Thu, 24 May 2018 17:06:56 +0800
Merge
1 /*
2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 /*
26 * This file has been modified by Loongson Technology in 2015. These
27 * modifications are Copyright (c) 2015 Loongson Technology, and are made
28 * available on the same license terms set forth above.
29 */
31 #include "precompiled.hpp"
32 #include "asm/macroAssembler.hpp"
33 #include "asm/macroAssembler.inline.hpp"
34 #include "ci/ciReplay.hpp"
35 #include "classfile/systemDictionary.hpp"
36 #include "code/exceptionHandlerTable.hpp"
37 #include "code/nmethod.hpp"
38 #include "compiler/compileLog.hpp"
39 #include "compiler/disassembler.hpp"
40 #include "compiler/oopMap.hpp"
41 #include "opto/addnode.hpp"
42 #include "opto/block.hpp"
43 #include "opto/c2compiler.hpp"
44 #include "opto/callGenerator.hpp"
45 #include "opto/callnode.hpp"
46 #include "opto/cfgnode.hpp"
47 #include "opto/chaitin.hpp"
48 #include "opto/compile.hpp"
49 #include "opto/connode.hpp"
50 #include "opto/divnode.hpp"
51 #include "opto/escape.hpp"
52 #include "opto/idealGraphPrinter.hpp"
53 #include "opto/loopnode.hpp"
54 #include "opto/machnode.hpp"
55 #include "opto/macro.hpp"
56 #include "opto/matcher.hpp"
57 #include "opto/mathexactnode.hpp"
58 #include "opto/memnode.hpp"
59 #include "opto/mulnode.hpp"
60 #include "opto/node.hpp"
61 #include "opto/opcodes.hpp"
62 #include "opto/output.hpp"
63 #include "opto/parse.hpp"
64 #include "opto/phaseX.hpp"
65 #include "opto/rootnode.hpp"
66 #include "opto/runtime.hpp"
67 #include "opto/stringopts.hpp"
68 #include "opto/type.hpp"
69 #include "opto/vectornode.hpp"
70 #include "runtime/arguments.hpp"
71 #include "runtime/signature.hpp"
72 #include "runtime/stubRoutines.hpp"
73 #include "runtime/timer.hpp"
74 #include "trace/tracing.hpp"
75 #include "utilities/copy.hpp"
76 #if defined AD_MD_HPP
77 # include AD_MD_HPP
78 #elif defined TARGET_ARCH_MODEL_x86_32
79 # include "adfiles/ad_x86_32.hpp"
80 #elif defined TARGET_ARCH_MODEL_x86_64
81 # include "adfiles/ad_x86_64.hpp"
82 #elif defined TARGET_ARCH_MODEL_sparc
83 # include "adfiles/ad_sparc.hpp"
84 #elif defined TARGET_ARCH_MODEL_zero
85 # include "adfiles/ad_zero.hpp"
86 #elif defined TARGET_ARCH_MODEL_ppc_64
87 # include "adfiles/ad_ppc_64.hpp"
88 #elif defined TARGET_ARCH_MODEL_mips_64
89 # include "adfiles/ad_mips_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 = live_nodes();
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 CastII nodes with range check dependency
424 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
425 Node* cast = range_check_cast_node(i);
426 if (!useful.member(cast)) {
427 remove_range_check_cast(cast);
428 }
429 }
430 // Remove useless expensive node
431 for (int i = C->expensive_count()-1; i >= 0; i--) {
432 Node* n = C->expensive_node(i);
433 if (!useful.member(n)) {
434 remove_expensive_node(n);
435 }
436 }
437 // clean up the late inline lists
438 remove_useless_late_inlines(&_string_late_inlines, useful);
439 remove_useless_late_inlines(&_boxing_late_inlines, useful);
440 remove_useless_late_inlines(&_late_inlines, useful);
441 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
442 }
444 //------------------------------frame_size_in_words-----------------------------
445 // frame_slots in units of words
446 int Compile::frame_size_in_words() const {
447 // shift is 0 in LP32 and 1 in LP64
448 const int shift = (LogBytesPerWord - LogBytesPerInt);
449 int words = _frame_slots >> shift;
450 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
451 return words;
452 }
454 // To bang the stack of this compiled method we use the stack size
455 // that the interpreter would need in case of a deoptimization. This
456 // removes the need to bang the stack in the deoptimization blob which
457 // in turn simplifies stack overflow handling.
458 int Compile::bang_size_in_bytes() const {
459 return MAX2(_interpreter_frame_size, frame_size_in_bytes());
460 }
462 // ============================================================================
463 //------------------------------CompileWrapper---------------------------------
464 class CompileWrapper : public StackObj {
465 Compile *const _compile;
466 public:
467 CompileWrapper(Compile* compile);
469 ~CompileWrapper();
470 };
472 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
473 // the Compile* pointer is stored in the current ciEnv:
474 ciEnv* env = compile->env();
475 assert(env == ciEnv::current(), "must already be a ciEnv active");
476 assert(env->compiler_data() == NULL, "compile already active?");
477 env->set_compiler_data(compile);
478 assert(compile == Compile::current(), "sanity");
480 compile->set_type_dict(NULL);
481 compile->set_type_hwm(NULL);
482 compile->set_type_last_size(0);
483 compile->set_last_tf(NULL, NULL);
484 compile->set_indexSet_arena(NULL);
485 compile->set_indexSet_free_block_list(NULL);
486 compile->init_type_arena();
487 Type::Initialize(compile);
488 _compile->set_scratch_buffer_blob(NULL);
489 _compile->begin_method();
490 }
491 CompileWrapper::~CompileWrapper() {
492 _compile->end_method();
493 if (_compile->scratch_buffer_blob() != NULL)
494 BufferBlob::free(_compile->scratch_buffer_blob());
495 _compile->env()->set_compiler_data(NULL);
496 }
499 //----------------------------print_compile_messages---------------------------
500 void Compile::print_compile_messages() {
501 #ifndef PRODUCT
502 // Check if recompiling
503 if (_subsume_loads == false && PrintOpto) {
504 // Recompiling without allowing machine instructions to subsume loads
505 tty->print_cr("*********************************************************");
506 tty->print_cr("** Bailout: Recompile without subsuming loads **");
507 tty->print_cr("*********************************************************");
508 }
509 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
510 // Recompiling without escape analysis
511 tty->print_cr("*********************************************************");
512 tty->print_cr("** Bailout: Recompile without escape analysis **");
513 tty->print_cr("*********************************************************");
514 }
515 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
516 // Recompiling without boxing elimination
517 tty->print_cr("*********************************************************");
518 tty->print_cr("** Bailout: Recompile without boxing elimination **");
519 tty->print_cr("*********************************************************");
520 }
521 if (env()->break_at_compile()) {
522 // Open the debugger when compiling this method.
523 tty->print("### Breaking when compiling: ");
524 method()->print_short_name();
525 tty->cr();
526 BREAKPOINT;
527 }
529 if( PrintOpto ) {
530 if (is_osr_compilation()) {
531 tty->print("[OSR]%3d", _compile_id);
532 } else {
533 tty->print("%3d", _compile_id);
534 }
535 }
536 #endif
537 }
540 //-----------------------init_scratch_buffer_blob------------------------------
541 // Construct a temporary BufferBlob and cache it for this compile.
542 void Compile::init_scratch_buffer_blob(int const_size) {
543 // If there is already a scratch buffer blob allocated and the
544 // constant section is big enough, use it. Otherwise free the
545 // current and allocate a new one.
546 BufferBlob* blob = scratch_buffer_blob();
547 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
548 // Use the current blob.
549 } else {
550 if (blob != NULL) {
551 BufferBlob::free(blob);
552 }
554 ResourceMark rm;
555 _scratch_const_size = const_size;
556 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
557 blob = BufferBlob::create("Compile::scratch_buffer", size);
558 // Record the buffer blob for next time.
559 set_scratch_buffer_blob(blob);
560 // Have we run out of code space?
561 if (scratch_buffer_blob() == NULL) {
562 // Let CompilerBroker disable further compilations.
563 record_failure("Not enough space for scratch buffer in CodeCache");
564 return;
565 }
566 }
568 // Initialize the relocation buffers
569 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
570 set_scratch_locs_memory(locs_buf);
571 }
574 //-----------------------scratch_emit_size-------------------------------------
575 // Helper function that computes size by emitting code
576 uint Compile::scratch_emit_size(const Node* n) {
577 // Start scratch_emit_size section.
578 set_in_scratch_emit_size(true);
580 // Emit into a trash buffer and count bytes emitted.
581 // This is a pretty expensive way to compute a size,
582 // but it works well enough if seldom used.
583 // All common fixed-size instructions are given a size
584 // method by the AD file.
585 // Note that the scratch buffer blob and locs memory are
586 // allocated at the beginning of the compile task, and
587 // may be shared by several calls to scratch_emit_size.
588 // The allocation of the scratch buffer blob is particularly
589 // expensive, since it has to grab the code cache lock.
590 BufferBlob* blob = this->scratch_buffer_blob();
591 assert(blob != NULL, "Initialize BufferBlob at start");
592 assert(blob->size() > MAX_inst_size, "sanity");
593 relocInfo* locs_buf = scratch_locs_memory();
594 address blob_begin = blob->content_begin();
595 address blob_end = (address)locs_buf;
596 assert(blob->content_contains(blob_end), "sanity");
597 CodeBuffer buf(blob_begin, blob_end - blob_begin);
598 buf.initialize_consts_size(_scratch_const_size);
599 buf.initialize_stubs_size(MAX_stubs_size);
600 assert(locs_buf != NULL, "sanity");
601 int lsize = MAX_locs_size / 3;
602 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
603 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
604 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
606 // Do the emission.
608 Label fakeL; // Fake label for branch instructions.
609 Label* saveL = NULL;
610 uint save_bnum = 0;
611 bool is_branch = n->is_MachBranch();
612 if (is_branch) {
613 MacroAssembler masm(&buf);
614 masm.bind(fakeL);
615 n->as_MachBranch()->save_label(&saveL, &save_bnum);
616 n->as_MachBranch()->label_set(&fakeL, 0);
617 }
618 n->emit(buf, this->regalloc());
620 // Emitting into the scratch buffer should not fail
621 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason()));
623 if (is_branch) // Restore label.
624 n->as_MachBranch()->label_set(saveL, save_bnum);
626 // End scratch_emit_size section.
627 set_in_scratch_emit_size(false);
629 return buf.insts_size();
630 }
633 // ============================================================================
634 //------------------------------Compile standard-------------------------------
635 debug_only( int Compile::_debug_idx = 100000; )
637 // Compile a method. entry_bci is -1 for normal compilations and indicates
638 // the continuation bci for on stack replacement.
641 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
642 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
643 : Phase(Compiler),
644 _env(ci_env),
645 _log(ci_env->log()),
646 _compile_id(ci_env->compile_id()),
647 _save_argument_registers(false),
648 _stub_name(NULL),
649 _stub_function(NULL),
650 _stub_entry_point(NULL),
651 _method(target),
652 _entry_bci(osr_bci),
653 _initial_gvn(NULL),
654 _for_igvn(NULL),
655 _warm_calls(NULL),
656 _subsume_loads(subsume_loads),
657 _do_escape_analysis(do_escape_analysis),
658 _eliminate_boxing(eliminate_boxing),
659 _failure_reason(NULL),
660 _code_buffer("Compile::Fill_buffer"),
661 _orig_pc_slot(0),
662 _orig_pc_slot_offset_in_bytes(0),
663 _has_method_handle_invokes(false),
664 _mach_constant_base_node(NULL),
665 _node_bundling_limit(0),
666 _node_bundling_base(NULL),
667 _java_calls(0),
668 _inner_loops(0),
669 _scratch_const_size(-1),
670 _in_scratch_emit_size(false),
671 _dead_node_list(comp_arena()),
672 _dead_node_count(0),
673 #ifndef PRODUCT
674 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
675 _in_dump_cnt(0),
676 _printer(IdealGraphPrinter::printer()),
677 #endif
678 _congraph(NULL),
679 _comp_arena(mtCompiler),
680 _node_arena(mtCompiler),
681 _old_arena(mtCompiler),
682 _Compile_types(mtCompiler),
683 _replay_inline_data(NULL),
684 _late_inlines(comp_arena(), 2, 0, NULL),
685 _string_late_inlines(comp_arena(), 2, 0, NULL),
686 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
687 _late_inlines_pos(0),
688 _number_of_mh_late_inlines(0),
689 _inlining_progress(false),
690 _inlining_incrementally(false),
691 _print_inlining_list(NULL),
692 _print_inlining_idx(0),
693 _interpreter_frame_size(0),
694 _max_node_limit(MaxNodeLimit) {
695 C = this;
697 CompileWrapper cw(this);
698 #ifndef PRODUCT
699 if (TimeCompiler2) {
700 tty->print(" ");
701 target->holder()->name()->print();
702 tty->print(".");
703 target->print_short_name();
704 tty->print(" ");
705 }
706 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
707 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
708 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
709 if (!print_opto_assembly) {
710 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
711 if (print_assembly && !Disassembler::can_decode()) {
712 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
713 print_opto_assembly = true;
714 }
715 }
716 set_print_assembly(print_opto_assembly);
717 set_parsed_irreducible_loop(false);
719 if (method()->has_option("ReplayInline")) {
720 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
721 }
722 #endif
723 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
724 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
725 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
727 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
728 // Make sure the method being compiled gets its own MDO,
729 // so we can at least track the decompile_count().
