Fri, 04 Mar 2016 16:15:48 +0300
8130150: Implement BigInteger.montgomeryMultiply intrinsic
Reviewed-by: kvn, mdoerr
1 /*
2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "compiler/disassembler.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/block.hpp"
37 #include "opto/c2compiler.hpp"
38 #include "opto/callGenerator.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/cfgnode.hpp"
41 #include "opto/chaitin.hpp"
42 #include "opto/compile.hpp"
43 #include "opto/connode.hpp"
44 #include "opto/divnode.hpp"
45 #include "opto/escape.hpp"
46 #include "opto/idealGraphPrinter.hpp"
47 #include "opto/loopnode.hpp"
48 #include "opto/machnode.hpp"
49 #include "opto/macro.hpp"
50 #include "opto/matcher.hpp"
51 #include "opto/mathexactnode.hpp"
52 #include "opto/memnode.hpp"
53 #include "opto/mulnode.hpp"
54 #include "opto/node.hpp"
55 #include "opto/opcodes.hpp"
56 #include "opto/output.hpp"
57 #include "opto/parse.hpp"
58 #include "opto/phaseX.hpp"
59 #include "opto/rootnode.hpp"
60 #include "opto/runtime.hpp"
61 #include "opto/stringopts.hpp"
62 #include "opto/type.hpp"
63 #include "opto/vectornode.hpp"
64 #include "runtime/arguments.hpp"
65 #include "runtime/signature.hpp"
66 #include "runtime/stubRoutines.hpp"
67 #include "runtime/timer.hpp"
68 #include "trace/tracing.hpp"
69 #include "utilities/copy.hpp"
70 #if defined AD_MD_HPP
71 # include AD_MD_HPP
72 #elif defined TARGET_ARCH_MODEL_x86_32
73 # include "adfiles/ad_x86_32.hpp"
74 #elif defined TARGET_ARCH_MODEL_x86_64
75 # include "adfiles/ad_x86_64.hpp"
76 #elif defined TARGET_ARCH_MODEL_sparc
77 # include "adfiles/ad_sparc.hpp"
78 #elif defined TARGET_ARCH_MODEL_zero
79 # include "adfiles/ad_zero.hpp"
80 #elif defined TARGET_ARCH_MODEL_ppc_64
81 # include "adfiles/ad_ppc_64.hpp"
82 #endif
85 // -------------------- Compile::mach_constant_base_node -----------------------
86 // Constant table base node singleton.
87 MachConstantBaseNode* Compile::mach_constant_base_node() {
88 if (_mach_constant_base_node == NULL) {
89 _mach_constant_base_node = new (C) MachConstantBaseNode();
90 _mach_constant_base_node->add_req(C->root());
91 }
92 return _mach_constant_base_node;
93 }
96 /// Support for intrinsics.
98 // Return the index at which m must be inserted (or already exists).
99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
100 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
101 #ifdef ASSERT
102 for (int i = 1; i < _intrinsics->length(); i++) {
103 CallGenerator* cg1 = _intrinsics->at(i-1);
104 CallGenerator* cg2 = _intrinsics->at(i);
105 assert(cg1->method() != cg2->method()
106 ? cg1->method() < cg2->method()
107 : cg1->is_virtual() < cg2->is_virtual(),
108 "compiler intrinsics list must stay sorted");
109 }
110 #endif
111 // Binary search sorted list, in decreasing intervals [lo, hi].
112 int lo = 0, hi = _intrinsics->length()-1;
113 while (lo <= hi) {
114 int mid = (uint)(hi + lo) / 2;
115 ciMethod* mid_m = _intrinsics->at(mid)->method();
116 if (m < mid_m) {
117 hi = mid-1;
118 } else if (m > mid_m) {
119 lo = mid+1;
120 } else {
121 // look at minor sort key
122 bool mid_virt = _intrinsics->at(mid)->is_virtual();
123 if (is_virtual < mid_virt) {
124 hi = mid-1;
125 } else if (is_virtual > mid_virt) {
126 lo = mid+1;
127 } else {
128 return mid; // exact match
129 }
130 }
131 }
132 return lo; // inexact match
133 }
135 void Compile::register_intrinsic(CallGenerator* cg) {
136 if (_intrinsics == NULL) {
137 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
138 }
139 // This code is stolen from ciObjectFactory::insert.
140 // Really, GrowableArray should have methods for
141 // insert_at, remove_at, and binary_search.
142 int len = _intrinsics->length();
143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
144 if (index == len) {
145 _intrinsics->append(cg);
146 } else {
147 #ifdef ASSERT
148 CallGenerator* oldcg = _intrinsics->at(index);
149 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
150 #endif
151 _intrinsics->append(_intrinsics->at(len-1));
152 int pos;
153 for (pos = len-2; pos >= index; pos--) {
154 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
155 }
156 _intrinsics->at_put(index, cg);
157 }
158 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
159 }
161 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
162 assert(m->is_loaded(), "don't try this on unloaded methods");
163 if (_intrinsics != NULL) {
164 int index = intrinsic_insertion_index(m, is_virtual);
165 if (index < _intrinsics->length()
166 && _intrinsics->at(index)->method() == m
167 && _intrinsics->at(index)->is_virtual() == is_virtual) {
168 return _intrinsics->at(index);
169 }
170 }
171 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
172 if (m->intrinsic_id() != vmIntrinsics::_none &&
173 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
174 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
175 if (cg != NULL) {
176 // Save it for next time:
177 register_intrinsic(cg);
178 return cg;
179 } else {
180 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
181 }
182 }
183 return NULL;
184 }
186 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
187 // in library_call.cpp.
190 #ifndef PRODUCT
191 // statistics gathering...
193 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
194 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
196 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
197 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
198 int oflags = _intrinsic_hist_flags[id];
199 assert(flags != 0, "what happened?");
200 if (is_virtual) {
201 flags |= _intrinsic_virtual;
202 }
203 bool changed = (flags != oflags);
204 if ((flags & _intrinsic_worked) != 0) {
205 juint count = (_intrinsic_hist_count[id] += 1);
206 if (count == 1) {
207 changed = true; // first time
208 }
209 // increment the overall count also:
210 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
211 }
212 if (changed) {
213 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
214 // Something changed about the intrinsic's virtuality.
215 if ((flags & _intrinsic_virtual) != 0) {
216 // This is the first use of this intrinsic as a virtual call.
217 if (oflags != 0) {
218 // We already saw it as a non-virtual, so note both cases.
219 flags |= _intrinsic_both;
220 }
221 } else if ((oflags & _intrinsic_both) == 0) {
222 // This is the first use of this intrinsic as a non-virtual
223 flags |= _intrinsic_both;
224 }
225 }
226 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
227 }
228 // update the overall flags also:
229 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
230 return changed;
231 }
233 static char* format_flags(int flags, char* buf) {
234 buf[0] = 0;
235 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
236 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
237 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
238 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
239 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
240 if (buf[0] == 0) strcat(buf, ",");
241 assert(buf[0] == ',', "must be");
242 return &buf[1];
243 }
245 void Compile::print_intrinsic_statistics() {
246 char flagsbuf[100];
247 ttyLocker ttyl;
248 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
249 tty->print_cr("Compiler intrinsic usage:");
250 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
251 if (total == 0) total = 1; // avoid div0 in case of no successes
252 #define PRINT_STAT_LINE(name, c, f) \
253 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
254 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
255 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
256 int flags = _intrinsic_hist_flags[id];
257 juint count = _intrinsic_hist_count[id];
258 if ((flags | count) != 0) {
259 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
260 }
261 }
262 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
263 if (xtty != NULL) xtty->tail("statistics");
264 }
266 void Compile::print_statistics() {
267 { ttyLocker ttyl;
268 if (xtty != NULL) xtty->head("statistics type='opto'");
269 Parse::print_statistics();
270 PhaseCCP::print_statistics();
271 PhaseRegAlloc::print_statistics();
272 Scheduling::print_statistics();
273 PhasePeephole::print_statistics();
274 PhaseIdealLoop::print_statistics();
275 if (xtty != NULL) xtty->tail("statistics");
276 }
277 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
278 // put this under its own <statistics> element.
279 print_intrinsic_statistics();
280 }
281 }
282 #endif //PRODUCT
284 // Support for bundling info
285 Bundle* Compile::node_bundling(const Node *n) {
286 assert(valid_bundle_info(n), "oob");
287 return &_node_bundling_base[n->_idx];
288 }
290 bool Compile::valid_bundle_info(const Node *n) {
291 return (_node_bundling_limit > n->_idx);
292 }
295 void Compile::gvn_replace_by(Node* n, Node* nn) {
296 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
297 Node* use = n->last_out(i);
298 bool is_in_table = initial_gvn()->hash_delete(use);
299 uint uses_found = 0;
300 for (uint j = 0; j < use->len(); j++) {
301 if (use->in(j) == n) {
302 if (j < use->req())
303 use->set_req(j, nn);
304 else
305 use->set_prec(j, nn);
306 uses_found++;
307 }
308 }
309 if (is_in_table) {
310 // reinsert into table
311 initial_gvn()->hash_find_insert(use);
312 }
313 record_for_igvn(use);
314 i -= uses_found; // we deleted 1 or more copies of this edge
315 }
316 }
319 static inline bool not_a_node(const Node* n) {
320 if (n == NULL) return true;
321 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
322 if (*(address*)n == badAddress) return true; // kill by Node::destruct
323 return false;
324 }
326 // Identify all nodes that are reachable from below, useful.
327 // Use breadth-first pass that records state in a Unique_Node_List,
328 // recursive traversal is slower.
329 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
330 int estimated_worklist_size = live_nodes();
331 useful.map( estimated_worklist_size, NULL ); // preallocate space
333 // Initialize worklist
334 if (root() != NULL) { useful.push(root()); }
335 // If 'top' is cached, declare it useful to preserve cached node
336 if( cached_top_node() ) { useful.push(cached_top_node()); }
338 // Push all useful nodes onto the list, breadthfirst
339 for( uint next = 0; next < useful.size(); ++next ) {
340 assert( next < unique(), "Unique useful nodes < total nodes");
341 Node *n = useful.at(next);
342 uint max = n->len();
343 for( uint i = 0; i < max; ++i ) {
344 Node *m = n->in(i);
345 if (not_a_node(m)) continue;
346 useful.push(m);
347 }
348 }
349 }
351 // Update dead_node_list with any missing dead nodes using useful
352 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
353 void Compile::update_dead_node_list(Unique_Node_List &useful) {
354 uint max_idx = unique();
355 VectorSet& useful_node_set = useful.member_set();
357 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
358 // If node with index node_idx is not in useful set,
359 // mark it as dead in dead node list.
360 if (! useful_node_set.test(node_idx) ) {
361 record_dead_node(node_idx);
362 }
363 }
364 }
366 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
367 int shift = 0;
368 for (int i = 0; i < inlines->length(); i++) {
369 CallGenerator* cg = inlines->at(i);
370 CallNode* call = cg->call_node();
371 if (shift > 0) {
372 inlines->at_put(i-shift, cg);
373 }
374 if (!useful.member(call)) {
375 shift++;
376 }
377 }
378 inlines->trunc_to(inlines->length()-shift);
379 }
381 // Disconnect all useless nodes by disconnecting those at the boundary.
382 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
383 uint next = 0;
384 while (next < useful.size()) {
385 Node *n = useful.at(next++);
386 if (n->is_SafePoint()) {
387 // We're done with a parsing phase. Replaced nodes are not valid
388 // beyond that point.
389 n->as_SafePoint()->delete_replaced_nodes();
390 }
391 // Use raw traversal of out edges since this code removes out edges
392 int max = n->outcnt();
393 for (int j = 0; j < max; ++j) {
394 Node* child = n->raw_out(j);
395 if (! useful.member(child)) {
396 assert(!child->is_top() || child != top(),
397 "If top is cached in Compile object it is in useful list");
398 // Only need to remove this out-edge to the useless node
399 n->raw_del_out(j);
400 --j;
401 --max;
402 }
403 }
404 if (n->outcnt() == 1 && n->has_special_unique_user()) {
405 record_for_igvn(n->unique_out());
406 }
407 }
408 // Remove useless macro and predicate opaq nodes
409 for (int i = C->macro_count()-1; i >= 0; i--) {
410 Node* n = C->macro_node(i);
411 if (!useful.member(n)) {
412 remove_macro_node(n);
413 }
414 }
415 // Remove useless CastII nodes with range check dependency
416 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
417 Node* cast = range_check_cast_node(i);
418 if (!useful.member(cast)) {
419 remove_range_check_cast(cast);
420 }
421 }
422 // Remove useless expensive node
423 for (int i = C->expensive_count()-1; i >= 0; i--) {
424 Node* n = C->expensive_node(i);
425 if (!useful.member(n)) {
426 remove_expensive_node(n);
427 }
428 }
429 // clean up the late inline lists
430 remove_useless_late_inlines(&_string_late_inlines, useful);
431 remove_useless_late_inlines(&_boxing_late_inlines, useful);
432 remove_useless_late_inlines(&_late_inlines, useful);
433 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
434 }
436 //------------------------------frame_size_in_words-----------------------------
437 // frame_slots in units of words
438 int Compile::frame_size_in_words() const {
439 // shift is 0 in LP32 and 1 in LP64
440 const int shift = (LogBytesPerWord - LogBytesPerInt);
441 int words = _frame_slots >> shift;
442 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
443 return words;
444 }
446 // To bang the stack of this compiled method we use the stack size
447 // that the interpreter would need in case of a deoptimization. This
448 // removes the need to bang the stack in the deoptimization blob which
449 // in turn simplifies stack overflow handling.
450 int Compile::bang_size_in_bytes() const {
451 return MAX2(_interpreter_frame_size, frame_size_in_bytes());
452 }
454 // ============================================================================
455 //------------------------------CompileWrapper---------------------------------
456 class CompileWrapper : public StackObj {
457 Compile *const _compile;
458 public:
459 CompileWrapper(Compile* compile);
461 ~CompileWrapper();
462 };
464 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
465 // the Compile* pointer is stored in the current ciEnv:
466 ciEnv* env = compile->env();
467 assert(env == ciEnv::current(), "must already be a ciEnv active");
468 assert(env->compiler_data() == NULL, "compile already active?");
469 env->set_compiler_data(compile);
470 assert(compile == Compile::current(), "sanity");
472 compile->set_type_dict(NULL);
473 compile->set_type_hwm(NULL);
474 compile->set_type_last_size(0);
475 compile->set_last_tf(NULL, NULL);
476 compile->set_indexSet_arena(NULL);
477 compile->set_indexSet_free_block_list(NULL);
478 compile->init_type_arena();
479 Type::Initialize(compile);
480 _compile->set_scratch_buffer_blob(NULL);
481 _compile->begin_method();
482 }
483 CompileWrapper::~CompileWrapper() {
484 _compile->end_method();
485 if (_compile->scratch_buffer_blob() != NULL)
486 BufferBlob::free(_compile->scratch_buffer_blob());
487 _compile->env()->set_compiler_data(NULL);
488 }
491 //----------------------------print_compile_messages---------------------------
492 void Compile::print_compile_messages() {
493 #ifndef PRODUCT
494 // Check if recompiling
495 if (_subsume_loads == false && PrintOpto) {
496 // Recompiling without allowing machine instructions to subsume loads
497 tty->print_cr("*********************************************************");
498 tty->print_cr("** Bailout: Recompile without subsuming loads **");
499 tty->print_cr("*********************************************************");
500 }
501 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
502 // Recompiling without escape analysis
503 tty->print_cr("*********************************************************");
504 tty->print_cr("** Bailout: Recompile without escape analysis **");
505 tty->print_cr("*********************************************************");
506 }
507 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
508 // Recompiling without boxing elimination
509 tty->print_cr("*********************************************************");
510 tty->print_cr("** Bailout: Recompile without boxing elimination **");
511 tty->print_cr("*********************************************************");
512 }
513 if (env()->break_at_compile()) {
514 // Open the debugger when compiling this method.
515 tty->print("### Breaking when compiling: ");
516 method()->print_short_name();
517 tty->cr();
518 BREAKPOINT;
519 }
521 if( PrintOpto ) {
522 if (is_osr_compilation()) {
523 tty->print("[OSR]%3d", _compile_id);
524 } else {
525 tty->print("%3d", _compile_id);
526 }
527 }
528 #endif
529 }
532 //-----------------------init_scratch_buffer_blob------------------------------
533 // Construct a temporary BufferBlob and cache it for this compile.
