:
2633 case Op_AndI: case Op_XorI: case Op_OrI: {
2634 // Move "last use" input to left by swapping inputs
2635 n->swap_edges(1, 2);
2636 break;
2637 }
2638 default:
2639 break;
2640 }
2641 }
2643 #ifdef ASSERT
2644 if( n->is_Mem() ) {
2645 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2646 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2647 // oop will be recorded in oop map if load crosses safepoint
2648 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2649 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2650 "raw memory operations should have control edge");
2651 }
2652 #endif
2653 // Count FPU ops and common calls, implements item (3)
2654 switch( nop ) {
2655 // Count all float operations that may use FPU
2656 case Op_AddF:
2657 case Op_SubF:
2658 case Op_MulF:
2659 case Op_DivF:
2660 case Op_NegF:
2661 case Op_ModF:
2662 case Op_ConvI2F:
2663 case Op_ConF:
2664 case Op_CmpF:
2665 case Op_CmpF3:
2666 // case Op_ConvL2F: // longs are split into 32-bit halves
2667 frc.inc_float_count();
2668 break;
2670 case Op_ConvF2D:
2671 case Op_ConvD2F:
2672 frc.inc_float_count();
2673 frc.inc_double_count();
2674 break;
2676 // Count all double operations that may use FPU
2677 case Op_AddD:
2678 case Op_SubD:
2679 case Op_MulD:
2680 case Op_DivD:
2681 case Op_NegD:
2682 case Op_ModD:
2683 case Op_ConvI2D:
2684 case Op_ConvD2I:
2685 // case Op_ConvL2D: // handled by leaf call
2686 // case Op_ConvD2L: // handled by leaf call
2687 case Op_ConD:
2688 case Op_CmpD:
2689 case Op_CmpD3:
2690 frc.inc_double_count();
2691 break;
2692 case Op_Opaque1: // Remove Opaque Nodes before matching
2693 case Op_Opaque2: // Remove Opaque Nodes before matching
2694 case Op_Opaque3:
2695 n->subsume_by(n->in(1), this);
2696 break;
2697 case Op_CallStaticJava:
2698 case Op_CallJava:
2699 case Op_CallDynamicJava:
2700 frc.inc_java_call_count(); // Count java call site;
2701 case Op_CallRuntime:
2702 case Op_CallLeaf:
2703 case Op_CallLeafNoFP: {
2704 assert( n->is_Call(), "" );
2705 CallNode *call = n->as_Call();
2706 // Count call sites where the FP mode bit would have to be flipped.
2707 // Do not count uncommon runtime calls:
2708 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2709 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2710 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2711 frc.inc_call_count(); // Count the call site
2712 } else { // See if uncommon argument is shared
2713 Node *n = call->in(TypeFunc::Parms);
2714 int nop = n->Opcode();
2715 // Clone shared simple arguments to uncommon calls, item (1).
2716 if( n->outcnt() > 1 &&
2717 !n->is_Proj() &&
2718 nop != Op_CreateEx &&
2719 nop != Op_CheckCastPP &&
2720 nop != Op_DecodeN &&
2721 nop != Op_DecodeNKlass &&
2722 !n->is_Mem() ) {
2723 Node *x = n->clone();
2724 call->set_req( TypeFunc::Parms, x );
2725 }
2726 }
2727 break;
2728 }
2730 case Op_StoreD:
2731 case Op_LoadD:
2732 case Op_LoadD_unaligned:
2733 frc.inc_double_count();
2734 goto handle_mem;
2735 case Op_StoreF:
2736 case Op_LoadF:
2737 frc.inc_float_count();
2738 goto handle_mem;
2740 case Op_StoreCM:
2741 {
2742 // Convert OopStore dependence into precedence edge
2743 Node* prec = n->in(MemNode::OopStore);
2744 n->del_req(MemNode::OopStore);
2745 n->add_prec(prec);
2746 eliminate_redundant_card_marks(n);
2747 }
2749 // fall through
2751 case Op_StoreB:
2752 case Op_StoreC:
2753 case Op_StorePConditional:
2754 case Op_StoreI:
2755 case Op_StoreL:
2756 case Op_StoreIConditional:
2757 case Op_StoreLConditional:
2758 case Op_CompareAndSwapI:
2759 case Op_CompareAndSwapL:
2760 case Op_CompareAndSwapP:
2761 case Op_CompareAndSwapN:
2762 case Op_GetAndAddI:
2763 case Op_GetAndAddL:
2764 case Op_GetAndSetI:
2765 case Op_GetAndSetL:
2766 case Op_GetAndSetP:
2767 case Op_GetAndSetN:
2768 case Op_StoreP:
2769 case Op_StoreN:
2770 case Op_StoreNKlass:
2771 case Op_LoadB:
2772 case Op_LoadUB:
2773 case Op_LoadUS:
2774 case Op_LoadI:
2775 case Op_LoadKlass:
2776 case Op_LoadNKlass:
2777 case Op_LoadL:
2778 case Op_LoadL_unaligned:
2779 case Op_LoadPLocked:
2780 case Op_LoadP:
2781 case Op_LoadN:
2782 case Op_LoadRange:
2783 case Op_LoadS: {
2784 handle_mem:
2785 #ifdef ASSERT
2786 if( VerifyOptoOopOffsets ) {
2787 assert( n->is_Mem(), "" );
2788 MemNode *mem = (MemNode*)n;
2789 // Check to see if address types have grounded out somehow.
2790 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2791 assert( !tp || oop_offset_is_sane(tp), "" );
2792 }
2793 #endif
2794 break;
2795 }
2797 case Op_AddP: { // Assert sane base pointers
2798 Node *addp = n->in(AddPNode::Address);
2799 assert( !addp->is_AddP() ||
2800 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2801 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2802 "Base pointers must match" );
2803 #ifdef _LP64
2804 if ((UseCompressedOops || UseCompressedClassPointers) &&
2805 addp->Opcode() == Op_ConP &&
2806 addp == n->in(AddPNode::Base) &&
2807 n->in(AddPNode::Offset)->is_Con()) {
2808 // Use addressing with narrow klass to load with offset on x86.
2809 // On sparc loading 32-bits constant and decoding it have less
2810 // instructions (4) then load 64-bits constant (7).
2811 // Do this transformation here since IGVN will convert ConN back to ConP.