730 // Need MDO to record RTM code generation state.
731 method()->ensure_method_data();
732 }
734 Init(::AliasLevel);
737 print_compile_messages();
739 _ilt = InlineTree::build_inline_tree_root();
741 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
742 assert(num_alias_types() >= AliasIdxRaw, "");
744 #define MINIMUM_NODE_HASH 1023
745 // Node list that Iterative GVN will start with
746 Unique_Node_List for_igvn(comp_arena());
747 set_for_igvn(&for_igvn);
749 // GVN that will be run immediately on new nodes
750 uint estimated_size = method()->code_size()*4+64;
751 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
752 PhaseGVN gvn(node_arena(), estimated_size);
753 set_initial_gvn(&gvn);
755 if (print_inlining() || print_intrinsics()) {
756 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
757 }
758 { // Scope for timing the parser
759 TracePhase t3("parse", &_t_parser, true);
761 // Put top into the hash table ASAP.
762 initial_gvn()->transform_no_reclaim(top());
764 // Set up tf(), start(), and find a CallGenerator.
765 CallGenerator* cg = NULL;
766 if (is_osr_compilation()) {
767 const TypeTuple *domain = StartOSRNode::osr_domain();
768 const TypeTuple *range = TypeTuple::make_range(method()->signature());
769 init_tf(TypeFunc::make(domain, range));
770 StartNode* s = new (this) StartOSRNode(root(), domain);
771 initial_gvn()->set_type_bottom(s);
772 init_start(s);
773 cg = CallGenerator::for_osr(method(), entry_bci());
774 } else {
775 // Normal case.
776 init_tf(TypeFunc::make(method()));
777 StartNode* s = new (this) StartNode(root(), tf()->domain());
778 initial_gvn()->set_type_bottom(s);
779 init_start(s);
780 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
781 // With java.lang.ref.reference.get() we must go through the
782 // intrinsic when G1 is enabled - even when get() is the root
783 // method of the compile - so that, if necessary, the value in
784 // the referent field of the reference object gets recorded by
785 // the pre-barrier code.
786 // Specifically, if G1 is enabled, the value in the referent
787 // field is recorded by the G1 SATB pre barrier. This will
788 // result in the referent being marked live and the reference
789 // object removed from the list of discovered references during
790 // reference processing.
791 cg = find_intrinsic(method(), false);
792 }
793 if (cg == NULL) {
794 float past_uses = method()->interpreter_invocation_count();
795 float expected_uses = past_uses;
796 cg = CallGenerator::for_inline(method(), expected_uses);
797 }
798 }
799 if (failing()) return;
800 if (cg == NULL) {
801 record_method_not_compilable_all_tiers("cannot parse method");
802 return;
803 }
804 JVMState* jvms = build_start_state(start(), tf());
805 if ((jvms = cg->generate(jvms)) == NULL) {
806 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
807 record_method_not_compilable("method parse failed");
808 }
809 return;
810 }
811 GraphKit kit(jvms);
813 if (!kit.stopped()) {
814 // Accept return values, and transfer control we know not where.
815 // This is done by a special, unique ReturnNode bound to root.
816 return_values(kit.jvms());
817 }
819 if (kit.has_exceptions()) {
820 // Any exceptions that escape from this call must be rethrown
821 // to whatever caller is dynamically above us on the stack.
822 // This is done by a special, unique RethrowNode bound to root.
823 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
824 }
826 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
828 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
829 inline_string_calls(true);
830 }
832 if (failing()) return;
834 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
836 // Remove clutter produced by parsing.
837 if (!failing()) {
838 ResourceMark rm;
839 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
840 }
841 }
843 // Note: Large methods are capped off in do_one_bytecode().
844 if (failing()) return;
846 // After parsing, node notes are no longer automagic.
847 // They must be propagated by register_new_node_with_optimizer(),
848 // clone(), or the like.
849 set_default_node_notes(NULL);
851 for (;;) {
852 int successes = Inline_Warm();
853 if (failing()) return;
854 if (successes == 0) break;
855 }
857 // Drain the list.
858 Finish_Warm();
859 #ifndef PRODUCT
860 if (_printer) {
861 _printer->print_inlining(this);
862 }
863 #endif
865 if (failing()) return;
866 NOT_PRODUCT( verify_graph_edges(); )
868 // Now optimize
869 Optimize();
870 if (failing()) return;
871 NOT_PRODUCT( verify_graph_edges(); )
873 #ifndef PRODUCT
874 if (PrintIdeal) {
875 ttyLocker ttyl; // keep the following output all in one block
876 // This output goes directly to the tty, not the compiler log.
877 // To enable tools to match it up with the compilation activity,
878 // be sure to tag this tty output with the compile ID.
879 if (xtty != NULL) {
880 xtty->head("ideal compile_id='%d'%s", compile_id(),
881 is_osr_compilation() ? " compile_kind='osr'" :
882 "");
883 }
884 root()->dump(9999);
885 if (xtty != NULL) {
886 xtty->tail("ideal");
887 }
888 }
889 #endif
891 NOT_PRODUCT( verify_barriers(); )
893 // Dump compilation data to replay it.
894 if (method()->has_option("DumpReplay")) {
895 env()->dump_replay_data(_compile_id);
896 }
897 if (method()->has_option("DumpInline") && (ilt() != NULL)) {
898 env()->dump_inline_data(_compile_id);
899 }
901 // Now that we know the size of all the monitors we can add a fixed slot
902 // for the original deopt pc.
904 _orig_pc_slot = fixed_slots();
905 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
906 set_fixed_slots(next_slot);
908 // Compute when to use implicit null checks. Used by matching trap based
909 // nodes and NullCheck optimization.
910 set_allowed_deopt_reasons();
912 // Now generate code
913 Code_Gen();
914 if (failing()) return;
916 // Check if we want to skip execution of all compiled code.
917 {
918 #ifndef PRODUCT
919 if (OptoNoExecute) {
920 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
921 return;
922 }
923 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
924 #endif
926 if (is_osr_compilation()) {
927 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
928 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
929 } else {
930 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
931 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
932 }
934 env()->register_method(_method, _entry_bci,
935 &_code_offsets,
936 _orig_pc_slot_offset_in_bytes,
937 code_buffer(),
938 frame_size_in_words(), _oop_map_set,
939 &_handler_table, &_inc_table,
940 compiler,
941 env()->comp_level(),
942 has_unsafe_access(),
943 SharedRuntime::is_wide_vector(max_vector_size()),
944 rtm_state()
945 );
947 if (log() != NULL) // Print code cache state into compiler log
948 log()->code_cache_state();
949 }
950 }
952 //------------------------------Compile----------------------------------------
953 // Compile a runtime stub
954 Compile::Compile( ciEnv* ci_env,
955 TypeFunc_generator generator,
956 address stub_function,
957 const char *stub_name,
958 int is_fancy_jump,
959 bool pass_tls,
960 bool save_arg_registers,
961 bool return_pc )
962 : Phase(Compiler),
963 _env(ci_env),
964 _log(ci_env->log()),
965 _compile_id(0),
966 _save_argument_registers(save_arg_registers),
967 _method(NULL),
968 _stub_name(stub_name),
969 _stub_function(stub_function),
970 _stub_entry_point(NULL),
971 _entry_bci(InvocationEntryBci),
972 _initial_gvn(NULL),
973 _for_igvn(NULL),
974 _warm_calls(NULL),
975 _orig_pc_slot(0),
976 _orig_pc_slot_offset_in_bytes(0),
977 _subsume_loads(true),
978 _do_escape_analysis(false),
979 _eliminate_boxing(false),
980 _failure_reason(NULL),
981 _code_buffer("Compile::Fill_buffer"),
982 _has_method_handle_invokes(false),
983 _mach_constant_base_node(NULL),
984 _node_bundling_limit(0),
985 _node_bundling_base(NULL),
986 _java_calls(0),
987 _inner_loops(0),
988 #ifndef PRODUCT
989 _trace_opto_output(TraceOptoOutput),
990 _in_dump_cnt(0),
991 _printer(NULL),
992 #endif
993 _comp_arena(mtCompiler),
994 _node_arena(mtCompiler),
995 _old_arena(mtCompiler),
996 _Compile_types(mtCompiler),
997 _dead_node_list(comp_arena()),
998 _dead_node_count(0),
999 _congraph(NULL),
1000 _replay_inline_data(NULL),
1001 _number_of_mh_late_inlines(0),
1002 _inlining_progress(false),
1003 _inlining_incrementally(false),
1004 _print_inlining_list(NULL),
1005 _print_inlining_idx(0),
1006 _allowed_reasons(0),
1007 _interpreter_frame_size(0),
1008 _max_node_limit(MaxNodeLimit) {
1009 C = this;
1011 #ifndef PRODUCT
1012 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
1013 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
1014 set_print_assembly(PrintFrameConverterAssembly);
1015 set_parsed_irreducible_loop(false);
1016 #endif
1017 set_has_irreducible_loop(false); // no loops
1019 CompileWrapper cw(this);
1020 Init(/*AliasLevel=*/ 0);
1021 init_tf((*generator)());
1023 {
1024 // The following is a dummy for the sake of GraphKit::gen_stub
1025 Unique_Node_List for_igvn(comp_arena());
1026 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1027 PhaseGVN gvn(Thread::current()->resource_area(),255);
1028 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1029 gvn.transform_no_reclaim(top());
1031 GraphKit kit;
1032 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1033 }
1035 NOT_PRODUCT( verify_graph_edges(); )
1036 Code_Gen();
1037 if (failing()) return;
1040 // Entry point will be accessed using compile->stub_entry_point();
1041 if (code_buffer() == NULL) {
1042 Matcher::soft_match_failure();
1043 } else {
1044 if (PrintAssembly && (WizardMode || Verbose))
1045 tty->print_cr("### Stub::%s", stub_name);
1047 if (!failing()) {
1048 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1050 // Make the NMethod
1051 // For now we mark the frame as never safe for profile stackwalking
1052 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1053 code_buffer(),
1054 CodeOffsets::frame_never_safe,
1055 // _code_offsets.value(CodeOffsets::Frame_Complete),
1056 frame_size_in_words(),
1057 _oop_map_set,
1058 save_arg_registers);
1059 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1061 _stub_entry_point = rs->entry_point();
1062 }
1063 }
1064 }
1066 //------------------------------Init-------------------------------------------
1067 // Prepare for a single compilation
1068 void Compile::Init(int aliaslevel) {
1069 _unique = 0;
1070 _regalloc = NULL;
1072 _tf = NULL; // filled in later
1073 _top = NULL; // cached later
1074 _matcher = NULL; // filled in later
1075 _cfg = NULL; // filled in later
1077 set_24_bit_selection_and_mode(Use24BitFP, false);
1079 _node_note_array = NULL;
1080 _default_node_notes = NULL;
1082 _immutable_memory = NULL; // filled in at first inquiry
1084 // Globally visible Nodes
1085 // First set TOP to NULL to give safe behavior during creation of RootNode
1086 set_cached_top_node(NULL);
1087 set_root(new (this) RootNode());
1088 // Now that you have a Root to point to, create the real TOP
1089 set_cached_top_node( new (this) ConNode(Type::TOP) );
1090 set_recent_alloc(NULL, NULL);
1092 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1093 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1094 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1095 env()->set_dependencies(new Dependencies(env()));
1097 _fixed_slots = 0;
1098 set_has_split_ifs(false);
1099 set_has_loops(has_method() && method()->has_loops()); // first approximation
1100 set_has_stringbuilder(false);
1101 set_has_boxed_value(false);
1102 _trap_can_recompile = false; // no traps emitted yet
1103 _major_progress = true; // start out assuming good things will happen
1104 set_has_unsafe_access(false);
1105 set_max_vector_size(0);
1106 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1107 set_decompile_count(0);
1109 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1110 set_num_loop_opts(LoopOptsCount);
1111 set_do_inlining(Inline);
1112 set_max_inline_size(MaxInlineSize);
1113 set_freq_inline_size(FreqInlineSize);
1114 set_do_scheduling(OptoScheduling);
1115 set_do_count_invocations(false);
1116 set_do_method_data_update(false);
1117 set_rtm_state(NoRTM); // No RTM lock eliding by default
1118 method_has_option_value("MaxNodeLimit", _max_node_limit);
1119 #if INCLUDE_RTM_OPT
1120 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1121 int rtm_state = method()->method_data()->rtm_state();
1122 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1123 // Don't generate RTM lock eliding code.
1124 set_rtm_state(NoRTM);
1125 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1126 // Generate RTM lock eliding code without abort ratio calculation code.
1127 set_rtm_state(UseRTM);
1128 } else if (UseRTMDeopt) {
1129 // Generate RTM lock eliding code and include abort ratio calculation
1130 // code if UseRTMDeopt is on.