534 void Compile::init_scratch_buffer_blob(int const_size) {
535 // If there is already a scratch buffer blob allocated and the
536 // constant section is big enough, use it. Otherwise free the
537 // current and allocate a new one.
538 BufferBlob* blob = scratch_buffer_blob();
539 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
540 // Use the current blob.
541 } else {
542 if (blob != NULL) {
543 BufferBlob::free(blob);
544 }
546 ResourceMark rm;
547 _scratch_const_size = const_size;
548 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
549 blob = BufferBlob::create("Compile::scratch_buffer", size);
550 // Record the buffer blob for next time.
551 set_scratch_buffer_blob(blob);
552 // Have we run out of code space?
553 if (scratch_buffer_blob() == NULL) {
554 // Let CompilerBroker disable further compilations.
555 record_failure("Not enough space for scratch buffer in CodeCache");
556 return;
557 }
558 }
560 // Initialize the relocation buffers
561 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
562 set_scratch_locs_memory(locs_buf);
563 }
566 //-----------------------scratch_emit_size-------------------------------------
567 // Helper function that computes size by emitting code
568 uint Compile::scratch_emit_size(const Node* n) {
569 // Start scratch_emit_size section.
570 set_in_scratch_emit_size(true);
572 // Emit into a trash buffer and count bytes emitted.
573 // This is a pretty expensive way to compute a size,
574 // but it works well enough if seldom used.
575 // All common fixed-size instructions are given a size
576 // method by the AD file.
577 // Note that the scratch buffer blob and locs memory are
578 // allocated at the beginning of the compile task, and
579 // may be shared by several calls to scratch_emit_size.
580 // The allocation of the scratch buffer blob is particularly
581 // expensive, since it has to grab the code cache lock.
582 BufferBlob* blob = this->scratch_buffer_blob();
583 assert(blob != NULL, "Initialize BufferBlob at start");
584 assert(blob->size() > MAX_inst_size, "sanity");
585 relocInfo* locs_buf = scratch_locs_memory();
586 address blob_begin = blob->content_begin();
587 address blob_end = (address)locs_buf;
588 assert(blob->content_contains(blob_end), "sanity");
589 CodeBuffer buf(blob_begin, blob_end - blob_begin);
590 buf.initialize_consts_size(_scratch_const_size);
591 buf.initialize_stubs_size(MAX_stubs_size);
592 assert(locs_buf != NULL, "sanity");
593 int lsize = MAX_locs_size / 3;
594 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
595 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
596 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
598 // Do the emission.
600 Label fakeL; // Fake label for branch instructions.
601 Label* saveL = NULL;
602 uint save_bnum = 0;
603 bool is_branch = n->is_MachBranch();
604 if (is_branch) {
605 MacroAssembler masm(&buf);
606 masm.bind(fakeL);
607 n->as_MachBranch()->save_label(&saveL, &save_bnum);
608 n->as_MachBranch()->label_set(&fakeL, 0);
609 }
610 n->emit(buf, this->regalloc());
611 if (is_branch) // Restore label.
612 n->as_MachBranch()->label_set(saveL, save_bnum);
614 // End scratch_emit_size section.
615 set_in_scratch_emit_size(false);
617 return buf.insts_size();
618 }
621 // ============================================================================
622 //------------------------------Compile standard-------------------------------
623 debug_only( int Compile::_debug_idx = 100000; )
625 // Compile a method. entry_bci is -1 for normal compilations and indicates
626 // the continuation bci for on stack replacement.
629 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
630 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
631 : Phase(Compiler),
632 _env(ci_env),
633 _log(ci_env->log()),
634 _compile_id(ci_env->compile_id()),
635 _save_argument_registers(false),
636 _stub_name(NULL),
637 _stub_function(NULL),
638 _stub_entry_point(NULL),
639 _method(target),
640 _entry_bci(osr_bci),
641 _initial_gvn(NULL),
642 _for_igvn(NULL),
643 _warm_calls(NULL),
644 _subsume_loads(subsume_loads),
645 _do_escape_analysis(do_escape_analysis),
646 _eliminate_boxing(eliminate_boxing),
647 _failure_reason(NULL),
648 _code_buffer("Compile::Fill_buffer"),
649 _orig_pc_slot(0),
650 _orig_pc_slot_offset_in_bytes(0),
651 _has_method_handle_invokes(false),
652 _mach_constant_base_node(NULL),
653 _node_bundling_limit(0),
654 _node_bundling_base(NULL),
655 _java_calls(0),
656 _inner_loops(0),
657 _scratch_const_size(-1),
658 _in_scratch_emit_size(false),
659 _dead_node_list(comp_arena()),
660 _dead_node_count(0),
661 #ifndef PRODUCT
662 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
663 _in_dump_cnt(0),
664 _printer(IdealGraphPrinter::printer()),
665 #endif
666 _congraph(NULL),
667 _comp_arena(mtCompiler),
668 _node_arena(mtCompiler),
669 _old_arena(mtCompiler),
670 _Compile_types(mtCompiler),
671 _replay_inline_data(NULL),
672 _late_inlines(comp_arena(), 2, 0, NULL),
673 _string_late_inlines(comp_arena(), 2, 0, NULL),
674 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
675 _late_inlines_pos(0),
676 _number_of_mh_late_inlines(0),
677 _inlining_progress(false),
678 _inlining_incrementally(false),
679 _print_inlining_list(NULL),
680 _print_inlining_idx(0),
681 _interpreter_frame_size(0),
682 _max_node_limit(MaxNodeLimit) {
683 C = this;
685 CompileWrapper cw(this);
686 #ifndef PRODUCT
687 if (TimeCompiler2) {
688 tty->print(" ");
689 target->holder()->name()->print();
690 tty->print(".");
691 target->print_short_name();
692 tty->print(" ");
693 }
694 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
695 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
696 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
697 if (!print_opto_assembly) {
698 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
699 if (print_assembly && !Disassembler::can_decode()) {
700 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
701 print_opto_assembly = true;
702 }
703 }
704 set_print_assembly(print_opto_assembly);
705 set_parsed_irreducible_loop(false);
707 if (method()->has_option("ReplayInline")) {
708 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
709 }
710 #endif
711 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
712 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
713 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
715 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
716 // Make sure the method being compiled gets its own MDO,
717 // so we can at least track the decompile_count().
718 // Need MDO to record RTM code generation state.
719 method()->ensure_method_data();
720 }
722 Init(::AliasLevel);
725 print_compile_messages();
727 _ilt = InlineTree::build_inline_tree_root();
729 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
730 assert(num_alias_types() >= AliasIdxRaw, "");
732 #define MINIMUM_NODE_HASH 1023
733 // Node list that Iterative GVN will start with
734 Unique_Node_List for_igvn(comp_arena());
735 set_for_igvn(&for_igvn);
737 // GVN that will be run immediately on new nodes
738 uint estimated_size = method()->code_size()*4+64;
739 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
740 PhaseGVN gvn(node_arena(), estimated_size);
741 set_initial_gvn(&gvn);
743 if (print_inlining() || print_intrinsics()) {
744 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
745 }
746 { // Scope for timing the parser
747 TracePhase t3("parse", &_t_parser, true);
749 // Put top into the hash table ASAP.
750 initial_gvn()->transform_no_reclaim(top());
752 // Set up tf(), start(), and find a CallGenerator.
753 CallGenerator* cg = NULL;
754 if (is_osr_compilation()) {
755 const TypeTuple *domain = StartOSRNode::osr_domain();
756 const TypeTuple *range = TypeTuple::make_range(method()->signature());
757 init_tf(TypeFunc::make(domain, range));
758 StartNode* s = new (this) StartOSRNode(root(), domain);
759 initial_gvn()->set_type_bottom(s);
760 init_start(s);
761 cg = CallGenerator::for_osr(method(), entry_bci());
762 } else {
763 // Normal case.
764 init_tf(TypeFunc::make(method()));
765 StartNode* s = new (this) StartNode(root(), tf()->domain());
766 initial_gvn()->set_type_bottom(s);
767 init_start(s);
768 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
769 // With java.lang.ref.reference.get() we must go through the
770 // intrinsic when G1 is enabled - even when get() is the root
771 // method of the compile - so that, if necessary, the value in
772 // the referent field of the reference object gets recorded by
773 // the pre-barrier code.
774 // Specifically, if G1 is enabled, the value in the referent
775 // field is recorded by the G1 SATB pre barrier. This will
776 // result in the referent being marked live and the reference
777 // object removed from the list of discovered references during
778 // reference processing.
779 cg = find_intrinsic(method(), false);
780 }
781 if (cg == NULL) {
782 float past_uses = method()->interpreter_invocation_count();
783 float expected_uses = past_uses;
784 cg = CallGenerator::for_inline(method(), expected_uses);
785 }
786 }
787 if (failing()) return;
788 if (cg == NULL) {
789 record_method_not_compilable_all_tiers("cannot parse method");
790 return;
791 }
792 JVMState* jvms = build_start_state(start(), tf());
793 if ((jvms = cg->generate(jvms)) == NULL) {
794 record_method_not_compilable("method parse failed");
795 return;
796 }
797 GraphKit kit(jvms);
799 if (!kit.stopped()) {
800 // Accept return values, and transfer control we know not where.
801 // This is done by a special, unique ReturnNode bound to root.
802 return_values(kit.jvms());
803 }
805 if (kit.has_exceptions()) {
806 // Any exceptions that escape from this call must be rethrown
807 // to whatever caller is dynamically above us on the stack.
808 // This is done by a special, unique RethrowNode bound to root.
809 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
810 }
812 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
814 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
815 inline_string_calls(true);
816 }
818 if (failing()) return;
820 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
822 // Remove clutter produced by parsing.
823 if (!failing()) {
824 ResourceMark rm;
825 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
826 }
827 }
829 // Note: Large methods are capped off in do_one_bytecode().
830 if (failing()) return;
832 // After parsing, node notes are no longer automagic.
833 // They must be propagated by register_new_node_with_optimizer(),
834 // clone(), or the like.
835 set_default_node_notes(NULL);
837 for (;;) {
838 int successes = Inline_Warm();
839 if (failing()) return;
840 if (successes == 0) break;
841 }
843 // Drain the list.
844 Finish_Warm();
845 #ifndef PRODUCT
846 if (_printer) {
847 _printer->print_inlining(this);
848 }
849 #endif
851 if (failing()) return;
852 NOT_PRODUCT( verify_graph_edges(); )
854 // Now optimize
855 Optimize();
856 if (failing()) return;
857 NOT_PRODUCT( verify_graph_edges(); )
859 #ifndef PRODUCT
860 if (PrintIdeal) {
861 ttyLocker ttyl; // keep the following output all in one block
862 // This output goes directly to the tty, not the compiler log.
863 // To enable tools to match it up with the compilation activity,
864 // be sure to tag this tty output with the compile ID.
865 if (xtty != NULL) {
866 xtty->head("ideal compile_id='%d'%s", compile_id(),
867 is_osr_compilation() ? " compile_kind='osr'" :
868 "");
869 }
870 root()->dump(9999);
871 if (xtty != NULL) {
872 xtty->tail("ideal");
873 }
874 }
875 #endif
877 NOT_PRODUCT( verify_barriers(); )
879 // Dump compilation data to replay it.
880 if (method()->has_option("DumpReplay")) {
881 env()->dump_replay_data(_compile_id);
882 }
883 if (method()->has_option("DumpInline") && (ilt() != NULL)) {
884 env()->dump_inline_data(_compile_id);
885 }
887 // Now that we know the size of all the monitors we can add a fixed slot
888 // for the original deopt pc.
890 _orig_pc_slot = fixed_slots();
891 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
892 set_fixed_slots(next_slot);
894 // Compute when to use implicit null checks. Used by matching trap based
895 // nodes and NullCheck optimization.
896 set_allowed_deopt_reasons();
898 // Now generate code
899 Code_Gen();
900 if (failing()) return;
902 // Check if we want to skip execution of all compiled code.
903 {
904 #ifndef PRODUCT
905 if (OptoNoExecute) {
906 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
907 return;
908 }
909 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
910 #endif
912 if (is_osr_compilation()) {
913 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
914 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
915 } else {
916 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
917 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
918 }
920 env()->register_method(_method, _entry_bci,
921 &_code_offsets,
922 _orig_pc_slot_offset_in_bytes,
923 code_buffer(),
924 frame_size_in_words(), _oop_map_set,
925 &_handler_table, &_inc_table,
926 compiler,
927 env()->comp_level(),
928 has_unsafe_access(),
929 SharedRuntime::is_wide_vector(max_vector_size()),
930 rtm_state()
931 );
933 if (log() != NULL) // Print code cache state into compiler log
934 log()->code_cache_state();
935 }
936 }
938 //------------------------------Compile----------------------------------------
939 // Compile a runtime stub
940 Compile::Compile( ciEnv* ci_env,
941 TypeFunc_generator generator,
942 address stub_function,
943 const char *stub_name,
944 int is_fancy_jump,
945 bool pass_tls,
946 bool save_arg_registers,
947 bool return_pc )
948 : Phase(Compiler),
949 _env(ci_env),
950 _log(ci_env->log()),
951 _compile_id(0),
952 _save_argument_registers(save_arg_registers),
953 _method(NULL),
954 _stub_name(stub_name),
955 _stub_function(stub_function),
956 _stub_entry_point(NULL),
957 _entry_bci(InvocationEntryBci),
958 _initial_gvn(NULL),
959 _for_igvn(NULL),
960 _warm_calls(NULL),
961 _orig_pc_slot(0),
962 _orig_pc_slot_offset_in_bytes(0),
963 _subsume_loads(true),
964 _do_escape_analysis(false),
965 _eliminate_boxing(false),
966 _failure_reason(NULL),
967 _code_buffer("Compile::Fill_buffer"),
968 _has_method_handle_invokes(false),
969 _mach_constant_base_node(NULL),
970 _node_bundling_limit(0),
971 _node_bundling_base(NULL),
972 _java_calls(0),
973 _inner_loops(0),
974 #ifndef PRODUCT
975 _trace_opto_output(TraceOptoOutput),
976 _in_dump_cnt(0),
977 _printer(NULL),
978 #endif
979 _comp_arena(mtCompiler),
980 _node_arena(mtCompiler),
981 _old_arena(mtCompiler),
982 _Compile_types(mtCompiler),
983 _dead_node_list(comp_arena()),
984 _dead_node_count(0),
985 _congraph(NULL),
986 _replay_inline_data(NULL),
987 _number_of_mh_late_inlines(0),
988 _inlining_progress(false),
989 _inlining_incrementally(false),
990 _print_inlining_list(NULL),
991 _print_inlining_idx(0),
992 _allowed_reasons(0),
993 _interpreter_frame_size(0),
994 _max_node_limit(MaxNodeLimit) {
995 C = this;
997 #ifndef PRODUCT
998 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
999 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
1000 set_print_assembly(PrintFrameConverterAssembly);
1001 set_parsed_irreducible_loop(false);
1002 #endif
1003 set_has_irreducible_loop(false); // no loops
1005 CompileWrapper cw(this);
1006 Init(/*AliasLevel=*/ 0);
1007 init_tf((*generator)());
1009 {
1010 // The following is a dummy for the sake of GraphKit::gen_stub
1011 Unique_Node_List for_igvn(comp_arena());
1012 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1013 PhaseGVN gvn(Thread::current()->resource_area(),255);
1014 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1015 gvn.transform_no_reclaim(top());
1017 GraphKit kit;
1018 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1019 }
1021 NOT_PRODUCT( verify_graph_edges(); )
1022 Code_Gen();
1023 if (failing()) return;
1026 // Entry point will be accessed using compile->stub_entry_point();
1027 if (code_buffer() == NULL) {
1028 Matcher::soft_match_failure();
1029 } else {
1030 if (PrintAssembly && (WizardMode || Verbose))
1031 tty->print_cr("### Stub::%s", stub_name);
1033 if (!failing()) {
1034 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1036 // Make the NMethod
1037 // For now we mark the frame as never safe for profile stackwalking
1038 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1039 code_buffer(),
1040 CodeOffsets::frame_never_safe,
1041 // _code_offsets.value(CodeOffsets::Frame_Complete),
1042 frame_size_in_words(),
1043 _oop_map_set,
1044 save_arg_registers);
1045 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1047 _stub_entry_point = rs->entry_point();
1048 }
1049 }
1050 }
1052 //------------------------------Init-------------------------------------------
1053 // Prepare for a single compilation
1054 void Compile::Init(int aliaslevel) {
1055 _unique = 0;
1056 _regalloc = NULL;
1058 _tf = NULL; // filled in later
1059 _top = NULL; // cached later
1060 _matcher = NULL; // filled in later
1061 _cfg = NULL; // filled in later
1063 set_24_bit_selection_and_mode(Use24BitFP, false);
1065 _node_note_array = NULL;
1066 _default_node_notes = NULL;
1068 _immutable_memory = NULL; // filled in at first inquiry
1070 // Globally visible Nodes
1071 // First set TOP to NULL to give safe behavior during creation of RootNode
1072 set_cached_top_node(NULL);
1073 set_root(new (this) RootNode());
1074 // Now that you have a Root to point to, create the real TOP
1075 set_cached_top_node( new (this) ConNode(Type::TOP) );
1076 set_recent_alloc(NULL, NULL);
1078 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1079 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1080 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1081 env()->set_dependencies(new Dependencies(env()));
1083 _fixed_slots = 0;
1084 set_has_split_ifs(false);
1085 set_has_loops(has_method() && method()->has_loops()); // first approximation
1086 set_has_stringbuilder(false);
1087 set_has_boxed_value(false);
1088 _trap_can_recompile = false; // no traps emitted yet
1089 _major_progress = true; // start out assuming good things will happen
1090 set_has_unsafe_access(false);
1091 set_max_vector_size(0);
1092 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1093 set_decompile_count(0);
1095 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1096 set_num_loop_opts(LoopOptsCount);
1097 set_do_inlining(Inline);
1098 set_max_inline_size(MaxInlineSize);
1099 set_freq_inline_size(FreqInlineSize);
1100 set_do_scheduling(OptoScheduling);
1101 set_do_count_invocations(false);
1102 set_do_method_data_update(false);
1103 set_rtm_state(NoRTM); // No RTM lock eliding by default
1104 method_has_option_value("MaxNodeLimit", _max_node_limit);
1105 #if INCLUDE_RTM_OPT
1106 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1107 int rtm_state = method()->method_data()->rtm_state();
1108 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1109 // Don't generate RTM lock eliding code.
