Mon, 23 Jun 2014 13:33:23 +0200
8046289: compiler/6340864/TestLongVect.java timeout with
Reviewed-by: iveresov, vlivanov
1 /*
2 * Copyright (c) 1997, 2014, 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 #ifdef TARGET_ARCH_MODEL_x86_32
71 # include "adfiles/ad_x86_32.hpp"
72 #endif
73 #ifdef TARGET_ARCH_MODEL_x86_64
74 # include "adfiles/ad_x86_64.hpp"
75 #endif
76 #ifdef TARGET_ARCH_MODEL_sparc
77 # include "adfiles/ad_sparc.hpp"
78 #endif
79 #ifdef TARGET_ARCH_MODEL_zero
80 # include "adfiles/ad_zero.hpp"
81 #endif
82 #ifdef TARGET_ARCH_MODEL_arm
83 # include "adfiles/ad_arm.hpp"
84 #endif
85 #ifdef TARGET_ARCH_MODEL_ppc_32
86 # include "adfiles/ad_ppc_32.hpp"
87 #endif
88 #ifdef TARGET_ARCH_MODEL_ppc_64
89 # include "adfiles/ad_ppc_64.hpp"
90 #endif
93 // -------------------- Compile::mach_constant_base_node -----------------------
94 // Constant table base node singleton.
95 MachConstantBaseNode* Compile::mach_constant_base_node() {
96 if (_mach_constant_base_node == NULL) {
97 _mach_constant_base_node = new (C) MachConstantBaseNode();
98 _mach_constant_base_node->add_req(C->root());
99 }
100 return _mach_constant_base_node;
101 }
104 /// Support for intrinsics.
106 // Return the index at which m must be inserted (or already exists).
107 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
108 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
109 #ifdef ASSERT
110 for (int i = 1; i < _intrinsics->length(); i++) {
111 CallGenerator* cg1 = _intrinsics->at(i-1);
112 CallGenerator* cg2 = _intrinsics->at(i);
113 assert(cg1->method() != cg2->method()
114 ? cg1->method() < cg2->method()
115 : cg1->is_virtual() < cg2->is_virtual(),
116 "compiler intrinsics list must stay sorted");
117 }
118 #endif
119 // Binary search sorted list, in decreasing intervals [lo, hi].
120 int lo = 0, hi = _intrinsics->length()-1;
121 while (lo <= hi) {
122 int mid = (uint)(hi + lo) / 2;
123 ciMethod* mid_m = _intrinsics->at(mid)->method();
124 if (m < mid_m) {
125 hi = mid-1;
126 } else if (m > mid_m) {
127 lo = mid+1;
128 } else {
129 // look at minor sort key
130 bool mid_virt = _intrinsics->at(mid)->is_virtual();
131 if (is_virtual < mid_virt) {
132 hi = mid-1;
133 } else if (is_virtual > mid_virt) {
134 lo = mid+1;
135 } else {
136 return mid; // exact match
137 }
138 }
139 }
140 return lo; // inexact match
141 }
143 void Compile::register_intrinsic(CallGenerator* cg) {
144 if (_intrinsics == NULL) {
145 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
146 }
147 // This code is stolen from ciObjectFactory::insert.
148 // Really, GrowableArray should have methods for
149 // insert_at, remove_at, and binary_search.
150 int len = _intrinsics->length();
151 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
152 if (index == len) {
153 _intrinsics->append(cg);
154 } else {
155 #ifdef ASSERT
156 CallGenerator* oldcg = _intrinsics->at(index);
157 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
158 #endif
159 _intrinsics->append(_intrinsics->at(len-1));
160 int pos;
161 for (pos = len-2; pos >= index; pos--) {
162 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
163 }
164 _intrinsics->at_put(index, cg);
165 }
166 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
167 }
169 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
170 assert(m->is_loaded(), "don't try this on unloaded methods");
171 if (_intrinsics != NULL) {
172 int index = intrinsic_insertion_index(m, is_virtual);
173 if (index < _intrinsics->length()
174 && _intrinsics->at(index)->method() == m
175 && _intrinsics->at(index)->is_virtual() == is_virtual) {
176 return _intrinsics->at(index);
177 }
178 }
179 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
180 if (m->intrinsic_id() != vmIntrinsics::_none &&
181 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
182 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
183 if (cg != NULL) {
184 // Save it for next time:
185 register_intrinsic(cg);
186 return cg;
187 } else {
188 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
189 }
190 }
191 return NULL;
192 }
194 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
195 // in library_call.cpp.
198 #ifndef PRODUCT
199 // statistics gathering...
201 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
202 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
204 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
205 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
206 int oflags = _intrinsic_hist_flags[id];
207 assert(flags != 0, "what happened?");
208 if (is_virtual) {
209 flags |= _intrinsic_virtual;
210 }
211 bool changed = (flags != oflags);
212 if ((flags & _intrinsic_worked) != 0) {
213 juint count = (_intrinsic_hist_count[id] += 1);
214 if (count == 1) {
215 changed = true; // first time
216 }
217 // increment the overall count also:
218 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
219 }
220 if (changed) {
221 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
222 // Something changed about the intrinsic's virtuality.
223 if ((flags & _intrinsic_virtual) != 0) {
224 // This is the first use of this intrinsic as a virtual call.
225 if (oflags != 0) {
226 // We already saw it as a non-virtual, so note both cases.
227 flags |= _intrinsic_both;
228 }
229 } else if ((oflags & _intrinsic_both) == 0) {
230 // This is the first use of this intrinsic as a non-virtual
231 flags |= _intrinsic_both;
232 }
233 }
234 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
235 }
236 // update the overall flags also:
237 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
238 return changed;
239 }
241 static char* format_flags(int flags, char* buf) {
242 buf[0] = 0;
243 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
244 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
245 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
246 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
247 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
248 if (buf[0] == 0) strcat(buf, ",");
249 assert(buf[0] == ',', "must be");
250 return &buf[1];
251 }
253 void Compile::print_intrinsic_statistics() {
254 char flagsbuf[100];
255 ttyLocker ttyl;
256 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
257 tty->print_cr("Compiler intrinsic usage:");
258 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
259 if (total == 0) total = 1; // avoid div0 in case of no successes
260 #define PRINT_STAT_LINE(name, c, f) \
261 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
262 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
263 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
264 int flags = _intrinsic_hist_flags[id];
265 juint count = _intrinsic_hist_count[id];
266 if ((flags | count) != 0) {
267 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
268 }
269 }
270 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
271 if (xtty != NULL) xtty->tail("statistics");
272 }
274 void Compile::print_statistics() {
275 { ttyLocker ttyl;
276 if (xtty != NULL) xtty->head("statistics type='opto'");
277 Parse::print_statistics();
278 PhaseCCP::print_statistics();
279 PhaseRegAlloc::print_statistics();
280 Scheduling::print_statistics();
281 PhasePeephole::print_statistics();
282 PhaseIdealLoop::print_statistics();
283 if (xtty != NULL) xtty->tail("statistics");
284 }
285 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
286 // put this under its own <statistics> element.
287 print_intrinsic_statistics();
288 }
289 }
290 #endif //PRODUCT
292 // Support for bundling info
293 Bundle* Compile::node_bundling(const Node *n) {
294 assert(valid_bundle_info(n), "oob");
295 return &_node_bundling_base[n->_idx];
296 }
298 bool Compile::valid_bundle_info(const Node *n) {
299 return (_node_bundling_limit > n->_idx);
300 }
303 void Compile::gvn_replace_by(Node* n, Node* nn) {
304 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
305 Node* use = n->last_out(i);
306 bool is_in_table = initial_gvn()->hash_delete(use);
307 uint uses_found = 0;
308 for (uint j = 0; j < use->len(); j++) {
309 if (use->in(j) == n) {
310 if (j < use->req())
311 use->set_req(j, nn);
312 else
313 use->set_prec(j, nn);
314 uses_found++;
315 }
316 }
317 if (is_in_table) {
318 // reinsert into table
319 initial_gvn()->hash_find_insert(use);
320 }
321 record_for_igvn(use);
322 i -= uses_found; // we deleted 1 or more copies of this edge
323 }
324 }
327 static inline bool not_a_node(const Node* n) {
328 if (n == NULL) return true;
329 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
330 if (*(address*)n == badAddress) return true; // kill by Node::destruct
331 return false;
332 }
334 // Identify all nodes that are reachable from below, useful.
335 // Use breadth-first pass that records state in a Unique_Node_List,
336 // recursive traversal is slower.
337 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
338 int estimated_worklist_size = unique();
339 useful.map( estimated_worklist_size, NULL ); // preallocate space
341 // Initialize worklist
342 if (root() != NULL) { useful.push(root()); }
343 // If 'top' is cached, declare it useful to preserve cached node
344 if( cached_top_node() ) { useful.push(cached_top_node()); }
346 // Push all useful nodes onto the list, breadthfirst
347 for( uint next = 0; next < useful.size(); ++next ) {
348 assert( next < unique(), "Unique useful nodes < total nodes");
349 Node *n = useful.at(next);
350 uint max = n->len();
351 for( uint i = 0; i < max; ++i ) {
352 Node *m = n->in(i);
353 if (not_a_node(m)) continue;
354 useful.push(m);
355 }
356 }
357 }
359 // Update dead_node_list with any missing dead nodes using useful
360 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
361 void Compile::update_dead_node_list(Unique_Node_List &useful) {
362 uint max_idx = unique();
363 VectorSet& useful_node_set = useful.member_set();
365 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
366 // If node with index node_idx is not in useful set,
367 // mark it as dead in dead node list.
368 if (! useful_node_set.test(node_idx) ) {
369 record_dead_node(node_idx);
370 }
371 }
372 }
374 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
375 int shift = 0;
376 for (int i = 0; i < inlines->length(); i++) {
377 CallGenerator* cg = inlines->at(i);
378 CallNode* call = cg->call_node();
379 if (shift > 0) {
380 inlines->at_put(i-shift, cg);
381 }
382 if (!useful.member(call)) {
383 shift++;
384 }
385 }
386 inlines->trunc_to(inlines->length()-shift);
387 }
389 // Disconnect all useless nodes by disconnecting those at the boundary.
390 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
391 uint next = 0;
392 while (next < useful.size()) {
393 Node *n = useful.at(next++);
394 if (n->is_SafePoint()) {
395 // We're done with a parsing phase. Replaced nodes are not valid
396 // beyond that point.
397 n->as_SafePoint()->delete_replaced_nodes();
398 }
399 // Use raw traversal of out edges since this code removes out edges
400 int max = n->outcnt();
401 for (int j = 0; j < max; ++j) {
402 Node* child = n->raw_out(j);
403 if (! useful.member(child)) {
404 assert(!child->is_top() || child != top(),
405 "If top is cached in Compile object it is in useful list");
406 // Only need to remove this out-edge to the useless node
407 n->raw_del_out(j);
408 --j;
409 --max;
410 }
411 }
412 if (n->outcnt() == 1 && n->has_special_unique_user()) {
413 record_for_igvn(n->unique_out());
414 }
415 }
416 // Remove useless macro and predicate opaq nodes
417 for (int i = C->macro_count()-1; i >= 0; i--) {
418 Node* n = C->macro_node(i);
419 if (!useful.member(n)) {
420 remove_macro_node(n);
421 }
422 }
423 // Remove useless expensive node
424 for (int i = C->expensive_count()-1; i >= 0; i--) {
425 Node* n = C->expensive_node(i);
426 if (!useful.member(n)) {
427 remove_expensive_node(n);
428 }
429 }
430 // clean up the late inline lists
431 remove_useless_late_inlines(&_string_late_inlines, useful);
432 remove_useless_late_inlines(&_boxing_late_inlines, useful);
433 remove_useless_late_inlines(&_late_inlines, useful);
434 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
435 }
437 //------------------------------frame_size_in_words-----------------------------
438 // frame_slots in units of words
439 int Compile::frame_size_in_words() const {
440 // shift is 0 in LP32 and 1 in LP64
441 const int shift = (LogBytesPerWord - LogBytesPerInt);
442 int words = _frame_slots >> shift;
443 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
444 return words;
445 }
447 // To bang the stack of this compiled method we use the stack size
448 // that the interpreter would need in case of a deoptimization. This
449 // removes the need to bang the stack in the deoptimization blob which
450 // in turn simplifies stack overflow handling.
451 int Compile::bang_size_in_bytes() const {
452 return MAX2(_interpreter_frame_size, frame_size_in_bytes());
453 }
455 // ============================================================================
456 //------------------------------CompileWrapper---------------------------------
457 class CompileWrapper : public StackObj {
458 Compile *const _compile;
459 public:
460 CompileWrapper(Compile* compile);
462 ~CompileWrapper();
463 };
465 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
466 // the Compile* pointer is stored in the current ciEnv:
467 ciEnv* env = compile->env();
468 assert(env == ciEnv::current(), "must already be a ciEnv active");
469 assert(env->compiler_data() == NULL, "compile already active?");
470 env->set_compiler_data(compile);
471 assert(compile == Compile::current(), "sanity");
473 compile->set_type_dict(NULL);
474 compile->set_type_hwm(NULL);
475 compile->set_type_last_size(0);
476 compile->set_last_tf(NULL, NULL);
477 compile->set_indexSet_arena(NULL);
478 compile->set_indexSet_free_block_list(NULL);
479 compile->init_type_arena();
480 Type::Initialize(compile);
481 _compile->set_scratch_buffer_blob(NULL);
482 _compile->begin_method();
483 }
484 CompileWrapper::~CompileWrapper() {
485 _compile->end_method();
486 if (_compile->scratch_buffer_blob() != NULL)
487 BufferBlob::free(_compile->scratch_buffer_blob());
488 _compile->env()->set_compiler_data(NULL);
489 }
492 //----------------------------print_compile_messages---------------------------
493 void Compile::print_compile_messages() {
494 #ifndef PRODUCT
495 // Check if recompiling
496 if (_subsume_loads == false && PrintOpto) {
497 // Recompiling without allowing machine instructions to subsume loads
498 tty->print_cr("*********************************************************");
499 tty->print_cr("** Bailout: Recompile without subsuming loads **");
500 tty->print_cr("*********************************************************");
501 }
502 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
503 // Recompiling without escape analysis
504 tty->print_cr("*********************************************************");
505 tty->print_cr("** Bailout: Recompile without escape analysis **");
506 tty->print_cr("*********************************************************");
507 }
508 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
509 // Recompiling without boxing elimination
510 tty->print_cr("*********************************************************");
511 tty->print_cr("** Bailout: Recompile without boxing elimination **");
512 tty->print_cr("*********************************************************");
513 }
514 if (env()->break_at_compile()) {
515 // Open the debugger when compiling this method.