1131 set_rtm_state(ProfileRTM);
1132 }
1133 }
1134 #endif
1135 if (debug_info()->recording_non_safepoints()) {
1136 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1137 (comp_arena(), 8, 0, NULL));
1138 set_default_node_notes(Node_Notes::make(this));
1139 }
1141 // // -- Initialize types before each compile --
1142 // // Update cached type information
1143 // if( _method && _method->constants() )
1144 // Type::update_loaded_types(_method, _method->constants());
1146 // Init alias_type map.
1147 if (!_do_escape_analysis && aliaslevel == 3)
1148 aliaslevel = 2; // No unique types without escape analysis
1149 _AliasLevel = aliaslevel;
1150 const int grow_ats = 16;
1151 _max_alias_types = grow_ats;
1152 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1153 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1154 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1155 {
1156 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1157 }
1158 // Initialize the first few types.
1159 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1160 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1161 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1162 _num_alias_types = AliasIdxRaw+1;
1163 // Zero out the alias type cache.
1164 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1165 // A NULL adr_type hits in the cache right away. Preload the right answer.
1166 probe_alias_cache(NULL)->_index = AliasIdxTop;
1168 _intrinsics = NULL;
1169 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1170 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1171 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1172 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1173 register_library_intrinsics();
1174 }
1176 //---------------------------init_start----------------------------------------
1177 // Install the StartNode on this compile object.
1178 void Compile::init_start(StartNode* s) {
1179 if (failing())
1180 return; // already failing
1181 assert(s == start(), "");
1182 }
1184 StartNode* Compile::start() const {
1185 assert(!failing(), "");
1186 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1187 Node* start = root()->fast_out(i);
1188 if( start->is_Start() )
1189 return start->as_Start();
1190 }
1191 fatal("Did not find Start node!");
1192 return NULL;
1193 }
1195 //-------------------------------immutable_memory-------------------------------------
1196 // Access immutable memory
1197 Node* Compile::immutable_memory() {
1198 if (_immutable_memory != NULL) {
1199 return _immutable_memory;
1200 }
1201 StartNode* s = start();
1202 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1203 Node *p = s->fast_out(i);
1204 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1205 _immutable_memory = p;
1206 return _immutable_memory;
1207 }
1208 }
1209 ShouldNotReachHere();
1210 return NULL;
1211 }
1213 //----------------------set_cached_top_node------------------------------------
1214 // Install the cached top node, and make sure Node::is_top works correctly.
1215 void Compile::set_cached_top_node(Node* tn) {
1216 if (tn != NULL) verify_top(tn);
1217 Node* old_top = _top;
1218 _top = tn;
1219 // Calling Node::setup_is_top allows the nodes the chance to adjust
1220 // their _out arrays.
1221 if (_top != NULL) _top->setup_is_top();
1222 if (old_top != NULL) old_top->setup_is_top();
1223 assert(_top == NULL || top()->is_top(), "");
1224 }
1226 #ifdef ASSERT
1227 uint Compile::count_live_nodes_by_graph_walk() {
1228 Unique_Node_List useful(comp_arena());
1229 // Get useful node list by walking the graph.
1230 identify_useful_nodes(useful);
1231 return useful.size();
1232 }
1234 void Compile::print_missing_nodes() {
1236 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1237 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1238 return;
1239 }
1241 // This is an expensive function. It is executed only when the user
1242 // specifies VerifyIdealNodeCount option or otherwise knows the
1243 // additional work that needs to be done to identify reachable nodes
1244 // by walking the flow graph and find the missing ones using
1245 // _dead_node_list.
1247 Unique_Node_List useful(comp_arena());
1248 // Get useful node list by walking the graph.
1249 identify_useful_nodes(useful);
1251 uint l_nodes = C->live_nodes();
1252 uint l_nodes_by_walk = useful.size();
1254 if (l_nodes != l_nodes_by_walk) {
1255 if (_log != NULL) {
1256 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1257 _log->stamp();
1258 _log->end_head();
1259 }
1260 VectorSet& useful_member_set = useful.member_set();
1261 int last_idx = l_nodes_by_walk;
1262 for (int i = 0; i < last_idx; i++) {
1263 if (useful_member_set.test(i)) {
1264 if (_dead_node_list.test(i)) {
1265 if (_log != NULL) {
1266 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1267 }
1268 if (PrintIdealNodeCount) {
1269 // Print the log message to tty
1270 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1271 useful.at(i)->dump();
1272 }
1273 }
1274 }
1275 else if (! _dead_node_list.test(i)) {
1276 if (_log != NULL) {
1277 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1278 }
1279 if (PrintIdealNodeCount) {
1280 // Print the log message to tty
1281 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1282 }
1283 }
1284 }
1285 if (_log != NULL) {
1286 _log->tail("mismatched_nodes");
1287 }
1288 }
1289 }
1290 #endif
1292 #ifndef PRODUCT
1293 void Compile::verify_top(Node* tn) const {
1294 if (tn != NULL) {
1295 assert(tn->is_Con(), "top node must be a constant");
1296 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1297 assert(tn->in(0) != NULL, "must have live top node");
1298 }
1299 }
1300 #endif
1303 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1305 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1306 guarantee(arr != NULL, "");
1307 int num_blocks = arr->length();
1308 if (grow_by < num_blocks) grow_by = num_blocks;
1309 int num_notes = grow_by * _node_notes_block_size;
1310 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1311 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1312 while (num_notes > 0) {
1313 arr->append(notes);
1314 notes += _node_notes_block_size;
1315 num_notes -= _node_notes_block_size;
1316 }
1317 assert(num_notes == 0, "exact multiple, please");
1318 }
1320 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1321 if (source == NULL || dest == NULL) return false;
1323 if (dest->is_Con())
1324 return false; // Do not push debug info onto constants.
1326 #ifdef ASSERT
1327 // Leave a bread crumb trail pointing to the original node:
1328 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1329 dest->set_debug_orig(source);
1330 }
1331 #endif
1333 if (node_note_array() == NULL)
1334 return false; // Not collecting any notes now.
1336 // This is a copy onto a pre-existing node, which may already have notes.
1337 // If both nodes have notes, do not overwrite any pre-existing notes.
1338 Node_Notes* source_notes = node_notes_at(source->_idx);
1339 if (source_notes == NULL || source_notes->is_clear()) return false;
1340 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1341 if (dest_notes == NULL || dest_notes->is_clear()) {
1342 return set_node_notes_at(dest->_idx, source_notes);
1343 }
1345 Node_Notes merged_notes = (*source_notes);
1346 // The order of operations here ensures that dest notes will win...
1347 merged_notes.update_from(dest_notes);
1348 return set_node_notes_at(dest->_idx, &merged_notes);
1349 }
1352 //--------------------------allow_range_check_smearing-------------------------
1353 // Gating condition for coalescing similar range checks.
1354 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1355 // single covering check that is at least as strong as any of them.
1356 // If the optimization succeeds, the simplified (strengthened) range check
1357 // will always succeed. If it fails, we will deopt, and then give up
1358 // on the optimization.
1359 bool Compile::allow_range_check_smearing() const {
1360 // If this method has already thrown a range-check,
1361 // assume it was because we already tried range smearing
1362 // and it failed.
1363 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1364 return !already_trapped;
1365 }
1368 //------------------------------flatten_alias_type-----------------------------
1369 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1370 int offset = tj->offset();
1371 TypePtr::PTR ptr = tj->ptr();
1373 // Known instance (scalarizable allocation) alias only with itself.
1374 bool is_known_inst = tj->isa_oopptr() != NULL &&
1375 tj->is_oopptr()->is_known_instance();
1377 // Process weird unsafe references.
1378 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1379 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1380 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1381 tj = TypeOopPtr::BOTTOM;
1382 ptr = tj->ptr();
1383 offset = tj->offset();
1384 }
1386 // Array pointers need some flattening
1387 const TypeAryPtr *ta = tj->isa_aryptr();
1388 if (ta && ta->is_stable()) {
1389 // Erase stability property for alias analysis.
1390 tj = ta = ta->cast_to_stable(false);
1391 }
1392 if( ta && is_known_inst ) {
1393 if ( offset != Type::OffsetBot &&
1394 offset > arrayOopDesc::length_offset_in_bytes() ) {
1395 offset = Type::OffsetBot; // Flatten constant access into array body only
1396 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1397 }
1398 } else if( ta && _AliasLevel >= 2 ) {
1399 // For arrays indexed by constant indices, we flatten the alias
1400 // space to include all of the array body. Only the header, klass
1401 // and array length can be accessed un-aliased.
1402 if( offset != Type::OffsetBot ) {
1403 if( ta->const_oop() ) { // MethodData* or Method*
1404 offset = Type::OffsetBot; // Flatten constant access into array body
1405 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1406 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1407 // range is OK as-is.
1408 tj = ta = TypeAryPtr::RANGE;
1409 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1410 tj = TypeInstPtr::KLASS; // all klass loads look alike
1411 ta = TypeAryPtr::RANGE; // generic ignored junk
1412 ptr = TypePtr::BotPTR;
1413 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1414 tj = TypeInstPtr::MARK;
1415 ta = TypeAryPtr::RANGE; // generic ignored junk
1416 ptr = TypePtr::BotPTR;
1417 } else { // Random constant offset into array body
1418 offset = Type::OffsetBot; // Flatten constant access into array body
1419 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1420 }
1421 }
1422 // Arrays of fixed size alias with arrays of unknown size.
1423 if (ta->size() != TypeInt::POS) {
1424 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1425 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1426 }
1427 // Arrays of known objects become arrays of unknown objects.
1428 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1429 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1430 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1431 }
1432 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1433 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1434 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1435 }
1436 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1437 // cannot be distinguished by bytecode alone.
1438 if (ta->elem() == TypeInt::BOOL) {
1439 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1440 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1441 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1442 }
1443 // During the 2nd round of IterGVN, NotNull castings are removed.
1444 // Make sure the Bottom and NotNull variants alias the same.
1445 // Also, make sure exact and non-exact variants alias the same.
1446 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1447 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1448 }
1449 }
1451 // Oop pointers need some flattening
1452 const TypeInstPtr *to = tj->isa_instptr();
1453 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1454 ciInstanceKlass *k = to->klass()->as_instance_klass();
1455 if( ptr == TypePtr::Constant ) {
1456 if (to->klass() != ciEnv::current()->Class_klass() ||
1457 offset < k->size_helper() * wordSize) {
1458 // No constant oop pointers (such as Strings); they alias with
1459 // unknown strings.
1460 assert(!is_known_inst, "not scalarizable allocation");
1461 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1462 }
1463 } else if( is_known_inst ) {
1464 tj = to; // Keep NotNull and klass_is_exact for instance type
1465 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1466 // During the 2nd round of IterGVN, NotNull castings are removed.
1467 // Make sure the Bottom and NotNull variants alias the same.
1468 // Also, make sure exact and non-exact variants alias the same.
1469 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1470 }
1471 if (to->speculative() != NULL) {
1472 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1473 }
1474 // Canonicalize the holder of this field
1475 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1476 // First handle header references such as a LoadKlassNode, even if the
1477 // object's klass is unloaded at compile time (4965979).
1478 if (!is_known_inst) { // Do it only for non-instance types
1479 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1480 }
1481 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1482 // Static fields are in the space above the normal instance
1483 // fields in the java.lang.Class instance.
1484 if (to->klass() != ciEnv::current()->Class_klass()) {
1485 to = NULL;
1486 tj = TypeOopPtr::BOTTOM;
1487 offset = tj->offset();
1488 }
1489 } else {
1490 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1491 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1492 if( is_known_inst ) {
1493 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1494 } else {
1495 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1496 }
1497 }
1498 }
1499 }
1501 // Klass pointers to object array klasses need some flattening
1502 const TypeKlassPtr *tk = tj->isa_klassptr();
1503 if( tk ) {
1504 // If we are referencing a field within a Klass, we need
1505 // to assume the worst case of an Object. Both exact and
1506 // inexact types must flatten to the same alias class so
1507 // use NotNull as the PTR.
1508 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1510 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1511 TypeKlassPtr::OBJECT->klass(),
1512 offset);
1513 }
1515 ciKlass* klass = tk->klass();
1516 if( klass->is_obj_array_klass() ) {
1517 ciKlass* k = TypeAryPtr::OOPS->klass();
1518 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1519 k = TypeInstPtr::BOTTOM->klass();
1520 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1521 }
1523 // Check for precise loads from the primary supertype array and force them
1524 // to the supertype cache alias index. Check for generic array loads from
1525 // the primary supertype array and also force them to the supertype cache
1526 // alias index. Since the same load can reach both, we need to merge
1527 // these 2 disparate memories into the same alias class. Since the
1528 // primary supertype array is read-only, there's no chance of confusion
1529 // where we bypass an array load and an array store.
1530 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1531 if (offset == Type::OffsetBot ||
1532 (offset >= primary_supers_offset &&
1533 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1534 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1535 offset = in_bytes(Klass::secondary_super_cache_offset());
1536 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1537 }
1538 }
1540 // Flatten all Raw pointers together.