1110 set_rtm_state(NoRTM);
1111 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1112 // Generate RTM lock eliding code without abort ratio calculation code.
1113 set_rtm_state(UseRTM);
1114 } else if (UseRTMDeopt) {
1115 // Generate RTM lock eliding code and include abort ratio calculation
1116 // code if UseRTMDeopt is on.
1117 set_rtm_state(ProfileRTM);
1118 }
1119 }
1120 #endif
1121 if (debug_info()->recording_non_safepoints()) {
1122 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1123 (comp_arena(), 8, 0, NULL));
1124 set_default_node_notes(Node_Notes::make(this));
1125 }
1127 // // -- Initialize types before each compile --
1128 // // Update cached type information
1129 // if( _method && _method->constants() )
1130 // Type::update_loaded_types(_method, _method->constants());
1132 // Init alias_type map.
1133 if (!_do_escape_analysis && aliaslevel == 3)
1134 aliaslevel = 2; // No unique types without escape analysis
1135 _AliasLevel = aliaslevel;
1136 const int grow_ats = 16;
1137 _max_alias_types = grow_ats;
1138 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1139 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1140 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1141 {
1142 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1143 }
1144 // Initialize the first few types.
1145 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1146 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1147 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1148 _num_alias_types = AliasIdxRaw+1;
1149 // Zero out the alias type cache.
1150 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1151 // A NULL adr_type hits in the cache right away. Preload the right answer.
1152 probe_alias_cache(NULL)->_index = AliasIdxTop;
1154 _intrinsics = NULL;
1155 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1156 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1157 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1158 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1159 register_library_intrinsics();
1160 }
1162 //---------------------------init_start----------------------------------------
1163 // Install the StartNode on this compile object.
1164 void Compile::init_start(StartNode* s) {
1165 if (failing())
1166 return; // already failing
1167 assert(s == start(), "");
1168 }
1170 StartNode* Compile::start() const {
1171 assert(!failing(), "");
1172 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1173 Node* start = root()->fast_out(i);
1174 if( start->is_Start() )
1175 return start->as_Start();
1176 }
1177 fatal("Did not find Start node!");
1178 return NULL;
1179 }
1181 //-------------------------------immutable_memory-------------------------------------
1182 // Access immutable memory
1183 Node* Compile::immutable_memory() {
1184 if (_immutable_memory != NULL) {
1185 return _immutable_memory;
1186 }
1187 StartNode* s = start();
1188 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1189 Node *p = s->fast_out(i);
1190 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1191 _immutable_memory = p;
1192 return _immutable_memory;
1193 }
1194 }
1195 ShouldNotReachHere();
1196 return NULL;
1197 }
1199 //----------------------set_cached_top_node------------------------------------
1200 // Install the cached top node, and make sure Node::is_top works correctly.
1201 void Compile::set_cached_top_node(Node* tn) {
1202 if (tn != NULL) verify_top(tn);
1203 Node* old_top = _top;
1204 _top = tn;
1205 // Calling Node::setup_is_top allows the nodes the chance to adjust
1206 // their _out arrays.
1207 if (_top != NULL) _top->setup_is_top();
1208 if (old_top != NULL) old_top->setup_is_top();
1209 assert(_top == NULL || top()->is_top(), "");
1210 }
1212 #ifdef ASSERT
1213 uint Compile::count_live_nodes_by_graph_walk() {
1214 Unique_Node_List useful(comp_arena());
1215 // Get useful node list by walking the graph.
1216 identify_useful_nodes(useful);
1217 return useful.size();
1218 }
1220 void Compile::print_missing_nodes() {
1222 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1223 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1224 return;
1225 }
1227 // This is an expensive function. It is executed only when the user
1228 // specifies VerifyIdealNodeCount option or otherwise knows the
1229 // additional work that needs to be done to identify reachable nodes
1230 // by walking the flow graph and find the missing ones using
1231 // _dead_node_list.
1233 Unique_Node_List useful(comp_arena());
1234 // Get useful node list by walking the graph.
1235 identify_useful_nodes(useful);
1237 uint l_nodes = C->live_nodes();
1238 uint l_nodes_by_walk = useful.size();
1240 if (l_nodes != l_nodes_by_walk) {
1241 if (_log != NULL) {
1242 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1243 _log->stamp();
1244 _log->end_head();
1245 }
1246 VectorSet& useful_member_set = useful.member_set();
1247 int last_idx = l_nodes_by_walk;
1248 for (int i = 0; i < last_idx; i++) {
1249 if (useful_member_set.test(i)) {
1250 if (_dead_node_list.test(i)) {
1251 if (_log != NULL) {
1252 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1253 }
1254 if (PrintIdealNodeCount) {
1255 // Print the log message to tty
1256 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1257 useful.at(i)->dump();
1258 }
1259 }
1260 }
1261 else if (! _dead_node_list.test(i)) {
1262 if (_log != NULL) {
1263 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1264 }
1265 if (PrintIdealNodeCount) {
1266 // Print the log message to tty
1267 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1268 }
1269 }
1270 }
1271 if (_log != NULL) {
1272 _log->tail("mismatched_nodes");
1273 }
1274 }
1275 }
1276 #endif
1278 #ifndef PRODUCT
1279 void Compile::verify_top(Node* tn) const {
1280 if (tn != NULL) {
1281 assert(tn->is_Con(), "top node must be a constant");
1282 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1283 assert(tn->in(0) != NULL, "must have live top node");
1284 }
1285 }
1286 #endif
1289 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1291 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1292 guarantee(arr != NULL, "");
1293 int num_blocks = arr->length();
1294 if (grow_by < num_blocks) grow_by = num_blocks;
1295 int num_notes = grow_by * _node_notes_block_size;
1296 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1297 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1298 while (num_notes > 0) {
1299 arr->append(notes);
1300 notes += _node_notes_block_size;
1301 num_notes -= _node_notes_block_size;
1302 }
1303 assert(num_notes == 0, "exact multiple, please");
1304 }
1306 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1307 if (source == NULL || dest == NULL) return false;
1309 if (dest->is_Con())
1310 return false; // Do not push debug info onto constants.
1312 #ifdef ASSERT
1313 // Leave a bread crumb trail pointing to the original node:
1314 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1315 dest->set_debug_orig(source);
1316 }
1317 #endif
1319 if (node_note_array() == NULL)
1320 return false; // Not collecting any notes now.
1322 // This is a copy onto a pre-existing node, which may already have notes.
1323 // If both nodes have notes, do not overwrite any pre-existing notes.
1324 Node_Notes* source_notes = node_notes_at(source->_idx);
1325 if (source_notes == NULL || source_notes->is_clear()) return false;
1326 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1327 if (dest_notes == NULL || dest_notes->is_clear()) {
1328 return set_node_notes_at(dest->_idx, source_notes);
1329 }
1331 Node_Notes merged_notes = (*source_notes);
1332 // The order of operations here ensures that dest notes will win...
1333 merged_notes.update_from(dest_notes);
1334 return set_node_notes_at(dest->_idx, &merged_notes);
1335 }
1338 //--------------------------allow_range_check_smearing-------------------------
1339 // Gating condition for coalescing similar range checks.
1340 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1341 // single covering check that is at least as strong as any of them.
1342 // If the optimization succeeds, the simplified (strengthened) range check
1343 // will always succeed. If it fails, we will deopt, and then give up
1344 // on the optimization.
1345 bool Compile::allow_range_check_smearing() const {
1346 // If this method has already thrown a range-check,
1347 // assume it was because we already tried range smearing
1348 // and it failed.
1349 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1350 return !already_trapped;
1351 }
1354 //------------------------------flatten_alias_type-----------------------------
1355 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1356 int offset = tj->offset();
1357 TypePtr::PTR ptr = tj->ptr();
1359 // Known instance (scalarizable allocation) alias only with itself.
1360 bool is_known_inst = tj->isa_oopptr() != NULL &&
1361 tj->is_oopptr()->is_known_instance();
1363 // Process weird unsafe references.
1364 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1365 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1366 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1367 tj = TypeOopPtr::BOTTOM;
1368 ptr = tj->ptr();
1369 offset = tj->offset();
1370 }
1372 // Array pointers need some flattening
1373 const TypeAryPtr *ta = tj->isa_aryptr();
1374 if (ta && ta->is_stable()) {
1375 // Erase stability property for alias analysis.
1376 tj = ta = ta->cast_to_stable(false);
1377 }
1378 if( ta && is_known_inst ) {
1379 if ( offset != Type::OffsetBot &&
1380 offset > arrayOopDesc::length_offset_in_bytes() ) {
1381 offset = Type::OffsetBot; // Flatten constant access into array body only
1382 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1383 }
1384 } else if( ta && _AliasLevel >= 2 ) {
1385 // For arrays indexed by constant indices, we flatten the alias
1386 // space to include all of the array body. Only the header, klass
1387 // and array length can be accessed un-aliased.
1388 if( offset != Type::OffsetBot ) {
1389 if( ta->const_oop() ) { // MethodData* or Method*
1390 offset = Type::OffsetBot; // Flatten constant access into array body
1391 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1392 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1393 // range is OK as-is.
1394 tj = ta = TypeAryPtr::RANGE;
1395 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1396 tj = TypeInstPtr::KLASS; // all klass loads look alike
1397 ta = TypeAryPtr::RANGE; // generic ignored junk
1398 ptr = TypePtr::BotPTR;
1399 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1400 tj = TypeInstPtr::MARK;
1401 ta = TypeAryPtr::RANGE; // generic ignored junk
1402 ptr = TypePtr::BotPTR;
1403 } else { // Random constant offset into array body
1404 offset = Type::OffsetBot; // Flatten constant access into array body
1405 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1406 }
1407 }
1408 // Arrays of fixed size alias with arrays of unknown size.
1409 if (ta->size() != TypeInt::POS) {
1410 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1411 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1412 }
1413 // Arrays of known objects become arrays of unknown objects.
1414 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1415 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1416 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1417 }
1418 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1419 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1420 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1421 }
1422 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1423 // cannot be distinguished by bytecode alone.
1424 if (ta->elem() == TypeInt::BOOL) {
1425 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1426 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1427 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1428 }
1429 // During the 2nd round of IterGVN, NotNull castings are removed.
1430 // Make sure the Bottom and NotNull variants alias the same.
1431 // Also, make sure exact and non-exact variants alias the same.
1432 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1433 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1434 }
1435 }
1437 // Oop pointers need some flattening
1438 const TypeInstPtr *to = tj->isa_instptr();
1439 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1440 ciInstanceKlass *k = to->klass()->as_instance_klass();
1441 if( ptr == TypePtr::Constant ) {
1442 if (to->klass() != ciEnv::current()->Class_klass() ||
1443 offset < k->size_helper() * wordSize) {
1444 // No constant oop pointers (such as Strings); they alias with
1445 // unknown strings.
1446 assert(!is_known_inst, "not scalarizable allocation");
1447 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1448 }
1449 } else if( is_known_inst ) {
1450 tj = to; // Keep NotNull and klass_is_exact for instance type
1451 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1452 // During the 2nd round of IterGVN, NotNull castings are removed.
1453 // Make sure the Bottom and NotNull variants alias the same.
1454 // Also, make sure exact and non-exact variants alias the same.
1455 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1456 }
1457 if (to->speculative() != NULL) {
1458 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1459 }
1460 // Canonicalize the holder of this field
1461 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1462 // First handle header references such as a LoadKlassNode, even if the
1463 // object's klass is unloaded at compile time (4965979).
1464 if (!is_known_inst) { // Do it only for non-instance types
1465 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1466 }
1467 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1468 // Static fields are in the space above the normal instance
1469 // fields in the java.lang.Class instance.
1470 if (to->klass() != ciEnv::current()->Class_klass()) {
1471 to = NULL;
1472 tj = TypeOopPtr::BOTTOM;
1473 offset = tj->offset();
1474 }
1475 } else {
1476 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1477 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1478 if( is_known_inst ) {
1479 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1480 } else {
1481 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1482 }
1483 }
1484 }
1485 }
1487 // Klass pointers to object array klasses need some flattening
1488 const TypeKlassPtr *tk = tj->isa_klassptr();
1489 if( tk ) {
1490 // If we are referencing a field within a Klass, we need
1491 // to assume the worst case of an Object. Both exact and
1492 // inexact types must flatten to the same alias class so
1493 // use NotNull as the PTR.
1494 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1496 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1497 TypeKlassPtr::OBJECT->klass(),
1498 offset);
1499 }
1501 ciKlass* klass = tk->klass();
1502 if( klass->is_obj_array_klass() ) {
1503 ciKlass* k = TypeAryPtr::OOPS->klass();
1504 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1505 k = TypeInstPtr::BOTTOM->klass();
1506 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1507 }
1509 // Check for precise loads from the primary supertype array and force them
1510 // to the supertype cache alias index. Check for generic array loads from
1511 // the primary supertype array and also force them to the supertype cache
1512 // alias index. Since the same load can reach both, we need to merge
1513 // these 2 disparate memories into the same alias class. Since the
1514 // primary supertype array is read-only, there's no chance of confusion
1515 // where we bypass an array load and an array store.
1516 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1517 if (offset == Type::OffsetBot ||
1518 (offset >= primary_supers_offset &&
1519 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1520 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1521 offset = in_bytes(Klass::secondary_super_cache_offset());
1522 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1523 }
1524 }
1526 // Flatten all Raw pointers together.