516 tty->print("### Breaking when compiling: ");
517 method()->print_short_name();
518 tty->cr();
519 BREAKPOINT;
520 }
522 if( PrintOpto ) {
523 if (is_osr_compilation()) {
524 tty->print("[OSR]%3d", _compile_id);
525 } else {
526 tty->print("%3d", _compile_id);
527 }
528 }
529 #endif
530 }
533 //-----------------------init_scratch_buffer_blob------------------------------
534 // Construct a temporary BufferBlob and cache it for this compile.
535 void Compile::init_scratch_buffer_blob(int const_size) {
536 // If there is already a scratch buffer blob allocated and the
537 // constant section is big enough, use it. Otherwise free the
538 // current and allocate a new one.
539 BufferBlob* blob = scratch_buffer_blob();
540 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
541 // Use the current blob.
542 } else {
543 if (blob != NULL) {
544 BufferBlob::free(blob);
545 }
547 ResourceMark rm;
548 _scratch_const_size = const_size;
549 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
550 blob = BufferBlob::create("Compile::scratch_buffer", size);
551 // Record the buffer blob for next time.
552 set_scratch_buffer_blob(blob);
553 // Have we run out of code space?
554 if (scratch_buffer_blob() == NULL) {
555 // Let CompilerBroker disable further compilations.
556 record_failure("Not enough space for scratch buffer in CodeCache");
557 return;
558 }
559 }
561 // Initialize the relocation buffers
562 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
563 set_scratch_locs_memory(locs_buf);
564 }
567 //-----------------------scratch_emit_size-------------------------------------
568 // Helper function that computes size by emitting code
569 uint Compile::scratch_emit_size(const Node* n) {
570 // Start scratch_emit_size section.
571 set_in_scratch_emit_size(true);
573 // Emit into a trash buffer and count bytes emitted.
574 // This is a pretty expensive way to compute a size,
575 // but it works well enough if seldom used.
576 // All common fixed-size instructions are given a size
577 // method by the AD file.
578 // Note that the scratch buffer blob and locs memory are
579 // allocated at the beginning of the compile task, and
580 // may be shared by several calls to scratch_emit_size.
581 // The allocation of the scratch buffer blob is particularly
582 // expensive, since it has to grab the code cache lock.
583 BufferBlob* blob = this->scratch_buffer_blob();
584 assert(blob != NULL, "Initialize BufferBlob at start");
585 assert(blob->size() > MAX_inst_size, "sanity");
586 relocInfo* locs_buf = scratch_locs_memory();
587 address blob_begin = blob->content_begin();
588 address blob_end = (address)locs_buf;
589 assert(blob->content_contains(blob_end), "sanity");
590 CodeBuffer buf(blob_begin, blob_end - blob_begin);
591 buf.initialize_consts_size(_scratch_const_size);
592 buf.initialize_stubs_size(MAX_stubs_size);
593 assert(locs_buf != NULL, "sanity");
594 int lsize = MAX_locs_size / 3;
595 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
596 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
597 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
599 // Do the emission.
601 Label fakeL; // Fake label for branch instructions.
602 Label* saveL = NULL;
603 uint save_bnum = 0;
604 bool is_branch = n->is_MachBranch();
605 if (is_branch) {
606 MacroAssembler masm(&buf);
607 masm.bind(fakeL);
608 n->as_MachBranch()->save_label(&saveL, &save_bnum);
609 n->as_MachBranch()->label_set(&fakeL, 0);
610 }
611 n->emit(buf, this->regalloc());
612 if (is_branch) // Restore label.
613 n->as_MachBranch()->label_set(saveL, save_bnum);
615 // End scratch_emit_size section.
616 set_in_scratch_emit_size(false);
618 return buf.insts_size();
619 }
622 // ============================================================================
623 //------------------------------Compile standard-------------------------------
624 debug_only( int Compile::_debug_idx = 100000; )
626 // Compile a method. entry_bci is -1 for normal compilations and indicates
627 // the continuation bci for on stack replacement.
630 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
631 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
632 : Phase(Compiler),
633 _env(ci_env),
634 _log(ci_env->log()),
635 _compile_id(ci_env->compile_id()),
636 _save_argument_registers(false),
637 _stub_name(NULL),
638 _stub_function(NULL),
639 _stub_entry_point(NULL),
640 _method(target),
641 _entry_bci(osr_bci),
642 _initial_gvn(NULL),
643 _for_igvn(NULL),
644 _warm_calls(NULL),
645 _subsume_loads(subsume_loads),
646 _do_escape_analysis(do_escape_analysis),
647 _eliminate_boxing(eliminate_boxing),
648 _failure_reason(NULL),
649 _code_buffer("Compile::Fill_buffer"),
650 _orig_pc_slot(0),
651 _orig_pc_slot_offset_in_bytes(0),
652 _has_method_handle_invokes(false),
653 _mach_constant_base_node(NULL),
654 _node_bundling_limit(0),
655 _node_bundling_base(NULL),
656 _java_calls(0),
657 _inner_loops(0),
658 _scratch_const_size(-1),
659 _in_scratch_emit_size(false),
660 _dead_node_list(comp_arena()),
661 _dead_node_count(0),
662 #ifndef PRODUCT
663 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
664 _in_dump_cnt(0),
665 _printer(IdealGraphPrinter::printer()),
666 #endif
667 _congraph(NULL),
668 _comp_arena(mtCompiler),
669 _node_arena(mtCompiler),
670 _old_arena(mtCompiler),
671 _Compile_types(mtCompiler),
672 _replay_inline_data(NULL),
673 _late_inlines(comp_arena(), 2, 0, NULL),
674 _string_late_inlines(comp_arena(), 2, 0, NULL),
675 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
676 _late_inlines_pos(0),
677 _number_of_mh_late_inlines(0),
678 _inlining_progress(false),
679 _inlining_incrementally(false),
680 _print_inlining_list(NULL),
681 _print_inlining_idx(0),
682 _interpreter_frame_size(0) {
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 C = this;
996 #ifndef PRODUCT
997 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
998 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
999 set_print_assembly(PrintFrameConverterAssembly);
1000 set_parsed_irreducible_loop(false);
1001 #endif
1002 set_has_irreducible_loop(false); // no loops
1004 CompileWrapper cw(this);
1005 Init(/*AliasLevel=*/ 0);
1006 init_tf((*generator)());
1008 {
1009 // The following is a dummy for the sake of GraphKit::gen_stub
1010 Unique_Node_List for_igvn(comp_arena());
1011 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1012 PhaseGVN gvn(Thread::current()->resource_area(),255);
1013 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1014 gvn.transform_no_reclaim(top());
1016 GraphKit kit;
1017 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1018 }
1020 NOT_PRODUCT( verify_graph_edges(); )
1021 Code_Gen();
1022 if (failing()) return;
1025 // Entry point will be accessed using compile->stub_entry_point();
1026 if (code_buffer() == NULL) {
1027 Matcher::soft_match_failure();
1028 } else {
1029 if (PrintAssembly && (WizardMode || Verbose))
1030 tty->print_cr("### Stub::%s", stub_name);
1032 if (!failing()) {
1033 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1035 // Make the NMethod
1036 // For now we mark the frame as never safe for profile stackwalking
1037 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1038 code_buffer(),
1039 CodeOffsets::frame_never_safe,
1040 // _code_offsets.value(CodeOffsets::Frame_Complete),
1041 frame_size_in_words(),
1042 _oop_map_set,
1043 save_arg_registers);
1044 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1046 _stub_entry_point = rs->entry_point();
1047 }
1048 }
1049 }
1051 //------------------------------Init-------------------------------------------
1052 // Prepare for a single compilation
1053 void Compile::Init(int aliaslevel) {
1054 _unique = 0;
1055 _regalloc = NULL;
1057 _tf = NULL; // filled in later
1058 _top = NULL; // cached later
1059 _matcher = NULL; // filled in later
1060 _cfg = NULL; // filled in later
1062 set_24_bit_selection_and_mode(Use24BitFP, false);
1064 _node_note_array = NULL;
1065 _default_node_notes = NULL;
1067 _immutable_memory = NULL; // filled in at first inquiry
1069 // Globally visible Nodes
1070 // First set TOP to NULL to give safe behavior during creation of RootNode
1071 set_cached_top_node(NULL);
1072 set_root(new (this) RootNode());
1073 // Now that you have a Root to point to, create the real TOP
1074 set_cached_top_node( new (this) ConNode(Type::TOP) );
1075 set_recent_alloc(NULL, NULL);
1077 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1078 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1079 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1080 env()->set_dependencies(new Dependencies(env()));
1082 _fixed_slots = 0;
1083 set_has_split_ifs(false);
1084 set_has_loops(has_method() && method()->has_loops()); // first approximation
1085 set_has_stringbuilder(false);
1086 set_has_boxed_value(false);
1087 _trap_can_recompile = false; // no traps emitted yet
1088 _major_progress = true; // start out assuming good things will happen
1089 set_has_unsafe_access(false);
1090 set_max_vector_size(0);
1091 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1092 set_decompile_count(0);
1094 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1095 set_num_loop_opts(LoopOptsCount);
1096 set_do_inlining(Inline);
1097 set_max_inline_size(MaxInlineSize);
1098 set_freq_inline_size(FreqInlineSize);
1099 set_do_scheduling(OptoScheduling);
1100 set_do_count_invocations(false);
1101 set_do_method_data_update(false);
1102 set_rtm_state(NoRTM); // No RTM lock eliding by default
1103 #if INCLUDE_RTM_OPT
1104 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1105 int rtm_state = method()->method_data()->rtm_state();
1106 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1107 // Don't generate RTM lock eliding code.
1108 set_rtm_state(NoRTM);
1109 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1110 // Generate RTM lock eliding code without abort ratio calculation code.
1111 set_rtm_state(UseRTM);
1112 } else if (UseRTMDeopt) {
1113 // Generate RTM lock eliding code and include abort ratio calculation
1114 // code if UseRTMDeopt is on.
1115 set_rtm_state(ProfileRTM);
1116 }
1117 }
1118 #endif
1119 if (debug_info()->recording_non_safepoints()) {
1120 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1121 (comp_arena(), 8, 0, NULL));
1122 set_default_node_notes(Node_Notes::make(this));
1123 }
1125 // // -- Initialize types before each compile --
1126 // // Update cached type information
1127 // if( _method && _method->constants() )
1128 // Type::update_loaded_types(_method, _method->constants());
1130 // Init alias_type map.
1131 if (!_do_escape_analysis && aliaslevel == 3)
1132 aliaslevel = 2; // No unique types without escape analysis
1133 _AliasLevel = aliaslevel;
1134 const int grow_ats = 16;
1135 _max_alias_types = grow_ats;
1136 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1137 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1138 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1139 {
1140 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1141 }
1142 // Initialize the first few types.
1143 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1144 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1145 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1146 _num_alias_types = AliasIdxRaw+1;
1147 // Zero out the alias type cache.
1148 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1149 // A NULL adr_type hits in the cache right away. Preload the right answer.
1150 probe_alias_cache(NULL)->_index = AliasIdxTop;
1152 _intrinsics = NULL;
1153 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1154 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1155 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1156 register_library_intrinsics();
1157 }
1159 //---------------------------init_start----------------------------------------
1160 // Install the StartNode on this compile object.
1161 void Compile::init_start(StartNode* s) {
1162 if (failing())
1163 return; // already failing
1164 assert(s == start(), "");
1165 }
1167 StartNode* Compile::start() const {
1168 assert(!failing(), "");
1169 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1170 Node* start = root()->fast_out(i);
1171 if( start->is_Start() )
1172 return start->as_Start();
1173 }
1174 fatal("Did not find Start node!");
1175 return NULL;
1176 }
1178 //-------------------------------immutable_memory-------------------------------------
1179 // Access immutable memory
1180 Node* Compile::immutable_memory() {
1181 if (_immutable_memory != NULL) {
1182 return _immutable_memory;
1183 }
1184 StartNode* s = start();
1185 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1186 Node *p = s->fast_out(i);
1187 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1188 _immutable_memory = p;
1189 return _immutable_memory;
1190 }
1191 }
1192 ShouldNotReachHere();
1193 return NULL;
1194 }
1196 //----------------------set_cached_top_node------------------------------------
1197 // Install the cached top node, and make sure Node::is_top works correctly.
1198 void Compile::set_cached_top_node(Node* tn) {
1199 if (tn != NULL) verify_top(tn);
1200 Node* old_top = _top;
1201 _top = tn;
1202 // Calling Node::setup_is_top allows the nodes the chance to adjust
1203 // their _out arrays.
1204 if (_top != NULL) _top->setup_is_top();
1205 if (old_top != NULL) old_top->setup_is_top();
1206 assert(_top == NULL || top()->is_top(), "");
1207 }
1209 #ifdef ASSERT
1210 uint Compile::count_live_nodes_by_graph_walk() {
1211 Unique_Node_List useful(comp_arena());
1212 // Get useful node list by walking the graph.
1213 identify_useful_nodes(useful);
1214 return useful.size();
1215 }
1217 void Compile::print_missing_nodes() {
1219 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1220 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1221 return;
1222 }
1224 // This is an expensive function. It is executed only when the user
1225 // specifies VerifyIdealNodeCount option or otherwise knows the
1226 // additional work that needs to be done to identify reachable nodes
1227 // by walking the flow graph and find the missing ones using
1228 // _dead_node_list.
1230 Unique_Node_List useful(comp_arena());
1231 // Get useful node list by walking the graph.
1232 identify_useful_nodes(useful);
1234 uint l_nodes = C->live_nodes();
1235 uint l_nodes_by_walk = useful.size();
1237 if (l_nodes != l_nodes_by_walk) {
1238 if (_log != NULL) {
1239 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1240 _log->stamp();
1241 _log->end_head();
1242 }
1243 VectorSet& useful_member_set = useful.member_set();
1244 int last_idx = l_nodes_by_walk;
1245 for (int i = 0; i < last_idx; i++) {
1246 if (useful_member_set.test(i)) {
1247 if (_dead_node_list.test(i)) {
1248 if (_log != NULL) {
1249 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1250 }
1251 if (PrintIdealNodeCount) {
1252 // Print the log message to tty
1253 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1254 useful.at(i)->dump();
1255 }
1256 }
1257 }
1258 else if (! _dead_node_list.test(i)) {
1259 if (_log != NULL) {
1260 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1261 }
1262 if (PrintIdealNodeCount) {
1263 // Print the log message to tty
1264 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1265 }
1266 }
1267 }
1268 if (_log != NULL) {
1269 _log->tail("mismatched_nodes");
1270 }
1271 }
1272 }
1273 #endif
1275 #ifndef PRODUCT
1276 void Compile::verify_top(Node* tn) const {
1277 if (tn != NULL) {
1278 assert(tn->is_Con(), "top node must be a constant");
1279 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1280 assert(tn->in(0) != NULL, "must have live top node");
1281 }
1282 }
1283 #endif
1286 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1288 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1289 guarantee(arr != NULL, "");
1290 int num_blocks = arr->length();
1291 if (grow_by < num_blocks) grow_by = num_blocks;
1292 int num_notes = grow_by * _node_notes_block_size;
1293 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1294 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1295 while (num_notes > 0) {
1296 arr->append(notes);
1297 notes += _node_notes_block_size;
1298 num_notes -= _node_notes_block_size;
1299 }
1300 assert(num_notes == 0, "exact multiple, please");
1301 }
1303 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1304 if (source == NULL || dest == NULL) return false;
1306 if (dest->is_Con())
1307 return false; // Do not push debug info onto constants.