1541 if (tj->base() == Type::RawPtr)
1542 tj = TypeRawPtr::BOTTOM;
1544 if (tj->base() == Type::AnyPtr)
1545 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1547 // Flatten all to bottom for now
1548 switch( _AliasLevel ) {
1549 case 0:
1550 tj = TypePtr::BOTTOM;
1551 break;
1552 case 1: // Flatten to: oop, static, field or array
1553 switch (tj->base()) {
1554 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1555 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1556 case Type::AryPtr: // do not distinguish arrays at all
1557 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1558 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1559 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1560 default: ShouldNotReachHere();
1561 }
1562 break;
1563 case 2: // No collapsing at level 2; keep all splits
1564 case 3: // No collapsing at level 3; keep all splits
1565 break;
1566 default:
1567 Unimplemented();
1568 }
1570 offset = tj->offset();
1571 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1573 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1574 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1575 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1576 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1577 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1578 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1579 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1580 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1581 assert( tj->ptr() != TypePtr::TopPTR &&
1582 tj->ptr() != TypePtr::AnyNull &&
1583 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1584 // assert( tj->ptr() != TypePtr::Constant ||
1585 // tj->base() == Type::RawPtr ||
1586 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1588 return tj;
1589 }
1591 void Compile::AliasType::Init(int i, const TypePtr* at) {
1592 _index = i;
1593 _adr_type = at;
1594 _field = NULL;
1595 _element = NULL;
1596 _is_rewritable = true; // default
1597 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1598 if (atoop != NULL && atoop->is_known_instance()) {
1599 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1600 _general_index = Compile::current()->get_alias_index(gt);
1601 } else {
1602 _general_index = 0;
1603 }
1604 }
1606 //---------------------------------print_on------------------------------------
1607 #ifndef PRODUCT
1608 void Compile::AliasType::print_on(outputStream* st) {
1609 if (index() < 10)
1610 st->print("@ <%d> ", index());
1611 else st->print("@ <%d>", index());
1612 st->print(is_rewritable() ? " " : " RO");
1613 int offset = adr_type()->offset();
1614 if (offset == Type::OffsetBot)
1615 st->print(" +any");
1616 else st->print(" +%-3d", offset);
1617 st->print(" in ");
1618 adr_type()->dump_on(st);
1619 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1620 if (field() != NULL && tjp) {
1621 if (tjp->klass() != field()->holder() ||
1622 tjp->offset() != field()->offset_in_bytes()) {
1623 st->print(" != ");
1624 field()->print();
1625 st->print(" ***");
1626 }
1627 }
1628 }
1630 void print_alias_types() {
1631 Compile* C = Compile::current();
1632 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1633 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1634 C->alias_type(idx)->print_on(tty);
1635 tty->cr();
1636 }
1637 }
1638 #endif
1641 //----------------------------probe_alias_cache--------------------------------
1642 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1643 intptr_t key = (intptr_t) adr_type;
1644 key ^= key >> logAliasCacheSize;
1645 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1646 }
1649 //-----------------------------grow_alias_types--------------------------------
1650 void Compile::grow_alias_types() {
1651 const int old_ats = _max_alias_types; // how many before?
1652 const int new_ats = old_ats; // how many more?
1653 const int grow_ats = old_ats+new_ats; // how many now?
1654 _max_alias_types = grow_ats;
1655 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1656 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1657 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1658 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1659 }
1662 //--------------------------------find_alias_type------------------------------
1663 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1664 if (_AliasLevel == 0)
1665 return alias_type(AliasIdxBot);
1667 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1668 if (ace->_adr_type == adr_type) {
1669 return alias_type(ace->_index);
1670 }
1672 // Handle special cases.
1673 if (adr_type == NULL) return alias_type(AliasIdxTop);
1674 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1676 // Do it the slow way.
1677 const TypePtr* flat = flatten_alias_type(adr_type);
1679 #ifdef ASSERT
1680 assert(flat == flatten_alias_type(flat), "idempotent");
1681 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1682 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1683 const TypeOopPtr* foop = flat->is_oopptr();
1684 // Scalarizable allocations have exact klass always.
1685 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1686 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1687 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1688 }
1689 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1690 #endif
1692 int idx = AliasIdxTop;
1693 for (int i = 0; i < num_alias_types(); i++) {
1694 if (alias_type(i)->adr_type() == flat) {
1695 idx = i;
1696 break;
1697 }
1698 }
1700 if (idx == AliasIdxTop) {
1701 if (no_create) return NULL;
1702 // Grow the array if necessary.
1703 if (_num_alias_types == _max_alias_types) grow_alias_types();
1704 // Add a new alias type.
1705 idx = _num_alias_types++;
1706 _alias_types[idx]->Init(idx, flat);
1707 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1708 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1709 if (flat->isa_instptr()) {
1710 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1711 && flat->is_instptr()->klass() == env()->Class_klass())
1712 alias_type(idx)->set_rewritable(false);
1713 }
1714 if (flat->isa_aryptr()) {
1715 #ifdef ASSERT
1716 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1717 // (T_BYTE has the weakest alignment and size restrictions...)
1718 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1719 #endif
1720 if (flat->offset() == TypePtr::OffsetBot) {
1721 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1722 }
1723 }
1724 if (flat->isa_klassptr()) {
1725 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1726 alias_type(idx)->set_rewritable(false);
1727 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1728 alias_type(idx)->set_rewritable(false);
1729 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1730 alias_type(idx)->set_rewritable(false);
1731 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1732 alias_type(idx)->set_rewritable(false);
1733 }
1734 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1735 // but the base pointer type is not distinctive enough to identify
1736 // references into JavaThread.)
1738 // Check for final fields.
1739 const TypeInstPtr* tinst = flat->isa_instptr();
1740 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1741 ciField* field;
1742 if (tinst->const_oop() != NULL &&
1743 tinst->klass() == ciEnv::current()->Class_klass() &&
1744 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1745 // static field
1746 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1747 field = k->get_field_by_offset(tinst->offset(), true);
1748 } else {
1749 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1750 field = k->get_field_by_offset(tinst->offset(), false);
1751 }
1752 assert(field == NULL ||
1753 original_field == NULL ||
1754 (field->holder() == original_field->holder() &&
1755 field->offset() == original_field->offset() &&
1756 field->is_static() == original_field->is_static()), "wrong field?");
1757 // Set field() and is_rewritable() attributes.
1758 if (field != NULL) alias_type(idx)->set_field(field);
1759 }
1760 }
1762 // Fill the cache for next time.
1763 ace->_adr_type = adr_type;
1764 ace->_index = idx;
1765 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1767 // Might as well try to fill the cache for the flattened version, too.
1768 AliasCacheEntry* face = probe_alias_cache(flat);
1769 if (face->_adr_type == NULL) {
1770 face->_adr_type = flat;
1771 face->_index = idx;
1772 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1773 }
1775 return alias_type(idx);
1776 }
1779 Compile::AliasType* Compile::alias_type(ciField* field) {
1780 const TypeOopPtr* t;
1781 if (field->is_static())
1782 t = TypeInstPtr::make(field->holder()->java_mirror());
1783 else
1784 t = TypeOopPtr::make_from_klass_raw(field->holder());
1785 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1786 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1787 return atp;
1788 }
1791 //------------------------------have_alias_type--------------------------------
1792 bool Compile::have_alias_type(const TypePtr* adr_type) {
1793 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1794 if (ace->_adr_type == adr_type) {
1795 return true;
1796 }
1798 // Handle special cases.
1799 if (adr_type == NULL) return true;
1800 if (adr_type == TypePtr::BOTTOM) return true;
1802 return find_alias_type(adr_type, true, NULL) != NULL;
1803 }
1805 //-----------------------------must_alias--------------------------------------
1806 // True if all values of the given address type are in the given alias category.
1807 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1808 if (alias_idx == AliasIdxBot) return true; // the universal category
1809 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1810 if (alias_idx == AliasIdxTop) return false; // the empty category
1811 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1813 // the only remaining possible overlap is identity
1814 int adr_idx = get_alias_index(adr_type);
1815 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1816 assert(adr_idx == alias_idx ||
1817 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1818 && adr_type != TypeOopPtr::BOTTOM),
1819 "should not be testing for overlap with an unsafe pointer");
1820 return adr_idx == alias_idx;
1821 }
1823 //------------------------------can_alias--------------------------------------
1824 // True if any values of the given address type are in the given alias category.
1825 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1826 if (alias_idx == AliasIdxTop) return false; // the empty category
1827 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1828 if (alias_idx == AliasIdxBot) return true; // the universal category
1829 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1831 // the only remaining possible overlap is identity
1832 int adr_idx = get_alias_index(adr_type);
1833 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1834 return adr_idx == alias_idx;
1835 }
1839 //---------------------------pop_warm_call-------------------------------------
1840 WarmCallInfo* Compile::pop_warm_call() {
1841 WarmCallInfo* wci = _warm_calls;
1842 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1843 return wci;
1844 }
1846 //----------------------------Inline_Warm--------------------------------------
1847 int Compile::Inline_Warm() {
1848 // If there is room, try to inline some more warm call sites.
1849 // %%% Do a graph index compaction pass when we think we're out of space?
1850 if (!InlineWarmCalls) return 0;
1852 int calls_made_hot = 0;
1853 int room_to_grow = NodeCountInliningCutoff - unique();
1854 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1855 int amount_grown = 0;
1856 WarmCallInfo* call;
1857 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1858 int est_size = (int)call->size();
1859 if (est_size > (room_to_grow - amount_grown)) {
1860 // This one won't fit anyway. Get rid of it.
1861 call->make_cold();
1862 continue;
1863 }
1864 call->make_hot();
1865 calls_made_hot++;
1866 amount_grown += est_size;
1867 amount_to_grow -= est_size;
1868 }
1870 if (calls_made_hot > 0) set_major_progress();
1871 return calls_made_hot;
1872 }
1875 //----------------------------Finish_Warm--------------------------------------
1876 void Compile::Finish_Warm() {
1877 if (!InlineWarmCalls) return;
1878 if (failing()) return;
1879 if (warm_calls() == NULL) return;
1881 // Clean up loose ends, if we are out of space for inlining.
1882 WarmCallInfo* call;
1883 while ((call = pop_warm_call()) != NULL) {
1884 call->make_cold();
1885 }
1886 }
1888 //---------------------cleanup_loop_predicates-----------------------
1889 // Remove the opaque nodes that protect the predicates so that all unused
1890 // checks and uncommon_traps will be eliminated from the ideal graph
1891 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1892 if (predicate_count()==0) return;
1893 for (int i = predicate_count(); i > 0; i--) {
1894 Node * n = predicate_opaque1_node(i-1);
1895 assert(n->Opcode() == Op_Opaque1, "must be");
1896 igvn.replace_node(n, n->in(1));
1897 }
1898 assert(predicate_count()==0, "should be clean!");
1899 }
1901 void Compile::add_range_check_cast(Node* n) {
1902 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1903 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1904 _range_check_casts->append(n);
1905 }
1907 // Remove all range check dependent CastIINodes.