2812 const Type* t = addp->bottom_type();
2813 if (t->isa_oopptr() || t->isa_klassptr()) {
2814 Node* nn = NULL;
2816 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2818 // Look for existing ConN node of the same exact type.
2819 Node* r = root();
2820 uint cnt = r->outcnt();
2821 for (uint i = 0; i < cnt; i++) {
2822 Node* m = r->raw_out(i);
2823 if (m!= NULL && m->Opcode() == op &&
2824 m->bottom_type()->make_ptr() == t) {
2825 nn = m;
2826 break;
2827 }
2828 }
2829 if (nn != NULL) {
2830 // Decode a narrow oop to match address
2831 // [R12 + narrow_oop_reg<<3 + offset]
2832 if (t->isa_oopptr()) {
2833 nn = new (this) DecodeNNode(nn, t);
2834 } else {
2835 nn = new (this) DecodeNKlassNode(nn, t);
2836 }
2837 n->set_req(AddPNode::Base, nn);
2838 n->set_req(AddPNode::Address, nn);
2839 if (addp->outcnt() == 0) {
2840 addp->disconnect_inputs(NULL, this);
2841 }
2842 }
2843 }
2844 }
2845 #endif
2846 break;
2847 }
2849 #ifdef _LP64
2850 case Op_CastPP:
2851 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2852 Node* in1 = n->in(1);
2853 const Type* t = n->bottom_type();
2854 Node* new_in1 = in1->clone();
2855 new_in1->as_DecodeN()->set_type(t);
2857 if (!Matcher::narrow_oop_use_complex_address()) {
2858 //
2859 // x86, ARM and friends can handle 2 adds in addressing mode
2860 // and Matcher can fold a DecodeN node into address by using
2861 // a narrow oop directly and do implicit NULL check in address:
2862 //
2863 // [R12 + narrow_oop_reg<<3 + offset]
2864 // NullCheck narrow_oop_reg
2865 //
2866 // On other platforms (Sparc) we have to keep new DecodeN node and
2867 // use it to do implicit NULL check in address:
2868 //
2869 // decode_not_null narrow_oop_reg, base_reg
2870 // [base_reg + offset]
2871 // NullCheck base_reg
2872 //
2873 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2874 // to keep the information to which NULL check the new DecodeN node
2875 // corresponds to use it as value in implicit_null_check().
2876 //
2877 new_in1->set_req(0, n->in(0));
2878 }
2880 n->subsume_by(new_in1, this);
2881 if (in1->outcnt() == 0) {
2882 in1->disconnect_inputs(NULL, this);
2883 }
2884 }
2885 break;
2887 case Op_CmpP:
2888 // Do this transformation here to preserve CmpPNode::sub() and
2889 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2890 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2891 Node* in1 = n->in(1);
2892 Node* in2 = n->in(2);
2893 if (!in1->is_DecodeNarrowPtr()) {
2894 in2 = in1;
2895 in1 = n->in(2);
2896 }
2897 assert(in1->is_DecodeNarrowPtr(), "sanity");
2899 Node* new_in2 = NULL;
2900 if (in2->is_DecodeNarrowPtr()) {
2901 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2902 new_in2 = in2->in(1);
2903 } else if (in2->Opcode() == Op_ConP) {
2904 const Type* t = in2->bottom_type();
2905 if (t == TypePtr::NULL_PTR) {
2906 assert(in1->is_DecodeN(), "compare klass to null?");
2907 // Don't convert CmpP null check into CmpN if compressed
2908 // oops implicit null check is not generated.
2909 // This will allow to generate normal oop implicit null check.
2910 if (Matcher::gen_narrow_oop_implicit_null_checks())
2911 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2912 //
2913 // This transformation together with CastPP transformation above
2914 // will generated code for implicit NULL checks for compressed oops.
2915 //
2916 // The original code after Optimize()
2917 //
2918 // LoadN memory, narrow_oop_reg
2919 // decode narrow_oop_reg, base_reg
2920 // CmpP base_reg, NULL
2921 // CastPP base_reg // NotNull
2922 // Load [base_reg + offset], val_reg
2923 //
2924 // after these transformations will be
2925 //
2926 // LoadN memory, narrow_oop_reg
2927 // CmpN narrow_oop_reg, NULL
2928 // decode_not_null narrow_oop_reg, base_reg
2929 // Load [base_reg + offset], val_reg
2930 //
2931 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2932 // since narrow oops can be used in debug info now (see the code in
2933 // final_graph_reshaping_walk()).
2934 //
2935 // At the end the code will be matched to
2936 // on x86:
2937 //
2938 // Load_narrow_oop memory, narrow_oop_reg
2939 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2940 // NullCheck narrow_oop_reg
2941 //
2942 // and on sparc:
2943 //
2944 // Load_narrow_oop memory, narrow_oop_reg
2945 // decode_not_null narrow_oop_reg, base_reg
2946 // Load [base_reg + offset], val_reg
2947 // NullCheck base_reg
2948 //
2949 } else if (t->isa_oopptr()) {
2950 new_in2 = ConNode::make(this, t->make_narrowoop());
2951 } else if (t->isa_klassptr()) {
2952 new_in2 = ConNode::make(this, t->make_narrowklass());
2953 }
2954 }
2955 if (new_in2 != NULL) {
2956 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2957 n->subsume_by(cmpN, this);
2958 if (in1->outcnt() == 0) {
2959 in1->disconnect_inputs(NULL, this);
2960 }
2961 if (in2->outcnt() == 0) {
2962 in2->disconnect_inputs(NULL, this);
2963 }
2964 }
2965 }
2966 break;
2968 case Op_DecodeN:
2969 case Op_DecodeNKlass:
2970 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2971 // DecodeN could be pinned when it can't be fold into
2972 // an address expression, see the code for Op_CastPP above.
2973 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2974 break;
2976 case Op_EncodeP:
2977 case Op_EncodePKlass: {
2978 Node* in1 = n->in(1);
2979 if (in1->is_DecodeNarrowPtr()) {
2980 n->subsume_by(in1->in(1), this);
2981 } else if (in1->Opcode() == Op_ConP) {
2982 const Type* t = in1->bottom_type();
2983 if (t == TypePtr::NULL_PTR) {
2984 assert(t->isa_oopptr(), "null klass?");
2985 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2986 } else if (t->isa_oopptr()) {
2987 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2988 } else if (t->isa_klassptr()) {
2989 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2990 }
2991 }
2992 if (in1->outcnt() == 0) {
2993 in1->disconnect_inputs(NULL, this);
2994 }
2995 break;
2996 }
2998 case Op_Proj: {
2999 if (OptimizeStringConcat) {
3000 ProjNode* p = n->as_Proj();
3001 if (p->_is_io_use) {
3002 // Separate projections were used for the exception path which
3003 // are normally removed by a late inline. If it wasn't inlined
3004 // then they will hang around and should just be replaced with
3005 // the original one.