1527 if (tj->base() == Type::RawPtr)
1528 tj = TypeRawPtr::BOTTOM;
1530 if (tj->base() == Type::AnyPtr)
1531 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1533 // Flatten all to bottom for now
1534 switch( _AliasLevel ) {
1535 case 0:
1536 tj = TypePtr::BOTTOM;
1537 break;
1538 case 1: // Flatten to: oop, static, field or array
1539 switch (tj->base()) {
1540 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1541 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1542 case Type::AryPtr: // do not distinguish arrays at all
1543 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1544 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1545 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1546 default: ShouldNotReachHere();
1547 }
1548 break;
1549 case 2: // No collapsing at level 2; keep all splits
1550 case 3: // No collapsing at level 3; keep all splits
1551 break;
1552 default:
1553 Unimplemented();
1554 }
1556 offset = tj->offset();
1557 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1559 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1560 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1561 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1562 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1563 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1564 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1565 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1566 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1567 assert( tj->ptr() != TypePtr::TopPTR &&
1568 tj->ptr() != TypePtr::AnyNull &&
1569 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1570 // assert( tj->ptr() != TypePtr::Constant ||
1571 // tj->base() == Type::RawPtr ||
1572 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1574 return tj;
1575 }
1577 void Compile::AliasType::Init(int i, const TypePtr* at) {
1578 _index = i;
1579 _adr_type = at;
1580 _field = NULL;
1581 _element = NULL;
1582 _is_rewritable = true; // default
1583 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1584 if (atoop != NULL && atoop->is_known_instance()) {
1585 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1586 _general_index = Compile::current()->get_alias_index(gt);
1587 } else {
1588 _general_index = 0;
1589 }
1590 }
1592 //---------------------------------print_on------------------------------------
1593 #ifndef PRODUCT
1594 void Compile::AliasType::print_on(outputStream* st) {
1595 if (index() < 10)
1596 st->print("@ <%d> ", index());
1597 else st->print("@ <%d>", index());
1598 st->print(is_rewritable() ? " " : " RO");
1599 int offset = adr_type()->offset();
1600 if (offset == Type::OffsetBot)
1601 st->print(" +any");
1602 else st->print(" +%-3d", offset);
1603 st->print(" in ");
1604 adr_type()->dump_on(st);
1605 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1606 if (field() != NULL && tjp) {
1607 if (tjp->klass() != field()->holder() ||
1608 tjp->offset() != field()->offset_in_bytes()) {
1609 st->print(" != ");
1610 field()->print();
1611 st->print(" ***");
1612 }
1613 }
1614 }
1616 void print_alias_types() {
1617 Compile* C = Compile::current();
1618 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1619 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1620 C->alias_type(idx)->print_on(tty);
1621 tty->cr();
1622 }
1623 }
1624 #endif
1627 //----------------------------probe_alias_cache--------------------------------
1628 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1629 intptr_t key = (intptr_t) adr_type;
1630 key ^= key >> logAliasCacheSize;
1631 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1632 }
1635 //-----------------------------grow_alias_types--------------------------------
1636 void Compile::grow_alias_types() {
1637 const int old_ats = _max_alias_types; // how many before?
1638 const int new_ats = old_ats; // how many more?
1639 const int grow_ats = old_ats+new_ats; // how many now?
1640 _max_alias_types = grow_ats;
1641 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1642 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1643 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1644 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1645 }
1648 //--------------------------------find_alias_type------------------------------
1649 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1650 if (_AliasLevel == 0)
1651 return alias_type(AliasIdxBot);
1653 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1654 if (ace->_adr_type == adr_type) {
1655 return alias_type(ace->_index);
1656 }
1658 // Handle special cases.
1659 if (adr_type == NULL) return alias_type(AliasIdxTop);
1660 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1662 // Do it the slow way.
1663 const TypePtr* flat = flatten_alias_type(adr_type);
1665 #ifdef ASSERT
1666 assert(flat == flatten_alias_type(flat), "idempotent");
1667 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1668 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1669 const TypeOopPtr* foop = flat->is_oopptr();
1670 // Scalarizable allocations have exact klass always.
1671 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1672 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1673 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1674 }
1675 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1676 #endif
1678 int idx = AliasIdxTop;
1679 for (int i = 0; i < num_alias_types(); i++) {
1680 if (alias_type(i)->adr_type() == flat) {
1681 idx = i;
1682 break;
1683 }
1684 }
1686 if (idx == AliasIdxTop) {
1687 if (no_create) return NULL;
1688 // Grow the array if necessary.
1689 if (_num_alias_types == _max_alias_types) grow_alias_types();
1690 // Add a new alias type.
1691 idx = _num_alias_types++;
1692 _alias_types[idx]->Init(idx, flat);
1693 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1694 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1695 if (flat->isa_instptr()) {
1696 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1697 && flat->is_instptr()->klass() == env()->Class_klass())
1698 alias_type(idx)->set_rewritable(false);
1699 }
1700 if (flat->isa_aryptr()) {
1701 #ifdef ASSERT
1702 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1703 // (T_BYTE has the weakest alignment and size restrictions...)
1704 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1705 #endif
1706 if (flat->offset() == TypePtr::OffsetBot) {
1707 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1708 }
1709 }
1710 if (flat->isa_klassptr()) {
1711 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1712 alias_type(idx)->set_rewritable(false);
1713 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1714 alias_type(idx)->set_rewritable(false);
1715 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1716 alias_type(idx)->set_rewritable(false);
1717 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1718 alias_type(idx)->set_rewritable(false);
1719 }
1720 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1721 // but the base pointer type is not distinctive enough to identify
1722 // references into JavaThread.)
1724 // Check for final fields.
1725 const TypeInstPtr* tinst = flat->isa_instptr();
1726 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1727 ciField* field;
1728 if (tinst->const_oop() != NULL &&
1729 tinst->klass() == ciEnv::current()->Class_klass() &&
1730 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1731 // static field
1732 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1733 field = k->get_field_by_offset(tinst->offset(), true);
1734 } else {
1735 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1736 field = k->get_field_by_offset(tinst->offset(), false);
1737 }
1738 assert(field == NULL ||
1739 original_field == NULL ||
1740 (field->holder() == original_field->holder() &&
1741 field->offset() == original_field->offset() &&
1742 field->is_static() == original_field->is_static()), "wrong field?");
1743 // Set field() and is_rewritable() attributes.
1744 if (field != NULL) alias_type(idx)->set_field(field);
1745 }
1746 }
1748 // Fill the cache for next time.
1749 ace->_adr_type = adr_type;
1750 ace->_index = idx;
1751 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1753 // Might as well try to fill the cache for the flattened version, too.
1754 AliasCacheEntry* face = probe_alias_cache(flat);
1755 if (face->_adr_type == NULL) {
1756 face->_adr_type = flat;
1757 face->_index = idx;
1758 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1759 }
1761 return alias_type(idx);
1762 }
1765 Compile::AliasType* Compile::alias_type(ciField* field) {
1766 const TypeOopPtr* t;
1767 if (field->is_static())
1768 t = TypeInstPtr::make(field->holder()->java_mirror());
1769 else
1770 t = TypeOopPtr::make_from_klass_raw(field->holder());
1771 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1772 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1773 return atp;
1774 }
1777 //------------------------------have_alias_type--------------------------------
1778 bool Compile::have_alias_type(const TypePtr* adr_type) {
1779 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1780 if (ace->_adr_type == adr_type) {
1781 return true;
1782 }
1784 // Handle special cases.
1785 if (adr_type == NULL) return true;
1786 if (adr_type == TypePtr::BOTTOM) return true;
1788 return find_alias_type(adr_type, true, NULL) != NULL;
1789 }
1791 //-----------------------------must_alias--------------------------------------
1792 // True if all values of the given address type are in the given alias category.
1793 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1794 if (alias_idx == AliasIdxBot) return true; // the universal category
1795 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1796 if (alias_idx == AliasIdxTop) return false; // the empty category
1797 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1799 // the only remaining possible overlap is identity
1800 int adr_idx = get_alias_index(adr_type);
1801 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1802 assert(adr_idx == alias_idx ||
1803 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1804 && adr_type != TypeOopPtr::BOTTOM),
1805 "should not be testing for overlap with an unsafe pointer");
1806 return adr_idx == alias_idx;
1807 }
1809 //------------------------------can_alias--------------------------------------
1810 // True if any values of the given address type are in the given alias category.
1811 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1812 if (alias_idx == AliasIdxTop) return false; // the empty category
1813 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1814 if (alias_idx == AliasIdxBot) return true; // the universal category
1815 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1817 // the only remaining possible overlap is identity
1818 int adr_idx = get_alias_index(adr_type);
1819 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1820 return adr_idx == alias_idx;
1821 }
1825 //---------------------------pop_warm_call-------------------------------------
1826 WarmCallInfo* Compile::pop_warm_call() {
1827 WarmCallInfo* wci = _warm_calls;
1828 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1829 return wci;
1830 }
1832 //----------------------------Inline_Warm--------------------------------------
1833 int Compile::Inline_Warm() {
1834 // If there is room, try to inline some more warm call sites.
1835 // %%% Do a graph index compaction pass when we think we're out of space?
1836 if (!InlineWarmCalls) return 0;
1838 int calls_made_hot = 0;
1839 int room_to_grow = NodeCountInliningCutoff - unique();
1840 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1841 int amount_grown = 0;
1842 WarmCallInfo* call;
1843 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1844 int est_size = (int)call->size();
1845 if (est_size > (room_to_grow - amount_grown)) {
1846 // This one won't fit anyway. Get rid of it.
1847 call->make_cold();
1848 continue;
1849 }
1850 call->make_hot();
1851 calls_made_hot++;
1852 amount_grown += est_size;
1853 amount_to_grow -= est_size;
1854 }
1856 if (calls_made_hot > 0) set_major_progress();
1857 return calls_made_hot;
1858 }
1861 //----------------------------Finish_Warm--------------------------------------
1862 void Compile::Finish_Warm() {
1863 if (!InlineWarmCalls) return;
1864 if (failing()) return;
1865 if (warm_calls() == NULL) return;
1867 // Clean up loose ends, if we are out of space for inlining.
1868 WarmCallInfo* call;
1869 while ((call = pop_warm_call()) != NULL) {
1870 call->make_cold();
1871 }
1872 }
1874 //---------------------cleanup_loop_predicates-----------------------
1875 // Remove the opaque nodes that protect the predicates so that all unused
1876 // checks and uncommon_traps will be eliminated from the ideal graph
1877 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1878 if (predicate_count()==0) return;
1879 for (int i = predicate_count(); i > 0; i--) {
1880 Node * n = predicate_opaque1_node(i-1);
1881 assert(n->Opcode() == Op_Opaque1, "must be");
1882 igvn.replace_node(n, n->in(1));
1883 }
1884 assert(predicate_count()==0, "should be clean!");
1885 }
1887 void Compile::add_range_check_cast(Node* n) {
1888 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1889 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1890 _range_check_casts->append(n);
1891 }
1893 // Remove all range check dependent CastIINodes.
1894 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1895 for (int i = range_check_cast_count(); i > 0; i--) {
1896 Node* cast = range_check_cast_node(i-1);
1897 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1898 igvn.replace_node(cast, cast->in(1));
1899 }
1900 assert(range_check_cast_count() == 0, "should be empty");
1901 }
1903 // StringOpts and late inlining of string methods
1904 void Compile::inline_string_calls(bool parse_time) {
1905 {
1906 // remove useless nodes to make the usage analysis simpler
1907 ResourceMark rm;
1908 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1909 }
1911 {
1912 ResourceMark rm;
1913 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1914 PhaseStringOpts pso(initial_gvn(), for_igvn());
1915 print_method(PHASE_AFTER_STRINGOPTS, 3);
1916 }
1918 // now inline anything that we skipped the first time around
1919 if (!parse_time) {
1920 _late_inlines_pos = _late_inlines.length();
1921 }
1923 while (_string_late_inlines.length() > 0) {
1924 CallGenerator* cg = _string_late_inlines.pop();
1925 cg->do_late_inline();
1926 if (failing()) return;
1927 }
1928 _string_late_inlines.trunc_to(0);
1929 }
1931 // Late inlining of boxing methods
1932 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1933 if (_boxing_late_inlines.length() > 0) {
1934 assert(has_boxed_value(), "inconsistent");
1936 PhaseGVN* gvn = initial_gvn();
1937 set_inlining_incrementally(true);
1939 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1940 for_igvn()->clear();
1941 gvn->replace_with(&igvn);
1943 _late_inlines_pos = _late_inlines.length();
1945 while (_boxing_late_inlines.length() > 0) {
1946 CallGenerator* cg = _boxing_late_inlines.pop();
1947 cg->do_late_inline();
1948 if (failing()) return;
1949 }
1950 _boxing_late_inlines.trunc_to(0);
1952 {
1953 ResourceMark rm;
1954 PhaseRemoveUseless pru(gvn, for_igvn());
1955 }
1957 igvn = PhaseIterGVN(gvn);
1958 igvn.optimize();
1960 set_inlining_progress(false);
1961 set_inlining_incrementally(false);
1962 }
1963 }
1965 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1966 assert(IncrementalInline, "incremental inlining should be on");
1967 PhaseGVN* gvn = initial_gvn();
1969 set_inlining_progress(false);
1970 for_igvn()->clear();
1971 gvn->replace_with(&igvn);
1973 int i = 0;
1975 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1976 CallGenerator* cg = _late_inlines.at(i);
1977 _late_inlines_pos = i+1;
1978 cg->do_late_inline();
1979 if (failing()) return;
1980 }
1981 int j = 0;
1982 for (; i < _late_inlines.length(); i++, j++) {
1983 _late_inlines.at_put(j, _late_inlines.at(i));
1984 }
1985 _late_inlines.trunc_to(j);
1987 {
1988 ResourceMark rm;
1989 PhaseRemoveUseless pru(gvn, for_igvn());
1990 }
1992 igvn = PhaseIterGVN(gvn);
1993 }
1995 // Perform incremental inlining until bound on number of live nodes is reached
1996 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1997 PhaseGVN* gvn = initial_gvn();
1999 set_inlining_incrementally(true);
2000 set_inlining_progress(true);
2001 uint low_live_nodes = 0;
2003 while(inlining_progress() && _late_inlines.length() > 0) {
2005 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2006 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2007 // PhaseIdealLoop is expensive so we only try it once we are
2008 // out of live nodes and we only try it again if the previous
2009 // helped got the number of nodes down significantly
2010 PhaseIdealLoop ideal_loop( igvn, false, true );
2011 if (failing()) return;
2012 low_live_nodes = live_nodes();
2013 _major_progress = true;
2014 }
2016 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2017 break;
2018 }
2019 }
2021 inline_incrementally_one(igvn);
2023 if (failing()) return;
2025 igvn.optimize();
2027 if (failing()) return;
2028 }
2030 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2032 if (_string_late_inlines.length() > 0) {
2033 assert(has_stringbuilder(), "inconsistent");
2034 for_igvn()->clear();
2035 initial_gvn()->replace_with(&igvn);
2037 inline_string_calls(false);
2039 if (failing()) return;
2041 {
2042 ResourceMark rm;
2043 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2044 }
2046 igvn = PhaseIterGVN(gvn);
2048 igvn.optimize();
2049 }
2051 set_inlining_incrementally(false);
2052 }
2055 //------------------------------Optimize---------------------------------------
2056 // Given a graph, optimize it.
2057 void Compile::Optimize() {
2058 TracePhase t1("optimizer", &_t_optimizer, true);
2060 #ifndef PRODUCT
2061 if (env()->break_at_compile()) {
2062 BREAKPOINT;
2063 }
2065 #endif
2067 ResourceMark rm;
2068 int loop_opts_cnt;
2070 NOT_PRODUCT( verify_graph_edges(); )
2072 print_method(PHASE_AFTER_PARSING);
2074 {
2075 // Iterative Global Value Numbering, including ideal transforms
2076 // Initialize IterGVN with types and values from parse-time GVN
2077 PhaseIterGVN igvn(initial_gvn());
2078 {
2079 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2080 igvn.optimize();
2081 }
2083 print_method(PHASE_ITER_GVN1, 2);
2085 if (failing()) return;
2087 {
2088 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2089 inline_incrementally(igvn);
2090 }
2092 print_method(PHASE_INCREMENTAL_INLINE, 2);
2094 if (failing()) return;
2096 if (eliminate_boxing()) {
2097 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2098 // Inline valueOf() methods now.
2099 inline_boxing_calls(igvn);
2101 if (AlwaysIncrementalInline) {
2102 inline_incrementally(igvn);
2103 }
2105 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2107 if (failing()) return;
2108 }
2110 // Remove the speculative part of types and clean up the graph from
2111 // the extra CastPP nodes whose only purpose is to carry them. Do
2112 // that early so that optimizations are not disrupted by the extra
2113 // CastPP nodes.
2114 remove_speculative_types(igvn);
2116 // No more new expensive nodes will be added to the list from here
2117 // so keep only the actual candidates for optimizations.
2118 cleanup_expensive_nodes(igvn);
2120 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2121 NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);)
2122 initial_gvn()->replace_with(&igvn);
2123 for_igvn()->clear();
2124 Unique_Node_List new_worklist(C->comp_arena());
2125 {
2126 ResourceMark rm;
2127 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2128 }
2129 set_for_igvn(&new_worklist);
2130 igvn = PhaseIterGVN(initial_gvn());
2131 igvn.optimize();
2132 }
2134 // Perform escape analysis
2135 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2136 if (has_loops()) {
2137 // Cleanup graph (remove dead nodes).