1908 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1909 for (int i = range_check_cast_count(); i > 0; i--) {
1910 Node* cast = range_check_cast_node(i-1);
1911 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1912 igvn.replace_node(cast, cast->in(1));
1913 }
1914 assert(range_check_cast_count() == 0, "should be empty");
1915 }
1917 // StringOpts and late inlining of string methods
1918 void Compile::inline_string_calls(bool parse_time) {
1919 {
1920 // remove useless nodes to make the usage analysis simpler
1921 ResourceMark rm;
1922 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1923 }
1925 {
1926 ResourceMark rm;
1927 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1928 PhaseStringOpts pso(initial_gvn(), for_igvn());
1929 print_method(PHASE_AFTER_STRINGOPTS, 3);
1930 }
1932 // now inline anything that we skipped the first time around
1933 if (!parse_time) {
1934 _late_inlines_pos = _late_inlines.length();
1935 }
1937 while (_string_late_inlines.length() > 0) {
1938 CallGenerator* cg = _string_late_inlines.pop();
1939 cg->do_late_inline();
1940 if (failing()) return;
1941 }
1942 _string_late_inlines.trunc_to(0);
1943 }
1945 // Late inlining of boxing methods
1946 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1947 if (_boxing_late_inlines.length() > 0) {
1948 assert(has_boxed_value(), "inconsistent");
1950 PhaseGVN* gvn = initial_gvn();
1951 set_inlining_incrementally(true);
1953 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1954 for_igvn()->clear();
1955 gvn->replace_with(&igvn);
1957 _late_inlines_pos = _late_inlines.length();
1959 while (_boxing_late_inlines.length() > 0) {
1960 CallGenerator* cg = _boxing_late_inlines.pop();
1961 cg->do_late_inline();
1962 if (failing()) return;
1963 }
1964 _boxing_late_inlines.trunc_to(0);
1966 {
1967 ResourceMark rm;
1968 PhaseRemoveUseless pru(gvn, for_igvn());
1969 }
1971 igvn = PhaseIterGVN(gvn);
1972 igvn.optimize();
1974 set_inlining_progress(false);
1975 set_inlining_incrementally(false);
1976 }
1977 }
1979 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1980 assert(IncrementalInline, "incremental inlining should be on");
1981 PhaseGVN* gvn = initial_gvn();
1983 set_inlining_progress(false);
1984 for_igvn()->clear();
1985 gvn->replace_with(&igvn);
1987 int i = 0;
1989 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1990 CallGenerator* cg = _late_inlines.at(i);
1991 _late_inlines_pos = i+1;
1992 cg->do_late_inline();
1993 if (failing()) return;
1994 }
1995 int j = 0;
1996 for (; i < _late_inlines.length(); i++, j++) {
1997 _late_inlines.at_put(j, _late_inlines.at(i));
1998 }
1999 _late_inlines.trunc_to(j);
2001 {
2002 ResourceMark rm;
2003 PhaseRemoveUseless pru(gvn, for_igvn());
2004 }
2006 igvn = PhaseIterGVN(gvn);
2007 }
2009 // Perform incremental inlining until bound on number of live nodes is reached
2010 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2011 PhaseGVN* gvn = initial_gvn();
2013 set_inlining_incrementally(true);
2014 set_inlining_progress(true);
2015 uint low_live_nodes = 0;
2017 while(inlining_progress() && _late_inlines.length() > 0) {
2019 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2020 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2021 // PhaseIdealLoop is expensive so we only try it once we are
2022 // out of live nodes and we only try it again if the previous
2023 // helped got the number of nodes down significantly
2024 PhaseIdealLoop ideal_loop( igvn, false, true );
2025 if (failing()) return;
2026 low_live_nodes = live_nodes();
2027 _major_progress = true;
2028 }
2030 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2031 break;
2032 }
2033 }
2035 inline_incrementally_one(igvn);
2037 if (failing()) return;
2039 igvn.optimize();
2041 if (failing()) return;
2042 }
2044 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2046 if (_string_late_inlines.length() > 0) {
2047 assert(has_stringbuilder(), "inconsistent");
2048 for_igvn()->clear();
2049 initial_gvn()->replace_with(&igvn);
2051 inline_string_calls(false);
2053 if (failing()) return;
2055 {
2056 ResourceMark rm;
2057 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2058 }
2060 igvn = PhaseIterGVN(gvn);
2062 igvn.optimize();
2063 }
2065 set_inlining_incrementally(false);
2066 }
2069 //------------------------------Optimize---------------------------------------
2070 // Given a graph, optimize it.
2071 void Compile::Optimize() {
2072 TracePhase t1("optimizer", &_t_optimizer, true);
2074 #ifndef PRODUCT
2075 if (env()->break_at_compile()) {
2076 BREAKPOINT;
2077 }
2079 #endif
2081 ResourceMark rm;
2082 int loop_opts_cnt;
2084 NOT_PRODUCT( verify_graph_edges(); )
2086 print_method(PHASE_AFTER_PARSING);
2088 {
2089 // Iterative Global Value Numbering, including ideal transforms
2090 // Initialize IterGVN with types and values from parse-time GVN
2091 PhaseIterGVN igvn(initial_gvn());
2092 {
2093 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2094 igvn.optimize();
2095 }
2097 print_method(PHASE_ITER_GVN1, 2);
2099 if (failing()) return;
2101 {
2102 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2103 inline_incrementally(igvn);
2104 }
2106 print_method(PHASE_INCREMENTAL_INLINE, 2);
2108 if (failing()) return;
2110 if (eliminate_boxing()) {
2111 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2112 // Inline valueOf() methods now.
2113 inline_boxing_calls(igvn);
2115 if (AlwaysIncrementalInline) {
2116 inline_incrementally(igvn);
2117 }
2119 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2121 if (failing()) return;
2122 }
2124 // Remove the speculative part of types and clean up the graph from
2125 // the extra CastPP nodes whose only purpose is to carry them. Do
2126 // that early so that optimizations are not disrupted by the extra
2127 // CastPP nodes.
2128 remove_speculative_types(igvn);
2130 // No more new expensive nodes will be added to the list from here
2131 // so keep only the actual candidates for optimizations.
2132 cleanup_expensive_nodes(igvn);
2134 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2135 NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);)
2136 initial_gvn()->replace_with(&igvn);
2137 for_igvn()->clear();
2138 Unique_Node_List new_worklist(C->comp_arena());
2139 {
2140 ResourceMark rm;
2141 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2142 }
2143 set_for_igvn(&new_worklist);
2144 igvn = PhaseIterGVN(initial_gvn());
2145 igvn.optimize();
2146 }
2148 // Perform escape analysis
2149 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2150 if (has_loops()) {
2151 // Cleanup graph (remove dead nodes).
2152 TracePhase t2("idealLoop", &_t_idealLoop, true);
2153 PhaseIdealLoop ideal_loop( igvn, false, true );
2154 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2155 if (failing()) return;
2156 }
2157 ConnectionGraph::do_analysis(this, &igvn);
2159 if (failing()) return;
2161 // Optimize out fields loads from scalar replaceable allocations.
2162 igvn.optimize();
2163 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2165 if (failing()) return;
2167 if (congraph() != NULL && macro_count() > 0) {
2168 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2169 PhaseMacroExpand mexp(igvn);
2170 mexp.eliminate_macro_nodes();
2171 igvn.set_delay_transform(false);
2173 igvn.optimize();
2174 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2176 if (failing()) return;
2177 }
2178 }
2180 // Loop transforms on the ideal graph. Range Check Elimination,
2181 // peeling, unrolling, etc.
2183 // Set loop opts counter
2184 loop_opts_cnt = num_loop_opts();
2185 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2186 {
2187 TracePhase t2("idealLoop", &_t_idealLoop, true);
2188 PhaseIdealLoop ideal_loop( igvn, true );
2189 loop_opts_cnt--;
2190 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2191 if (failing()) return;
2192 }
2193 // Loop opts pass if partial peeling occurred in previous pass
2194 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2195 TracePhase t3("idealLoop", &_t_idealLoop, true);
2196 PhaseIdealLoop ideal_loop( igvn, false );
2197 loop_opts_cnt--;
2198 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2199 if (failing()) return;
2200 }
2201 // Loop opts pass for loop-unrolling before CCP
2202 if(major_progress() && (loop_opts_cnt > 0)) {
2203 TracePhase t4("idealLoop", &_t_idealLoop, true);
2204 PhaseIdealLoop ideal_loop( igvn, false );
2205 loop_opts_cnt--;
2206 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2207 }
2208 if (!failing()) {
2209 // Verify that last round of loop opts produced a valid graph
2210 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2211 PhaseIdealLoop::verify(igvn);
2212 }
2213 }
2214 if (failing()) return;
2216 // Conditional Constant Propagation;
2217 PhaseCCP ccp( &igvn );
2218 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2219 {
2220 TracePhase t2("ccp", &_t_ccp, true);
2221 ccp.do_transform();
2222 }
2223 print_method(PHASE_CPP1, 2);
2225 assert( true, "Break here to ccp.dump_old2new_map()");
2227 // Iterative Global Value Numbering, including ideal transforms
2228 {
2229 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2230 igvn = ccp;
2231 igvn.optimize();
2232 }
2234 print_method(PHASE_ITER_GVN2, 2);
2236 if (failing()) return;
2238 // Loop transforms on the ideal graph. Range Check Elimination,
2239 // peeling, unrolling, etc.
2240 if(loop_opts_cnt > 0) {
2241 debug_only( int cnt = 0; );
2242 while(major_progress() && (loop_opts_cnt > 0)) {
2243 TracePhase t2("idealLoop", &_t_idealLoop, true);
2244 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2245 PhaseIdealLoop ideal_loop( igvn, true);
2246 loop_opts_cnt--;
2247 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2248 if (failing()) return;
2249 }
2250 }
2252 {
2253 // Verify that all previous optimizations produced a valid graph
2254 // at least to this point, even if no loop optimizations were done.
2255 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2256 PhaseIdealLoop::verify(igvn);
2257 }
2259 if (range_check_cast_count() > 0) {
2260 // No more loop optimizations. Remove all range check dependent CastIINodes.
2261 C->remove_range_check_casts(igvn);
2262 igvn.optimize();
2263 }
2265 {
2266 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2267 PhaseMacroExpand mex(igvn);
2268 if (mex.expand_macro_nodes()) {
2269 assert(failing(), "must bail out w/ explicit message");
2270 return;
2271 }
2272 }
2274 } // (End scope of igvn; run destructor if necessary for asserts.)
2276 dump_inlining();
2277 // A method with only infinite loops has no edges entering loops from root
2278 {
2279 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2280 if (final_graph_reshaping()) {
2281 assert(failing(), "must bail out w/ explicit message");
2282 return;
2283 }
2284 }
2286 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2287 }
2290 //------------------------------Code_Gen---------------------------------------
2291 // Given a graph, generate code for it
2292 void Compile::Code_Gen() {
2293 if (failing()) {
2294 return;
2295 }
2297 // Perform instruction selection. You might think we could reclaim Matcher
2298 // memory PDQ, but actually the Matcher is used in generating spill code.
2299 // Internals of the Matcher (including some VectorSets) must remain live
2300 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2301 // set a bit in reclaimed memory.
2303 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2304 // nodes. Mapping is only valid at the root of each matched subtree.
2305 NOT_PRODUCT( verify_graph_edges(); )
2307 Matcher matcher;
2308 _matcher = &matcher;
2309 {
2310 TracePhase t2("matcher", &_t_matcher, true);
2311 matcher.match();
2312 }
2313 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2314 // nodes. Mapping is only valid at the root of each matched subtree.
2315 NOT_PRODUCT( verify_graph_edges(); )
2317 // If you have too many nodes, or if matching has failed, bail out
2318 check_node_count(0, "out of nodes matching instructions");
2319 if (failing()) {
2320 return;
2321 }
2323 // Build a proper-looking CFG
2324 PhaseCFG cfg(node_arena(), root(), matcher);
2325 _cfg = &cfg;
2326 {
2327 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2328 bool success = cfg.do_global_code_motion();
2329 if (!success) {
2330 return;
2331 }
2333 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2334 NOT_PRODUCT( verify_graph_edges(); )
2335 debug_only( cfg.verify(); )
2336 }
2338 PhaseChaitin regalloc(unique(), cfg, matcher);
2339 _regalloc = ®alloc;
2340 {
2341 TracePhase t2("regalloc", &_t_registerAllocation, true);
2342 // Perform register allocation. After Chaitin, use-def chains are
2343 // no longer accurate (at spill code) and so must be ignored.
2344 // Node->LRG->reg mappings are still accurate.
2345 _regalloc->Register_Allocate();
2347 // Bail out if the allocator builds too many nodes
2348 if (failing()) {
2349 return;
2350 }
2351 }
2353 // Prior to register allocation we kept empty basic blocks in case the
2354 // the allocator needed a place to spill. After register allocation we
2355 // are not adding any new instructions. If any basic block is empty, we
2356 // can now safely remove it.
2357 {
2358 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2359 cfg.remove_empty_blocks();
2360 if (do_freq_based_layout()) {
2361 PhaseBlockLayout layout(cfg);
2362 } else {
2363 cfg.set_loop_alignment();
2364 }
2365 cfg.fixup_flow();
2366 }
2368 // Apply peephole optimizations
2369 if( OptoPeephole ) {
2370 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2371 PhasePeephole peep( _regalloc, cfg);
2372 peep.do_transform();
2373 }
2375 // Do late expand if CPU requires this.
2376 if (Matcher::require_postalloc_expand) {
2377 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2378 cfg.postalloc_expand(_regalloc);
2379 }
2381 // Convert Nodes to instruction bits in a buffer
2382 {
2383 // %%%% workspace merge brought two timers together for one job
2384 TracePhase t2a("output", &_t_output, true);
2385 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2386 Output();
2387 }
2389 print_method(PHASE_FINAL_CODE);
2391 // He's dead, Jim.