3006 Node* proj = NULL;
3007 // Replace with just one
3008 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3009 Node *use = i.get();
3010 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3011 proj = use;
3012 break;
3013 }
3014 }
3015 assert(proj != NULL, "must be found");
3016 p->subsume_by(proj, this);
3017 }
3018 }
3019 break;
3020 }
3022 case Op_Phi:
3023 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3024 // The EncodeP optimization may create Phi with the same edges
3025 // for all paths. It is not handled well by Register Allocator.
3026 Node* unique_in = n->in(1);
3027 assert(unique_in != NULL, "");
3028 uint cnt = n->req();
3029 for (uint i = 2; i < cnt; i++) {
3030 Node* m = n->in(i);
3031 assert(m != NULL, "");
3032 if (unique_in != m)
3033 unique_in = NULL;
3034 }
3035 if (unique_in != NULL) {
3036 n->subsume_by(unique_in, this);
3037 }
3038 }
3039 break;
3041 #endif
3043 #ifdef ASSERT
3044 case Op_CastII:
3045 // Verify that all range check dependent CastII nodes were removed.
3046 if (n->isa_CastII()->has_range_check()) {
3047 n->dump(3);
3048 assert(false, "Range check dependent CastII node was not removed");
3049 }
3050 break;
3051 #endif
3053 case Op_ModI:
3054 if (UseDivMod) {
3055 // Check if a%b and a/b both exist
3056 Node* d = n->find_similar(Op_DivI);
3057 if (d) {
3058 // Replace them with a fused divmod if supported
3059 if (Matcher::has_match_rule(Op_DivModI)) {
3060 DivModINode* divmod = DivModINode::make(this, n);
3061 d->subsume_by(divmod->div_proj(), this);
3062 n->subsume_by(divmod->mod_proj(), this);
3063 } else {
3064 // replace a%b with a-((a/b)*b)
3065 Node* mult = new (this) MulINode(d, d->in(2));
3066 Node* sub = new (this) SubINode(d->in(1), mult);
3067 n->subsume_by(sub, this);
3068 }
3069 }
3070 }
3071 break;
3073 case Op_ModL:
3074 if (UseDivMod) {
3075 // Check if a%b and a/b both exist
3076 Node* d = n->find_similar(Op_DivL);
3077 if (d) {
3078 // Replace them with a fused divmod if supported
3079 if (Matcher::has_match_rule(Op_DivModL)) {
3080 DivModLNode* divmod = DivModLNode::make(this, n);
3081 d->subsume_by(divmod->div_proj(), this);
3082 n->subsume_by(divmod->mod_proj(), this);
3083 } else {
3084 // replace a%b with a-((a/b)*b)
3085 Node* mult = new (this) MulLNode(d, d->in(2));
3086 Node* sub = new (this) SubLNode(d->in(1), mult);
3087 n->subsume_by(sub, this);
3088 }
3089 }
3090 }
3091 break;
3093 case Op_LoadVector:
3094 case Op_StoreVector:
3095 break;
3097 case Op_PackB:
3098 case Op_PackS:
3099 case Op_PackI:
3100 case Op_PackF:
3101 case Op_PackL:
3102 case Op_PackD:
3103 if (n->req()-1 > 2) {
3104 // Replace many operand PackNodes with a binary tree for matching
3105 PackNode* p = (PackNode*) n;
3106 Node* btp = p->binary_tree_pack(this, 1, n->req());
3107 n->subsume_by(btp, this);
3108 }
3109 break;
3110 case Op_Loop:
3111 case Op_CountedLoop:
3112 if (n->as_Loop()->is_inner_loop()) {
3113 frc.inc_inner_loop_count();
3114 }
3115 break;
3116 case Op_LShiftI:
3117 case Op_RShiftI:
3118 case Op_URShiftI:
3119 case Op_LShiftL:
3120 case Op_RShiftL:
3121 case Op_URShiftL:
3122 if (Matcher::need_masked_shift_count) {
3123 // The cpu's shift instructions don't restrict the count to the
3124 // lower 5/6 bits. We need to do the masking ourselves.
3125 Node* in2 = n->in(2);
3126 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3127 const TypeInt* t = in2->find_int_type();
3128 if (t != NULL && t->is_con()) {
3129 juint shift = t->get_con();
3130 if (shift > mask) { // Unsigned cmp
3131 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
3132 }
3133 } else {
3134 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3135 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3136 n->set_req(2, shift);
3137 }
3138 }
3139 if (in2->outcnt() == 0) { // Remove dead node
3140 in2->disconnect_inputs(NULL, this);
3141 }
3142 }
3143 break;
3144 case Op_MemBarStoreStore:
3145 case Op_MemBarRelease:
3146 // Break the link with AllocateNode: it is no longer useful and
3147 // confuses register allocation.
3148 if (n->req() > MemBarNode::Precedent) {
3149 n->set_req(MemBarNode::Precedent, top());
3150 }
3151 break;
3152 default:
3153 assert( !n->is_Call(), "" );
3154 assert( !n->is_Mem(), "" );
3155 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3156 break;
3157 }
3159 // Collect CFG split points
3160 if (n->is_MultiBranch())
3161 frc._tests.push(n);
3162 }
3164 //------------------------------final_graph_reshaping_walk---------------------
3165 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3166 // requires that the walk visits a node's inputs before visiting the node.