2138 TracePhase t2("idealLoop", &_t_idealLoop, true);
2139 PhaseIdealLoop ideal_loop( igvn, false, true );
2140 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2141 if (failing()) return;
2142 }
2143 ConnectionGraph::do_analysis(this, &igvn);
2145 if (failing()) return;
2147 // Optimize out fields loads from scalar replaceable allocations.
2148 igvn.optimize();
2149 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2151 if (failing()) return;
2153 if (congraph() != NULL && macro_count() > 0) {
2154 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2155 PhaseMacroExpand mexp(igvn);
2156 mexp.eliminate_macro_nodes();
2157 igvn.set_delay_transform(false);
2159 igvn.optimize();
2160 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2162 if (failing()) return;
2163 }
2164 }
2166 // Loop transforms on the ideal graph. Range Check Elimination,
2167 // peeling, unrolling, etc.
2169 // Set loop opts counter
2170 loop_opts_cnt = num_loop_opts();
2171 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2172 {
2173 TracePhase t2("idealLoop", &_t_idealLoop, true);
2174 PhaseIdealLoop ideal_loop( igvn, true );
2175 loop_opts_cnt--;
2176 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2177 if (failing()) return;
2178 }
2179 // Loop opts pass if partial peeling occurred in previous pass
2180 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2181 TracePhase t3("idealLoop", &_t_idealLoop, true);
2182 PhaseIdealLoop ideal_loop( igvn, false );
2183 loop_opts_cnt--;
2184 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2185 if (failing()) return;
2186 }
2187 // Loop opts pass for loop-unrolling before CCP
2188 if(major_progress() && (loop_opts_cnt > 0)) {
2189 TracePhase t4("idealLoop", &_t_idealLoop, true);
2190 PhaseIdealLoop ideal_loop( igvn, false );
2191 loop_opts_cnt--;
2192 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2193 }
2194 if (!failing()) {
2195 // Verify that last round of loop opts produced a valid graph
2196 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2197 PhaseIdealLoop::verify(igvn);
2198 }
2199 }
2200 if (failing()) return;
2202 // Conditional Constant Propagation;
2203 PhaseCCP ccp( &igvn );
2204 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2205 {
2206 TracePhase t2("ccp", &_t_ccp, true);
2207 ccp.do_transform();
2208 }
2209 print_method(PHASE_CPP1, 2);
2211 assert( true, "Break here to ccp.dump_old2new_map()");
2213 // Iterative Global Value Numbering, including ideal transforms
2214 {
2215 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2216 igvn = ccp;
2217 igvn.optimize();
2218 }
2220 print_method(PHASE_ITER_GVN2, 2);
2222 if (failing()) return;
2224 // Loop transforms on the ideal graph. Range Check Elimination,
2225 // peeling, unrolling, etc.
2226 if(loop_opts_cnt > 0) {
2227 debug_only( int cnt = 0; );
2228 while(major_progress() && (loop_opts_cnt > 0)) {
2229 TracePhase t2("idealLoop", &_t_idealLoop, true);
2230 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2231 PhaseIdealLoop ideal_loop( igvn, true);
2232 loop_opts_cnt--;
2233 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2234 if (failing()) return;
2235 }
2236 }
2238 {
2239 // Verify that all previous optimizations produced a valid graph
2240 // at least to this point, even if no loop optimizations were done.
2241 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2242 PhaseIdealLoop::verify(igvn);
2243 }
2245 if (range_check_cast_count() > 0) {
2246 // No more loop optimizations. Remove all range check dependent CastIINodes.
2247 C->remove_range_check_casts(igvn);
2248 igvn.optimize();
2249 }
2251 {
2252 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2253 PhaseMacroExpand mex(igvn);
2254 if (mex.expand_macro_nodes()) {
2255 assert(failing(), "must bail out w/ explicit message");
2256 return;
2257 }
2258 }
2260 } // (End scope of igvn; run destructor if necessary for asserts.)
2262 dump_inlining();
2263 // A method with only infinite loops has no edges entering loops from root
2264 {
2265 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2266 if (final_graph_reshaping()) {
2267 assert(failing(), "must bail out w/ explicit message");
2268 return;
2269 }
2270 }
2272 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2273 }
2276 //------------------------------Code_Gen---------------------------------------
2277 // Given a graph, generate code for it
2278 void Compile::Code_Gen() {
2279 if (failing()) {
2280 return;
2281 }
2283 // Perform instruction selection. You might think we could reclaim Matcher
2284 // memory PDQ, but actually the Matcher is used in generating spill code.
2285 // Internals of the Matcher (including some VectorSets) must remain live
2286 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2287 // set a bit in reclaimed memory.
2289 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2290 // nodes. Mapping is only valid at the root of each matched subtree.
2291 NOT_PRODUCT( verify_graph_edges(); )
2293 Matcher matcher;
2294 _matcher = &matcher;
2295 {
2296 TracePhase t2("matcher", &_t_matcher, true);
2297 matcher.match();
2298 }
2299 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2300 // nodes. Mapping is only valid at the root of each matched subtree.
2301 NOT_PRODUCT( verify_graph_edges(); )
2303 // If you have too many nodes, or if matching has failed, bail out
2304 check_node_count(0, "out of nodes matching instructions");
2305 if (failing()) {
2306 return;
2307 }
2309 // Build a proper-looking CFG
2310 PhaseCFG cfg(node_arena(), root(), matcher);
2311 _cfg = &cfg;
2312 {
2313 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2314 bool success = cfg.do_global_code_motion();
2315 if (!success) {
2316 return;
2317 }
2319 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2320 NOT_PRODUCT( verify_graph_edges(); )
2321 debug_only( cfg.verify(); )
2322 }
2324 PhaseChaitin regalloc(unique(), cfg, matcher);
2325 _regalloc = ®alloc;
2326 {
2327 TracePhase t2("regalloc", &_t_registerAllocation, true);
2328 // Perform register allocation. After Chaitin, use-def chains are
2329 // no longer accurate (at spill code) and so must be ignored.
2330 // Node->LRG->reg mappings are still accurate.
2331 _regalloc->Register_Allocate();
2333 // Bail out if the allocator builds too many nodes
2334 if (failing()) {
2335 return;
2336 }
2337 }
2339 // Prior to register allocation we kept empty basic blocks in case the
2340 // the allocator needed a place to spill. After register allocation we
2341 // are not adding any new instructions. If any basic block is empty, we
2342 // can now safely remove it.
2343 {
2344 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2345 cfg.remove_empty_blocks();
2346 if (do_freq_based_layout()) {
2347 PhaseBlockLayout layout(cfg);
2348 } else {
2349 cfg.set_loop_alignment();
2350 }
2351 cfg.fixup_flow();
2352 }
2354 // Apply peephole optimizations
2355 if( OptoPeephole ) {
2356 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2357 PhasePeephole peep( _regalloc, cfg);
2358 peep.do_transform();
2359 }
2361 // Do late expand if CPU requires this.
2362 if (Matcher::require_postalloc_expand) {
2363 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2364 cfg.postalloc_expand(_regalloc);
2365 }
2367 // Convert Nodes to instruction bits in a buffer
2368 {
2369 // %%%% workspace merge brought two timers together for one job
2370 TracePhase t2a("output", &_t_output, true);
2371 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2372 Output();
2373 }
2375 print_method(PHASE_FINAL_CODE);
2377 // He's dead, Jim.
2378 _cfg = (PhaseCFG*)0xdeadbeef;
2379 _regalloc = (PhaseChaitin*)0xdeadbeef;
2380 }
2383 //------------------------------dump_asm---------------------------------------
2384 // Dump formatted assembly
2385 #ifndef PRODUCT
2386 void Compile::dump_asm(int *pcs, uint pc_limit) {
2387 bool cut_short = false;
2388 tty->print_cr("#");
2389 tty->print("# "); _tf->dump(); tty->cr();
2390 tty->print_cr("#");
2392 // For all blocks
2393 int pc = 0x0; // Program counter
2394 char starts_bundle = ' ';
2395 _regalloc->dump_frame();
2397 Node *n = NULL;
2398 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2399 if (VMThread::should_terminate()) {
2400 cut_short = true;
2401 break;
2402 }
2403 Block* block = _cfg->get_block(i);
2404 if (block->is_connector() && !Verbose) {
2405 continue;
2406 }
2407 n = block->head();
2408 if (pcs && n->_idx < pc_limit) {
2409 tty->print("%3.3x ", pcs[n->_idx]);
2410 } else {
2411 tty->print(" ");
2412 }
2413 block->dump_head(_cfg);
2414 if (block->is_connector()) {
2415 tty->print_cr(" # Empty connector block");
2416 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2417 tty->print_cr(" # Block is sole successor of call");
2418 }
2420 // For all instructions
2421 Node *delay = NULL;
2422 for (uint j = 0; j < block->number_of_nodes(); j++) {
2423 if (VMThread::should_terminate()) {
2424 cut_short = true;
2425 break;
2426 }
2427 n = block->get_node(j);
2428 if (valid_bundle_info(n)) {
2429 Bundle* bundle = node_bundling(n);
2430 if (bundle->used_in_unconditional_delay()) {
2431 delay = n;
2432 continue;
2433 }
2434 if (bundle->starts_bundle()) {
2435 starts_bundle = '+';
2436 }
2437 }
2439 if (WizardMode) {
2440 n->dump();
2441 }
2443 if( !n->is_Region() && // Dont print in the Assembly
2444 !n->is_Phi() && // a few noisely useless nodes
2445 !n->is_Proj() &&
2446 !n->is_MachTemp() &&
2447 !n->is_SafePointScalarObject() &&
2448 !n->is_Catch() && // Would be nice to print exception table targets
2449 !n->is_MergeMem() && // Not very interesting
2450 !n->is_top() && // Debug info table constants
2451 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2452 ) {
2453 if (pcs && n->_idx < pc_limit)
2454 tty->print("%3.3x", pcs[n->_idx]);
2455 else
2456 tty->print(" ");
2457 tty->print(" %c ", starts_bundle);
2458 starts_bundle = ' ';
2459 tty->print("\t");
2460 n->format(_regalloc, tty);
2461 tty->cr();
2462 }
2464 // If we have an instruction with a delay slot, and have seen a delay,
2465 // then back up and print it
2466 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2467 assert(delay != NULL, "no unconditional delay instruction");
2468 if (WizardMode) delay->dump();
2470 if (node_bundling(delay)->starts_bundle())
2471 starts_bundle = '+';
2472 if (pcs && n->_idx < pc_limit)
2473 tty->print("%3.3x", pcs[n->_idx]);
2474 else
2475 tty->print(" ");
2476 tty->print(" %c ", starts_bundle);
2477 starts_bundle = ' ';
2478 tty->print("\t");
2479 delay->format(_regalloc, tty);
2480 tty->cr();
2481 delay = NULL;
2482 }
2484 // Dump the exception table as well
2485 if( n->is_Catch() && (Verbose || WizardMode) ) {
2486 // Print the exception table for this offset
2487 _handler_table.print_subtable_for(pc);
2488 }
2489 }
2491 if (pcs && n->_idx < pc_limit)
2492 tty->print_cr("%3.3x", pcs[n->_idx]);
2493 else
2494 tty->cr();
2496 assert(cut_short || delay == NULL, "no unconditional delay branch");
2498 } // End of per-block dump
2499 tty->cr();
2501 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2502 }
2503 #endif
2505 //------------------------------Final_Reshape_Counts---------------------------
2506 // This class defines counters to help identify when a method
2507 // may/must be executed using hardware with only 24-bit precision.
2508 struct Final_Reshape_Counts : public StackObj {
2509 int _call_count; // count non-inlined 'common' calls
2510 int _float_count; // count float ops requiring 24-bit precision
2511 int _double_count; // count double ops requiring more precision
2512 int _java_call_count; // count non-inlined 'java' calls
2513 int _inner_loop_count; // count loops which need alignment
2514 VectorSet _visited; // Visitation flags
2515 Node_List _tests; // Set of IfNodes & PCTableNodes
2517 Final_Reshape_Counts() :
2518 _call_count(0), _float_count(0), _double_count(0),
2519 _java_call_count(0), _inner_loop_count(0),
2520 _visited( Thread::current()->resource_area() ) { }
2522 void inc_call_count () { _call_count ++; }
2523 void inc_float_count () { _float_count ++; }
2524 void inc_double_count() { _double_count++; }
2525 void inc_java_call_count() { _java_call_count++; }
2526 void inc_inner_loop_count() { _inner_loop_count++; }
2528 int get_call_count () const { return _call_count ; }
2529 int get_float_count () const { return _float_count ; }
2530 int get_double_count() const { return _double_count; }
2531 int get_java_call_count() const { return _java_call_count; }
2532 int get_inner_loop_count() const { return _inner_loop_count; }
2533 };
2535 #ifdef ASSERT
2536 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2537 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2538 // Make sure the offset goes inside the instance layout.
2539 return k->contains_field_offset(tp->offset());
2540 // Note that OffsetBot and OffsetTop are very negative.
2541 }
2542 #endif
2544 // Eliminate trivially redundant StoreCMs and accumulate their
2545 // precedence edges.
2546 void Compile::eliminate_redundant_card_marks(Node* n) {
2547 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2548 if (n->in(MemNode::Address)->outcnt() > 1) {
2549 // There are multiple users of the same address so it might be
2550 // possible to eliminate some of the StoreCMs
2551 Node* mem = n->in(MemNode::Memory);
2552 Node* adr = n->in(MemNode::Address);
2553 Node* val = n->in(MemNode::ValueIn);
2554 Node* prev = n;
2555 bool done = false;
2556 // Walk the chain of StoreCMs eliminating ones that match. As
2557 // long as it's a chain of single users then the optimization is
2558 // safe. Eliminating partially redundant StoreCMs would require
2559 // cloning copies down the other paths.
2560 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2561 if (adr == mem->in(MemNode::Address) &&
2562 val == mem->in(MemNode::ValueIn)) {
2563 // redundant StoreCM
2564 if (mem->req() > MemNode::OopStore) {
2565 // Hasn't been processed by this code yet.
2566 n->add_prec(mem->in(MemNode::OopStore));
2567 } else {
2568 // Already converted to precedence edge
2569 for (uint i = mem->req(); i < mem->len(); i++) {
2570 // Accumulate any precedence edges
2571 if (mem->in(i) != NULL) {
2572 n->add_prec(mem->in(i));
2573 }
2574 }
2575 // Everything above this point has been processed.
2576 done = true;
2577 }
2578 // Eliminate the previous StoreCM
2579 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2580 assert(mem->outcnt() == 0, "should be dead");
2581 mem->disconnect_inputs(NULL, this);
2582 } else {
2583 prev = mem;
2584 }
2585 mem = prev->in(MemNode::Memory);
2586 }
2587 }
2588 }
2590 //------------------------------final_graph_reshaping_impl----------------------
2591 // Implement items 1-5 from final_graph_reshaping below.
2592 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2594 if ( n->outcnt() == 0 ) return; // dead node
2595 uint nop = n->Opcode();
2597 // Check for 2-input instruction with "last use" on right input.
2598 // Swap to left input. Implements item (2).