2392 _cfg = (PhaseCFG*)0xdeadbeef;
2393 _regalloc = (PhaseChaitin*)0xdeadbeef;
2394 }
2397 //------------------------------dump_asm---------------------------------------
2398 // Dump formatted assembly
2399 #ifndef PRODUCT
2400 void Compile::dump_asm(int *pcs, uint pc_limit) {
2401 bool cut_short = false;
2402 tty->print_cr("#");
2403 tty->print("# "); _tf->dump(); tty->cr();
2404 tty->print_cr("#");
2406 // For all blocks
2407 int pc = 0x0; // Program counter
2408 char starts_bundle = ' ';
2409 _regalloc->dump_frame();
2411 Node *n = NULL;
2412 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2413 if (VMThread::should_terminate()) {
2414 cut_short = true;
2415 break;
2416 }
2417 Block* block = _cfg->get_block(i);
2418 if (block->is_connector() && !Verbose) {
2419 continue;
2420 }
2421 n = block->head();
2422 if (pcs && n->_idx < pc_limit) {
2423 tty->print("%3.3x ", pcs[n->_idx]);
2424 } else {
2425 tty->print(" ");
2426 }
2427 block->dump_head(_cfg);
2428 if (block->is_connector()) {
2429 tty->print_cr(" # Empty connector block");
2430 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2431 tty->print_cr(" # Block is sole successor of call");
2432 }
2434 // For all instructions
2435 Node *delay = NULL;
2436 for (uint j = 0; j < block->number_of_nodes(); j++) {
2437 if (VMThread::should_terminate()) {
2438 cut_short = true;
2439 break;
2440 }
2441 n = block->get_node(j);
2442 if (valid_bundle_info(n)) {
2443 Bundle* bundle = node_bundling(n);
2444 if (bundle->used_in_unconditional_delay()) {
2445 delay = n;
2446 continue;
2447 }
2448 if (bundle->starts_bundle()) {
2449 starts_bundle = '+';
2450 }
2451 }
2453 if (WizardMode) {
2454 n->dump();
2455 }
2457 if( !n->is_Region() && // Dont print in the Assembly
2458 !n->is_Phi() && // a few noisely useless nodes
2459 !n->is_Proj() &&
2460 !n->is_MachTemp() &&
2461 !n->is_SafePointScalarObject() &&
2462 !n->is_Catch() && // Would be nice to print exception table targets
2463 !n->is_MergeMem() && // Not very interesting
2464 !n->is_top() && // Debug info table constants
2465 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2466 ) {
2467 if (pcs && n->_idx < pc_limit)
2468 tty->print("%3.3x", pcs[n->_idx]);
2469 else
2470 tty->print(" ");
2471 tty->print(" %c ", starts_bundle);
2472 starts_bundle = ' ';
2473 tty->print("\t");
2474 n->format(_regalloc, tty);
2475 tty->cr();
2476 }
2478 // If we have an instruction with a delay slot, and have seen a delay,
2479 // then back up and print it
2480 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2481 assert(delay != NULL, "no unconditional delay instruction");
2482 if (WizardMode) delay->dump();
2484 if (node_bundling(delay)->starts_bundle())
2485 starts_bundle = '+';
2486 if (pcs && n->_idx < pc_limit)
2487 tty->print("%3.3x", pcs[n->_idx]);
2488 else
2489 tty->print(" ");
2490 tty->print(" %c ", starts_bundle);
2491 starts_bundle = ' ';
2492 tty->print("\t");
2493 delay->format(_regalloc, tty);
2494 tty->cr();
2495 delay = NULL;
2496 }
2498 // Dump the exception table as well
2499 if( n->is_Catch() && (Verbose || WizardMode) ) {
2500 // Print the exception table for this offset
2501 _handler_table.print_subtable_for(pc);
2502 }
2503 }
2505 if (pcs && n->_idx < pc_limit)
2506 tty->print_cr("%3.3x", pcs[n->_idx]);
2507 else
2508 tty->cr();
2510 assert(cut_short || delay == NULL, "no unconditional delay branch");
2512 } // End of per-block dump
2513 tty->cr();
2515 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2516 }
2517 #endif
2519 //------------------------------Final_Reshape_Counts---------------------------
2520 // This class defines counters to help identify when a method
2521 // may/must be executed using hardware with only 24-bit precision.
2522 struct Final_Reshape_Counts : public StackObj {
2523 int _call_count; // count non-inlined 'common' calls
2524 int _float_count; // count float ops requiring 24-bit precision
2525 int _double_count; // count double ops requiring more precision
2526 int _java_call_count; // count non-inlined 'java' calls
2527 int _inner_loop_count; // count loops which need alignment
2528 VectorSet _visited; // Visitation flags
2529 Node_List _tests; // Set of IfNodes & PCTableNodes
2531 Final_Reshape_Counts() :
2532 _call_count(0), _float_count(0), _double_count(0),
2533 _java_call_count(0), _inner_loop_count(0),
2534 _visited( Thread::current()->resource_area() ) { }
2536 void inc_call_count () { _call_count ++; }
2537 void inc_float_count () { _float_count ++; }
2538 void inc_double_count() { _double_count++; }
2539 void inc_java_call_count() { _java_call_count++; }
2540 void inc_inner_loop_count() { _inner_loop_count++; }
2542 int get_call_count () const { return _call_count ; }
2543 int get_float_count () const { return _float_count ; }
2544 int get_double_count() const { return _double_count; }
2545 int get_java_call_count() const { return _java_call_count; }
2546 int get_inner_loop_count() const { return _inner_loop_count; }
2547 };
2549 #ifdef ASSERT
2550 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2551 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2552 // Make sure the offset goes inside the instance layout.
2553 return k->contains_field_offset(tp->offset());
2554 // Note that OffsetBot and OffsetTop are very negative.
2555 }
2556 #endif
2558 // Eliminate trivially redundant StoreCMs and accumulate their
2559 // precedence edges.
2560 void Compile::eliminate_redundant_card_marks(Node* n) {
2561 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2562 if (n->in(MemNode::Address)->outcnt() > 1) {
2563 // There are multiple users of the same address so it might be
2564 // possible to eliminate some of the StoreCMs
2565 Node* mem = n->in(MemNode::Memory);
2566 Node* adr = n->in(MemNode::Address);
2567 Node* val = n->in(MemNode::ValueIn);
2568 Node* prev = n;
2569 bool done = false;
2570 // Walk the chain of StoreCMs eliminating ones that match. As
2571 // long as it's a chain of single users then the optimization is
2572 // safe. Eliminating partially redundant StoreCMs would require
2573 // cloning copies down the other paths.
2574 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2575 if (adr == mem->in(MemNode::Address) &&
2576 val == mem->in(MemNode::ValueIn)) {
2577 // redundant StoreCM
2578 if (mem->req() > MemNode::OopStore) {
2579 // Hasn't been processed by this code yet.
2580 n->add_prec(mem->in(MemNode::OopStore));
2581 } else {
2582 // Already converted to precedence edge
2583 for (uint i = mem->req(); i < mem->len(); i++) {
2584 // Accumulate any precedence edges
2585 if (mem->in(i) != NULL) {
2586 n->add_prec(mem->in(i));
2587 }
2588 }
2589 // Everything above this point has been processed.
2590 done = true;
2591 }
2592 // Eliminate the previous StoreCM
2593 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2594 assert(mem->outcnt() == 0, "should be dead");
2595 mem->disconnect_inputs(NULL, this);
2596 } else {
2597 prev = mem;
2598 }
2599 mem = prev->in(MemNode::Memory);
2600 }
2601 }
2602 }
2604 //------------------------------final_graph_reshaping_impl----------------------
2605 // Implement items 1-5 from final_graph_reshaping below.
2606 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2608 if ( n->outcnt() == 0 ) return; // dead node
2609 uint nop = n->Opcode();
2611 // Check for 2-input instruction with "last use" on right input.
2612 // Swap to left input. Implements item (2).
2613 if( n->req() == 3 && // two-input instruction
2614 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2615 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2616 n->in(2)->outcnt() == 1 &&// right use IS a last use
2617 !n->in(2)->is_Con() ) { // right use is not a constant
2618 // Check for commutative opcode
2619 switch( nop ) {
2620 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2621 case Op_MaxI: case Op_MinI:
2622 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2623 case Op_AndL: case Op_XorL: case Op_OrL:
2624 case Op_AndI: case Op_XorI: case Op_OrI: {
2625 // Move "last use" input to left by swapping inputs
2626 n->swap_edges(1, 2);
2627 break;
2628 }
2629 default:
2630 break;
2631 }
2632 }
2634 #ifdef ASSERT
2635 if( n->is_Mem() ) {
2636 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2637 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2638 // oop will be recorded in oop map if load crosses safepoint
2639 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2640 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2641 "raw memory operations should have control edge");
2642 }
2643 #endif
2644 // Count FPU ops and common calls, implements item (3)
2645 switch( nop ) {
2646 // Count all float operations that may use FPU
2647 case Op_AddF:
2648 case Op_SubF:
2649 case Op_MulF:
2650 case Op_DivF:
2651 case Op_NegF:
2652 case Op_ModF:
2653 case Op_ConvI2F:
2654 case Op_ConF:
2655 case Op_CmpF:
2656 case Op_CmpF3:
2657 // case Op_ConvL2F: // longs are split into 32-bit halves
2658 frc.inc_float_count();
2659 break;
2661 case Op_ConvF2D:
2662 case Op_ConvD2F:
2663 frc.inc_float_count();
2664 frc.inc_double_count();
2665 break;
2667 // Count all double operations that may use FPU
2668 case Op_AddD:
2669 case Op_SubD:
2670 case Op_MulD:
2671 case Op_DivD:
2672 case Op_NegD:
2673 case Op_ModD:
2674 case Op_ConvI2D:
2675 case Op_ConvD2I:
2676 // case Op_ConvL2D: // handled by leaf call
2677 // case Op_ConvD2L: // handled by leaf call
2678 case Op_ConD:
2679 case Op_CmpD:
2680 case Op_CmpD3:
2681 frc.inc_double_count();
2682 break;
2683 case Op_Opaque1: // Remove Opaque Nodes before matching
2684 case Op_Opaque2: // Remove Opaque Nodes before matching
2685 case Op_Opaque3:
2686 n->subsume_by(n->in(1), this);
2687 break;
2688 case Op_CallStaticJava:
2689 case Op_CallJava:
2690 case Op_CallDynamicJava:
2691 frc.inc_java_call_count(); // Count java call site;
2692 case Op_CallRuntime:
2693 case Op_CallLeaf:
2694 case Op_CallLeafNoFP: {
2695 assert( n->is_Call(), "" );
2696 CallNode *call = n->as_Call();
2697 // Count call sites where the FP mode bit would have to be flipped.
2698 // Do not count uncommon runtime calls:
2699 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2700 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2701 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2702 frc.inc_call_count(); // Count the call site
2703 } else { // See if uncommon argument is shared
2704 Node *n = call->in(TypeFunc::Parms);
2705 int nop = n->Opcode();
2706 // Clone shared simple arguments to uncommon calls, item (1).
2707 if( n->outcnt() > 1 &&
2708 !n->is_Proj() &&
2709 nop != Op_CreateEx &&
2710 nop != Op_CheckCastPP &&
2711 nop != Op_DecodeN &&
2712 nop != Op_DecodeNKlass &&
2713 !n->is_Mem() ) {
2714 Node *x = n->clone();
2715 call->set_req( TypeFunc::Parms, x );
2716 }
2717 }
2718 break;
2719 }
2721 case Op_StoreD:
2722 case Op_LoadD:
2723 case Op_LoadD_unaligned:
2724 frc.inc_double_count();
2725 goto handle_mem;
2726 case Op_StoreF:
2727 case Op_LoadF:
2728 frc.inc_float_count();
2729 goto handle_mem;
2731 case Op_StoreCM:
2732 {
2733 // Convert OopStore dependence into precedence edge
2734 Node* prec = n->in(MemNode::OopStore);
2735 n->del_req(MemNode::OopStore);
2736 n->add_prec(prec);
2737 eliminate_redundant_card_marks(n);
2738 }
2740 // fall through
2742 case Op_StoreB:
2743 case Op_StoreC:
2744 case Op_StorePConditional:
2745 case Op_StoreI:
2746 case Op_StoreL:
2747 case Op_StoreIConditional:
2748 case Op_StoreLConditional:
2749 case Op_CompareAndSwapI:
2750 case Op_CompareAndSwapL:
2751 case Op_CompareAndSwapP:
2752 case Op_CompareAndSwapN:
2753 case Op_GetAndAddI:
2754 case Op_GetAndAddL:
2755 case Op_GetAndSetI:
2756 case Op_GetAndSetL:
2757 case Op_GetAndSetP:
2758 case Op_GetAndSetN:
2759 case Op_StoreP:
2760 case Op_StoreN:
2761 case Op_StoreNKlass:
2762 case Op_LoadB:
2763 case Op_LoadUB:
2764 case Op_LoadUS:
2765 case Op_LoadI:
2766 case Op_LoadKlass:
2767 case Op_LoadNKlass:
2768 case Op_LoadL:
2769 case Op_LoadL_unaligned:
2770 case Op_LoadPLocked:
2771 case Op_LoadP:
2772 case Op_LoadN:
2773 case Op_LoadRange:
2774 case Op_LoadS: {
2775 handle_mem:
2776 #ifdef ASSERT
2777 if( VerifyOptoOopOffsets ) {
2778 assert( n->is_Mem(), "" );
2779 MemNode *mem = (MemNode*)n;
2780 // Check to see if address types have grounded out somehow.
2781 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2782 assert( !tp || oop_offset_is_sane(tp), "" );
2783 }
2784 #endif
2785 break;
2786 }
2788 case Op_AddP: { // Assert sane base pointers
2789 Node *addp = n->in(AddPNode::Address);
2790 assert( !addp->is_AddP() ||
2791 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2792 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2793 "Base pointers must match" );
2794 #ifdef _LP64
2795 if ((UseCompressedOops || UseCompressedClassPointers) &&
2796 addp->Opcode() == Op_ConP &&
2797 addp == n->in(AddPNode::Base) &&
2798 n->in(AddPNode::Offset)->is_Con()) {
2799 // Use addressing with narrow klass to load with offset on x86.
2800 // On sparc loading 32-bits constant and decoding it have less
2801 // instructions (4) then load 64-bits constant (7).
2802 // Do this transformation here since IGVN will convert ConN back to ConP.