3167 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3168 ResourceArea *area = Thread::current()->resource_area();
3169 Unique_Node_List sfpt(area);
3171 frc._visited.set(root->_idx); // first, mark node as visited
3172 uint cnt = root->req();
3173 Node *n = root;
3174 uint i = 0;
3175 while (true) {
3176 if (i < cnt) {
3177 // Place all non-visited non-null inputs onto stack
3178 Node* m = n->in(i);
3179 ++i;
3180 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3181 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3182 // compute worst case interpreter size in case of a deoptimization
3183 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3185 sfpt.push(m);
3186 }
3187 cnt = m->req();
3188 nstack.push(n, i); // put on stack parent and next input's index
3189 n = m;
3190 i = 0;
3191 }
3192 } else {
3193 // Now do post-visit work
3194 final_graph_reshaping_impl( n, frc );
3195 if (nstack.is_empty())
3196 break; // finished
3197 n = nstack.node(); // Get node from stack
3198 cnt = n->req();
3199 i = nstack.index();
3200 nstack.pop(); // Shift to the next node on stack
3201 }
3202 }
3204 // Skip next transformation if compressed oops are not used.
3205 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3206 (!UseCompressedOops && !UseCompressedClassPointers))
3207 return;
3209 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3210 // It could be done for an uncommon traps or any safepoints/calls
3211 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3212 while (sfpt.size() > 0) {
3213 n = sfpt.pop();
3214 JVMState *jvms = n->as_SafePoint()->jvms();
3215 assert(jvms != NULL, "sanity");
3216 int start = jvms->debug_start();
3217 int end = n->req();
3218 bool is_uncommon = (n->is_CallStaticJava() &&
3219 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3220 for (int j = start; j < end; j++) {
3221 Node* in = n->in(j);
3222 if (in->is_DecodeNarrowPtr()) {
3223 bool safe_to_skip = true;
3224 if (!is_uncommon ) {
3225 // Is it safe to skip?
3226 for (uint i = 0; i < in->outcnt(); i++) {
3227 Node* u = in->raw_out(i);
3228 if (!u->is_SafePoint() ||
3229 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3230 safe_to_skip = false;
3231 }
3232 }
3233 }
3234 if (safe_to_skip) {
3235 n->set_req(j, in->in(1));
3236 }
3237 if (in->outcnt() == 0) {
3238 in->disconnect_inputs(NULL, this);
3239 }
3240 }
3241 }
3242 }
3243 }
3245 //------------------------------final_graph_reshaping--------------------------
3246 // Final Graph Reshaping.
3247 //
3248 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3249 // and not commoned up and forced early. Must come after regular
3250 // optimizations to avoid GVN undoing the cloning. Clone constant
3251 // inputs to Loop Phis; these will be split by the allocator anyways.
3252 // Remove Opaque nodes.
3253 // (2) Move last-uses by commutative operations to the left input to encourage
3254 // Intel update-in-place two-address operations and better register usage
3255 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3256 // calls canonicalizing them back.
3257 // (3) Count the number of double-precision FP ops, single-precision FP ops
3258 // and call sites. On Intel, we can get correct rounding either by
3259 // forcing singles to memory (requires extra stores and loads after each
3260 // FP bytecode) or we can set a rounding mode bit (requires setting and
3261 // clearing the mode bit around call sites). The mode bit is only used
3262 // if the relative frequency of single FP ops to calls is low enough.
3263 // This is a key transform for SPEC mpeg_audio.
3264 // (4) Detect infinite loops; blobs of code reachable from above but not
3265 // below. Several of the Code_Gen algorithms fail on such code shapes,
3266 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3267 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3268 // Detection is by looking for IfNodes where only 1 projection is
3269 // reachable from below or CatchNodes missing some targets.
3270 // (5) Assert for insane oop offsets in debug mode.
3272 bool Compile::final_graph_reshaping() {
3273 // an infinite loop may have been eliminated by the optimizer,
3274 // in which case the graph will be empty.
3275 if (root()->req() == 1) {
3276 record_method_not_compilable("trivial infinite loop");
3277 return true;
3278 }
3280 // Expensive nodes have their control input set to prevent the GVN
3281 // from freely commoning them. There's no GVN beyond this point so
3282 // no need to keep the control input. We want the expensive nodes to
3283 // be freely moved to the least frequent code path by gcm.
3284 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3285 for (int i = 0; i < expensive_count(); i++) {
3286 _expensive_nodes->at(i)->set_req(0, NULL);
3287 }
3289 Final_Reshape_Counts frc;
3291 // Visit everybody reachable!
3292 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3293 Node_Stack nstack(live_nodes() >> 1);
3294 final_graph_reshaping_walk(nstack, root(), frc);
3296 // Check for unreachable (from below) code (i.e., infinite loops).
3297 for( uint i = 0; i < frc._tests.size(); i++ ) {
3298 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3299 // Get number of CFG targets.
3300 // Note that PCTables include exception targets after calls.
3301 uint required_outcnt = n->required_outcnt();
3302 if (n->outcnt() != required_outcnt) {
3303 // Check for a few special cases. Rethrow Nodes never take the
3304 // 'fall-thru' path, so expected kids is 1 less.
3305 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3306 if (n->in(0)->in(0)->is_Call()) {
3307 CallNode *call = n->in(0)->in(0)->as_Call();
3308 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3309 required_outcnt--; // Rethrow always has 1 less kid
3310 } else if (call->req() > TypeFunc::Parms &&
3311 call->is_CallDynamicJava()) {
3312 // Check for null receiver. In such case, the optimizer has
3313 // detected that the virtual call will always result in a null
3314 // pointer exception. The fall-through projection of this CatchNode
3315 // will not be populated.
3316 Node *arg0 = call->in(TypeFunc::Parms);
3317 if (arg0->is_Type() &&
3318 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3319 required_outcnt--;
3320 }
3321 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3322 call->req() > TypeFunc::Parms+1 &&
3323 call->is_CallStaticJava()) {
3324 // Check for negative array length. In such case, the optimizer has
3325 // detected that the allocation attempt will always result in an
3326 // exception. There is no fall-through projection of this CatchNode .
3327 Node *arg1 = call->in(TypeFunc::Parms+1);
3328 if (arg1->is_Type() &&
3329 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3330 required_outcnt--;
3331 }
3332 }
3333 }
3334 }
3335 // Recheck with a better notion of 'required_outcnt'
3336 if (n->outcnt() != required_outcnt) {
3337 record_method_not_compilable("malformed control flow");
3338 return true; // Not all targets reachable!
3339 }
3340 }
3341 // Check that I actually visited all kids. Unreached kids
3342 // must be infinite loops.
3343 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3344 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3345 record_method_not_compilable("infinite loop");
3346 return true; // Found unvisited kid; must be unreach
3347 }
3348 }
3350 // If original bytecodes contained a mixture of floats and doubles
3351 // check if the optimizer has made it homogenous, item (3).