2599 if( n->req() == 3 && // two-input instruction
2600 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2601 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2602 n->in(2)->outcnt() == 1 &&// right use IS a last use
2603 !n->in(2)->is_Con() ) { // right use is not a constant
2604 // Check for commutative opcode
2605 switch( nop ) {
2606 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2607 case Op_MaxI: case Op_MinI:
2608 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2609 case Op_AndL: case Op_XorL: case Op_OrL:
2610 case Op_AndI: case Op_XorI: case Op_OrI: {
2611 // Move "last use" input to left by swapping inputs
2612 n->swap_edges(1, 2);
2613 break;
2614 }
2615 default:
2616 break;
2617 }
2618 }
2620 #ifdef ASSERT
2621 if( n->is_Mem() ) {
2622 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2623 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2624 // oop will be recorded in oop map if load crosses safepoint
2625 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2626 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2627 "raw memory operations should have control edge");
2628 }
2629 #endif
2630 // Count FPU ops and common calls, implements item (3)
2631 switch( nop ) {
2632 // Count all float operations that may use FPU
2633 case Op_AddF:
2634 case Op_SubF:
2635 case Op_MulF:
2636 case Op_DivF:
2637 case Op_NegF:
2638 case Op_ModF:
2639 case Op_ConvI2F:
2640 case Op_ConF:
2641 case Op_CmpF:
2642 case Op_CmpF3:
2643 // case Op_ConvL2F: // longs are split into 32-bit halves
2644 frc.inc_float_count();
2645 break;
2647 case Op_ConvF2D:
2648 case Op_ConvD2F:
2649 frc.inc_float_count();
2650 frc.inc_double_count();
2651 break;
2653 // Count all double operations that may use FPU
2654 case Op_AddD:
2655 case Op_SubD:
2656 case Op_MulD:
2657 case Op_DivD:
2658 case Op_NegD:
2659 case Op_ModD:
2660 case Op_ConvI2D:
2661 case Op_ConvD2I:
2662 // case Op_ConvL2D: // handled by leaf call
2663 // case Op_ConvD2L: // handled by leaf call
2664 case Op_ConD:
2665 case Op_CmpD:
2666 case Op_CmpD3:
2667 frc.inc_double_count();
2668 break;
2669 case Op_Opaque1: // Remove Opaque Nodes before matching
2670 case Op_Opaque2: // Remove Opaque Nodes before matching
2671 case Op_Opaque3:
2672 n->subsume_by(n->in(1), this);
2673 break;
2674 case Op_CallStaticJava:
2675 case Op_CallJava:
2676 case Op_CallDynamicJava:
2677 frc.inc_java_call_count(); // Count java call site;
2678 case Op_CallRuntime:
2679 case Op_CallLeaf:
2680 case Op_CallLeafNoFP: {
2681 assert( n->is_Call(), "" );
2682 CallNode *call = n->as_Call();
2683 // Count call sites where the FP mode bit would have to be flipped.
2684 // Do not count uncommon runtime calls:
2685 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2686 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2687 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2688 frc.inc_call_count(); // Count the call site
2689 } else { // See if uncommon argument is shared
2690 Node *n = call->in(TypeFunc::Parms);
2691 int nop = n->Opcode();
2692 // Clone shared simple arguments to uncommon calls, item (1).
2693 if( n->outcnt() > 1 &&
2694 !n->is_Proj() &&
2695 nop != Op_CreateEx &&
2696 nop != Op_CheckCastPP &&
2697 nop != Op_DecodeN &&
2698 nop != Op_DecodeNKlass &&
2699 !n->is_Mem() ) {
2700 Node *x = n->clone();
2701 call->set_req( TypeFunc::Parms, x );
2702 }
2703 }
2704 break;
2705 }
2707 case Op_StoreD:
2708 case Op_LoadD:
2709 case Op_LoadD_unaligned:
2710 frc.inc_double_count();
2711 goto handle_mem;
2712 case Op_StoreF:
2713 case Op_LoadF:
2714 frc.inc_float_count();
2715 goto handle_mem;
2717 case Op_StoreCM:
2718 {
2719 // Convert OopStore dependence into precedence edge
2720 Node* prec = n->in(MemNode::OopStore);
2721 n->del_req(MemNode::OopStore);
2722 n->add_prec(prec);
2723 eliminate_redundant_card_marks(n);
2724 }
2726 // fall through
2728 case Op_StoreB:
2729 case Op_StoreC:
2730 case Op_StorePConditional:
2731 case Op_StoreI:
2732 case Op_StoreL:
2733 case Op_StoreIConditional:
2734 case Op_StoreLConditional:
2735 case Op_CompareAndSwapI:
2736 case Op_CompareAndSwapL:
2737 case Op_CompareAndSwapP:
2738 case Op_CompareAndSwapN:
2739 case Op_GetAndAddI:
2740 case Op_GetAndAddL:
2741 case Op_GetAndSetI:
2742 case Op_GetAndSetL:
2743 case Op_GetAndSetP:
2744 case Op_GetAndSetN:
2745 case Op_StoreP:
2746 case Op_StoreN:
2747 case Op_StoreNKlass:
2748 case Op_LoadB:
2749 case Op_LoadUB:
2750 case Op_LoadUS:
2751 case Op_LoadI:
2752 case Op_LoadKlass:
2753 case Op_LoadNKlass:
2754 case Op_LoadL:
2755 case Op_LoadL_unaligned:
2756 case Op_LoadPLocked:
2757 case Op_LoadP:
2758 case Op_LoadN:
2759 case Op_LoadRange:
2760 case Op_LoadS: {
2761 handle_mem:
2762 #ifdef ASSERT
2763 if( VerifyOptoOopOffsets ) {
2764 assert( n->is_Mem(), "" );
2765 MemNode *mem = (MemNode*)n;
2766 // Check to see if address types have grounded out somehow.
2767 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2768 assert( !tp || oop_offset_is_sane(tp), "" );
2769 }
2770 #endif
2771 break;
2772 }
2774 case Op_AddP: { // Assert sane base pointers
2775 Node *addp = n->in(AddPNode::Address);
2776 assert( !addp->is_AddP() ||
2777 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2778 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2779 "Base pointers must match" );
2780 #ifdef _LP64
2781 if ((UseCompressedOops || UseCompressedClassPointers) &&
2782 addp->Opcode() == Op_ConP &&
2783 addp == n->in(AddPNode::Base) &&
2784 n->in(AddPNode::Offset)->is_Con()) {
2785 // Use addressing with narrow klass to load with offset on x86.
2786 // On sparc loading 32-bits constant and decoding it have less
2787 // instructions (4) then load 64-bits constant (7).
2788 // Do this transformation here since IGVN will convert ConN back to ConP.
2789 const Type* t = addp->bottom_type();
2790 if (t->isa_oopptr() || t->isa_klassptr()) {
2791 Node* nn = NULL;
2793 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2795 // Look for existing ConN node of the same exact type.
2796 Node* r = root();
2797 uint cnt = r->outcnt();
2798 for (uint i = 0; i < cnt; i++) {
2799 Node* m = r->raw_out(i);
2800 if (m!= NULL && m->Opcode() == op &&
2801 m->bottom_type()->make_ptr() == t) {
2802 nn = m;
2803 break;
2804 }
2805 }
2806 if (nn != NULL) {
2807 // Decode a narrow oop to match address
2808 // [R12 + narrow_oop_reg<<3 + offset]
2809 if (t->isa_oopptr()) {
2810 nn = new (this) DecodeNNode(nn, t);
2811 } else {
2812 nn = new (this) DecodeNKlassNode(nn, t);
2813 }
2814 n->set_req(AddPNode::Base, nn);
2815 n->set_req(AddPNode::Address, nn);
2816 if (addp->outcnt() == 0) {
2817 addp->disconnect_inputs(NULL, this);
2818 }
2819 }
2820 }
2821 }
2822 #endif
2823 break;
2824 }
2826 #ifdef _LP64
2827 case Op_CastPP:
2828 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2829 Node* in1 = n->in(1);
2830 const Type* t = n->bottom_type();
2831 Node* new_in1 = in1->clone();
2832 new_in1->as_DecodeN()->set_type(t);
2834 if (!Matcher::narrow_oop_use_complex_address()) {
2835 //
2836 // x86, ARM and friends can handle 2 adds in addressing mode
2837 // and Matcher can fold a DecodeN node into address by using
2838 // a narrow oop directly and do implicit NULL check in address:
2839 //
2840 // [R12 + narrow_oop_reg<<3 + offset]
2841 // NullCheck narrow_oop_reg
2842 //
2843 // On other platforms (Sparc) we have to keep new DecodeN node and
2844 // use it to do implicit NULL check in address:
2845 //
2846 // decode_not_null narrow_oop_reg, base_reg
2847 // [base_reg + offset]
2848 // NullCheck base_reg
2849 //
2850 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2851 // to keep the information to which NULL check the new DecodeN node
2852 // corresponds to use it as value in implicit_null_check().
2853 //
2854 new_in1->set_req(0, n->in(0));
2855 }
2857 n->subsume_by(new_in1, this);
2858 if (in1->outcnt() == 0) {
2859 in1->disconnect_inputs(NULL, this);
2860 }
2861 }
2862 break;
2864 case Op_CmpP:
2865 // Do this transformation here to preserve CmpPNode::sub() and
2866 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2867 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2868 Node* in1 = n->in(1);
2869 Node* in2 = n->in(2);
2870 if (!in1->is_DecodeNarrowPtr()) {
2871 in2 = in1;
2872 in1 = n->in(2);
2873 }
2874 assert(in1->is_DecodeNarrowPtr(), "sanity");
2876 Node* new_in2 = NULL;
2877 if (in2->is_DecodeNarrowPtr()) {
2878 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2879 new_in2 = in2->in(1);
2880 } else if (in2->Opcode() == Op_ConP) {
2881 const Type* t = in2->bottom_type();
2882 if (t == TypePtr::NULL_PTR) {
2883 assert(in1->is_DecodeN(), "compare klass to null?");
2884 // Don't convert CmpP null check into CmpN if compressed
2885 // oops implicit null check is not generated.
2886 // This will allow to generate normal oop implicit null check.
2887 if (Matcher::gen_narrow_oop_implicit_null_checks())
2888 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2889 //
2890 // This transformation together with CastPP transformation above
2891 // will generated code for implicit NULL checks for compressed oops.
2892 //
2893 // The original code after Optimize()
2894 //
2895 // LoadN memory, narrow_oop_reg
2896 // decode narrow_oop_reg, base_reg
2897 // CmpP base_reg, NULL
2898 // CastPP base_reg // NotNull
2899 // Load [base_reg + offset], val_reg
2900 //
2901 // after these transformations will be
2902 //
2903 // LoadN memory, narrow_oop_reg
2904 // CmpN narrow_oop_reg, NULL
2905 // decode_not_null narrow_oop_reg, base_reg
2906 // Load [base_reg + offset], val_reg
2907 //
2908 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2909 // since narrow oops can be used in debug info now (see the code in
2910 // final_graph_reshaping_walk()).
2911 //
2912 // At the end the code will be matched to
2913 // on x86:
2914 //
2915 // Load_narrow_oop memory, narrow_oop_reg
2916 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2917 // NullCheck narrow_oop_reg
2918 //
2919 // and on sparc:
2920 //
2921 // Load_narrow_oop memory, narrow_oop_reg
2922 // decode_not_null narrow_oop_reg, base_reg
2923 // Load [base_reg + offset], val_reg
2924 // NullCheck base_reg
2925 //
2926 } else if (t->isa_oopptr()) {
2927 new_in2 = ConNode::make(this, t->make_narrowoop());
2928 } else if (t->isa_klassptr()) {
2929 new_in2 = ConNode::make(this, t->make_narrowklass());
2930 }
2931 }
2932 if (new_in2 != NULL) {
2933 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2934 n->subsume_by(cmpN, this);
2935 if (in1->outcnt() == 0) {
2936 in1->disconnect_inputs(NULL, this);
2937 }
2938 if (in2->outcnt() == 0) {
2939 in2->disconnect_inputs(NULL, this);
2940 }
2941 }
2942 }
2943 break;
2945 case Op_DecodeN:
2946 case Op_DecodeNKlass:
2947 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2948 // DecodeN could be pinned when it can't be fold into
2949 // an address expression, see the code for Op_CastPP above.
2950 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2951 break;
2953 case Op_EncodeP:
2954 case Op_EncodePKlass: {
2955 Node* in1 = n->in(1);
2956 if (in1->is_DecodeNarrowPtr()) {
2957 n->subsume_by(in1->in(1), this);
2958 } else if (in1->Opcode() == Op_ConP) {
2959 const Type* t = in1->bottom_type();
2960 if (t == TypePtr::NULL_PTR) {
2961 assert(t->isa_oopptr(), "null klass?");
2962 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2963 } else if (t->isa_oopptr()) {
2964 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2965 } else if (t->isa_klassptr()) {
2966 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2967 }
2968 }
2969 if (in1->outcnt() == 0) {
2970 in1->disconnect_inputs(NULL, this);
2971 }
2972 break;
2973 }
2975 case Op_Proj: {
2976 if (OptimizeStringConcat) {
2977 ProjNode* p = n->as_Proj();
2978 if (p->_is_io_use) {
2979 // Separate projections were used for the exception path which
2980 // are normally removed by a late inline. If it wasn't inlined
2981 // then they will hang around and should just be replaced with
2982 // the original one.
2983 Node* proj = NULL;
2984 // Replace with just one
2985 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2986 Node *use = i.get();
2987 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2988 proj = use;
2989 break;
2990 }
2991 }
2992 assert(proj != NULL, "must be found");
2993 p->subsume_by(proj, this);
2994 }
2995 }
2996 break;
2997 }
2999 case Op_Phi:
3000 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3001 // The EncodeP optimization may create Phi with the same edges
3002 // for all paths. It is not handled well by Register Allocator.
3003 Node* unique_in = n->in(1);
3004 assert(unique_in != NULL, "");
3005 uint cnt = n->req();
3006 for (uint i = 2; i < cnt; i++) {
3007 Node* m = n->in(i);
3008 assert(m != NULL, "");
3009 if (unique_in != m)
3010 unique_in = NULL;
3011 }
3012 if (unique_in != NULL) {
3013 n->subsume_by(unique_in, this);
3014 }
3015 }
3016 break;
3018 #endif
3020 #ifdef ASSERT
3021 case Op_CastII:
3022 // Verify that all range check dependent CastII nodes were removed.
3023 if (n->isa_CastII()->has_range_check()) {
3024 n->dump(3);
3025 assert(false, "Range check dependent CastII node was not removed");
3026 }
3027 break;
3028 #endif
3030 case Op_ModI:
3031 if (UseDivMod) {
3032 // Check if a%b and a/b both exist
3033 Node* d = n->find_similar(Op_DivI);
3034 if (d) {
3035 // Replace them with a fused divmod if supported
3036 if (Matcher::has_match_rule(Op_DivModI)) {
3037 DivModINode* divmod = DivModINode::make(this, n);
3038 d->subsume_by(divmod->div_proj(), this);
3039 n->subsume_by(divmod->mod_proj(), this);
3040 } else {
3041 // replace a%b with a-((a/b)*b)
3042 Node* mult = new (this) MulINode(d, d->in(2));
3043 Node* sub = new (this) SubINode(d->in(1), mult);
3044 n->subsume_by(sub, this);
3045 }
3046 }
3047 }
3048 break;
3050 case Op_ModL:
3051 if (UseDivMod) {
3052 // Check if a%b and a/b both exist
3053 Node* d = n->find_similar(Op_DivL);
3054 if (d) {
3055 // Replace them with a fused divmod if supported
3056 if (Matcher::has_match_rule(Op_DivModL)) {
3057 DivModLNode* divmod = DivModLNode::make(this, n);
3058 d->subsume_by(divmod->div_proj(), this);
3059 n->subsume_by(divmod->mod_proj(), this);
3060 } else {
3061 // replace a%b with a-((a/b)*b)
3062 Node* mult = new (this) MulLNode(d, d->in(2));
3063 Node* sub = new (this) SubLNode(d->in(1), mult);
3064 n->subsume_by(sub, this);
3065 }
3066 }
3067 }
3068 break;
3070 case Op_LoadVector:
3071 case Op_StoreVector:
3072 break;
3074 case Op_PackB:
3075 case Op_PackS:
3076 case Op_PackI:
3077 case Op_PackF:
3078 case Op_PackL:
3079 case Op_PackD:
3080 if (n->req()-1 > 2) {
3081 // Replace many operand PackNodes with a binary tree for matching
3082 PackNode* p = (PackNode*) n;
3083 Node* btp = p->binary_tree_pack(this, 1, n->req());
3084 n->subsume_by(btp, this);
3085 }
3086 break;
3087 case Op_Loop:
3088 case Op_CountedLoop:
3089 if (n->as_Loop()->is_inner_loop()) {
3090 frc.inc_inner_loop_count();
3091 }
3092 break;
3093 case Op_LShiftI:
3094 case Op_RShiftI:
3095 case Op_URShiftI:
3096 case Op_LShiftL:
3097 case Op_RShiftL:
3098 case Op_URShiftL:
3099 if (Matcher::need_masked_shift_count) {
3100 // The cpu's shift instructions don't restrict the count to the
3101 // lower 5/6 bits. We need to do the masking ourselves.