2803 const Type* t = addp->bottom_type();
2804 if (t->isa_oopptr() || t->isa_klassptr()) {
2805 Node* nn = NULL;
2807 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2809 // Look for existing ConN node of the same exact type.
2810 Node* r = root();
2811 uint cnt = r->outcnt();
2812 for (uint i = 0; i < cnt; i++) {
2813 Node* m = r->raw_out(i);
2814 if (m!= NULL && m->Opcode() == op &&
2815 m->bottom_type()->make_ptr() == t) {
2816 nn = m;
2817 break;
2818 }
2819 }
2820 if (nn != NULL) {
2821 // Decode a narrow oop to match address
2822 // [R12 + narrow_oop_reg<<3 + offset]
2823 if (t->isa_oopptr()) {
2824 nn = new (this) DecodeNNode(nn, t);
2825 } else {
2826 nn = new (this) DecodeNKlassNode(nn, t);
2827 }
2828 n->set_req(AddPNode::Base, nn);
2829 n->set_req(AddPNode::Address, nn);
2830 if (addp->outcnt() == 0) {
2831 addp->disconnect_inputs(NULL, this);
2832 }
2833 }
2834 }
2835 }
2836 #endif
2837 break;
2838 }
2840 #ifdef _LP64
2841 case Op_CastPP:
2842 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2843 Node* in1 = n->in(1);
2844 const Type* t = n->bottom_type();
2845 Node* new_in1 = in1->clone();
2846 new_in1->as_DecodeN()->set_type(t);
2848 if (!Matcher::narrow_oop_use_complex_address()) {
2849 //
2850 // x86, ARM and friends can handle 2 adds in addressing mode
2851 // and Matcher can fold a DecodeN node into address by using
2852 // a narrow oop directly and do implicit NULL check in address:
2853 //
2854 // [R12 + narrow_oop_reg<<3 + offset]
2855 // NullCheck narrow_oop_reg
2856 //
2857 // On other platforms (Sparc) we have to keep new DecodeN node and
2858 // use it to do implicit NULL check in address:
2859 //
2860 // decode_not_null narrow_oop_reg, base_reg
2861 // [base_reg + offset]
2862 // NullCheck base_reg
2863 //
2864 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2865 // to keep the information to which NULL check the new DecodeN node
2866 // corresponds to use it as value in implicit_null_check().
2867 //
2868 new_in1->set_req(0, n->in(0));
2869 }
2871 n->subsume_by(new_in1, this);
2872 if (in1->outcnt() == 0) {
2873 in1->disconnect_inputs(NULL, this);
2874 }
2875 }
2876 break;
2878 case Op_CmpP:
2879 // Do this transformation here to preserve CmpPNode::sub() and
2880 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2881 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2882 Node* in1 = n->in(1);
2883 Node* in2 = n->in(2);
2884 if (!in1->is_DecodeNarrowPtr()) {
2885 in2 = in1;
2886 in1 = n->in(2);
2887 }
2888 assert(in1->is_DecodeNarrowPtr(), "sanity");
2890 Node* new_in2 = NULL;
2891 if (in2->is_DecodeNarrowPtr()) {
2892 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2893 new_in2 = in2->in(1);
2894 } else if (in2->Opcode() == Op_ConP) {
2895 const Type* t = in2->bottom_type();
2896 if (t == TypePtr::NULL_PTR) {
2897 assert(in1->is_DecodeN(), "compare klass to null?");
2898 // Don't convert CmpP null check into CmpN if compressed
2899 // oops implicit null check is not generated.
2900 // This will allow to generate normal oop implicit null check.
2901 if (Matcher::gen_narrow_oop_implicit_null_checks())
2902 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2903 //
2904 // This transformation together with CastPP transformation above
2905 // will generated code for implicit NULL checks for compressed oops.
2906 //
2907 // The original code after Optimize()
2908 //
2909 // LoadN memory, narrow_oop_reg
2910 // decode narrow_oop_reg, base_reg
2911 // CmpP base_reg, NULL
2912 // CastPP base_reg // NotNull
2913 // Load [base_reg + offset], val_reg
2914 //
2915 // after these transformations will be
2916 //
2917 // LoadN memory, narrow_oop_reg
2918 // CmpN narrow_oop_reg, NULL
2919 // decode_not_null narrow_oop_reg, base_reg
2920 // Load [base_reg + offset], val_reg
2921 //
2922 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2923 // since narrow oops can be used in debug info now (see the code in
2924 // final_graph_reshaping_walk()).
2925 //
2926 // At the end the code will be matched to
2927 // on x86:
2928 //
2929 // Load_narrow_oop memory, narrow_oop_reg
2930 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2931 // NullCheck narrow_oop_reg
2932 //
2933 // and on sparc:
2934 //
2935 // Load_narrow_oop memory, narrow_oop_reg
2936 // decode_not_null narrow_oop_reg, base_reg
2937 // Load [base_reg + offset], val_reg
2938 // NullCheck base_reg
2939 //
2940 } else if (t->isa_oopptr()) {
2941 new_in2 = ConNode::make(this, t->make_narrowoop());
2942 } else if (t->isa_klassptr()) {
2943 new_in2 = ConNode::make(this, t->make_narrowklass());
2944 }
2945 }
2946 if (new_in2 != NULL) {
2947 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2948 n->subsume_by(cmpN, this);
2949 if (in1->outcnt() == 0) {
2950 in1->disconnect_inputs(NULL, this);
2951 }
2952 if (in2->outcnt() == 0) {
2953 in2->disconnect_inputs(NULL, this);
2954 }
2955 }
2956 }
2957 break;
2959 case Op_DecodeN:
2960 case Op_DecodeNKlass:
2961 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2962 // DecodeN could be pinned when it can't be fold into
2963 // an address expression, see the code for Op_CastPP above.
2964 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2965 break;
2967 case Op_EncodeP:
2968 case Op_EncodePKlass: {
2969 Node* in1 = n->in(1);
2970 if (in1->is_DecodeNarrowPtr()) {
2971 n->subsume_by(in1->in(1), this);
2972 } else if (in1->Opcode() == Op_ConP) {
2973 const Type* t = in1->bottom_type();
2974 if (t == TypePtr::NULL_PTR) {
2975 assert(t->isa_oopptr(), "null klass?");
2976 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2977 } else if (t->isa_oopptr()) {
2978 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2979 } else if (t->isa_klassptr()) {
2980 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2981 }
2982 }
2983 if (in1->outcnt() == 0) {
2984 in1->disconnect_inputs(NULL, this);
2985 }
2986 break;
2987 }
2989 case Op_Proj: {
2990 if (OptimizeStringConcat) {
2991 ProjNode* p = n->as_Proj();
2992 if (p->_is_io_use) {
2993 // Separate projections were used for the exception path which
2994 // are normally removed by a late inline. If it wasn't inlined
2995 // then they will hang around and should just be replaced with
2996 // the original one.
2997 Node* proj = NULL;
2998 // Replace with just one
2999 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3000 Node *use = i.get();
3001 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3002 proj = use;
3003 break;
3004 }
3005 }
3006 assert(proj != NULL, "must be found");
3007 p->subsume_by(proj, this);
3008 }
3009 }
3010 break;
3011 }
3013 case Op_Phi:
3014 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3015 // The EncodeP optimization may create Phi with the same edges
3016 // for all paths. It is not handled well by Register Allocator.
3017 Node* unique_in = n->in(1);
3018 assert(unique_in != NULL, "");
3019 uint cnt = n->req();
3020 for (uint i = 2; i < cnt; i++) {
3021 Node* m = n->in(i);
3022 assert(m != NULL, "");
3023 if (unique_in != m)
3024 unique_in = NULL;
3025 }
3026 if (unique_in != NULL) {
3027 n->subsume_by(unique_in, this);
3028 }
3029 }
3030 break;
3032 #endif
3034 #ifdef ASSERT
3035 case Op_CastII:
3036 // Verify that all range check dependent CastII nodes were removed.
3037 if (n->isa_CastII()->has_range_check()) {
3038 n->dump(3);
3039 assert(false, "Range check dependent CastII node was not removed");
3040 }
3041 break;
3042 #endif
3044 case Op_ModI:
3045 if (UseDivMod) {
3046 // Check if a%b and a/b both exist
3047 Node* d = n->find_similar(Op_DivI);
3048 if (d) {
3049 // Replace them with a fused divmod if supported
3050 if (Matcher::has_match_rule(Op_DivModI)) {
3051 DivModINode* divmod = DivModINode::make(this, n);
3052 d->subsume_by(divmod->div_proj(), this);
3053 n->subsume_by(divmod->mod_proj(), this);
3054 } else {
3055 // replace a%b with a-((a/b)*b)
3056 Node* mult = new (this) MulINode(d, d->in(2));
3057 Node* sub = new (this) SubINode(d->in(1), mult);
3058 n->subsume_by(sub, this);
3059 }
3060 }
3061 }
3062 break;
3064 case Op_ModL:
3065 if (UseDivMod) {
3066 // Check if a%b and a/b both exist
3067 Node* d = n->find_similar(Op_DivL);
3068 if (d) {
3069 // Replace them with a fused divmod if supported
3070 if (Matcher::has_match_rule(Op_DivModL)) {
3071 DivModLNode* divmod = DivModLNode::make(this, n);
3072 d->subsume_by(divmod->div_proj(), this);
3073 n->subsume_by(divmod->mod_proj(), this);
3074 } else {
3075 // replace a%b with a-((a/b)*b)
3076 Node* mult = new (this) MulLNode(d, d->in(2));
3077 Node* sub = new (this) SubLNode(d->in(1), mult);
3078 n->subsume_by(sub, this);
3079 }
3080 }
3081 }
3082 break;
3084 case Op_LoadVector:
3085 case Op_StoreVector:
3086 break;
3088 case Op_PackB:
3089 case Op_PackS:
3090 case Op_PackI:
3091 case Op_PackF:
3092 case Op_PackL:
3093 case Op_PackD:
3094 if (n->req()-1 > 2) {
3095 // Replace many operand PackNodes with a binary tree for matching
3096 PackNode* p = (PackNode*) n;
3097 Node* btp = p->binary_tree_pack(this, 1, n->req());
3098 n->subsume_by(btp, this);
3099 }
3100 break;
3101 case Op_Loop:
3102 case Op_CountedLoop:
3103 if (n->as_Loop()->is_inner_loop()) {
3104 frc.inc_inner_loop_count();
3105 }
3106 break;
3107 case Op_LShiftI:
3108 case Op_RShiftI:
3109 case Op_URShiftI:
3110 case Op_LShiftL:
3111 case Op_RShiftL:
3112 case Op_URShiftL:
3113 if (Matcher::need_masked_shift_count) {
3114 // The cpu's shift instructions don't restrict the count to the
3115 // lower 5/6 bits. We need to do the masking ourselves.
3116 Node* in2 = n->in(2);
3117 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3118 const TypeInt* t = in2->find_int_type();
3119 if (t != NULL && t->is_con()) {
3120 juint shift = t->get_con();
3121 if (shift > mask) { // Unsigned cmp
3122 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
3123 }
3124 } else {
3125 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3126 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3127 n->set_req(2, shift);
3128 }
3129 }
3130 if (in2->outcnt() == 0) { // Remove dead node
3131 in2->disconnect_inputs(NULL, this);
3132 }
3133 }
3134 break;
3135 case Op_MemBarStoreStore:
3136 case Op_MemBarRelease:
3137 // Break the link with AllocateNode: it is no longer useful and
3138 // confuses register allocation.
3139 if (n->req() > MemBarNode::Precedent) {
3140 n->set_req(MemBarNode::Precedent, top());
3141 }
3142 break;
3143 default:
3144 assert( !n->is_Call(), "" );
3145 assert( !n->is_Mem(), "" );
3146 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3147 break;
3148 }
3150 // Collect CFG split points
3151 if (n->is_MultiBranch())
3152 frc._tests.push(n);
3153 }
3155 //------------------------------final_graph_reshaping_walk---------------------
3156 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3157 // requires that the walk visits a node's inputs before visiting the node.
3158 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3159 ResourceArea *area = Thread::current()->resource_area();
3160 Unique_Node_List sfpt(area);
3162 frc._visited.set(root->_idx); // first, mark node as visited
3163 uint cnt = root->req();
3164 Node *n = root;
3165 uint i = 0;
3166 while (true) {
3167 if (i < cnt) {
3168 // Place all non-visited non-null inputs onto stack
3169 Node* m = n->in(i);
3170 ++i;
3171 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3172 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3173 // compute worst case interpreter size in case of a deoptimization
3174 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3176 sfpt.push(m);
3177 }
3178 cnt = m->req();
3179 nstack.push(n, i); // put on stack parent and next input's index
3180 n = m;
3181 i = 0;
3182 }
3183 } else {
3184 // Now do post-visit work
3185 final_graph_reshaping_impl( n, frc );
3186 if (nstack.is_empty())
3187 break; // finished
3188 n = nstack.node(); // Get node from stack
3189 cnt = n->req();
3190 i = nstack.index();
3191 nstack.pop(); // Shift to the next node on stack
3192 }
3193 }
3195 // Skip next transformation if compressed oops are not used.