1309 #ifdef ASSERT
1310 // Leave a bread crumb trail pointing to the original node:
1311 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1312 dest->set_debug_orig(source);
1313 }
1314 #endif
1316 if (node_note_array() == NULL)
1317 return false; // Not collecting any notes now.
1319 // This is a copy onto a pre-existing node, which may already have notes.
1320 // If both nodes have notes, do not overwrite any pre-existing notes.
1321 Node_Notes* source_notes = node_notes_at(source->_idx);
1322 if (source_notes == NULL || source_notes->is_clear()) return false;
1323 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1324 if (dest_notes == NULL || dest_notes->is_clear()) {
1325 return set_node_notes_at(dest->_idx, source_notes);
1326 }
1328 Node_Notes merged_notes = (*source_notes);
1329 // The order of operations here ensures that dest notes will win...
1330 merged_notes.update_from(dest_notes);
1331 return set_node_notes_at(dest->_idx, &merged_notes);
1332 }
1335 //--------------------------allow_range_check_smearing-------------------------
1336 // Gating condition for coalescing similar range checks.
1337 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1338 // single covering check that is at least as strong as any of them.
1339 // If the optimization succeeds, the simplified (strengthened) range check
1340 // will always succeed. If it fails, we will deopt, and then give up
1341 // on the optimization.
1342 bool Compile::allow_range_check_smearing() const {
1343 // If this method has already thrown a range-check,
1344 // assume it was because we already tried range smearing
1345 // and it failed.
1346 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1347 return !already_trapped;
1348 }
1351 //------------------------------flatten_alias_type-----------------------------
1352 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1353 int offset = tj->offset();
1354 TypePtr::PTR ptr = tj->ptr();
1356 // Known instance (scalarizable allocation) alias only with itself.
1357 bool is_known_inst = tj->isa_oopptr() != NULL &&
1358 tj->is_oopptr()->is_known_instance();
1360 // Process weird unsafe references.
1361 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1362 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1363 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1364 tj = TypeOopPtr::BOTTOM;
1365 ptr = tj->ptr();
1366 offset = tj->offset();
1367 }
1369 // Array pointers need some flattening
1370 const TypeAryPtr *ta = tj->isa_aryptr();
1371 if (ta && ta->is_stable()) {
1372 // Erase stability property for alias analysis.
1373 tj = ta = ta->cast_to_stable(false);
1374 }
1375 if( ta && is_known_inst ) {
1376 if ( offset != Type::OffsetBot &&
1377 offset > arrayOopDesc::length_offset_in_bytes() ) {
1378 offset = Type::OffsetBot; // Flatten constant access into array body only
1379 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1380 }
1381 } else if( ta && _AliasLevel >= 2 ) {
1382 // For arrays indexed by constant indices, we flatten the alias
1383 // space to include all of the array body. Only the header, klass
1384 // and array length can be accessed un-aliased.
1385 if( offset != Type::OffsetBot ) {
1386 if( ta->const_oop() ) { // MethodData* or Method*
1387 offset = Type::OffsetBot; // Flatten constant access into array body
1388 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1389 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1390 // range is OK as-is.
1391 tj = ta = TypeAryPtr::RANGE;
1392 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1393 tj = TypeInstPtr::KLASS; // all klass loads look alike
1394 ta = TypeAryPtr::RANGE; // generic ignored junk
1395 ptr = TypePtr::BotPTR;
1396 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1397 tj = TypeInstPtr::MARK;
1398 ta = TypeAryPtr::RANGE; // generic ignored junk
1399 ptr = TypePtr::BotPTR;
1400 } else { // Random constant offset into array body
1401 offset = Type::OffsetBot; // Flatten constant access into array body
1402 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1403 }
1404 }
1405 // Arrays of fixed size alias with arrays of unknown size.
1406 if (ta->size() != TypeInt::POS) {
1407 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1408 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1409 }
1410 // Arrays of known objects become arrays of unknown objects.
1411 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1412 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1413 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1414 }
1415 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1416 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1417 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1418 }
1419 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1420 // cannot be distinguished by bytecode alone.
1421 if (ta->elem() == TypeInt::BOOL) {
1422 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1423 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1424 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1425 }
1426 // During the 2nd round of IterGVN, NotNull castings are removed.
1427 // Make sure the Bottom and NotNull variants alias the same.
1428 // Also, make sure exact and non-exact variants alias the same.
1429 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1430 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1431 }
1432 }
1434 // Oop pointers need some flattening
1435 const TypeInstPtr *to = tj->isa_instptr();
1436 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1437 ciInstanceKlass *k = to->klass()->as_instance_klass();
1438 if( ptr == TypePtr::Constant ) {
1439 if (to->klass() != ciEnv::current()->Class_klass() ||
1440 offset < k->size_helper() * wordSize) {
1441 // No constant oop pointers (such as Strings); they alias with
1442 // unknown strings.
1443 assert(!is_known_inst, "not scalarizable allocation");
1444 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1445 }
1446 } else if( is_known_inst ) {
1447 tj = to; // Keep NotNull and klass_is_exact for instance type
1448 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1449 // During the 2nd round of IterGVN, NotNull castings are removed.
1450 // Make sure the Bottom and NotNull variants alias the same.
1451 // Also, make sure exact and non-exact variants alias the same.
1452 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1453 }
1454 if (to->speculative() != NULL) {
1455 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1456 }
1457 // Canonicalize the holder of this field
1458 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1459 // First handle header references such as a LoadKlassNode, even if the
1460 // object's klass is unloaded at compile time (4965979).
1461 if (!is_known_inst) { // Do it only for non-instance types
1462 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1463 }
1464 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1465 // Static fields are in the space above the normal instance
1466 // fields in the java.lang.Class instance.
1467 if (to->klass() != ciEnv::current()->Class_klass()) {
1468 to = NULL;
1469 tj = TypeOopPtr::BOTTOM;
1470 offset = tj->offset();
1471 }
1472 } else {
1473 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1474 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1475 if( is_known_inst ) {
1476 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1477 } else {
1478 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1479 }
1480 }
1481 }
1482 }
1484 // Klass pointers to object array klasses need some flattening
1485 const TypeKlassPtr *tk = tj->isa_klassptr();
1486 if( tk ) {
1487 // If we are referencing a field within a Klass, we need
1488 // to assume the worst case of an Object. Both exact and
1489 // inexact types must flatten to the same alias class so
1490 // use NotNull as the PTR.
1491 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1493 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1494 TypeKlassPtr::OBJECT->klass(),
1495 offset);
1496 }
1498 ciKlass* klass = tk->klass();
1499 if( klass->is_obj_array_klass() ) {
1500 ciKlass* k = TypeAryPtr::OOPS->klass();
1501 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1502 k = TypeInstPtr::BOTTOM->klass();
1503 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1504 }
1506 // Check for precise loads from the primary supertype array and force them
1507 // to the supertype cache alias index. Check for generic array loads from
1508 // the primary supertype array and also force them to the supertype cache
1509 // alias index. Since the same load can reach both, we need to merge
1510 // these 2 disparate memories into the same alias class. Since the
1511 // primary supertype array is read-only, there's no chance of confusion
1512 // where we bypass an array load and an array store.
1513 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1514 if (offset == Type::OffsetBot ||
1515 (offset >= primary_supers_offset &&
1516 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1517 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1518 offset = in_bytes(Klass::secondary_super_cache_offset());
1519 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1520 }
1521 }
1523 // Flatten all Raw pointers together.
1524 if (tj->base() == Type::RawPtr)
1525 tj = TypeRawPtr::BOTTOM;
1527 if (tj->base() == Type::AnyPtr)
1528 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1530 // Flatten all to bottom for now
1531 switch( _AliasLevel ) {
1532 case 0:
1533 tj = TypePtr::BOTTOM;
1534 break;
1535 case 1: // Flatten to: oop, static, field or array
1536 switch (tj->base()) {
1537 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1538 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1539 case Type::AryPtr: // do not distinguish arrays at all
1540 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1541 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1542 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1543 default: ShouldNotReachHere();
1544 }
1545 break;
1546 case 2: // No collapsing at level 2; keep all splits
1547 case 3: // No collapsing at level 3; keep all splits
1548 break;
1549 default:
1550 Unimplemented();
1551 }
1553 offset = tj->offset();
1554 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1556 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1557 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1558 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1559 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1560 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1561 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1562 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1563 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1564 assert( tj->ptr() != TypePtr::TopPTR &&
1565 tj->ptr() != TypePtr::AnyNull &&
1566 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1567 // assert( tj->ptr() != TypePtr::Constant ||
1568 // tj->base() == Type::RawPtr ||
1569 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1571 return tj;
1572 }
1574 void Compile::AliasType::Init(int i, const TypePtr* at) {
1575 _index = i;
1576 _adr_type = at;
1577 _field = NULL;
1578 _element = NULL;
1579 _is_rewritable = true; // default
1580 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1581 if (atoop != NULL && atoop->is_known_instance()) {
1582 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1583 _general_index = Compile::current()->get_alias_index(gt);
1584 } else {
1585 _general_index = 0;
1586 }
1587 }
1589 //---------------------------------print_on------------------------------------
1590 #ifndef PRODUCT
1591 void Compile::AliasType::print_on(outputStream* st) {
1592 if (index() < 10)
1593 st->print("@ <%d> ", index());
1594 else st->print("@ <%d>", index());
1595 st->print(is_rewritable() ? " " : " RO");
1596 int offset = adr_type()->offset();
1597 if (offset == Type::OffsetBot)
1598 st->print(" +any");
1599 else st->print(" +%-3d", offset);
1600 st->print(" in ");
1601 adr_type()->dump_on(st);
1602 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1603 if (field() != NULL && tjp) {
1604 if (tjp->klass() != field()->holder() ||
1605 tjp->offset() != field()->offset_in_bytes()) {
1606 st->print(" != ");
1607 field()->print();
1608 st->print(" ***");
1609 }
1610 }
1611 }
1613 void print_alias_types() {
1614 Compile* C = Compile::current();
1615 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1616 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1617 C->alias_type(idx)->print_on(tty);
1618 tty->cr();
1619 }
1620 }
1621 #endif
1624 //----------------------------probe_alias_cache--------------------------------
1625 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1626 intptr_t key = (intptr_t) adr_type;
1627 key ^= key >> logAliasCacheSize;
1628 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1629 }
1632 //-----------------------------grow_alias_types--------------------------------
1633 void Compile::grow_alias_types() {
1634 const int old_ats = _max_alias_types; // how many before?
1635 const int new_ats = old_ats; // how many more?
1636 const int grow_ats = old_ats+new_ats; // how many now?
1637 _max_alias_types = grow_ats;
1638 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1639 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1640 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1641 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1642 }
1645 //--------------------------------find_alias_type------------------------------
1646 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1647 if (_AliasLevel == 0)
1648 return alias_type(AliasIdxBot);
1650 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1651 if (ace->_adr_type == adr_type) {
1652 return alias_type(ace->_index);
1653 }
1655 // Handle special cases.
1656 if (adr_type == NULL) return alias_type(AliasIdxTop);
1657 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1659 // Do it the slow way.
1660 const TypePtr* flat = flatten_alias_type(adr_type);
1662 #ifdef ASSERT
1663 assert(flat == flatten_alias_type(flat), "idempotent");
1664 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1665 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1666 const TypeOopPtr* foop = flat->is_oopptr();
1667 // Scalarizable allocations have exact klass always.
1668 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1669 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1670 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1671 }
1672 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1673 #endif
1675 int idx = AliasIdxTop;
1676 for (int i = 0; i < num_alias_types(); i++) {
1677 if (alias_type(i)->adr_type() == flat) {
1678 idx = i;
1679 break;
1680 }
1681 }
1683 if (idx == AliasIdxTop) {
1684 if (no_create) return NULL;
1685 // Grow the array if necessary.
1686 if (_num_alias_types == _max_alias_types) grow_alias_types();
1687 // Add a new alias type.
1688 idx = _num_alias_types++;
1689 _alias_types[idx]->Init(idx, flat);
1690 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1691 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1692 if (flat->isa_instptr()) {
1693 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1694 && flat->is_instptr()->klass() == env()->Class_klass())
1695 alias_type(idx)->set_rewritable(false);
1696 }
1697 if (flat->isa_aryptr()) {
1698 #ifdef ASSERT
1699 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1700 // (T_BYTE has the weakest alignment and size restrictions...)
1701 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1702 #endif
1703 if (flat->offset() == TypePtr::OffsetBot) {
1704 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1705 }
1706 }
1707 if (flat->isa_klassptr()) {
1708 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1709 alias_type(idx)->set_rewritable(false);
1710 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1711 alias_type(idx)->set_rewritable(false);
1712 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1713 alias_type(idx)->set_rewritable(false);
1714 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1715 alias_type(idx)->set_rewritable(false);
1716 }
1717 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1718 // but the base pointer type is not distinctive enough to identify
1719 // references into JavaThread.)
1721 // Check for final fields.
1722 const TypeInstPtr* tinst = flat->isa_instptr();
1723 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1724 ciField* field;
1725 if (tinst->const_oop() != NULL &&
1726 tinst->klass() == ciEnv::current()->Class_klass() &&
1727 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1728 // static field
1729 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1730 field = k->get_field_by_offset(tinst->offset(), true);
1731 } else {
1732 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1733 field = k->get_field_by_offset(tinst->offset(), false);
1734 }
1735 assert(field == NULL ||
1736 original_field == NULL ||
1737 (field->holder() == original_field->holder() &&
1738 field->offset() == original_field->offset() &&
1739 field->is_static() == original_field->is_static()), "wrong field?");
1740 // Set field() and is_rewritable() attributes.
1741 if (field != NULL) alias_type(idx)->set_field(field);
1742 }
1743 }
1745 // Fill the cache for next time.
1746 ace->_adr_type = adr_type;
1747 ace->_index = idx;
1748 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1750 // Might as well try to fill the cache for the flattened version, too.