3352 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3353 frc.get_float_count() > 32 &&
3354 frc.get_double_count() == 0 &&
3355 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3356 set_24_bit_selection_and_mode( false, true );
3357 }
3359 set_java_calls(frc.get_java_call_count());
3360 set_inner_loops(frc.get_inner_loop_count());
3362 // No infinite loops, no reason to bail out.
3363 return false;
3364 }
3366 //-----------------------------too_many_traps----------------------------------
3367 // Report if there are too many traps at the current method and bci.
3368 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3369 bool Compile::too_many_traps(ciMethod* method,
3370 int bci,
3371 Deoptimization::DeoptReason reason) {
3372 ciMethodData* md = method->method_data();
3373 if (md->is_empty()) {
3374 // Assume the trap has not occurred, or that it occurred only
3375 // because of a transient condition during start-up in the interpreter.
3376 return false;
3377 }
3378 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3379 if (md->has_trap_at(bci, m, reason) != 0) {
3380 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3381 // Also, if there are multiple reasons, or if there is no per-BCI record,
3382 // assume the worst.
3383 if (log())
3384 log()->elem("observe trap='%s' count='%d'",
3385 Deoptimization::trap_reason_name(reason),
3386 md->trap_count(reason));
3387 return true;
3388 } else {
3389 // Ignore method/bci and see if there have been too many globally.
3390 return too_many_traps(reason, md);
3391 }
3392 }
3394 // Less-accurate variant which does not require a method and bci.
3395 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3396 ciMethodData* logmd) {
3397 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3398 // Too many traps globally.
3399 // Note that we use cumulative trap_count, not just md->trap_count.
3400 if (log()) {
3401 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3402 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3403 Deoptimization::trap_reason_name(reason),
3404 mcount, trap_count(reason));
3405 }
3406 return true;
3407 } else {
3408 // The coast is clear.
3409 return false;
3410 }
3411 }
3413 //--------------------------too_many_recompiles--------------------------------
3414 // Report if there are too many recompiles at the current method and bci.
3415 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3416 // Is not eager to return true, since this will cause the compiler to use
3417 // Action_none for a trap point, to avoid too many recompilations.
3418 bool Compile::too_many_recompiles(ciMethod* method,
3419 int bci,
3420 Deoptimization::DeoptReason reason) {
3421 ciMethodData* md = method->method_data();
3422 if (md->is_empty()) {
3423 // Assume the trap has not occurred, or that it occurred only
3424 // because of a transient condition during start-up in the interpreter.
3425 return false;
3426 }
3427 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3428 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3429 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3430 Deoptimization::DeoptReason per_bc_reason
3431 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3432 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3433 if ((per_bc_reason == Deoptimization::Reason_none
3434 || md->has_trap_at(bci, m, reason) != 0)
3435 // The trap frequency measure we care about is the recompile count:
3436 && md->trap_recompiled_at(bci, m)
3437 && md->overflow_recompile_count() >= bc_cutoff) {
3438 // Do not emit a trap here if it has already caused recompilations.
3439 // Also, if there are multiple reasons, or if there is no per-BCI record,
3440 // assume the worst.
3441 if (log())
3442 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3443 Deoptimization::trap_reason_name(reason),
3444 md->trap_count(reason),
3445 md->overflow_recompile_count());
3446 return true;
3447 } else if (trap_count(reason) != 0
3448 && decompile_count() >= m_cutoff) {
3449 // Too many recompiles globally, and we have seen this sort of trap.
3450 // Use cumulative decompile_count, not just md->decompile_count.
3451 if (log())
3452 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3453 Deoptimization::trap_reason_name(reason),
3454 md->trap_count(reason), trap_count(reason),
3455 md->decompile_count(), decompile_count());
3456 return true;
3457 } else {
3458 // The coast is clear.
3459 return false;
3460 }
3461 }
3463 // Compute when not to trap. Used by matching trap based nodes and
3464 // NullCheck optimization.
3465 void Compile::set_allowed_deopt_reasons() {
3466 _allowed_reasons = 0;
3467 if (is_method_compilation()) {
3468 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3469 assert(rs < BitsPerInt, "recode bit map");
3470 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3471 _allowed_reasons |= nth_bit(rs);
3472 }
3473 }
3474 }
3475 }
3477 #ifndef PRODUCT
3478 //------------------------------verify_graph_edges---------------------------
3479 // Walk the Graph and verify that there is a one-to-one correspondence
3480 // between Use-Def edges and Def-Use edges in the graph.
3481 void Compile::verify_graph_edges(bool no_dead_code) {
3482 if (VerifyGraphEdges) {
3483 ResourceArea *area = Thread::current()->resource_area();
3484 Unique_Node_List visited(area);
3485 // Call recursive graph walk to check edges
3486 _root->verify_edges(visited);
3487 if (no_dead_code) {
3488 // Now make sure that no visited node is used by an unvisited node.
3489 bool dead_nodes = false;
3490 Unique_Node_List checked(area);
3491 while (visited.size() > 0) {
3492 Node* n = visited.pop();
3493 checked.push(n);
3494 for (uint i = 0; i < n->outcnt(); i++) {
3495 Node* use = n->raw_out(i);
3496 if (checked.member(use)) continue; // already checked
3497 if (visited.member(use)) continue; // already in the graph
3498 if (use->is_Con()) continue; // a dead ConNode is OK
3499 // At this point, we have found a dead node which is DU-reachable.
3500 if (!dead_nodes) {
3501 tty->print_cr("*** Dead nodes reachable via DU edges:");
3502 dead_nodes = true;
3503 }
3504 use->dump(2);
3505 tty->print_cr("---");
3506 checked.push(use); // No repeats; pretend it is now checked.