3102 Node* in2 = n->in(2);
3103 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3104 const TypeInt* t = in2->find_int_type();
3105 if (t != NULL && t->is_con()) {
3106 juint shift = t->get_con();
3107 if (shift > mask) { // Unsigned cmp
3108 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
3109 }
3110 } else {
3111 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3112 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3113 n->set_req(2, shift);
3114 }
3115 }
3116 if (in2->outcnt() == 0) { // Remove dead node
3117 in2->disconnect_inputs(NULL, this);
3118 }
3119 }
3120 break;
3121 case Op_MemBarStoreStore:
3122 case Op_MemBarRelease:
3123 // Break the link with AllocateNode: it is no longer useful and
3124 // confuses register allocation.
3125 if (n->req() > MemBarNode::Precedent) {
3126 n->set_req(MemBarNode::Precedent, top());
3127 }
3128 break;
3129 default:
3130 assert( !n->is_Call(), "" );
3131 assert( !n->is_Mem(), "" );
3132 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3133 break;
3134 }
3136 // Collect CFG split points
3137 if (n->is_MultiBranch())
3138 frc._tests.push(n);
3139 }
3141 //------------------------------final_graph_reshaping_walk---------------------
3142 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3143 // requires that the walk visits a node's inputs before visiting the node.
3144 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3145 ResourceArea *area = Thread::current()->resource_area();
3146 Unique_Node_List sfpt(area);
3148 frc._visited.set(root->_idx); // first, mark node as visited
3149 uint cnt = root->req();
3150 Node *n = root;
3151 uint i = 0;
3152 while (true) {
3153 if (i < cnt) {
3154 // Place all non-visited non-null inputs onto stack
3155 Node* m = n->in(i);
3156 ++i;
3157 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3158 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3159 // compute worst case interpreter size in case of a deoptimization
3160 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3162 sfpt.push(m);
3163 }
3164 cnt = m->req();
3165 nstack.push(n, i); // put on stack parent and next input's index
3166 n = m;
3167 i = 0;
3168 }
3169 } else {
3170 // Now do post-visit work
3171 final_graph_reshaping_impl( n, frc );
3172 if (nstack.is_empty())
3173 break; // finished
3174 n = nstack.node(); // Get node from stack
3175 cnt = n->req();
3176 i = nstack.index();
3177 nstack.pop(); // Shift to the next node on stack
3178 }
3179 }
3181 // Skip next transformation if compressed oops are not used.
3182 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3183 (!UseCompressedOops && !UseCompressedClassPointers))
3184 return;
3186 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3187 // It could be done for an uncommon traps or any safepoints/calls
3188 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3189 while (sfpt.size() > 0) {
3190 n = sfpt.pop();
3191 JVMState *jvms = n->as_SafePoint()->jvms();
3192 assert(jvms != NULL, "sanity");
3193 int start = jvms->debug_start();
3194 int end = n->req();
3195 bool is_uncommon = (n->is_CallStaticJava() &&
3196 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3197 for (int j = start; j < end; j++) {
3198 Node* in = n->in(j);
3199 if (in->is_DecodeNarrowPtr()) {
3200 bool safe_to_skip = true;
3201 if (!is_uncommon ) {
3202 // Is it safe to skip?
3203 for (uint i = 0; i < in->outcnt(); i++) {
3204 Node* u = in->raw_out(i);
3205 if (!u->is_SafePoint() ||
3206 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3207 safe_to_skip = false;
3208 }
3209 }
3210 }
3211 if (safe_to_skip) {
3212 n->set_req(j, in->in(1));
3213 }
3214 if (in->outcnt() == 0) {
3215 in->disconnect_inputs(NULL, this);
3216 }
3217 }
3218 }
3219 }
3220 }
3222 //------------------------------final_graph_reshaping--------------------------
3223 // Final Graph Reshaping.
3224 //
3225 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3226 // and not commoned up and forced early. Must come after regular
3227 // optimizations to avoid GVN undoing the cloning. Clone constant
3228 // inputs to Loop Phis; these will be split by the allocator anyways.
3229 // Remove Opaque nodes.
3230 // (2) Move last-uses by commutative operations to the left input to encourage
3231 // Intel update-in-place two-address operations and better register usage
3232 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3233 // calls canonicalizing them back.
3234 // (3) Count the number of double-precision FP ops, single-precision FP ops
3235 // and call sites. On Intel, we can get correct rounding either by
3236 // forcing singles to memory (requires extra stores and loads after each
3237 // FP bytecode) or we can set a rounding mode bit (requires setting and
3238 // clearing the mode bit around call sites). The mode bit is only used
3239 // if the relative frequency of single FP ops to calls is low enough.
3240 // This is a key transform for SPEC mpeg_audio.
3241 // (4) Detect infinite loops; blobs of code reachable from above but not
3242 // below. Several of the Code_Gen algorithms fail on such code shapes,
3243 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3244 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3245 // Detection is by looking for IfNodes where only 1 projection is
3246 // reachable from below or CatchNodes missing some targets.
3247 // (5) Assert for insane oop offsets in debug mode.
3249 bool Compile::final_graph_reshaping() {
3250 // an infinite loop may have been eliminated by the optimizer,
3251 // in which case the graph will be empty.
3252 if (root()->req() == 1) {
3253 record_method_not_compilable("trivial infinite loop");
3254 return true;
3255 }
3257 // Expensive nodes have their control input set to prevent the GVN
3258 // from freely commoning them. There's no GVN beyond this point so
3259 // no need to keep the control input. We want the expensive nodes to
3260 // be freely moved to the least frequent code path by gcm.
3261 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3262 for (int i = 0; i < expensive_count(); i++) {
3263 _expensive_nodes->at(i)->set_req(0, NULL);
3264 }
3266 Final_Reshape_Counts frc;
3268 // Visit everybody reachable!
3269 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3270 Node_Stack nstack(live_nodes() >> 1);
3271 final_graph_reshaping_walk(nstack, root(), frc);
3273 // Check for unreachable (from below) code (i.e., infinite loops).
3274 for( uint i = 0; i < frc._tests.size(); i++ ) {
3275 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3276 // Get number of CFG targets.
3277 // Note that PCTables include exception targets after calls.
3278 uint required_outcnt = n->required_outcnt();
3279 if (n->outcnt() != required_outcnt) {
3280 // Check for a few special cases. Rethrow Nodes never take the
3281 // 'fall-thru' path, so expected kids is 1 less.
3282 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3283 if (n->in(0)->in(0)->is_Call()) {
3284 CallNode *call = n->in(0)->in(0)->as_Call();
3285 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3286 required_outcnt--; // Rethrow always has 1 less kid
3287 } else if (call->req() > TypeFunc::Parms &&
3288 call->is_CallDynamicJava()) {
3289 // Check for null receiver. In such case, the optimizer has
3290 // detected that the virtual call will always result in a null
3291 // pointer exception. The fall-through projection of this CatchNode
3292 // will not be populated.
3293 Node *arg0 = call->in(TypeFunc::Parms);
3294 if (arg0->is_Type() &&
3295 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3296 required_outcnt--;
3297 }
3298 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3299 call->req() > TypeFunc::Parms+1 &&
3300 call->is_CallStaticJava()) {
3301 // Check for negative array length. In such case, the optimizer has
3302 // detected that the allocation attempt will always result in an
3303 // exception. There is no fall-through projection of this CatchNode .
3304 Node *arg1 = call->in(TypeFunc::Parms+1);
3305 if (arg1->is_Type() &&
3306 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3307 required_outcnt--;
3308 }
3309 }
3310 }
3311 }
3312 // Recheck with a better notion of 'required_outcnt'
3313 if (n->outcnt() != required_outcnt) {
3314 record_method_not_compilable("malformed control flow");
3315 return true; // Not all targets reachable!
3316 }
3317 }
3318 // Check that I actually visited all kids. Unreached kids
3319 // must be infinite loops.
3320 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3321 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3322 record_method_not_compilable("infinite loop");
3323 return true; // Found unvisited kid; must be unreach
3324 }
3325 }
3327 // If original bytecodes contained a mixture of floats and doubles
3328 // check if the optimizer has made it homogenous, item (3).
3329 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3330 frc.get_float_count() > 32 &&
3331 frc.get_double_count() == 0 &&
3332 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3333 set_24_bit_selection_and_mode( false, true );
3334 }
3336 set_java_calls(frc.get_java_call_count());
3337 set_inner_loops(frc.get_inner_loop_count());
3339 // No infinite loops, no reason to bail out.
3340 return false;
3341 }
3343 //-----------------------------too_many_traps----------------------------------
3344 // Report if there are too many traps at the current method and bci.
3345 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3346 bool Compile::too_many_traps(ciMethod* method,
3347 int bci,
3348 Deoptimization::DeoptReason reason) {
3349 ciMethodData* md = method->method_data();
3350 if (md->is_empty()) {
3351 // Assume the trap has not occurred, or that it occurred only
3352 // because of a transient condition during start-up in the interpreter.
3353 return false;
3354 }
3355 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3356 if (md->has_trap_at(bci, m, reason) != 0) {
3357 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3358 // Also, if there are multiple reasons, or if there is no per-BCI record,
3359 // assume the worst.
3360 if (log())
3361 log()->elem("observe trap='%s' count='%d'",
3362 Deoptimization::trap_reason_name(reason),
3363 md->trap_count(reason));
3364 return true;
3365 } else {
3366 // Ignore method/bci and see if there have been too many globally.
3367 return too_many_traps(reason, md);
3368 }
3369 }
3371 // Less-accurate variant which does not require a method and bci.
3372 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3373 ciMethodData* logmd) {
3374 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3375 // Too many traps globally.
3376 // Note that we use cumulative trap_count, not just md->trap_count.
3377 if (log()) {
3378 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3379 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3380 Deoptimization::trap_reason_name(reason),
3381 mcount, trap_count(reason));
3382 }
3383 return true;
3384 } else {
3385 // The coast is clear.
3386 return false;
3387 }
3388 }
3390 //--------------------------too_many_recompiles--------------------------------
3391 // Report if there are too many recompiles at the current method and bci.
3392 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3393 // Is not eager to return true, since this will cause the compiler to use
3394 // Action_none for a trap point, to avoid too many recompilations.
3395 bool Compile::too_many_recompiles(ciMethod* method,
3396 int bci,
3397 Deoptimization::DeoptReason reason) {
3398 ciMethodData* md = method->method_data();
3399 if (md->is_empty()) {
3400 // Assume the trap has not occurred, or that it occurred only
3401 // because of a transient condition during start-up in the interpreter.
3402 return false;
3403 }
3404 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3405 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3406 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3407 Deoptimization::DeoptReason per_bc_reason
3408 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3409 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3410 if ((per_bc_reason == Deoptimization::Reason_none
3411 || md->has_trap_at(bci, m, reason) != 0)
3412 // The trap frequency measure we care about is the recompile count:
3413 && md->trap_recompiled_at(bci, m)
3414 && md->overflow_recompile_count() >= bc_cutoff) {
3415 // Do not emit a trap here if it has already caused recompilations.
3416 // Also, if there are multiple reasons, or if there is no per-BCI record,
3417 // assume the worst.
3418 if (log())
3419 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3420 Deoptimization::trap_reason_name(reason),
3421 md->trap_count(reason),
3422 md->overflow_recompile_count());
3423 return true;
3424 } else if (trap_count(reason) != 0
3425 && decompile_count() >= m_cutoff) {
3426 // Too many recompiles globally, and we have seen this sort of trap.
3427 // Use cumulative decompile_count, not just md->decompile_count.
3428 if (log())
3429 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3430 Deoptimization::trap_reason_name(reason),
3431 md->trap_count(reason), trap_count(reason),
3432 md->decompile_count(), decompile_count());
3433 return true;
3434 } else {
3435 // The coast is clear.
3436 return false;
3437 }
3438 }
3440 // Compute when not to trap. Used by matching trap based nodes and
3441 // NullCheck optimization.
3442 void Compile::set_allowed_deopt_reasons() {
3443 _allowed_reasons = 0;
3444 if (is_method_compilation()) {
3445 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3446 assert(rs < BitsPerInt, "recode bit map");
3447 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3448 _allowed_reasons |= nth_bit(rs);
3449 }
3450 }
3451 }
3452 }
3454 #ifndef PRODUCT
3455 //------------------------------verify_graph_edges---------------------------
3456 // Walk the Graph and verify that there is a one-to-one correspondence
3457 // between Use-Def edges and Def-Use edges in the graph.
3458 void Compile::verify_graph_edges(bool no_dead_code) {
3459 if (VerifyGraphEdges) {
3460 ResourceArea *area = Thread::current()->resource_area();
3461 Unique_Node_List visited(area);
3462 // Call recursive graph walk to check edges
3463 _root->verify_edges(visited);
3464 if (no_dead_code) {
3465 // Now make sure that no visited node is used by an unvisited node.
3466 bool dead_nodes = 0;
3467 Unique_Node_List checked(area);
3468 while (visited.size() > 0) {
3469 Node* n = visited.pop();
3470 checked.push(n);
3471 for (uint i = 0; i < n->outcnt(); i++) {
3472 Node* use = n->raw_out(i);
3473 if (checked.member(use)) continue; // already checked
3474 if (visited.member(use)) continue; // already in the graph
3475 if (use->is_Con()) continue; // a dead ConNode is OK
3476 // At this point, we have found a dead node which is DU-reachable.
3477 if (dead_nodes++ == 0)
3478 tty->print_cr("*** Dead nodes reachable via DU edges:");
3479 use->dump(2);
3480 tty->print_cr("---");
3481 checked.push(use); // No repeats; pretend it is now checked.
3482 }
3483 }
3484 assert(dead_nodes == 0, "using nodes must be reachable from root");
3485 }
3486 }
3487 }
3489 // Verify GC barriers consistency
3490 // Currently supported:
3491 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3492 void Compile::verify_barriers() {
3493 if (UseG1GC) {
3494 // Verify G1 pre-barriers
3495 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3497 ResourceArea *area = Thread::current()->resource_area();
3498 Unique_Node_List visited(area);
3499 Node_List worklist(area);
3500 // We're going to walk control flow backwards starting from the Root
3501 worklist.push(_root);
3502 while (worklist.size() > 0) {
3503 Node* x = worklist.pop();
3504 if (x == NULL || x == top()) continue;
3505 if (visited.member(x)) {
3506 continue;
3507 } else {
3508 visited.push(x);
3509 }
3511 if (x->is_Region()) {
3512 for (uint i = 1; i < x->req(); i++) {
3513 worklist.push(x->in(i));
3514 }
3515 } else {
3516 worklist.push(x->in(0));
3517 // We are looking for the pattern:
3518 // /->ThreadLocal
3519 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3520 // \->ConI(0)
3521 // We want to verify that the If and the LoadB have the same control
3522 // See GraphKit::g1_write_barrier_pre()
3523 if (x->is_If()) {
3524 IfNode *iff = x->as_If();
3525 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3526 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3527 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3528 && cmp->in(1)->is_Load()) {
3529 LoadNode* load = cmp->in(1)->as_Load();
3530 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3531 && load->in(2)->in(3)->is_Con()
3532 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3534 Node* if_ctrl = iff->in(0);
3535 Node* load_ctrl = load->in(0);
3537 if (if_ctrl != load_ctrl) {
3538 // Skip possible CProj->NeverBranch in infinite loops
3539 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3540 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3541 if_ctrl = if_ctrl->in(0)->in(0);
3542 }
3543 }
3544 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3545 }
3546 }
3547 }
3548 }
3549 }
3550 }
3551 }
3552 }
3554 #endif
3556 // The Compile object keeps track of failure reasons separately from the ciEnv.
3557 // This is required because there is not quite a 1-1 relation between the
3558 // ciEnv and its compilation task and the Compile object. Note that one
3559 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3560 // to backtrack and retry without subsuming loads. Other than this backtracking
3561 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3562 // by the logic in C2Compiler.
3563 void Compile::record_failure(const char* reason) {
3564 if (log() != NULL) {
3565 log()->elem("failure reason='%s' phase='compile'", reason);
3566 }
3567 if (_failure_reason == NULL) {
3568 // Record the first failure reason.