1751 AliasCacheEntry* face = probe_alias_cache(flat);
1752 if (face->_adr_type == NULL) {
1753 face->_adr_type = flat;
1754 face->_index = idx;
1755 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1756 }
1758 return alias_type(idx);
1759 }
1762 Compile::AliasType* Compile::alias_type(ciField* field) {
1763 const TypeOopPtr* t;
1764 if (field->is_static())
1765 t = TypeInstPtr::make(field->holder()->java_mirror());
1766 else
1767 t = TypeOopPtr::make_from_klass_raw(field->holder());
1768 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1769 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1770 return atp;
1771 }
1774 //------------------------------have_alias_type--------------------------------
1775 bool Compile::have_alias_type(const TypePtr* adr_type) {
1776 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1777 if (ace->_adr_type == adr_type) {
1778 return true;
1779 }
1781 // Handle special cases.
1782 if (adr_type == NULL) return true;
1783 if (adr_type == TypePtr::BOTTOM) return true;
1785 return find_alias_type(adr_type, true, NULL) != NULL;
1786 }
1788 //-----------------------------must_alias--------------------------------------
1789 // True if all values of the given address type are in the given alias category.
1790 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1791 if (alias_idx == AliasIdxBot) return true; // the universal category
1792 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1793 if (alias_idx == AliasIdxTop) return false; // the empty category
1794 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1796 // the only remaining possible overlap is identity
1797 int adr_idx = get_alias_index(adr_type);
1798 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1799 assert(adr_idx == alias_idx ||
1800 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1801 && adr_type != TypeOopPtr::BOTTOM),
1802 "should not be testing for overlap with an unsafe pointer");
1803 return adr_idx == alias_idx;
1804 }
1806 //------------------------------can_alias--------------------------------------
1807 // True if any values of the given address type are in the given alias category.
1808 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1809 if (alias_idx == AliasIdxTop) return false; // the empty category
1810 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1811 if (alias_idx == AliasIdxBot) return true; // the universal category
1812 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1814 // the only remaining possible overlap is identity
1815 int adr_idx = get_alias_index(adr_type);
1816 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1817 return adr_idx == alias_idx;
1818 }
1822 //---------------------------pop_warm_call-------------------------------------
1823 WarmCallInfo* Compile::pop_warm_call() {
1824 WarmCallInfo* wci = _warm_calls;
1825 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1826 return wci;
1827 }
1829 //----------------------------Inline_Warm--------------------------------------
1830 int Compile::Inline_Warm() {
1831 // If there is room, try to inline some more warm call sites.
1832 // %%% Do a graph index compaction pass when we think we're out of space?
1833 if (!InlineWarmCalls) return 0;
1835 int calls_made_hot = 0;
1836 int room_to_grow = NodeCountInliningCutoff - unique();
1837 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1838 int amount_grown = 0;
1839 WarmCallInfo* call;
1840 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1841 int est_size = (int)call->size();
1842 if (est_size > (room_to_grow - amount_grown)) {
1843 // This one won't fit anyway. Get rid of it.
1844 call->make_cold();
1845 continue;
1846 }
1847 call->make_hot();
1848 calls_made_hot++;
1849 amount_grown += est_size;
1850 amount_to_grow -= est_size;
1851 }
1853 if (calls_made_hot > 0) set_major_progress();
1854 return calls_made_hot;
1855 }
1858 //----------------------------Finish_Warm--------------------------------------
1859 void Compile::Finish_Warm() {
1860 if (!InlineWarmCalls) return;
1861 if (failing()) return;
1862 if (warm_calls() == NULL) return;
1864 // Clean up loose ends, if we are out of space for inlining.
1865 WarmCallInfo* call;
1866 while ((call = pop_warm_call()) != NULL) {
1867 call->make_cold();
1868 }
1869 }
1871 //---------------------cleanup_loop_predicates-----------------------
1872 // Remove the opaque nodes that protect the predicates so that all unused
1873 // checks and uncommon_traps will be eliminated from the ideal graph
1874 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1875 if (predicate_count()==0) return;
1876 for (int i = predicate_count(); i > 0; i--) {
1877 Node * n = predicate_opaque1_node(i-1);
1878 assert(n->Opcode() == Op_Opaque1, "must be");
1879 igvn.replace_node(n, n->in(1));
1880 }
1881 assert(predicate_count()==0, "should be clean!");
1882 }
1884 // StringOpts and late inlining of string methods
1885 void Compile::inline_string_calls(bool parse_time) {
1886 {
1887 // remove useless nodes to make the usage analysis simpler
1888 ResourceMark rm;
1889 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1890 }
1892 {
1893 ResourceMark rm;
1894 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1895 PhaseStringOpts pso(initial_gvn(), for_igvn());
1896 print_method(PHASE_AFTER_STRINGOPTS, 3);
1897 }
1899 // now inline anything that we skipped the first time around
1900 if (!parse_time) {
1901 _late_inlines_pos = _late_inlines.length();
1902 }
1904 while (_string_late_inlines.length() > 0) {
1905 CallGenerator* cg = _string_late_inlines.pop();
1906 cg->do_late_inline();
1907 if (failing()) return;
1908 }
1909 _string_late_inlines.trunc_to(0);
1910 }
1912 // Late inlining of boxing methods
1913 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1914 if (_boxing_late_inlines.length() > 0) {
1915 assert(has_boxed_value(), "inconsistent");
1917 PhaseGVN* gvn = initial_gvn();
1918 set_inlining_incrementally(true);
1920 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1921 for_igvn()->clear();
1922 gvn->replace_with(&igvn);
1924 _late_inlines_pos = _late_inlines.length();
1926 while (_boxing_late_inlines.length() > 0) {
1927 CallGenerator* cg = _boxing_late_inlines.pop();
1928 cg->do_late_inline();
1929 if (failing()) return;
1930 }
1931 _boxing_late_inlines.trunc_to(0);
1933 {
1934 ResourceMark rm;
1935 PhaseRemoveUseless pru(gvn, for_igvn());
1936 }
1938 igvn = PhaseIterGVN(gvn);
1939 igvn.optimize();
1941 set_inlining_progress(false);
1942 set_inlining_incrementally(false);
1943 }
1944 }
1946 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1947 assert(IncrementalInline, "incremental inlining should be on");
1948 PhaseGVN* gvn = initial_gvn();
1950 set_inlining_progress(false);
1951 for_igvn()->clear();
1952 gvn->replace_with(&igvn);
1954 int i = 0;
1956 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1957 CallGenerator* cg = _late_inlines.at(i);
1958 _late_inlines_pos = i+1;
1959 cg->do_late_inline();
1960 if (failing()) return;
1961 }
1962 int j = 0;
1963 for (; i < _late_inlines.length(); i++, j++) {
1964 _late_inlines.at_put(j, _late_inlines.at(i));
1965 }
1966 _late_inlines.trunc_to(j);
1968 {
1969 ResourceMark rm;
1970 PhaseRemoveUseless pru(gvn, for_igvn());
1971 }
1973 igvn = PhaseIterGVN(gvn);
1974 }
1976 // Perform incremental inlining until bound on number of live nodes is reached
1977 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1978 PhaseGVN* gvn = initial_gvn();
1980 set_inlining_incrementally(true);
1981 set_inlining_progress(true);
1982 uint low_live_nodes = 0;
1984 while(inlining_progress() && _late_inlines.length() > 0) {
1986 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1987 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1988 // PhaseIdealLoop is expensive so we only try it once we are
1989 // out of live nodes and we only try it again if the previous
1990 // helped got the number of nodes down significantly
1991 PhaseIdealLoop ideal_loop( igvn, false, true );
1992 if (failing()) return;
1993 low_live_nodes = live_nodes();
1994 _major_progress = true;
1995 }
1997 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1998 break;
1999 }
2000 }
2002 inline_incrementally_one(igvn);
2004 if (failing()) return;
2006 igvn.optimize();
2008 if (failing()) return;
2009 }
2011 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2013 if (_string_late_inlines.length() > 0) {
2014 assert(has_stringbuilder(), "inconsistent");
2015 for_igvn()->clear();
2016 initial_gvn()->replace_with(&igvn);
2018 inline_string_calls(false);
2020 if (failing()) return;
2022 {
2023 ResourceMark rm;
2024 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2025 }
2027 igvn = PhaseIterGVN(gvn);
2029 igvn.optimize();
2030 }
2032 set_inlining_incrementally(false);
2033 }
2036 //------------------------------Optimize---------------------------------------
2037 // Given a graph, optimize it.
2038 void Compile::Optimize() {
2039 TracePhase t1("optimizer", &_t_optimizer, true);
2041 #ifndef PRODUCT
2042 if (env()->break_at_compile()) {
2043 BREAKPOINT;
2044 }
2046 #endif
2048 ResourceMark rm;
2049 int loop_opts_cnt;
2051 NOT_PRODUCT( verify_graph_edges(); )
2053 print_method(PHASE_AFTER_PARSING);
2055 {
2056 // Iterative Global Value Numbering, including ideal transforms
2057 // Initialize IterGVN with types and values from parse-time GVN
2058 PhaseIterGVN igvn(initial_gvn());
2059 {
2060 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2061 igvn.optimize();
2062 }
2064 print_method(PHASE_ITER_GVN1, 2);
2066 if (failing()) return;
2068 {
2069 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2070 inline_incrementally(igvn);
2071 }
2073 print_method(PHASE_INCREMENTAL_INLINE, 2);
2075 if (failing()) return;
2077 if (eliminate_boxing()) {
2078 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2079 // Inline valueOf() methods now.
2080 inline_boxing_calls(igvn);
2082 if (AlwaysIncrementalInline) {
2083 inline_incrementally(igvn);
2084 }
2086 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2088 if (failing()) return;
2089 }
2091 // Remove the speculative part of types and clean up the graph from
2092 // the extra CastPP nodes whose only purpose is to carry them. Do
2093 // that early so that optimizations are not disrupted by the extra
2094 // CastPP nodes.
2095 remove_speculative_types(igvn);
2097 // No more new expensive nodes will be added to the list from here
2098 // so keep only the actual candidates for optimizations.
2099 cleanup_expensive_nodes(igvn);
2101 // Perform escape analysis
2102 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2103 if (has_loops()) {
2104 // Cleanup graph (remove dead nodes).
2105 TracePhase t2("idealLoop", &_t_idealLoop, true);
2106 PhaseIdealLoop ideal_loop( igvn, false, true );
2107 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2108 if (failing()) return;
2109 }
2110 ConnectionGraph::do_analysis(this, &igvn);
2112 if (failing()) return;
2114 // Optimize out fields loads from scalar replaceable allocations.
2115 igvn.optimize();
2116 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2118 if (failing()) return;
2120 if (congraph() != NULL && macro_count() > 0) {
2121 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2122 PhaseMacroExpand mexp(igvn);
2123 mexp.eliminate_macro_nodes();
2124 igvn.set_delay_transform(false);
2126 igvn.optimize();
2127 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2129 if (failing()) return;
2130 }
2131 }
2133 // Loop transforms on the ideal graph. Range Check Elimination,
2134 // peeling, unrolling, etc.
2136 // Set loop opts counter
2137 loop_opts_cnt = num_loop_opts();
2138 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2139 {
2140 TracePhase t2("idealLoop", &_t_idealLoop, true);
2141 PhaseIdealLoop ideal_loop( igvn, true );
2142 loop_opts_cnt--;
2143 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2144 if (failing()) return;
2145 }
2146 // Loop opts pass if partial peeling occurred in previous pass
2147 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2148 TracePhase t3("idealLoop", &_t_idealLoop, true);
2149 PhaseIdealLoop ideal_loop( igvn, false );
2150 loop_opts_cnt--;
2151 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2152 if (failing()) return;
2153 }
2154 // Loop opts pass for loop-unrolling before CCP
2155 if(major_progress() && (loop_opts_cnt > 0)) {
2156 TracePhase t4("idealLoop", &_t_idealLoop, true);
2157 PhaseIdealLoop ideal_loop( igvn, false );
2158 loop_opts_cnt--;
2159 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2160 }
2161 if (!failing()) {
2162 // Verify that last round of loop opts produced a valid graph
2163 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2164 PhaseIdealLoop::verify(igvn);
2165 }
2166 }
2167 if (failing()) return;
2169 // Conditional Constant Propagation;
2170 PhaseCCP ccp( &igvn );
2171 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2172 {
2173 TracePhase t2("ccp", &_t_ccp, true);
2174 ccp.do_transform();
2175 }
2176 print_method(PHASE_CPP1, 2);
2178 assert( true, "Break here to ccp.dump_old2new_map()");
2180 // Iterative Global Value Numbering, including ideal transforms
2181 {
2182 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2183 igvn = ccp;
2184 igvn.optimize();
2185 }
2187 print_method(PHASE_ITER_GVN2, 2);
2189 if (failing()) return;
2191 // Loop transforms on the ideal graph. Range Check Elimination,
2192 // peeling, unrolling, etc.
2193 if(loop_opts_cnt > 0) {
2194 debug_only( int cnt = 0; );
2195 while(major_progress() && (loop_opts_cnt > 0)) {
2196 TracePhase t2("idealLoop", &_t_idealLoop, true);
2197 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2198 PhaseIdealLoop ideal_loop( igvn, true);
2199 loop_opts_cnt--;
2200 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2201 if (failing()) return;
2202 }
2203 }
2205 {
2206 // Verify that all previous optimizations produced a valid graph
2207 // at least to this point, even if no loop optimizations were done.
2208 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2209 PhaseIdealLoop::verify(igvn);
2210 }
2212 {
2213 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2214 PhaseMacroExpand mex(igvn);
2215 if (mex.expand_macro_nodes()) {
2216 assert(failing(), "must bail out w/ explicit message");
2217 return;
2218 }
2219 }
2221 } // (End scope of igvn; run destructor if necessary for asserts.)
2223 dump_inlining();
2224 // A method with only infinite loops has no edges entering loops from root
2225 {
2226 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2227 if (final_graph_reshaping()) {
2228 assert(failing(), "must bail out w/ explicit message");
2229 return;
2230 }
2231 }
2233 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2234 }
2237 //------------------------------Code_Gen---------------------------------------
2238 // Given a graph, generate code for it
2239 void Compile::Code_Gen() {
2240 if (failing()) {
2241 return;
2242 }
2244 // Perform instruction selection. You might think we could reclaim Matcher
2245 // memory PDQ, but actually the Matcher is used in generating spill code.
2246 // Internals of the Matcher (including some VectorSets) must remain live
2247 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2248 // set a bit in reclaimed memory.
2250 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2251 // nodes. Mapping is only valid at the root of each matched subtree.
2252 NOT_PRODUCT( verify_graph_edges(); )
2254 Matcher matcher;
2255 _matcher = &matcher;
2256 {
2257 TracePhase t2("matcher", &_t_matcher, true);
2258 matcher.match();
2259 }
2260 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2261 // nodes. Mapping is only valid at the root of each matched subtree.