3507 }
3508 }
3509 assert(!dead_nodes, "using nodes must be reachable from root");
3510 }
3511 }
3512 }
3514 // Verify GC barriers consistency
3515 // Currently supported:
3516 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3517 void Compile::verify_barriers() {
3518 if (UseG1GC) {
3519 // Verify G1 pre-barriers
3520 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3522 ResourceArea *area = Thread::current()->resource_area();
3523 Unique_Node_List visited(area);
3524 Node_List worklist(area);
3525 // We're going to walk control flow backwards starting from the Root
3526 worklist.push(_root);
3527 while (worklist.size() > 0) {
3528 Node* x = worklist.pop();
3529 if (x == NULL || x == top()) continue;
3530 if (visited.member(x)) {
3531 continue;
3532 } else {
3533 visited.push(x);
3534 }
3536 if (x->is_Region()) {
3537 for (uint i = 1; i < x->req(); i++) {
3538 worklist.push(x->in(i));
3539 }
3540 } else {
3541 worklist.push(x->in(0));
3542 // We are looking for the pattern:
3543 // /->ThreadLocal
3544 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3545 // \->ConI(0)
3546 // We want to verify that the If and the LoadB have the same control
3547 // See GraphKit::g1_write_barrier_pre()
3548 if (x->is_If()) {
3549 IfNode *iff = x->as_If();
3550 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3551 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3552 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3553 && cmp->in(1)->is_Load()) {
3554 LoadNode* load = cmp->in(1)->as_Load();
3555 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3556 && load->in(2)->in(3)->is_Con()
3557 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3559 Node* if_ctrl = iff->in(0);
3560 Node* load_ctrl = load->in(0);
3562 if (if_ctrl != load_ctrl) {
3563 // Skip possible CProj->NeverBranch in infinite loops
3564 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3565 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3566 if_ctrl = if_ctrl->in(0)->in(0);
3567 }
3568 }
3569 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3570 }
3571 }
3572 }
3573 }
3574 }
3575 }
3576 }
3577 }
3579 #endif
3581 // The Compile object keeps track of failure reasons separately from the ciEnv.
3582 // This is required because there is not quite a 1-1 relation between the
3583 // ciEnv and its compilation task and the Compile object. Note that one
3584 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3585 // to backtrack and retry without subsuming loads. Other than this backtracking
3586 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3587 // by the logic in C2Compiler.
3588 void Compile::record_failure(const char* reason) {
3589 if (log() != NULL) {
3590 log()->elem("failure reason='%s' phase='compile'", reason);
3591 }
3592 if (_failure_reason == NULL) {
3593 // Record the first failure reason.
3594 _failure_reason = reason;
3595 }
3597 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3598 C->print_method(PHASE_FAILURE);
3599 }
3600 _root = NULL; // flush the graph, too
3601 }
3603 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3604 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3605 _phase_name(name), _dolog(dolog)
3606 {
3607 if (dolog) {
3608 C = Compile::current();
3609 _log = C->log();
3610 } else {
3611 C = NULL;
3612 _log = NULL;
3613 }
3614 if (_log != NULL) {
3615 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3616 _log->stamp();
3617 _log->end_head();
3618 }
3619 }
3621 Compile::TracePhase::~TracePhase() {
3623 C = Compile::current();
3624 if (_dolog) {
3625 _log = C->log();
3626 } else {
3627 _log = NULL;
3628 }
3630 #ifdef ASSERT
3631 if (PrintIdealNodeCount) {
3632 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3633 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3634 }
3636 if (VerifyIdealNodeCount) {
3637 Compile::current()->print_missing_nodes();
3638 }
3639 #endif
3641 if (_log != NULL) {
3642 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3643 }
3644 }
3646 //=============================================================================
3647 // Two Constant's are equal when the type and the value are equal.
3648 bool Compile::Constant::operator==(const Constant& other) {
3649 if (type() != other.type() ) return false;
3650 if (can_be_reused() != other.can_be_reused()) return false;
3651 // For floating point values we compare the bit pattern.
3652 switch (type()) {
3653 case T_FLOAT: return (_v._value.i == other._v._value.i);
3654 case T_LONG:
3655 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3656 case T_OBJECT:
3657 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3658 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3659 case T_METADATA: return (_v._metadata == other._v._metadata);
3660 default: ShouldNotReachHere();
3661 }
3662 return false;
3663 }
3665 static int type_to_size_in_bytes(BasicType t) {
3666 switch (t) {
3667 case T_LONG: return sizeof(jlong );
3668 case T_FLOAT: return sizeof(jfloat );
3669 case T_DOUBLE: return sizeof(jdouble);
3670 case T_METADATA: return sizeof(Metadata*);
3671 // We use T_VOID as marker for jump-table entries (labels) which
3672 // need an internal word relocation.
3673 case T_VOID:
3674 case T_ADDRESS:
3675 case T_OBJECT: return sizeof(jobject);
3676 }
3678 ShouldNotReachHere();
3679 return -1;
3680 }
3682 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3683 // sort descending
3684 if (a->freq() > b->freq()) return -1;
3685 if (a->freq() < b->freq()) return 1;
3686 return 0;
3687 }
3689 void Compile::ConstantTable::calculate_offsets_and_size() {
3690 // First, sort the array by frequencies.
3691 _constants.sort(qsort_comparator);
3693 #ifdef ASSERT
3694 // Make sure all jump-table entries were sorted to the end of the
3695 // array (they have a negative frequency).
3696 bool found_void = false;
3697 for (int i = 0; i < _constants.length(); i++) {
3698 Constant con = _constants.at(i);
3699 if (con.type() == T_VOID)
3700 found_void = true; // jump-tables
3701 else
3702 assert(!found_void, "wrong sorting");
3703 }
3704 #endif
3706 int offset = 0;
3707 for (int i = 0; i < _constants.length(); i++) {
3708 Constant* con = _constants.adr_at(i);
3710 // Align offset for type.
3711 int typesize = type_to_size_in_bytes(con->type());
3712 offset = align_size_up(offset, typesize);
3713 con->set_offset(offset); // set constant's offset
3715 if (con->type() == T_VOID) {
3716 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3717 offset = offset + typesize * n->outcnt(); // expand jump-table
3718 } else {
3719 offset = offset + typesize;
3720 }
3721 }
3723 // Align size up to the next section start (which is insts; see
3724 // CodeBuffer::align_at_start).
3725 assert(_size == -1, "already set?");
3726 _size = align_size_up(offset, CodeEntryAlignment);
3727 }
3729 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3730 MacroAssembler _masm(&cb);
3731 for (int i = 0; i < _constants.length(); i++) {
3732 Constant con = _constants.at(i);
3733 address constant_addr = NULL;
3734 switch (con.type()) {
3735 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3736 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3737 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3738 case T_OBJECT: {
3739 jobject obj = con.get_jobject();
3740 int oop_index = _masm.oop_recorder()->find_index(obj);
3741 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3742 break;
3743 }
3744 case T_ADDRESS: {
3745 address addr = (address) con.get_jobject();
3746 constant_addr = _masm.address_constant(addr);
3747 break;
3748 }
3749 // We use T_VOID as marker for jump-table entries (labels) which
3750 // need an internal word relocation.