3569 _failure_reason = reason;
3570 }
3572 EventCompilerFailure event;
3573 if (event.should_commit()) {
3574 event.set_compileID(Compile::compile_id());
3575 event.set_failure(reason);
3576 event.commit();
3577 }
3579 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3580 C->print_method(PHASE_FAILURE);
3581 }
3582 _root = NULL; // flush the graph, too
3583 }
3585 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3586 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3587 _phase_name(name), _dolog(dolog)
3588 {
3589 if (dolog) {
3590 C = Compile::current();
3591 _log = C->log();
3592 } else {
3593 C = NULL;
3594 _log = NULL;
3595 }
3596 if (_log != NULL) {
3597 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3598 _log->stamp();
3599 _log->end_head();
3600 }
3601 }
3603 Compile::TracePhase::~TracePhase() {
3605 C = Compile::current();
3606 if (_dolog) {
3607 _log = C->log();
3608 } else {
3609 _log = NULL;
3610 }
3612 #ifdef ASSERT
3613 if (PrintIdealNodeCount) {
3614 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3615 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3616 }
3618 if (VerifyIdealNodeCount) {
3619 Compile::current()->print_missing_nodes();
3620 }
3621 #endif
3623 if (_log != NULL) {
3624 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3625 }
3626 }
3628 //=============================================================================
3629 // Two Constant's are equal when the type and the value are equal.
3630 bool Compile::Constant::operator==(const Constant& other) {
3631 if (type() != other.type() ) return false;
3632 if (can_be_reused() != other.can_be_reused()) return false;
3633 // For floating point values we compare the bit pattern.
3634 switch (type()) {
3635 case T_FLOAT: return (_v._value.i == other._v._value.i);
3636 case T_LONG:
3637 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3638 case T_OBJECT:
3639 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3640 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3641 case T_METADATA: return (_v._metadata == other._v._metadata);
3642 default: ShouldNotReachHere();
3643 }
3644 return false;
3645 }
3647 static int type_to_size_in_bytes(BasicType t) {
3648 switch (t) {
3649 case T_LONG: return sizeof(jlong );
3650 case T_FLOAT: return sizeof(jfloat );
3651 case T_DOUBLE: return sizeof(jdouble);
3652 case T_METADATA: return sizeof(Metadata*);
3653 // We use T_VOID as marker for jump-table entries (labels) which
3654 // need an internal word relocation.
3655 case T_VOID:
3656 case T_ADDRESS:
3657 case T_OBJECT: return sizeof(jobject);
3658 }
3660 ShouldNotReachHere();
3661 return -1;
3662 }
3664 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3665 // sort descending
3666 if (a->freq() > b->freq()) return -1;
3667 if (a->freq() < b->freq()) return 1;
3668 return 0;
3669 }
3671 void Compile::ConstantTable::calculate_offsets_and_size() {
3672 // First, sort the array by frequencies.
3673 _constants.sort(qsort_comparator);
3675 #ifdef ASSERT
3676 // Make sure all jump-table entries were sorted to the end of the
3677 // array (they have a negative frequency).
3678 bool found_void = false;
3679 for (int i = 0; i < _constants.length(); i++) {
3680 Constant con = _constants.at(i);
3681 if (con.type() == T_VOID)
3682 found_void = true; // jump-tables
3683 else
3684 assert(!found_void, "wrong sorting");
3685 }
3686 #endif
3688 int offset = 0;
3689 for (int i = 0; i < _constants.length(); i++) {
3690 Constant* con = _constants.adr_at(i);
3692 // Align offset for type.
3693 int typesize = type_to_size_in_bytes(con->type());
3694 offset = align_size_up(offset, typesize);
3695 con->set_offset(offset); // set constant's offset
3697 if (con->type() == T_VOID) {
3698 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3699 offset = offset + typesize * n->outcnt(); // expand jump-table
3700 } else {
3701 offset = offset + typesize;
3702 }
3703 }
3705 // Align size up to the next section start (which is insts; see
3706 // CodeBuffer::align_at_start).
3707 assert(_size == -1, "already set?");
3708 _size = align_size_up(offset, CodeEntryAlignment);
3709 }
3711 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3712 MacroAssembler _masm(&cb);
3713 for (int i = 0; i < _constants.length(); i++) {
3714 Constant con = _constants.at(i);
3715 address constant_addr = NULL;
3716 switch (con.type()) {
3717 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3718 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3719 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3720 case T_OBJECT: {
3721 jobject obj = con.get_jobject();
3722 int oop_index = _masm.oop_recorder()->find_index(obj);
3723 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3724 break;
3725 }
3726 case T_ADDRESS: {
3727 address addr = (address) con.get_jobject();
3728 constant_addr = _masm.address_constant(addr);
3729 break;
3730 }
3731 // We use T_VOID as marker for jump-table entries (labels) which
3732 // need an internal word relocation.
3733 case T_VOID: {
3734 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3735 // Fill the jump-table with a dummy word. The real value is
3736 // filled in later in fill_jump_table.
3737 address dummy = (address) n;
3738 constant_addr = _masm.address_constant(dummy);
3739 // Expand jump-table
3740 for (uint i = 1; i < n->outcnt(); i++) {
3741 address temp_addr = _masm.address_constant(dummy + i);
3742 assert(temp_addr, "consts section too small");
3743 }
3744 break;
3745 }
3746 case T_METADATA: {
3747 Metadata* obj = con.get_metadata();
3748 int metadata_index = _masm.oop_recorder()->find_index(obj);
3749 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3750 break;
3751 }
3752 default: ShouldNotReachHere();
3753 }
3754 assert(constant_addr, "consts section too small");
3755 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3756 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())));
3757 }
3758 }
3760 int Compile::ConstantTable::find_offset(Constant& con) const {
3761 int idx = _constants.find(con);
3762 assert(idx != -1, "constant must be in constant table");
3763 int offset = _constants.at(idx).offset();
3764 assert(offset != -1, "constant table not emitted yet?");
3765 return offset;
3766 }
3768 void Compile::ConstantTable::add(Constant& con) {
3769 if (con.can_be_reused()) {
3770 int idx = _constants.find(con);
3771 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3772 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3773 return;
3774 }
3775 }
3776 (void) _constants.append(con);
3777 }
3779 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3780 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3781 Constant con(type, value, b->_freq);
3782 add(con);
3783 return con;
3784 }
3786 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3787 Constant con(metadata);
3788 add(con);
3789 return con;
3790 }
3792 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3793 jvalue value;
3794 BasicType type = oper->type()->basic_type();
3795 switch (type) {
3796 case T_LONG: value.j = oper->constantL(); break;
3797 case T_FLOAT: value.f = oper->constantF(); break;
3798 case T_DOUBLE: value.d = oper->constantD(); break;
3799 case T_OBJECT:
3800 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3801 case T_METADATA: return add((Metadata*)oper->constant()); break;
3802 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3803 }
3804 return add(n, type, value);
3805 }
3807 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3808 jvalue value;
3809 // We can use the node pointer here to identify the right jump-table
3810 // as this method is called from Compile::Fill_buffer right before
3811 // the MachNodes are emitted and the jump-table is filled (means the
3812 // MachNode pointers do not change anymore).
3813 value.l = (jobject) n;
3814 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3815 add(con);
3816 return con;
3817 }
3819 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3820 // If called from Compile::scratch_emit_size do nothing.
3821 if (Compile::current()->in_scratch_emit_size()) return;
3823 assert(labels.is_nonempty(), "must be");
3824 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3826 // Since MachConstantNode::constant_offset() also contains
3827 // table_base_offset() we need to subtract the table_base_offset()
3828 // to get the plain offset into the constant table.
3829 int offset = n->constant_offset() - table_base_offset();
3831 MacroAssembler _masm(&cb);
3832 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3834 for (uint i = 0; i < n->outcnt(); i++) {
3835 address* constant_addr = &jump_table_base[i];
3836 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)));
3837 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3838 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3839 }
3840 }
3842 void Compile::dump_inlining() {
3843 if (print_inlining() || print_intrinsics()) {
3844 // Print inlining message for candidates that we couldn't inline
3845 // for lack of space or non constant receiver
3846 for (int i = 0; i < _late_inlines.length(); i++) {
3847 CallGenerator* cg = _late_inlines.at(i);
3848 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3849 }
3850 Unique_Node_List useful;
3851 useful.push(root());
3852 for (uint next = 0; next < useful.size(); ++next) {
3853 Node* n = useful.at(next);
3854 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3855 CallNode* call = n->as_Call();
3856 CallGenerator* cg = call->generator();
3857 cg->print_inlining_late("receiver not constant");
3858 }
3859 uint max = n->len();
3860 for ( uint i = 0; i < max; ++i ) {
3861 Node *m = n->in(i);
3862 if ( m == NULL ) continue;
3863 useful.push(m);
3864 }
3865 }
3866 for (int i = 0; i < _print_inlining_list->length(); i++) {
3867 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
3868 }
3869 }
3870 }
3872 // Dump inlining replay data to the stream.
3873 // Don't change thread state and acquire any locks.
3874 void Compile::dump_inline_data(outputStream* out) {
3875 InlineTree* inl_tree = ilt();
3876 if (inl_tree != NULL) {
3877 out->print(" inline %d", inl_tree->count());
3878 inl_tree->dump_replay_data(out);
3879 }
3880 }
3882 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3883 if (n1->Opcode() < n2->Opcode()) return -1;
3884 else if (n1->Opcode() > n2->Opcode()) return 1;
3886 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()));
3887 for (uint i = 1; i < n1->req(); i++) {
3888 if (n1->in(i) < n2->in(i)) return -1;
3889 else if (n1->in(i) > n2->in(i)) return 1;
3890 }
3892 return 0;
3893 }
3895 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3896 Node* n1 = *n1p;
3897 Node* n2 = *n2p;
3899 return cmp_expensive_nodes(n1, n2);
3900 }
3902 void Compile::sort_expensive_nodes() {
3903 if (!expensive_nodes_sorted()) {
3904 _expensive_nodes->sort(cmp_expensive_nodes);
3905 }
3906 }
3908 bool Compile::expensive_nodes_sorted() const {
3909 for (int i = 1; i < _expensive_nodes->length(); i++) {
3910 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3911 return false;
3912 }
3913 }
3914 return true;
3915 }
3917 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3918 if (_expensive_nodes->length() == 0) {
3919 return false;
3920 }
3922 assert(OptimizeExpensiveOps, "optimization off?");
3924 // Take this opportunity to remove dead nodes from the list
3925 int j = 0;
3926 for (int i = 0; i < _expensive_nodes->length(); i++) {
3927 Node* n = _expensive_nodes->at(i);
3928 if (!n->is_unreachable(igvn)) {
3929 assert(n->is_expensive(), "should be expensive");
3930 _expensive_nodes->at_put(j, n);
3931 j++;
3932 }
3933 }
3934 _expensive_nodes->trunc_to(j);
3936 // Then sort the list so that similar nodes are next to each other
3937 // and check for at least two nodes of identical kind with same data
3938 // inputs.
3939 sort_expensive_nodes();
3941 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3942 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3943 return true;
3944 }
3945 }
3947 return false;
3948 }
3950 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3951 if (_expensive_nodes->length() == 0) {
3952 return;
3953 }
3955 assert(OptimizeExpensiveOps, "optimization off?");
3957 // Sort to bring similar nodes next to each other and clear the
3958 // control input of nodes for which there's only a single copy.
3959 sort_expensive_nodes();
3961 int j = 0;
3962 int identical = 0;
3963 int i = 0;
3964 for (; i < _expensive_nodes->length()-1; i++) {
3965 assert(j <= i, "can't write beyond current index");
3966 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3967 identical++;
3968 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3969 continue;
3970 }
3971 if (identical > 0) {
3972 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3973 identical = 0;
3974 } else {
3975 Node* n = _expensive_nodes->at(i);
3976 igvn.hash_delete(n);
3977 n->set_req(0, NULL);
3978 igvn.hash_insert(n);
3979 }
3980 }
3981 if (identical > 0) {
3982 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3983 } else if (_expensive_nodes->length() >= 1) {
3984 Node* n = _expensive_nodes->at(i);
3985 igvn.hash_delete(n);
3986 n->set_req(0, NULL);
3987 igvn.hash_insert(n);
3988 }
3989 _expensive_nodes->trunc_to(j);
3990 }
3992 void Compile::add_expensive_node(Node * n) {
3993 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3994 assert(n->is_expensive(), "expensive nodes with non-null control here only");
3995 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
3996 if (OptimizeExpensiveOps) {
3997 _expensive_nodes->append(n);
3998 } else {
3999 // Clear control input and let IGVN optimize expensive nodes if
4000 // OptimizeExpensiveOps is off.
4001 n->set_req(0, NULL);
4002 }
4003 }
4005 /**
4006 * Remove the speculative part of types and clean up the graph
4007 */
4008 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4009 if (UseTypeSpeculation) {
4010 Unique_Node_List worklist;
4011 worklist.push(root());
4012 int modified = 0;
4013 // Go over all type nodes that carry a speculative type, drop the
4014 // speculative part of the type and enqueue the node for an igvn
4015 // which may optimize it out.
4016 for (uint next = 0; next < worklist.size(); ++next) {
4017 Node *n = worklist.at(next);
4018 if (n->is_Type()) {
4019 TypeNode* tn = n->as_Type();
4020 const Type* t = tn->type();
4021 const Type* t_no_spec = t->remove_speculative();
4022 if (t_no_spec != t) {
4023 bool in_hash = igvn.hash_delete(n);
4024 assert(in_hash, "node should be in igvn hash table");
4025 tn->set_type(t_no_spec);
4026 igvn.hash_insert(n);
4027 igvn._worklist.push(n); // give it a chance to go away
4028 modified++;
4029 }
4030 }
4031 uint max = n->len();
4032 for( uint i = 0; i < max; ++i ) {
4033 Node *m = n->in(i);
4034 if (not_a_node(m)) continue;
4035 worklist.push(m);
4036 }
4037 }
4038 // Drop the speculative part of all types in the igvn's type table
4039 igvn.remove_speculative_types();
4040 if (modified > 0) {
4041 igvn.optimize();
4042 }
4043 #ifdef ASSERT
4044 // Verify that after the IGVN is over no speculative type has resurfaced
4045 worklist.clear();
4046 worklist.push(root());
4047 for (uint next = 0; next < worklist.size(); ++next) {
4048 Node *n = worklist.at(next);
4049 const Type* t = igvn.type_or_null(n);
4050 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4051 if (n->is_Type()) {
4052 t = n->as_Type()->type();
4053 assert(t == t->remove_speculative(), "no more speculative types");
4054 }
4055 uint max = n->len();
4056 for( uint i = 0; i < max; ++i ) {
4057 Node *m = n->in(i);
4058 if (not_a_node(m)) continue;
4059 worklist.push(m);
4060 }
4061 }
4062 igvn.check_no_speculative_types();
4063 #endif
4064 }
4065 }
4067 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4068 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4069 if (ctrl != NULL) {
4070 // Express control dependency by a CastII node with a narrow type.
4071 value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */);
4072 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4073 // node from floating above the range check during loop optimizations. Otherwise, the
4074 // ConvI2L node may be eliminated independently of the range check, causing the data path
4075 // to become TOP while the control path is still there (although it's unreachable).
4076 value->set_req(0, ctrl);
4077 // Save CastII node to remove it after loop optimizations.
4078 phase->C->add_range_check_cast(value);
4079 value = phase->transform(value);
4080 }
4081 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4082 return phase->transform(new (phase->C) ConvI2LNode(value, ltype));
4083 }
4085 // Auxiliary method to support randomized stressing/fuzzing.
4086 //
4087 // This method can be called the arbitrary number of times, with current count
4088 // as the argument. The logic allows selecting a single candidate from the
4089 // running list of candidates as follows:
4090 // int count = 0;
4091 // Cand* selected = null;
4092 // while(cand = cand->next()) {
4093 // if (randomized_select(++count)) {
4094 // selected = cand;
4095 // }
4096 // }
4097 //
4098 // Including count equalizes the chances any candidate is "selected".
4099 // This is useful when we don't have the complete list of candidates to choose
4100 // from uniformly. In this case, we need to adjust the randomicity of the
4101 // selection, or else we will end up biasing the selection towards the latter
4102 // candidates.
4103 //
4104 // Quick back-envelope calculation shows that for the list of n candidates
4105 // the equal probability for the candidate to persist as "best" can be
4106 // achieved by replacing it with "next" k-th candidate with the probability
4107 // of 1/k. It can be easily shown that by the end of the run, the
4108 // probability for any candidate is converged to 1/n, thus giving the
4109 // uniform distribution among all the candidates.
4110 //
4111 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4112 #define RANDOMIZED_DOMAIN_POW 29
4113 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4114 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4115 bool Compile::randomized_select(int count) {
4116 assert(count > 0, "only positive");
4117 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4118 }