2262 NOT_PRODUCT( verify_graph_edges(); )
2264 // If you have too many nodes, or if matching has failed, bail out
2265 check_node_count(0, "out of nodes matching instructions");
2266 if (failing()) {
2267 return;
2268 }
2270 // Build a proper-looking CFG
2271 PhaseCFG cfg(node_arena(), root(), matcher);
2272 _cfg = &cfg;
2273 {
2274 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2275 bool success = cfg.do_global_code_motion();
2276 if (!success) {
2277 return;
2278 }
2280 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2281 NOT_PRODUCT( verify_graph_edges(); )
2282 debug_only( cfg.verify(); )
2283 }
2285 PhaseChaitin regalloc(unique(), cfg, matcher);
2286 _regalloc = ®alloc;
2287 {
2288 TracePhase t2("regalloc", &_t_registerAllocation, true);
2289 // Perform register allocation. After Chaitin, use-def chains are
2290 // no longer accurate (at spill code) and so must be ignored.
2291 // Node->LRG->reg mappings are still accurate.
2292 _regalloc->Register_Allocate();
2294 // Bail out if the allocator builds too many nodes
2295 if (failing()) {
2296 return;
2297 }
2298 }
2300 // Prior to register allocation we kept empty basic blocks in case the
2301 // the allocator needed a place to spill. After register allocation we
2302 // are not adding any new instructions. If any basic block is empty, we
2303 // can now safely remove it.
2304 {
2305 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2306 cfg.remove_empty_blocks();
2307 if (do_freq_based_layout()) {
2308 PhaseBlockLayout layout(cfg);
2309 } else {
2310 cfg.set_loop_alignment();
2311 }
2312 cfg.fixup_flow();
2313 }
2315 // Apply peephole optimizations
2316 if( OptoPeephole ) {
2317 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2318 PhasePeephole peep( _regalloc, cfg);
2319 peep.do_transform();
2320 }
2322 // Do late expand if CPU requires this.
2323 if (Matcher::require_postalloc_expand) {
2324 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2325 cfg.postalloc_expand(_regalloc);
2326 }
2328 // Convert Nodes to instruction bits in a buffer
2329 {
2330 // %%%% workspace merge brought two timers together for one job
2331 TracePhase t2a("output", &_t_output, true);
2332 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2333 Output();
2334 }
2336 print_method(PHASE_FINAL_CODE);
2338 // He's dead, Jim.
2339 _cfg = (PhaseCFG*)0xdeadbeef;
2340 _regalloc = (PhaseChaitin*)0xdeadbeef;
2341 }
2344 //------------------------------dump_asm---------------------------------------
2345 // Dump formatted assembly
2346 #ifndef PRODUCT
2347 void Compile::dump_asm(int *pcs, uint pc_limit) {
2348 bool cut_short = false;
2349 tty->print_cr("#");
2350 tty->print("# "); _tf->dump(); tty->cr();
2351 tty->print_cr("#");
2353 // For all blocks
2354 int pc = 0x0; // Program counter
2355 char starts_bundle = ' ';
2356 _regalloc->dump_frame();
2358 Node *n = NULL;
2359 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2360 if (VMThread::should_terminate()) {
2361 cut_short = true;
2362 break;
2363 }
2364 Block* block = _cfg->get_block(i);
2365 if (block->is_connector() && !Verbose) {
2366 continue;
2367 }
2368 n = block->head();
2369 if (pcs && n->_idx < pc_limit) {
2370 tty->print("%3.3x ", pcs[n->_idx]);
2371 } else {
2372 tty->print(" ");
2373 }
2374 block->dump_head(_cfg);
2375 if (block->is_connector()) {
2376 tty->print_cr(" # Empty connector block");
2377 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2378 tty->print_cr(" # Block is sole successor of call");
2379 }
2381 // For all instructions
2382 Node *delay = NULL;
2383 for (uint j = 0; j < block->number_of_nodes(); j++) {
2384 if (VMThread::should_terminate()) {
2385 cut_short = true;
2386 break;
2387 }
2388 n = block->get_node(j);
2389 if (valid_bundle_info(n)) {
2390 Bundle* bundle = node_bundling(n);
2391 if (bundle->used_in_unconditional_delay()) {
2392 delay = n;
2393 continue;
2394 }
2395 if (bundle->starts_bundle()) {
2396 starts_bundle = '+';
2397 }
2398 }
2400 if (WizardMode) {
2401 n->dump();
2402 }
2404 if( !n->is_Region() && // Dont print in the Assembly
2405 !n->is_Phi() && // a few noisely useless nodes
2406 !n->is_Proj() &&
2407 !n->is_MachTemp() &&
2408 !n->is_SafePointScalarObject() &&
2409 !n->is_Catch() && // Would be nice to print exception table targets
2410 !n->is_MergeMem() && // Not very interesting
2411 !n->is_top() && // Debug info table constants
2412 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2413 ) {
2414 if (pcs && n->_idx < pc_limit)
2415 tty->print("%3.3x", pcs[n->_idx]);
2416 else
2417 tty->print(" ");
2418 tty->print(" %c ", starts_bundle);
2419 starts_bundle = ' ';
2420 tty->print("\t");
2421 n->format(_regalloc, tty);
2422 tty->cr();
2423 }
2425 // If we have an instruction with a delay slot, and have seen a delay,
2426 // then back up and print it
2427 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2428 assert(delay != NULL, "no unconditional delay instruction");
2429 if (WizardMode) delay->dump();
2431 if (node_bundling(delay)->starts_bundle())
2432 starts_bundle = '+';
2433 if (pcs && n->_idx < pc_limit)
2434 tty->print("%3.3x", pcs[n->_idx]);
2435 else
2436 tty->print(" ");
2437 tty->print(" %c ", starts_bundle);
2438 starts_bundle = ' ';
2439 tty->print("\t");
2440 delay->format(_regalloc, tty);
2441 tty->cr();
2442 delay = NULL;
2443 }
2445 // Dump the exception table as well
2446 if( n->is_Catch() && (Verbose || WizardMode) ) {
2447 // Print the exception table for this offset
2448 _handler_table.print_subtable_for(pc);
2449 }
2450 }
2452 if (pcs && n->_idx < pc_limit)
2453 tty->print_cr("%3.3x", pcs[n->_idx]);
2454 else
2455 tty->cr();
2457 assert(cut_short || delay == NULL, "no unconditional delay branch");
2459 } // End of per-block dump
2460 tty->cr();
2462 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2463 }
2464 #endif
2466 //------------------------------Final_Reshape_Counts---------------------------
2467 // This class defines counters to help identify when a method
2468 // may/must be executed using hardware with only 24-bit precision.
2469 struct Final_Reshape_Counts : public StackObj {
2470 int _call_count; // count non-inlined 'common' calls
2471 int _float_count; // count float ops requiring 24-bit precision
2472 int _double_count; // count double ops requiring more precision
2473 int _java_call_count; // count non-inlined 'java' calls
2474 int _inner_loop_count; // count loops which need alignment
2475 VectorSet _visited; // Visitation flags
2476 Node_List _tests; // Set of IfNodes & PCTableNodes
2478 Final_Reshape_Counts() :
2479 _call_count(0), _float_count(0), _double_count(0),
2480 _java_call_count(0), _inner_loop_count(0),
2481 _visited( Thread::current()->resource_area() ) { }
2483 void inc_call_count () { _call_count ++; }
2484 void inc_float_count () { _float_count ++; }
2485 void inc_double_count() { _double_count++; }
2486 void inc_java_call_count() { _java_call_count++; }
2487 void inc_inner_loop_count() { _inner_loop_count++; }
2489 int get_call_count () const { return _call_count ; }
2490 int get_float_count () const { return _float_count ; }
2491 int get_double_count() const { return _double_count; }
2492 int get_java_call_count() const { return _java_call_count; }
2493 int get_inner_loop_count() const { return _inner_loop_count; }
2494 };
2496 #ifdef ASSERT
2497 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2498 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2499 // Make sure the offset goes inside the instance layout.
2500 return k->contains_field_offset(tp->offset());
2501 // Note that OffsetBot and OffsetTop are very negative.
2502 }
2503 #endif
2505 // Eliminate trivially redundant StoreCMs and accumulate their
2506 // precedence edges.
2507 void Compile::eliminate_redundant_card_marks(Node* n) {
2508 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2509 if (n->in(MemNode::Address)->outcnt() > 1) {
2510 // There are multiple users of the same address so it might be
2511 // possible to eliminate some of the StoreCMs
2512 Node* mem = n->in(MemNode::Memory);
2513 Node* adr = n->in(MemNode::Address);
2514 Node* val = n->in(MemNode::ValueIn);
2515 Node* prev = n;
2516 bool done = false;
2517 // Walk the chain of StoreCMs eliminating ones that match. As
2518 // long as it's a chain of single users then the optimization is
2519 // safe. Eliminating partially redundant StoreCMs would require
2520 // cloning copies down the other paths.
2521 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2522 if (adr == mem->in(MemNode::Address) &&
2523 val == mem->in(MemNode::ValueIn)) {
2524 // redundant StoreCM
2525 if (mem->req() > MemNode::OopStore) {
2526 // Hasn't been processed by this code yet.
2527 n->add_prec(mem->in(MemNode::OopStore));
2528 } else {
2529 // Already converted to precedence edge
2530 for (uint i = mem->req(); i < mem->len(); i++) {
2531 // Accumulate any precedence edges
2532 if (mem->in(i) != NULL) {
2533 n->add_prec(mem->in(i));
2534 }
2535 }
2536 // Everything above this point has been processed.
2537 done = true;
2538 }
2539 // Eliminate the previous StoreCM
2540 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2541 assert(mem->outcnt() == 0, "should be dead");
2542 mem->disconnect_inputs(NULL, this);
2543 } else {
2544 prev = mem;
2545 }
2546 mem = prev->in(MemNode::Memory);
2547 }
2548 }
2549 }
2551 //------------------------------final_graph_reshaping_impl----------------------
2552 // Implement items 1-5 from final_graph_reshaping below.
2553 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2555 if ( n->outcnt() == 0 ) return; // dead node
2556 uint nop = n->Opcode();
2558 // Check for 2-input instruction with "last use" on right input.
2559 // Swap to left input. Implements item (2).
2560 if( n->req() == 3 && // two-input instruction
2561 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2562 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2563 n->in(2)->outcnt() == 1 &&// right use IS a last use
2564 !n->in(2)->is_Con() ) { // right use is not a constant
2565 // Check for commutative opcode
2566 switch( nop ) {
2567 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2568 case Op_MaxI: case Op_MinI:
2569 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2570 case Op_AndL: case Op_XorL: case Op_OrL:
2571 case Op_AndI: case Op_XorI: case Op_OrI: {
2572 // Move "last use" input to left by swapping inputs
2573 n->swap_edges(1, 2);
2574 break;
2575 }
2576 default:
2577 break;
2578 }
2579 }
2581 #ifdef ASSERT
2582 if( n->is_Mem() ) {
2583 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2584 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2585 // oop will be recorded in oop map if load crosses safepoint
2586 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2587 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2588 "raw memory operations should have control edge");
2589 }
2590 #endif
2591 // Count FPU ops and common calls, implements item (3)
2592 switch( nop ) {
2593 // Count all float operations that may use FPU
2594 case Op_AddF:
2595 case Op_SubF:
2596 case Op_MulF:
2597 case Op_DivF:
2598 case Op_NegF:
2599 case Op_ModF:
2600 case Op_ConvI2F:
2601 case Op_ConF:
2602 case Op_CmpF:
2603 case Op_CmpF3:
2604 // case Op_ConvL2F: // longs are split into 32-bit halves
2605 frc.inc_float_count();
2606 break;
2608 case Op_ConvF2D:
2609 case Op_ConvD2F:
2610 frc.inc_float_count();
2611 frc.inc_double_count();
2612 break;
2614 // Count all double operations that may use FPU
2615 case Op_AddD:
2616 case Op_SubD:
2617 case Op_MulD:
2618 case Op_DivD:
2619 case Op_NegD:
2620 case Op_ModD:
2621 case Op_ConvI2D:
2622 case Op_ConvD2I:
2623 // case Op_ConvL2D: // handled by leaf call
2624 // case Op_ConvD2L: // handled by leaf call
2625 case Op_ConD:
2626 case Op_CmpD:
2627 case Op_CmpD3:
2628 frc.inc_double_count();
2629 break;
2630 case Op_Opaque1: // Remove Opaque Nodes before matching
2631 case Op_Opaque2: // Remove Opaque Nodes before matching
2632 case Op_Opaque3:
2633 n->subsume_by(n->in(1), this);
2634 break;
2635 case Op_CallStaticJava:
2636 case Op_CallJava:
2637 case Op_CallDynamicJava:
2638 frc.inc_java_call_count(); // Count java call site;
2639 case Op_CallRuntime:
2640 case Op_CallLeaf:
2641 case Op_CallLeafNoFP: {
2642 assert( n->is_Call(), "" );
2643 CallNode *call = n->as_Call();
2644 // Count call sites where the FP mode bit would have to be flipped.
2645 // Do not count uncommon runtime calls:
2646 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2647 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2648 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2649 frc.inc_call_count(); // Count the call site
2650 } else { // See if uncommon argument is shared
2651 Node *n = call->in(TypeFunc::Parms);
2652 int nop = n->Opcode();
2653 // Clone shared simple arguments to uncommon calls, item (1).
2654 if( n->outcnt() > 1 &&
2655 !n->is_Proj() &&
2656 nop != Op_CreateEx &&
2657 nop != Op_CheckCastPP &&
2658 nop != Op_DecodeN &&
2659 nop != Op_DecodeNKlass &&
2660 !n->is_Mem() ) {
2661 Node *x = n->clone();
2662 call->set_req( TypeFunc::Parms, x );
2663 }
2664 }
2665 break;
2666 }
2668 case Op_StoreD:
2669 case Op_LoadD:
2670 case Op_LoadD_unaligned:
2671 frc.inc_double_count();
2672 goto handle_mem;
2673 case Op_StoreF:
2674 case Op_LoadF:
2675 frc.inc_float_count();
2676 goto handle_mem;
2678 case Op_StoreCM:
2679 {
2680 // Convert OopStore dependence into precedence edge
2681 Node* prec = n->in(MemNode::OopStore);
2682 n->del_req(MemNode::OopStore);
2683 n->add_prec(prec);
2684 eliminate_redundant_card_marks(n);
2685 }
2687 // fall through
2689 case Op_StoreB:
2690 case Op_StoreC:
2691 case Op_StorePConditional:
2692 case Op_StoreI:
2693 case Op_StoreL:
2694 case Op_StoreIConditional:
2695 case Op_StoreLConditional:
2696 case Op_CompareAndSwapI:
2697 case Op_CompareAndSwapL:
2698 case Op_CompareAndSwapP:
2699 case Op_CompareAndSwapN:
2700 case Op_GetAndAddI:
2701 case Op_GetAndAddL:
2702 case Op_GetAndSetI:
2703 case Op_GetAndSetL:
2704 case Op_GetAndSetP:
2705 case Op_GetAndSetN:
2706 case Op_StoreP:
2707 case Op_StoreN:
2708 case Op_StoreNKlass:
2709 case Op_LoadB:
2710 case Op_LoadUB:
2711 case Op_LoadUS:
2712 case Op_LoadI:
2713 case Op_LoadKlass:
2714 case Op_LoadNKlass:
2715 case Op_LoadL:
2716 case Op_LoadL_unaligned:
2717 case Op_LoadPLocked:
2718 case Op_LoadP:
2719 case Op_LoadN:
2720 case Op_LoadRange:
2721 case Op_LoadS: {
2722 handle_mem:
2723 #ifdef ASSERT
2724 if( VerifyOptoOopOffsets ) {
2725 assert( n->is_Mem(), "" );
2726 MemNode *mem = (MemNode*)n;
2727 // Check to see if address types have grounded out somehow.