3751 case T_VOID: {
3752 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3753 // Fill the jump-table with a dummy word. The real value is
3754 // filled in later in fill_jump_table.
3755 address dummy = (address) n;
3756 constant_addr = _masm.address_constant(dummy);
3757 // Expand jump-table
3758 for (uint i = 1; i < n->outcnt(); i++) {
3759 address temp_addr = _masm.address_constant(dummy + i);
3760 assert(temp_addr, "consts section too small");
3761 }
3762 break;
3763 }
3764 case T_METADATA: {
3765 Metadata* obj = con.get_metadata();
3766 int metadata_index = _masm.oop_recorder()->find_index(obj);
3767 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3768 break;
3769 }
3770 default: ShouldNotReachHere();
3771 }
3772 assert(constant_addr, "consts section too small");
3773 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3774 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())));
3775 }
3776 }
3778 int Compile::ConstantTable::find_offset(Constant& con) const {
3779 int idx = _constants.find(con);
3780 assert(idx != -1, "constant must be in constant table");
3781 int offset = _constants.at(idx).offset();
3782 assert(offset != -1, "constant table not emitted yet?");
3783 return offset;
3784 }
3786 void Compile::ConstantTable::add(Constant& con) {
3787 if (con.can_be_reused()) {
3788 int idx = _constants.find(con);
3789 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3790 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3791 return;
3792 }
3793 }
3794 (void) _constants.append(con);
3795 }
3797 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3798 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3799 Constant con(type, value, b->_freq);
3800 add(con);
3801 return con;
3802 }
3804 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3805 Constant con(metadata);
3806 add(con);
3807 return con;
3808 }
3810 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3811 jvalue value;
3812 BasicType type = oper->type()->basic_type();
3813 switch (type) {
3814 case T_LONG: value.j = oper->constantL(); break;
3815 case T_FLOAT: value.f = oper->constantF(); break;
3816 case T_DOUBLE: value.d = oper->constantD(); break;
3817 case T_OBJECT:
3818 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3819 case T_METADATA: return add((Metadata*)oper->constant()); break;
3820 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3821 }
3822 return add(n, type, value);
3823 }
3825 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3826 jvalue value;
3827 // We can use the node pointer here to identify the right jump-table
3828 // as this method is called from Compile::Fill_buffer right before
3829 // the MachNodes are emitted and the jump-table is filled (means the
3830 // MachNode pointers do not change anymore).
3831 value.l = (jobject) n;
3832 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3833 add(con);
3834 return con;
3835 }
3837 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3838 // If called from Compile::scratch_emit_size do nothing.
3839 if (Compile::current()->in_scratch_emit_size()) return;
3841 assert(labels.is_nonempty(), "must be");
3842 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3844 // Since MachConstantNode::constant_offset() also contains
3845 // table_base_offset() we need to subtract the table_base_offset()
3846 // to get the plain offset into the constant table.
3847 int offset = n->constant_offset() - table_base_offset();
3849 MacroAssembler _masm(&cb);
3850 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3852 for (uint i = 0; i < n->outcnt(); i++) {
3853 address* constant_addr = &jump_table_base[i];
3854 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i)));
3855 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3856 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3857 }
3858 }
3860 void Compile::dump_inlining() {
3861 if (print_inlining() || print_intrinsics()) {
3862 // Print inlining message for candidates that we couldn't inline
3863 // for lack of space or non constant receiver
3864 for (int i = 0; i < _late_inlines.length(); i++) {
3865 CallGenerator* cg = _late_inlines.at(i);
3866 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3867 }
3868 Unique_Node_List useful;
3869 useful.push(root());
3870 for (uint next = 0; next < useful.size(); ++next) {
3871 Node* n = useful.at(next);
3872 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3873 CallNode* call = n->as_Call();
3874 CallGenerator* cg = call->generator();
3875 cg->print_inlining_late("receiver not constant");
3876 }
3877 uint max = n->len();
3878 for ( uint i = 0; i < max; ++i ) {
3879 Node *m = n->in(i);
3880 if ( m == NULL ) continue;
3881 useful.push(m);
3882 }
3883 }
3884 for (int i = 0; i < _print_inlining_list->length(); i++) {
3885 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
3886 }
3887 }
3888 }
3890 // Dump inlining replay data to the stream.
3891 // Don't change thread state and acquire any locks.
3892 void Compile::dump_inline_data(outputStream* out) {
3893 InlineTree* inl_tree = ilt();
3894 if (inl_tree != NULL) {
3895 out->print(" inline %d", inl_tree->count());
3896 inl_tree->dump_replay_data(out);
3897 }
3898 }
3900 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3901 if (n1->Opcode() < n2->Opcode()) return -1;
3902 else if (n1->Opcode() > n2->Opcode()) return 1;
3904 assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()));
3905 for (uint i = 1; i < n1->req(); i++) {
3906 if (n1->in(i) < n2->in(i)) return -1;
3907 else if (n1->in(i) > n2->in(i)) return 1;
3908 }
3910 return 0;
3911 }
3913 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3914 Node* n1 = *n1p;
3915 Node* n2 = *n2p;
3917 return cmp_expensive_nodes(n1, n2);
3918 }
3920 void Compile::sort_expensive_nodes() {
3921 if (!expensive_nodes_sorted()) {
3922 _expensive_nodes->sort(cmp_expensive_nodes);
3923 }
3924 }
3926 bool Compile::expensive_nodes_sorted() const {
3927 for (int i = 1; i < _expensive_nodes->length(); i++) {
3928 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3929 return false;
3930 }
3931 }
3932 return true;
3933 }
3935 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3936 if (_expensive_nodes->length() == 0) {
3937 return false;
3938 }
3940 assert(OptimizeExpensiveOps, "optimization off?");
3942 // Take this opportunity to remove dead nodes from the list
3943 int j = 0;
3944 for (int i = 0; i < _expensive_nodes->length(); i++) {
3945 Node* n = _expensive_nodes->at(i);
3946 if (!n->is_unreachable(igvn)) {
3947 assert(n->is_expensive(), "should be expensive");
3948 _expensive_nodes->at_put(j, n);
3949 j++;
3950 }
3951 }
3952 _expensive_nodes->trunc_to(j);
3954 // Then sort the list so that similar nodes are next to each other
3955 // and check for at least two nodes of identical kind with same data
3956 // inputs.