2728 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2729 assert( !tp || oop_offset_is_sane(tp), "" );
2730 }
2731 #endif
2732 break;
2733 }
2735 case Op_AddP: { // Assert sane base pointers
2736 Node *addp = n->in(AddPNode::Address);
2737 assert( !addp->is_AddP() ||
2738 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2739 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2740 "Base pointers must match" );
2741 #ifdef _LP64
2742 if ((UseCompressedOops || UseCompressedClassPointers) &&
2743 addp->Opcode() == Op_ConP &&
2744 addp == n->in(AddPNode::Base) &&
2745 n->in(AddPNode::Offset)->is_Con()) {
2746 // Use addressing with narrow klass to load with offset on x86.
2747 // On sparc loading 32-bits constant and decoding it have less
2748 // instructions (4) then load 64-bits constant (7).
2749 // Do this transformation here since IGVN will convert ConN back to ConP.
2750 const Type* t = addp->bottom_type();
2751 if (t->isa_oopptr() || t->isa_klassptr()) {
2752 Node* nn = NULL;
2754 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2756 // Look for existing ConN node of the same exact type.
2757 Node* r = root();
2758 uint cnt = r->outcnt();
2759 for (uint i = 0; i < cnt; i++) {
2760 Node* m = r->raw_out(i);
2761 if (m!= NULL && m->Opcode() == op &&
2762 m->bottom_type()->make_ptr() == t) {
2763 nn = m;
2764 break;
2765 }
2766 }
2767 if (nn != NULL) {
2768 // Decode a narrow oop to match address
2769 // [R12 + narrow_oop_reg<<3 + offset]
2770 if (t->isa_oopptr()) {
2771 nn = new (this) DecodeNNode(nn, t);
2772 } else {
2773 nn = new (this) DecodeNKlassNode(nn, t);
2774 }
2775 n->set_req(AddPNode::Base, nn);
2776 n->set_req(AddPNode::Address, nn);
2777 if (addp->outcnt() == 0) {
2778 addp->disconnect_inputs(NULL, this);
2779 }
2780 }
2781 }
2782 }
2783 #endif
2784 break;
2785 }
2787 #ifdef _LP64
2788 case Op_CastPP:
2789 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2790 Node* in1 = n->in(1);
2791 const Type* t = n->bottom_type();
2792 Node* new_in1 = in1->clone();
2793 new_in1->as_DecodeN()->set_type(t);
2795 if (!Matcher::narrow_oop_use_complex_address()) {
2796 //
2797 // x86, ARM and friends can handle 2 adds in addressing mode
2798 // and Matcher can fold a DecodeN node into address by using
2799 // a narrow oop directly and do implicit NULL check in address:
2800 //
2801 // [R12 + narrow_oop_reg<<3 + offset]
2802 // NullCheck narrow_oop_reg
2803 //
2804 // On other platforms (Sparc) we have to keep new DecodeN node and
2805 // use it to do implicit NULL check in address:
2806 //
2807 // decode_not_null narrow_oop_reg, base_reg
2808 // [base_reg + offset]
2809 // NullCheck base_reg
2810 //
2811 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2812 // to keep the information to which NULL check the new DecodeN node
2813 // corresponds to use it as value in implicit_null_check().
2814 //
2815 new_in1->set_req(0, n->in(0));
2816 }
2818 n->subsume_by(new_in1, this);
2819 if (in1->outcnt() == 0) {
2820 in1->disconnect_inputs(NULL, this);
2821 }
2822 }
2823 break;
2825 case Op_CmpP:
2826 // Do this transformation here to preserve CmpPNode::sub() and
2827 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2828 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2829 Node* in1 = n->in(1);
2830 Node* in2 = n->in(2);
2831 if (!in1->is_DecodeNarrowPtr()) {
2832 in2 = in1;
2833 in1 = n->in(2);
2834 }
2835 assert(in1->is_DecodeNarrowPtr(), "sanity");
2837 Node* new_in2 = NULL;
2838 if (in2->is_DecodeNarrowPtr()) {
2839 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2840 new_in2 = in2->in(1);
2841 } else if (in2->Opcode() == Op_ConP) {
2842 const Type* t = in2->bottom_type();
2843 if (t == TypePtr::NULL_PTR) {
2844 assert(in1->is_DecodeN(), "compare klass to null?");
2845 // Don't convert CmpP null check into CmpN if compressed
2846 // oops implicit null check is not generated.
2847 // This will allow to generate normal oop implicit null check.
2848 if (Matcher::gen_narrow_oop_implicit_null_checks())
2849 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2850 //
2851 // This transformation together with CastPP transformation above
2852 // will generated code for implicit NULL checks for compressed oops.
2853 //
2854 // The original code after Optimize()
2855 //
2856 // LoadN memory, narrow_oop_reg
2857 // decode narrow_oop_reg, base_reg
2858 // CmpP base_reg, NULL
2859 // CastPP base_reg // NotNull
2860 // Load [base_reg + offset], val_reg
2861 //
2862 // after these transformations will be
2863 //
2864 // LoadN memory, narrow_oop_reg
2865 // CmpN narrow_oop_reg, NULL
2866 // decode_not_null narrow_oop_reg, base_reg
2867 // Load [base_reg + offset], val_reg
2868 //
2869 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2870 // since narrow oops can be used in debug info now (see the code in
2871 // final_graph_reshaping_walk()).
2872 //
2873 // At the end the code will be matched to
2874 // on x86:
2875 //
2876 // Load_narrow_oop memory, narrow_oop_reg
2877 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2878 // NullCheck narrow_oop_reg
2879 //
2880 // and on sparc:
2881 //
2882 // Load_narrow_oop memory, narrow_oop_reg
2883 // decode_not_null narrow_oop_reg, base_reg
2884 // Load [base_reg + offset], val_reg
2885 // NullCheck base_reg
2886 //
2887 } else if (t->isa_oopptr()) {
2888 new_in2 = ConNode::make(this, t->make_narrowoop());
2889 } else if (t->isa_klassptr()) {
2890 new_in2 = ConNode::make(this, t->make_narrowklass());
2891 }
2892 }
2893 if (new_in2 != NULL) {
2894 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2895 n->subsume_by(cmpN, this);
2896 if (in1->outcnt() == 0) {
2897 in1->disconnect_inputs(NULL, this);
2898 }
2899 if (in2->outcnt() == 0) {
2900 in2->disconnect_inputs(NULL, this);
2901 }
2902 }
2903 }
2904 break;
2906 case Op_DecodeN:
2907 case Op_DecodeNKlass:
2908 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2909 // DecodeN could be pinned when it can't be fold into
2910 // an address expression, see the code for Op_CastPP above.
2911 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2912 break;
2914 case Op_EncodeP:
2915 case Op_EncodePKlass: {
2916 Node* in1 = n->in(1);
2917 if (in1->is_DecodeNarrowPtr()) {
2918 n->subsume_by(in1->in(1), this);
2919 } else if (in1->Opcode() == Op_ConP) {
2920 const Type* t = in1->bottom_type();
2921 if (t == TypePtr::NULL_PTR) {
2922 assert(t->isa_oopptr(), "null klass?");
2923 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2924 } else if (t->isa_oopptr()) {
2925 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2926 } else if (t->isa_klassptr()) {
2927 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2928 }
2929 }
2930 if (in1->outcnt() == 0) {
2931 in1->disconnect_inputs(NULL, this);
2932 }
2933 break;
2934 }
2936 case Op_Proj: {
2937 if (OptimizeStringConcat) {
2938 ProjNode* p = n->as_Proj();
2939 if (p->_is_io_use) {
2940 // Separate projections were used for the exception path which
2941 // are normally removed by a late inline. If it wasn't inlined
2942 // then they will hang around and should just be replaced with
2943 // the original one.
2944 Node* proj = NULL;
2945 // Replace with just one
2946 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2947 Node *use = i.get();
2948 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2949 proj = use;
2950 break;
2951 }
2952 }
2953 assert(proj != NULL, "must be found");
2954 p->subsume_by(proj, this);
2955 }
2956 }
2957 break;
2958 }
2960 case Op_Phi:
2961 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
2962 // The EncodeP optimization may create Phi with the same edges
2963 // for all paths. It is not handled well by Register Allocator.
2964 Node* unique_in = n->in(1);
2965 assert(unique_in != NULL, "");
2966 uint cnt = n->req();
2967 for (uint i = 2; i < cnt; i++) {
2968 Node* m = n->in(i);
2969 assert(m != NULL, "");
2970 if (unique_in != m)
2971 unique_in = NULL;
2972 }
2973 if (unique_in != NULL) {
2974 n->subsume_by(unique_in, this);
2975 }
2976 }
2977 break;
2979 #endif
2981 case Op_ModI:
2982 if (UseDivMod) {
2983 // Check if a%b and a/b both exist
2984 Node* d = n->find_similar(Op_DivI);
2985 if (d) {
2986 // Replace them with a fused divmod if supported
2987 if (Matcher::has_match_rule(Op_DivModI)) {
2988 DivModINode* divmod = DivModINode::make(this, n);
2989 d->subsume_by(divmod->div_proj(), this);
2990 n->subsume_by(divmod->mod_proj(), this);
2991 } else {
2992 // replace a%b with a-((a/b)*b)
2993 Node* mult = new (this) MulINode(d, d->in(2));
2994 Node* sub = new (this) SubINode(d->in(1), mult);
2995 n->subsume_by(sub, this);
2996 }
2997 }
2998 }
2999 break;
3001 case Op_ModL:
3002 if (UseDivMod) {
3003 // Check if a%b and a/b both exist
3004 Node* d = n->find_similar(Op_DivL);
3005 if (d) {
3006 // Replace them with a fused divmod if supported
3007 if (Matcher::has_match_rule(Op_DivModL)) {
3008 DivModLNode* divmod = DivModLNode::make(this, n);
3009 d->subsume_by(divmod->div_proj(), this);
3010 n->subsume_by(divmod->mod_proj(), this);
3011 } else {
3012 // replace a%b with a-((a/b)*b)
3013 Node* mult = new (this) MulLNode(d, d->in(2));
3014 Node* sub = new (this) SubLNode(d->in(1), mult);
3015 n->subsume_by(sub, this);
3016 }
3017 }
3018 }
3019 break;
3021 case Op_LoadVector:
3022 case Op_StoreVector:
3023 break;
3025 case Op_PackB:
3026 case Op_PackS:
3027 case Op_PackI:
3028 case Op_PackF:
3029 case Op_PackL:
3030 case Op_PackD:
3031 if (n->req()-1 > 2) {
3032 // Replace many operand PackNodes with a binary tree for matching
3033 PackNode* p = (PackNode*) n;
3034 Node* btp = p->binary_tree_pack(this, 1, n->req());
3035 n->subsume_by(btp, this);
3036 }
3037 break;
3038 case Op_Loop:
3039 case Op_CountedLoop:
3040 if (n->as_Loop()->is_inner_loop()) {
3041 frc.inc_inner_loop_count();
3042 }
3043 break;
3044 case Op_LShiftI:
3045 case Op_RShiftI:
3046 case Op_URShiftI:
3047 case Op_LShiftL:
3048 case Op_RShiftL:
3049 case Op_URShiftL:
3050 if (Matcher::need_masked_shift_count) {
3051 // The cpu's shift instructions don't restrict the count to the
3052 // lower 5/6 bits. We need to do the masking ourselves.
3053 Node* in2 = n->in(2);
3054 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3055 const TypeInt* t = in2->find_int_type();
3056 if (t != NULL && t->is_con()) {
3057 juint shift = t->get_con();
3058 if (shift > mask) { // Unsigned cmp
3059 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
3060 }
3061 } else {
3062 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3063 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3064 n->set_req(2, shift);
3065 }
3066 }
3067 if (in2->outcnt() == 0) { // Remove dead node
3068 in2->disconnect_inputs(NULL, this);
3069 }
3070 }
3071 break;
3072 case Op_MemBarStoreStore:
3073 case Op_MemBarRelease:
3074 // Break the link with AllocateNode: it is no longer useful and
3075 // confuses register allocation.
3076 if (n->req() > MemBarNode::Precedent) {
3077 n->set_req(MemBarNode::Precedent, top());
3078 }
3079 break;
3080 default:
3081 assert( !n->is_Call(), "" );
3082 assert( !n->is_Mem(), "" );
3083 break;
3084 }
3086 // Collect CFG split points
3087 if (n->is_MultiBranch())
3088 frc._tests.push(n);
3089 }
3091 //------------------------------final_graph_reshaping_walk---------------------
3092 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3093 // requires that the walk visits a node's inputs before visiting the node.
3094 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3095 ResourceArea *area = Thread::current()->resource_area();
3096 Unique_Node_List sfpt(area);
3098 frc._visited.set(root->_idx); // first, mark node as visited
3099 uint cnt = root->req();
3100 Node *n = root;
3101 uint i = 0;
3102 while (true) {
3103 if (i < cnt) {
3104 // Place all non-visited non-null inputs onto stack
3105 Node* m = n->in(i);
3106 ++i;
3107 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3108 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3109 // compute worst case interpreter size in case of a deoptimization
3110 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3112 sfpt.push(m);
3113 }
3114 cnt = m->req();
3115 nstack.push(n, i); // put on stack parent and next input's index
3116 n = m;
3117 i = 0;
3118 }
3119 } else {
3120 // Now do post-visit work
3121 final_graph_reshaping_impl( n, frc );
3122 if (nstack.is_empty())
3123 break; // finished
3124 n = nstack.node(); // Get node from stack
3125 cnt = n->req();
3126 i = nstack.index();
3127 nstack.pop(); // Shift to the next node on stack
3128 }
3129 }
3131 // Skip next transformation if compressed oops are not used.
3132 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3133 (!UseCompressedOops && !UseCompressedClassPointers))
3134 return;
3136 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3137 // It could be done for an uncommon traps or any safepoints/calls
3138 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3139 while (sfpt.size() > 0) {
3140 n = sfpt.pop();
3141 JVMState *jvms = n->as_SafePoint()->jvms();
3142 assert(jvms != NULL, "sanity");
3143 int start = jvms->debug_start();
3144 int end = n->req();
3145 bool is_uncommon = (n->is_CallStaticJava() &&
3146 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3147 for (int j = start; j < end; j++) {
3148 Node* in = n->in(j);
3149 if (in->is_DecodeNarrowPtr()) {
3150 bool safe_to_skip = true;
3151 if (!is_uncommon ) {
3152 // Is it safe to skip?
3153 for (uint i = 0; i < in->outcnt(); i++) {
3154 Node* u = in->raw_out(i);
3155 if (!u->is_SafePoint() ||
3156 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3157 safe_to_skip = false;
3158 }
3159 }
3160 }
3161 if (safe_to_skip) {
3162 n->set_req(j, in->in(1));
3163 }
3164 if (in->outcnt() == 0) {
3165 in->disconnect_inputs(NULL, this);
3166 }
3167 }
3168 }
3169 }
3170 }
3172 //------------------------------final_graph_reshaping--------------------------
3173 // Final Graph Reshaping.
3174 //
3175 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3176 // and not commoned up and forced early. Must come after regular
3177 // optimizations to avoid GVN undoing the cloning. Clone constant
3178 // inputs to Loop Phis; these will be split by the allocator anyways.
3179 // Remove Opaque nodes.
3180 // (2) Move last-uses by commutative operations to the left input to encourage
3181 // Intel update-in-place two-address operations and better register usage
3182 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3183 // calls canonicalizing them back.
3184 // (3) Count the number of double-precision FP ops, single-precision FP ops
3185 // and call sites. On Intel, we can get correct rounding either by
3186 // forcing singles to memory (requires extra stores and loads after each
3187 // FP bytecode) or we can set a rounding mode bit (requires setting and
3188 // clearing the mode bit around call sites). The mode bit is only used
3189 // if the relative frequency of single FP ops to calls is low enough.
3190 // This is a key transform for SPEC mpeg_audio.
3191 // (4) Detect infinite loops; blobs of code reachable from above but not
3192 // below. Several of the Code_Gen algorithms fail on such code shapes,
3193 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3194 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3195 // Detection is by looking for IfNodes where only 1 projection is
3196 // reachable from below or CatchNodes missing some targets.
3197 // (5) Assert for insane oop offsets in debug mode.
3199 bool Compile::final_graph_reshaping() {
3200 // an infinite loop may have been eliminated by the optimizer,
3201 // in which case the graph will be empty.
3202 if (root()->req() == 1) {
3203 record_method_not_compilable("trivial infinite loop");
3204 return true;
3205 }
3207 // Expensive nodes have their control input set to prevent the GVN
3208 // from freely commoning them. There's no GVN beyond this point so
3209 // no need to keep the control input. We want the expensive nodes to
3210 // be freely moved to the least frequent code path by gcm.
3211 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3212 for (int i = 0; i < expensive_count(); i++) {
3213 _expensive_nodes->at(i)->set_req(0, NULL);
3214 }
3216 Final_Reshape_Counts frc;
3218 // Visit everybody reachable!
3219 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3220 Node_Stack nstack(unique() >> 1);
3221 final_graph_reshaping_walk(nstack, root(), frc);
3223 // Check for unreachable (from below) code (i.e., infinite loops).
3224 for( uint i = 0; i < frc._tests.size(); i++ ) {
3225 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3226 // Get number of CFG targets.
3227 // Note that PCTables include exception targets after calls.
3228 uint required_outcnt = n->required_outcnt();
3229 if (n->outcnt() != required_outcnt) {
3230 // Check for a few special cases. Rethrow Nodes never take the
3231 // 'fall-thru' path, so expected kids is 1 less.
3232 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3233 if (n->in(0)->in(0)->is_Call()) {
3234 CallNode *call = n->in(0)->in(0)->as_Call();
3235 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3236 required_outcnt--; // Rethrow always has 1 less kid
3237 } else if (call->req() > TypeFunc::Parms &&
3238 call->is_CallDynamicJava()) {
3239 // Check for null receiver. In such case, the optimizer has
3240 // detected that the virtual call will always result in a null
3241 // pointer exception. The fall-through projection of this CatchNode
3242 // will not be populated.
3243 Node *arg0 = call->in(TypeFunc::Parms);
3244 if (arg0->is_Type() &&
3245 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3246 required_outcnt--;
3247 }
3248 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3249 call->req() > TypeFunc::Parms+1 &&
3250 call->is_CallStaticJava()) {
3251 // Check for negative array length. In such case, the optimizer has
3252 // detected that the allocation attempt will always result in an
3253 // exception. There is no fall-through projection of this CatchNode .
3254 Node *arg1 = call->in(TypeFunc::Parms+1);
3255 if (arg1->is_Type() &&
3256 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3257 required_outcnt--;
3258 }
3259 }
3260 }
3261 }
3262 // Recheck with a better notion of 'required_outcnt'
3263 if (n->outcnt() != required_outcnt) {
3264 record_method_not_compilable("malformed control flow");
3265 return true; // Not all targets reachable!
3266 }
3267 }
3268 // Check that I actually visited all kids. Unreached kids
3269 // must be infinite loops.
3270 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3271 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3272 record_method_not_compilable("infinite loop");
3273 return true; // Found unvisited kid; must be unreach
3274 }
3275 }
3277 // If original bytecodes contained a mixture of floats and doubles
3278 // check if the optimizer has made it homogenous, item (3).
3279 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3280 frc.get_float_count() > 32 &&
3281 frc.get_double_count() == 0 &&
3282 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3283 set_24_bit_selection_and_mode( false, true );
3284 }
3286 set_java_calls(frc.get_java_call_count());
3287 set_inner_loops(frc.get_inner_loop_count());
3289 // No infinite loops, no reason to bail out.
3290 return false;
3291 }
3293 //-----------------------------too_many_traps----------------------------------
3294 // Report if there are too many traps at the current method and bci.
3295 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3296 bool Compile::too_many_traps(ciMethod* method,
3297 int bci,
3298 Deoptimization::DeoptReason reason) {
3299 ciMethodData* md = method->method_data();
3300 if (md->is_empty()) {
3301 // Assume the trap has not occurred, or that it occurred only
3302 // because of a transient condition during start-up in the interpreter.
3303 return false;
3304 }
3305 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3306 if (md->has_trap_at(bci, m, reason) != 0) {
3307 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3308 // Also, if there are multiple reasons, or if there is no per-BCI record,
3309 // assume the worst.
3310 if (log())
3311 log()->elem("observe trap='%s' count='%d'",
3312 Deoptimization::trap_reason_name(reason),
3313 md->trap_count(reason));
3314 return true;
3315 } else {
3316 // Ignore method/bci and see if there have been too many globally.
3317 return too_many_traps(reason, md);
3318 }
3319 }
3321 // Less-accurate variant which does not require a method and bci.
3322 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3323 ciMethodData* logmd) {
3324 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3325 // Too many traps globally.
3326 // Note that we use cumulative trap_count, not just md->trap_count.
3327 if (log()) {
3328 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3329 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3330 Deoptimization::trap_reason_name(reason),
3331 mcount, trap_count(reason));
3332 }
3333 return true;
3334 } else {
3335 // The coast is clear.
3336 return false;
3337 }
3338 }
3340 //--------------------------too_many_recompiles--------------------------------
3341 // Report if there are too many recompiles at the current method and bci.
3342 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3343 // Is not eager to return true, since this will cause the compiler to use
3344 // Action_none for a trap point, to avoid too many recompilations.
3345 bool Compile::too_many_recompiles(ciMethod* method,
3346 int bci,
3347 Deoptimization::DeoptReason reason) {
3348 ciMethodData* md = method->method_data();
3349 if (md->is_empty()) {
3350 // Assume the trap has not occurred, or that it occurred only
3351 // because of a transient condition during start-up in the interpreter.
3352 return false;
3353 }
3354 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3355 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3356 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3357 Deoptimization::DeoptReason per_bc_reason
3358 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3359 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3360 if ((per_bc_reason == Deoptimization::Reason_none
3361 || md->has_trap_at(bci, m, reason) != 0)
3362 // The trap frequency measure we care about is the recompile count:
3363 && md->trap_recompiled_at(bci, m)
3364 && md->overflow_recompile_count() >= bc_cutoff) {
3365 // Do not emit a trap here if it has already caused recompilations.
3366 // Also, if there are multiple reasons, or if there is no per-BCI record,
3367 // assume the worst.
3368 if (log())
3369 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3370 Deoptimization::trap_reason_name(reason),
3371 md->trap_count(reason),
3372 md->overflow_recompile_count());
3373 return true;
3374 } else if (trap_count(reason) != 0
3375 && decompile_count() >= m_cutoff) {
3376 // Too many recompiles globally, and we have seen this sort of trap.
3377 // Use cumulative decompile_count, not just md->decompile_count.
3378 if (log())
3379 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3380 Deoptimization::trap_reason_name(reason),
3381 md->trap_count(reason), trap_count(reason),
3382 md->decompile_count(), decompile_count());
3383 return true;
3384 } else {
3385 // The coast is clear.
3386 return false;
3387 }
3388 }
3390 // Compute when not to trap. Used by matching trap based nodes and
3391 // NullCheck optimization.
3392 void Compile::set_allowed_deopt_reasons() {
3393 _allowed_reasons = 0;
3394 if (is_method_compilation()) {
3395 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3396 assert(rs < BitsPerInt, "recode bit map");
3397 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3398 _allowed_reasons |= nth_bit(rs);
3399 }
3400 }
3401 }
3402 }
3404 #ifndef PRODUCT
3405 //------------------------------verify_graph_edges---------------------------
3406 // Walk the Graph and verify that there is a one-to-one correspondence
3407 // between Use-Def edges and Def-Use edges in the graph.
3408 void Compile::verify_graph_edges(bool no_dead_code) {
3409 if (VerifyGraphEdges) {
3410 ResourceArea *area = Thread::current()->resource_area();
3411 Unique_Node_List visited(area);
3412 // Call recursive graph walk to check edges
3413 _root->verify_edges(visited);
3414 if (no_dead_code) {
3415 // Now make sure that no visited node is used by an unvisited node.
3416 bool dead_nodes = 0;
3417 Unique_Node_List checked(area);
3418 while (visited.size() > 0) {
3419 Node* n = visited.pop();
3420 checked.push(n);
3421 for (uint i = 0; i < n->outcnt(); i++) {
3422 Node* use = n->raw_out(i);
3423 if (checked.member(use)) continue; // already checked
3424 if (visited.member(use)) continue; // already in the graph
3425 if (use->is_Con()) continue; // a dead ConNode is OK
3426 // At this point, we have found a dead node which is DU-reachable.
3427 if (dead_nodes++ == 0)
3428 tty->print_cr("*** Dead nodes reachable via DU edges:");
3429 use->dump(2);
3430 tty->print_cr("---");
3431 checked.push(use); // No repeats; pretend it is now checked.
3432 }
3433 }
3434 assert(dead_nodes == 0, "using nodes must be reachable from root");
3435 }
3436 }
3437 }
3439 // Verify GC barriers consistency
3440 // Currently supported:
3441 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3442 void Compile::verify_barriers() {
3443 if (UseG1GC) {
3444 // Verify G1 pre-barriers
3445 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3447 ResourceArea *area = Thread::current()->resource_area();
3448 Unique_Node_List visited(area);
3449 Node_List worklist(area);
3450 // We're going to walk control flow backwards starting from the Root
3451 worklist.push(_root);
3452 while (worklist.size() > 0) {
3453 Node* x = worklist.pop();
3454 if (x == NULL || x == top()) continue;
3455 if (visited.member(x)) {
3456 continue;
3457 } else {
3458 visited.push(x);
3459 }
3461 if (x->is_Region()) {
3462 for (uint i = 1; i < x->req(); i++) {
3463 worklist.push(x->in(i));
3464 }
3465 } else {
3466 worklist.push(x->in(0));
3467 // We are looking for the pattern:
3468 // /->ThreadLocal
3469 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3470 // \->ConI(0)
3471 // We want to verify that the If and the LoadB have the same control
3472 // See GraphKit::g1_write_barrier_pre()
3473 if (x->is_If()) {
3474 IfNode *iff = x->as_If();
3475 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3476 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3477 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3478 && cmp->in(1)->is_Load()) {
3479 LoadNode* load = cmp->in(1)->as_Load();
3480 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3481 && load->in(2)->in(3)->is_Con()
3482 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3484 Node* if_ctrl = iff->in(0);
3485 Node* load_ctrl = load->in(0);
3487 if (if_ctrl != load_ctrl) {
3488 // Skip possible CProj->NeverBranch in infinite loops
3489 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3490 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3491 if_ctrl = if_ctrl->in(0)->in(0);
3492 }
3493 }
3494 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3495 }
3496 }
3497 }
3498 }
3499 }
3500 }
3501 }
3502 }
3504 #endif
3506 // The Compile object keeps track of failure reasons separately from the ciEnv.
3507 // This is required because there is not quite a 1-1 relation between the
3508 // ciEnv and its compilation task and the Compile object. Note that one
3509 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3510 // to backtrack and retry without subsuming loads. Other than this backtracking
3511 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3512 // by the logic in C2Compiler.
3513 void Compile::record_failure(const char* reason) {
3514 if (log() != NULL) {
3515 log()->elem("failure reason='%s' phase='compile'", reason);
3516 }
3517 if (_failure_reason == NULL) {
3518 // Record the first failure reason.
3519 _failure_reason = reason;
3520 }
3522 EventCompilerFailure event;
3523 if (event.should_commit()) {
3524 event.set_compileID(Compile::compile_id());
3525 event.set_failure(reason);
3526 event.commit();
3527 }
3529 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3530 C->print_method(PHASE_FAILURE);
3531 }
3532 _root = NULL; // flush the graph, too
3533 }
3535 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3536 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3537 _phase_name(name), _dolog(dolog)
3538 {
3539 if (dolog) {
3540 C = Compile::current();
3541 _log = C->log();
3542 } else {
3543 C = NULL;
3544 _log = NULL;
3545 }
3546 if (_log != NULL) {
3547 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3548 _log->stamp();
3549 _log->end_head();
3550 }
3551 }
3553 Compile::TracePhase::~TracePhase() {
3555 C = Compile::current();
3556 if (_dolog) {
3557 _log = C->log();
3558 } else {
3559 _log = NULL;
3560 }
3562 #ifdef ASSERT
3563 if (PrintIdealNodeCount) {
3564 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3565 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3566 }
3568 if (VerifyIdealNodeCount) {
3569 Compile::current()->print_missing_nodes();
3570 }
3571 #endif
3573 if (_log != NULL) {
3574 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3575 }
3576 }
3578 //=============================================================================
3579 // Two Constant's are equal when the type and the value are equal.