3957 sort_expensive_nodes();
3959 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3960 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3961 return true;
3962 }
3963 }
3965 return false;
3966 }
3968 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3969 if (_expensive_nodes->length() == 0) {
3970 return;
3971 }
3973 assert(OptimizeExpensiveOps, "optimization off?");
3975 // Sort to bring similar nodes next to each other and clear the
3976 // control input of nodes for which there's only a single copy.
3977 sort_expensive_nodes();
3979 int j = 0;
3980 int identical = 0;
3981 int i = 0;
3982 for (; i < _expensive_nodes->length()-1; i++) {
3983 assert(j <= i, "can't write beyond current index");
3984 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3985 identical++;
3986 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3987 continue;
3988 }
3989 if (identical > 0) {
3990 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3991 identical = 0;
3992 } else {
3993 Node* n = _expensive_nodes->at(i);
3994 igvn.hash_delete(n);
3995 n->set_req(0, NULL);
3996 igvn.hash_insert(n);
3997 }
3998 }
3999 if (identical > 0) {
4000 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4001 } else if (_expensive_nodes->length() >= 1) {
4002 Node* n = _expensive_nodes->at(i);
4003 igvn.hash_delete(n);
4004 n->set_req(0, NULL);
4005 igvn.hash_insert(n);
4006 }
4007 _expensive_nodes->trunc_to(j);
4008 }
4010 void Compile::add_expensive_node(Node * n) {
4011 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4012 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4013 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4014 if (OptimizeExpensiveOps) {
4015 _expensive_nodes->append(n);
4016 } else {
4017 // Clear control input and let IGVN optimize expensive nodes if
4018 // OptimizeExpensiveOps is off.
4019 n->set_req(0, NULL);
4020 }
4021 }
4023 /**
4024 * Remove the speculative part of types and clean up the graph
4025 */
4026 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4027 if (UseTypeSpeculation) {
4028 Unique_Node_List worklist;
4029 worklist.push(root());
4030 int modified = 0;
4031 // Go over all type nodes that carry a speculative type, drop the
4032 // speculative part of the type and enqueue the node for an igvn
4033 // which may optimize it out.
4034 for (uint next = 0; next < worklist.size(); ++next) {
4035 Node *n = worklist.at(next);
4036 if (n->is_Type()) {
4037 TypeNode* tn = n->as_Type();
4038 const Type* t = tn->type();
4039 const Type* t_no_spec = t->remove_speculative();
4040 if (t_no_spec != t) {
4041 bool in_hash = igvn.hash_delete(n);
4042 assert(in_hash, "node should be in igvn hash table");
4043 tn->set_type(t_no_spec);
4044 igvn.hash_insert(n);
4045 igvn._worklist.push(n); // give it a chance to go away
4046 modified++;
4047 }
4048 }
4049 uint max = n->len();
4050 for( uint i = 0; i < max; ++i ) {
4051 Node *m = n->in(i);
4052 if (not_a_node(m)) continue;
4053 worklist.push(m);
4054 }
4055 }
4056 // Drop the speculative part of all types in the igvn's type table
4057 igvn.remove_speculative_types();
4058 if (modified > 0) {
4059 igvn.optimize();
4060 }
4061 #ifdef ASSERT
4062 // Verify that after the IGVN is over no speculative type has resurfaced
4063 worklist.clear();
4064 worklist.push(root());
4065 for (uint next = 0; next < worklist.size(); ++next) {
4066 Node *n = worklist.at(next);
4067 const Type* t = igvn.type_or_null(n);
4068 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4069 if (n->is_Type()) {
4070 t = n->as_Type()->type();
4071 assert(t == t->remove_speculative(), "no more speculative types");
4072 }
4073 uint max = n->len();
4074 for( uint i = 0; i < max; ++i ) {
4075 Node *m = n->in(i);
4076 if (not_a_node(m)) continue;
4077 worklist.push(m);
4078 }
4079 }
4080 igvn.check_no_speculative_types();
4081 #endif
4082 }
4083 }
4085 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4086 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4087 if (ctrl != NULL) {
4088 // Express control dependency by a CastII node with a narrow type.
4089 value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */);
4090 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4091 // node from floating above the range check during loop optimizations. Otherwise, the
4092 // ConvI2L node may be eliminated independently of the range check, causing the data path
4093 // to become TOP while the control path is still there (although it's unreachable).
4094 value->set_req(0, ctrl);
4095 // Save CastII node to remove it after loop optimizations.
4096 phase->C->add_range_check_cast(value);
4097 value = phase->transform(value);
4098 }
4099 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4100 return phase->transform(new (phase->C) ConvI2LNode(value, ltype));
4101 }
4103 // Auxiliary method to support randomized stressing/fuzzing.
4104 //
4105 // This method can be called the arbitrary number of times, with current count
4106 // as the argument. The logic allows selecting a single candidate from the
4107 // running list of candidates as follows:
4108 // int count = 0;
4109 // Cand* selected = null;
4110 // while(cand = cand->next()) {
4111 // if (randomized_select(++count)) {
4112 // selected = cand;
4113 // }
4114 // }
4115 //
4116 // Including count equalizes the chances any candidate is "selected".
4117 // This is useful when we don't have the complete list of candidates to choose
4118 // from uniformly. In this case, we need to adjust the randomicity of the
4119 // selection, or else we will end up biasing the selection towards the latter
4120 // candidates.
4121 //
4122 // Quick back-envelope calculation shows that for the list of n candidates
4123 // the equal probability for the candidate to persist as "best" can be
4124 // achieved by replacing it with "next" k-th candidate with the probability
4125 // of 1/k. It can be easily shown that by the end of the run, the
4126 // probability for any candidate is converged to 1/n, thus giving the
4127 // uniform distribution among all the candidates.
4128 //
4129 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4130 #define RANDOMIZED_DOMAIN_POW 29
4131 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4132 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4133 bool Compile::randomized_select(int count) {
4134 assert(count > 0, "only positive");
4135 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4136 }