3580 bool Compile::Constant::operator==(const Constant& other) {
3581 if (type() != other.type() ) return false;
3582 if (can_be_reused() != other.can_be_reused()) return false;
3583 // For floating point values we compare the bit pattern.
3584 switch (type()) {
3585 case T_FLOAT: return (_v._value.i == other._v._value.i);
3586 case T_LONG:
3587 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3588 case T_OBJECT:
3589 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3590 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3591 case T_METADATA: return (_v._metadata == other._v._metadata);
3592 default: ShouldNotReachHere();
3593 }
3594 return false;
3595 }
3597 static int type_to_size_in_bytes(BasicType t) {
3598 switch (t) {
3599 case T_LONG: return sizeof(jlong );
3600 case T_FLOAT: return sizeof(jfloat );
3601 case T_DOUBLE: return sizeof(jdouble);
3602 case T_METADATA: return sizeof(Metadata*);
3603 // We use T_VOID as marker for jump-table entries (labels) which
3604 // need an internal word relocation.
3605 case T_VOID:
3606 case T_ADDRESS:
3607 case T_OBJECT: return sizeof(jobject);
3608 }
3610 ShouldNotReachHere();
3611 return -1;
3612 }
3614 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3615 // sort descending
3616 if (a->freq() > b->freq()) return -1;
3617 if (a->freq() < b->freq()) return 1;
3618 return 0;
3619 }
3621 void Compile::ConstantTable::calculate_offsets_and_size() {
3622 // First, sort the array by frequencies.
3623 _constants.sort(qsort_comparator);
3625 #ifdef ASSERT
3626 // Make sure all jump-table entries were sorted to the end of the
3627 // array (they have a negative frequency).
3628 bool found_void = false;
3629 for (int i = 0; i < _constants.length(); i++) {
3630 Constant con = _constants.at(i);
3631 if (con.type() == T_VOID)
3632 found_void = true; // jump-tables
3633 else
3634 assert(!found_void, "wrong sorting");
3635 }
3636 #endif
3638 int offset = 0;
3639 for (int i = 0; i < _constants.length(); i++) {
3640 Constant* con = _constants.adr_at(i);
3642 // Align offset for type.
3643 int typesize = type_to_size_in_bytes(con->type());
3644 offset = align_size_up(offset, typesize);
3645 con->set_offset(offset); // set constant's offset
3647 if (con->type() == T_VOID) {
3648 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3649 offset = offset + typesize * n->outcnt(); // expand jump-table
3650 } else {
3651 offset = offset + typesize;
3652 }
3653 }
3655 // Align size up to the next section start (which is insts; see
3656 // CodeBuffer::align_at_start).
3657 assert(_size == -1, "already set?");
3658 _size = align_size_up(offset, CodeEntryAlignment);
3659 }
3661 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3662 MacroAssembler _masm(&cb);
3663 for (int i = 0; i < _constants.length(); i++) {
3664 Constant con = _constants.at(i);
3665 address constant_addr;
3666 switch (con.type()) {
3667 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3668 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3669 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3670 case T_OBJECT: {
3671 jobject obj = con.get_jobject();
3672 int oop_index = _masm.oop_recorder()->find_index(obj);
3673 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3674 break;
3675 }
3676 case T_ADDRESS: {
3677 address addr = (address) con.get_jobject();
3678 constant_addr = _masm.address_constant(addr);
3679 break;
3680 }
3681 // We use T_VOID as marker for jump-table entries (labels) which
3682 // need an internal word relocation.
3683 case T_VOID: {
3684 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3685 // Fill the jump-table with a dummy word. The real value is
3686 // filled in later in fill_jump_table.
3687 address dummy = (address) n;
3688 constant_addr = _masm.address_constant(dummy);
3689 // Expand jump-table
3690 for (uint i = 1; i < n->outcnt(); i++) {
3691 address temp_addr = _masm.address_constant(dummy + i);
3692 assert(temp_addr, "consts section too small");
3693 }
3694 break;
3695 }
3696 case T_METADATA: {
3697 Metadata* obj = con.get_metadata();
3698 int metadata_index = _masm.oop_recorder()->find_index(obj);
3699 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3700 break;
3701 }
3702 default: ShouldNotReachHere();
3703 }
3704 assert(constant_addr, "consts section too small");
3705 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3706 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())));
3707 }
3708 }
3710 int Compile::ConstantTable::find_offset(Constant& con) const {
3711 int idx = _constants.find(con);
3712 assert(idx != -1, "constant must be in constant table");
3713 int offset = _constants.at(idx).offset();
3714 assert(offset != -1, "constant table not emitted yet?");
3715 return offset;
3716 }
3718 void Compile::ConstantTable::add(Constant& con) {
3719 if (con.can_be_reused()) {
3720 int idx = _constants.find(con);
3721 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3722 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3723 return;
3724 }
3725 }
3726 (void) _constants.append(con);
3727 }
3729 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3730 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3731 Constant con(type, value, b->_freq);
3732 add(con);
3733 return con;
3734 }
3736 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3737 Constant con(metadata);
3738 add(con);
3739 return con;
3740 }
3742 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3743 jvalue value;
3744 BasicType type = oper->type()->basic_type();
3745 switch (type) {
3746 case T_LONG: value.j = oper->constantL(); break;
3747 case T_FLOAT: value.f = oper->constantF(); break;
3748 case T_DOUBLE: value.d = oper->constantD(); break;
3749 case T_OBJECT:
3750 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3751 case T_METADATA: return add((Metadata*)oper->constant()); break;
3752 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3753 }
3754 return add(n, type, value);
3755 }
3757 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3758 jvalue value;
3759 // We can use the node pointer here to identify the right jump-table
3760 // as this method is called from Compile::Fill_buffer right before
3761 // the MachNodes are emitted and the jump-table is filled (means the
3762 // MachNode pointers do not change anymore).
3763 value.l = (jobject) n;
3764 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3765 add(con);
3766 return con;
3767 }
3769 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3770 // If called from Compile::scratch_emit_size do nothing.
3771 if (Compile::current()->in_scratch_emit_size()) return;
3773 assert(labels.is_nonempty(), "must be");
3774 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3776 // Since MachConstantNode::constant_offset() also contains
3777 // table_base_offset() we need to subtract the table_base_offset()
3778 // to get the plain offset into the constant table.
3779 int offset = n->constant_offset() - table_base_offset();
3781 MacroAssembler _masm(&cb);
3782 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3784 for (uint i = 0; i < n->outcnt(); i++) {
3785 address* constant_addr = &jump_table_base[i];
3786 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)));
3787 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3788 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3789 }
3790 }
3792 void Compile::dump_inlining() {
3793 if (print_inlining() || print_intrinsics()) {
3794 // Print inlining message for candidates that we couldn't inline
3795 // for lack of space or non constant receiver
3796 for (int i = 0; i < _late_inlines.length(); i++) {
3797 CallGenerator* cg = _late_inlines.at(i);
3798 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3799 }
3800 Unique_Node_List useful;
3801 useful.push(root());
3802 for (uint next = 0; next < useful.size(); ++next) {
3803 Node* n = useful.at(next);
3804 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3805 CallNode* call = n->as_Call();
3806 CallGenerator* cg = call->generator();
3807 cg->print_inlining_late("receiver not constant");
3808 }
3809 uint max = n->len();
3810 for ( uint i = 0; i < max; ++i ) {
3811 Node *m = n->in(i);
3812 if ( m == NULL ) continue;
3813 useful.push(m);
3814 }
3815 }
3816 for (int i = 0; i < _print_inlining_list->length(); i++) {
3817 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
3818 }
3819 }
3820 }
3822 // Dump inlining replay data to the stream.
3823 // Don't change thread state and acquire any locks.
3824 void Compile::dump_inline_data(outputStream* out) {
3825 InlineTree* inl_tree = ilt();
3826 if (inl_tree != NULL) {
3827 out->print(" inline %d", inl_tree->count());
3828 inl_tree->dump_replay_data(out);
3829 }
3830 }
3832 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3833 if (n1->Opcode() < n2->Opcode()) return -1;
3834 else if (n1->Opcode() > n2->Opcode()) return 1;
3836 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()));
3837 for (uint i = 1; i < n1->req(); i++) {
3838 if (n1->in(i) < n2->in(i)) return -1;
3839 else if (n1->in(i) > n2->in(i)) return 1;
3840 }
3842 return 0;
3843 }
3845 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3846 Node* n1 = *n1p;
3847 Node* n2 = *n2p;
3849 return cmp_expensive_nodes(n1, n2);
3850 }
3852 void Compile::sort_expensive_nodes() {
3853 if (!expensive_nodes_sorted()) {
3854 _expensive_nodes->sort(cmp_expensive_nodes);
3855 }
3856 }
3858 bool Compile::expensive_nodes_sorted() const {
3859 for (int i = 1; i < _expensive_nodes->length(); i++) {
3860 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3861 return false;
3862 }
3863 }
3864 return true;
3865 }
3867 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3868 if (_expensive_nodes->length() == 0) {
3869 return false;
3870 }
3872 assert(OptimizeExpensiveOps, "optimization off?");
3874 // Take this opportunity to remove dead nodes from the list
3875 int j = 0;
3876 for (int i = 0; i < _expensive_nodes->length(); i++) {
3877 Node* n = _expensive_nodes->at(i);
3878 if (!n->is_unreachable(igvn)) {
3879 assert(n->is_expensive(), "should be expensive");
3880 _expensive_nodes->at_put(j, n);
3881 j++;
3882 }
3883 }
3884 _expensive_nodes->trunc_to(j);
3886 // Then sort the list so that similar nodes are next to each other
3887 // and check for at least two nodes of identical kind with same data
3888 // inputs.
3889 sort_expensive_nodes();
3891 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3892 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3893 return true;
3894 }
3895 }
3897 return false;
3898 }
3900 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3901 if (_expensive_nodes->length() == 0) {
3902 return;
3903 }
3905 assert(OptimizeExpensiveOps, "optimization off?");
3907 // Sort to bring similar nodes next to each other and clear the
3908 // control input of nodes for which there's only a single copy.
3909 sort_expensive_nodes();
3911 int j = 0;
3912 int identical = 0;
3913 int i = 0;
3914 for (; i < _expensive_nodes->length()-1; i++) {
3915 assert(j <= i, "can't write beyond current index");
3916 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3917 identical++;
3918 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3919 continue;
3920 }
3921 if (identical > 0) {
3922 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3923 identical = 0;
3924 } else {
3925 Node* n = _expensive_nodes->at(i);
3926 igvn.hash_delete(n);
3927 n->set_req(0, NULL);
3928 igvn.hash_insert(n);
3929 }
3930 }
3931 if (identical > 0) {
3932 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3933 } else if (_expensive_nodes->length() >= 1) {
3934 Node* n = _expensive_nodes->at(i);
3935 igvn.hash_delete(n);
3936 n->set_req(0, NULL);
3937 igvn.hash_insert(n);
3938 }
3939 _expensive_nodes->trunc_to(j);
3940 }
3942 void Compile::add_expensive_node(Node * n) {
3943 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3944 assert(n->is_expensive(), "expensive nodes with non-null control here only");
3945 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
3946 if (OptimizeExpensiveOps) {
3947 _expensive_nodes->append(n);
3948 } else {
3949 // Clear control input and let IGVN optimize expensive nodes if
3950 // OptimizeExpensiveOps is off.
3951 n->set_req(0, NULL);
3952 }
3953 }
3955 /**
3956 * Remove the speculative part of types and clean up the graph
3957 */
3958 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
3959 if (UseTypeSpeculation) {
3960 Unique_Node_List worklist;
3961 worklist.push(root());
3962 int modified = 0;
3963 // Go over all type nodes that carry a speculative type, drop the
3964 // speculative part of the type and enqueue the node for an igvn
3965 // which may optimize it out.
3966 for (uint next = 0; next < worklist.size(); ++next) {
3967 Node *n = worklist.at(next);
3968 if (n->is_Type()) {
3969 TypeNode* tn = n->as_Type();
3970 const Type* t = tn->type();
3971 const Type* t_no_spec = t->remove_speculative();
3972 if (t_no_spec != t) {
3973 bool in_hash = igvn.hash_delete(n);
3974 assert(in_hash, "node should be in igvn hash table");
3975 tn->set_type(t_no_spec);
3976 igvn.hash_insert(n);
3977 igvn._worklist.push(n); // give it a chance to go away
3978 modified++;
3979 }
3980 }
3981 uint max = n->len();
3982 for( uint i = 0; i < max; ++i ) {
3983 Node *m = n->in(i);
3984 if (not_a_node(m)) continue;
3985 worklist.push(m);
3986 }
3987 }
3988 // Drop the speculative part of all types in the igvn's type table
3989 igvn.remove_speculative_types();
3990 if (modified > 0) {
3991 igvn.optimize();
3992 }
3993 #ifdef ASSERT
3994 // Verify that after the IGVN is over no speculative type has resurfaced
3995 worklist.clear();
3996 worklist.push(root());
3997 for (uint next = 0; next < worklist.size(); ++next) {
3998 Node *n = worklist.at(next);
3999 const Type* t = igvn.type_or_null(n);
4000 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4001 if (n->is_Type()) {
4002 t = n->as_Type()->type();
4003 assert(t == t->remove_speculative(), "no more speculative types");
4004 }
4005 uint max = n->len();
4006 for( uint i = 0; i < max; ++i ) {
4007 Node *m = n->in(i);
4008 if (not_a_node(m)) continue;
4009 worklist.push(m);
4010 }
4011 }
4012 igvn.check_no_speculative_types();
4013 #endif
4014 }
4015 }
4017 // Auxiliary method to support randomized stressing/fuzzing.
4018 //
4019 // This method can be called the arbitrary number of times, with current count
4020 // as the argument. The logic allows selecting a single candidate from the
4021 // running list of candidates as follows:
4022 // int count = 0;
4023 // Cand* selected = null;
4024 // while(cand = cand->next()) {
4025 // if (randomized_select(++count)) {
4026 // selected = cand;
4027 // }
4028 // }
4029 //
4030 // Including count equalizes the chances any candidate is "selected".
4031 // This is useful when we don't have the complete list of candidates to choose
4032 // from uniformly. In this case, we need to adjust the randomicity of the
4033 // selection, or else we will end up biasing the selection towards the latter
4034 // candidates.
4035 //
4036 // Quick back-envelope calculation shows that for the list of n candidates
4037 // the equal probability for the candidate to persist as "best" can be
4038 // achieved by replacing it with "next" k-th candidate with the probability
4039 // of 1/k. It can be easily shown that by the end of the run, the
4040 // probability for any candidate is converged to 1/n, thus giving the
4041 // uniform distribution among all the candidates.
4042 //
4043 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4044 #define RANDOMIZED_DOMAIN_POW 29
4045 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4046 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4047 bool Compile::randomized_select(int count) {
4048 assert(count > 0, "only positive");
4049 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4050 }