3196 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3197 (!UseCompressedOops && !UseCompressedClassPointers))
3198 return;
3200 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3201 // It could be done for an uncommon traps or any safepoints/calls
3202 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3203 while (sfpt.size() > 0) {
3204 n = sfpt.pop();
3205 JVMState *jvms = n->as_SafePoint()->jvms();
3206 assert(jvms != NULL, "sanity");
3207 int start = jvms->debug_start();
3208 int end = n->req();
3209 bool is_uncommon = (n->is_CallStaticJava() &&
3210 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3211 for (int j = start; j < end; j++) {
3212 Node* in = n->in(j);
3213 if (in->is_DecodeNarrowPtr()) {
3214 bool safe_to_skip = true;
3215 if (!is_uncommon ) {
3216 // Is it safe to skip?
3217 for (uint i = 0; i < in->outcnt(); i++) {
3218 Node* u = in->raw_out(i);
3219 if (!u->is_SafePoint() ||
3220 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3221 safe_to_skip = false;
3222 }
3223 }
3224 }
3225 if (safe_to_skip) {
3226 n->set_req(j, in->in(1));
3227 }
3228 if (in->outcnt() == 0) {
3229 in->disconnect_inputs(NULL, this);
3230 }
3231 }
3232 }
3233 }
3234 }
3236 //------------------------------final_graph_reshaping--------------------------
3237 // Final Graph Reshaping.
3238 //
3239 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3240 // and not commoned up and forced early. Must come after regular
3241 // optimizations to avoid GVN undoing the cloning. Clone constant
3242 // inputs to Loop Phis; these will be split by the allocator anyways.
3243 // Remove Opaque nodes.
3244 // (2) Move last-uses by commutative operations to the left input to encourage
3245 // Intel update-in-place two-address operations and better register usage
3246 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3247 // calls canonicalizing them back.
3248 // (3) Count the number of double-precision FP ops, single-precision FP ops
3249 // and call sites. On Intel, we can get correct rounding either by
3250 // forcing singles to memory (requires extra stores and loads after each
3251 // FP bytecode) or we can set a rounding mode bit (requires setting and
3252 // clearing the mode bit around call sites). The mode bit is only used
3253 // if the relative frequency of single FP ops to calls is low enough.
3254 // This is a key transform for SPEC mpeg_audio.
3255 // (4) Detect infinite loops; blobs of code reachable from above but not
3256 // below. Several of the Code_Gen algorithms fail on such code shapes,
3257 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3258 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3259 // Detection is by looking for IfNodes where only 1 projection is
3260 // reachable from below or CatchNodes missing some targets.
3261 // (5) Assert for insane oop offsets in debug mode.
3263 bool Compile::final_graph_reshaping() {
3264 // an infinite loop may have been eliminated by the optimizer,
3265 // in which case the graph will be empty.
3266 if (root()->req() == 1) {
3267 record_method_not_compilable("trivial infinite loop");
3268 return true;
3269 }
3271 // Expensive nodes have their control input set to prevent the GVN
3272 // from freely commoning them. There's no GVN beyond this point so
3273 // no need to keep the control input. We want the expensive nodes to
3274 // be freely moved to the least frequent code path by gcm.
3275 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3276 for (int i = 0; i < expensive_count(); i++) {
3277 _expensive_nodes->at(i)->set_req(0, NULL);
3278 }
3280 Final_Reshape_Counts frc;
3282 // Visit everybody reachable!
3283 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3284 Node_Stack nstack(live_nodes() >> 1);
3285 final_graph_reshaping_walk(nstack, root(), frc);
3287 // Check for unreachable (from below) code (i.e., infinite loops).
3288 for( uint i = 0; i < frc._tests.size(); i++ ) {
3289 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3290 // Get number of CFG targets.
3291 // Note that PCTables include exception targets after calls.
3292 uint required_outcnt = n->required_outcnt();
3293 if (n->outcnt() != required_outcnt) {
3294 // Check for a few special cases. Rethrow Nodes never take the
3295 // 'fall-thru' path, so expected kids is 1 less.
3296 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3297 if (n->in(0)->in(0)->is_Call()) {
3298 CallNode *call = n->in(0)->in(0)->as_Call();
3299 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3300 required_outcnt--; // Rethrow always has 1 less kid
3301 } else if (call->req() > TypeFunc::Parms &&
3302 call->is_CallDynamicJava()) {
3303 // Check for null receiver. In such case, the optimizer has
3304 // detected that the virtual call will always result in a null
3305 // pointer exception. The fall-through projection of this CatchNode
3306 // will not be populated.
3307 Node *arg0 = call->in(TypeFunc::Parms);
3308 if (arg0->is_Type() &&
3309 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3310 required_outcnt--;
3311 }
3312 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3313 call->req() > TypeFunc::Parms+1 &&
3314 call->is_CallStaticJava()) {
3315 // Check for negative array length. In such case, the optimizer has
3316 // detected that the allocation attempt will always result in an
3317 // exception. There is no fall-through projection of this CatchNode .
3318 Node *arg1 = call->in(TypeFunc::Parms+1);
3319 if (arg1->is_Type() &&
3320 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3321 required_outcnt--;
3322 }
3323 }
3324 }
3325 }
3326 // Recheck with a better notion of 'required_outcnt'
3327 if (n->outcnt() != required_outcnt) {
3328 record_method_not_compilable("malformed control flow");
3329 return true; // Not all targets reachable!
3330 }
3331 }
3332 // Check that I actually visited all kids. Unreached kids
3333 // must be infinite loops.
3334 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3335 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3336 record_method_not_compilable("infinite loop");
3337 return true; // Found unvisited kid; must be unreach
3338 }
3339 }
3341 // If original bytecodes contained a mixture of floats and doubles
3342 // check if the optimizer has made it homogenous, item (3).
3343 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3344 frc.get_float_count() > 32 &&
3345 frc.get_double_count() == 0 &&
3346 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3347 set_24_bit_selection_and_mode( false, true );
3348 }
3350 set_java_calls(frc.get_java_call_count());
3351 set_inner_loops(frc.get_inner_loop_count());
3353 // No infinite loops, no reason to bail out.
3354 return false;
3355 }
3357 //-----------------------------too_many_traps----------------------------------
3358 // Report if there are too many traps at the current method and bci.
3359 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3360 bool Compile::too_many_traps(ciMethod* method,
3361 int bci,
3362 Deoptimization::DeoptReason reason) {
3363 ciMethodData* md = method->method_data();
3364 if (md->is_empty()) {
3365 // Assume the trap has not occurred, or that it occurred only
3366 // because of a transient condition during start-up in the interpreter.
3367 return false;
3368 }
3369 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3370 if (md->has_trap_at(bci, m, reason) != 0) {
3371 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3372 // Also, if there are multiple reasons, or if there is no per-BCI record,
3373 // assume the worst.
3374 if (log())
3375 log()->elem("observe trap='%s' count='%d'",
3376 Deoptimization::trap_reason_name(reason),
3377 md->trap_count(reason));
3378 return true;
3379 } else {
3380 // Ignore method/bci and see if there have been too many globally.
3381 return too_many_traps(reason, md);
3382 }
3383 }
3385 // Less-accurate variant which does not require a method and bci.
3386 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3387 ciMethodData* logmd) {
3388 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3389 // Too many traps globally.
3390 // Note that we use cumulative trap_count, not just md->trap_count.
3391 if (log()) {
3392 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3393 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3394 Deoptimization::trap_reason_name(reason),
3395 mcount, trap_count(reason));
3396 }
3397 return true;
3398 } else {
3399 // The coast is clear.
3400 return false;
3401 }
3402 }
3404 //--------------------------too_many_recompiles--------------------------------
3405 // Report if there are too many recompiles at the current method and bci.
3406 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3407 // Is not eager to return true, since this will cause the compiler to use
3408 // Action_none for a trap point, to avoid too many recompilations.
3409 bool Compile::too_many_recompiles(ciMethod* method,
3410 int bci,
3411 Deoptimization::DeoptReason reason) {
3412 ciMethodData* md = method->method_data();
3413 if (md->is_empty()) {
3414 // Assume the trap has not occurred, or that it occurred only
3415 // because of a transient condition during start-up in the interpreter.
3416 return false;
3417 }
3418 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3419 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3420 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3421 Deoptimization::DeoptReason per_bc_reason
3422 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3423 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3424 if ((per_bc_reason == Deoptimization::Reason_none
3425 || md->has_trap_at(bci, m, reason) != 0)
3426 // The trap frequency measure we care about is the recompile count:
3427 && md->trap_recompiled_at(bci, m)
3428 && md->overflow_recompile_count() >= bc_cutoff) {
3429 // Do not emit a trap here if it has already caused recompilations.
3430 // Also, if there are multiple reasons, or if there is no per-BCI record,
3431 // assume the worst.
3432 if (log())
3433 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3434 Deoptimization::trap_reason_name(reason),
3435 md->trap_count(reason),
3436 md->overflow_recompile_count());
3437 return true;
3438 } else if (trap_count(reason) != 0
3439 && decompile_count() >= m_cutoff) {
3440 // Too many recompiles globally, and we have seen this sort of trap.
3441 // Use cumulative decompile_count, not just md->decompile_count.
3442 if (log())
3443 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3444 Deoptimization::trap_reason_name(reason),
3445 md->trap_count(reason), trap_count(reason),
3446 md->decompile_count(), decompile_count());
3447 return true;
3448 } else {
3449 // The coast is clear.
3450 return false;
3451 }
3452 }
3454 // Compute when not to trap. Used by matching trap based nodes and
3455 // NullCheck optimization.
3456 void Compile::set_allowed_deopt_reasons() {
3457 _allowed_reasons = 0;
3458 if (is_method_compilation()) {
3459 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3460 assert(rs < BitsPerInt, "recode bit map");
3461 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3462 _allowed_reasons |= nth_bit(rs);
3463 }
3464 }
3465 }
3466 }
3468 #ifndef PRODUCT
3469 //------------------------------verify_graph_edges---------------------------
3470 // Walk the Graph and verify that there is a one-to-one correspondence
3471 // between Use-Def edges and Def-Use edges in the graph.
3472 void Compile::verify_graph_edges(bool no_dead_code) {
3473 if (VerifyGraphEdges) {
3474 ResourceArea *area = Thread::current()->resource_area();
3475 Unique_Node_List visited(area);
3476 // Call recursive graph walk to check edges
3477 _root->verify_edges(visited);
3478 if (no_dead_code) {
3479 // Now make sure that no visited node is used by an unvisited node.
3480 bool dead_nodes = 0;
3481 Unique_Node_List checked(area);
3482 while (visited.size() > 0) {
3483 Node* n = visited.pop();
3484 checked.push(n);
3485 for (uint i = 0; i < n->outcnt(); i++) {
3486 Node* use = n->raw_out(i);
3487 if (checked.member(use)) continue; // already checked
3488 if (visited.member(use)) continue; // already in the graph
3489 if (use->is_Con()) continue; // a dead ConNode is OK
3490 // At this point, we have found a dead node which is DU-reachable.
3491 if (dead_nodes++ == 0)
3492 tty->print_cr("*** Dead nodes reachable via DU edges:");
3493 use->dump(2);
3494 tty->print_cr("---");
3495 checked.push(use); // No repeats; pretend it is now checked.
3496 }
3497 }
3498 assert(dead_nodes == 0, "using nodes must be reachable from root");
3499 }
3500 }
3501 }
3503 // Verify GC barriers consistency
3504 // Currently supported:
3505 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3506 void Compile::verify_barriers() {
3507 if (UseG1GC) {
3508 // Verify G1 pre-barriers
3509 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3511 ResourceArea *area = Thread::current()->resource_area();
3512 Unique_Node_List visited(area);
3513 Node_List worklist(area);
3514 // We're going to walk control flow backwards starting from the Root
3515 worklist.push(_root);
3516 while (worklist.size() > 0) {
3517 Node* x = worklist.pop();
3518 if (x == NULL || x == top()) continue;
3519 if (visited.member(x)) {
3520 continue;
3521 } else {
3522 visited.push(x);
3523 }
3525 if (x->is_Region()) {
3526 for (uint i = 1; i < x->req(); i++) {
3527 worklist.push(x->in(i));
3528 }
3529 } else {
3530 worklist.push(x->in(0));
3531 // We are looking for the pattern:
3532 // /->ThreadLocal
3533 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3534 // \->ConI(0)
3535 // We want to verify that the If and the LoadB have the same control
3536 // See GraphKit::g1_write_barrier_pre()
3537 if (x->is_If()) {
3538 IfNode *iff = x->as_If();
3539 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3540 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3541 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3542 && cmp->in(1)->is_Load()) {
3543 LoadNode* load = cmp->in(1)->as_Load();
3544 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3545 && load->in(2)->in(3)->is_Con()
3546 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3548 Node* if_ctrl = iff->in(0);
3549 Node* load_ctrl = load->in(0);
3551 if (if_ctrl != load_ctrl) {
3552 // Skip possible CProj->NeverBranch in infinite loops
3553 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3554 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3555 if_ctrl = if_ctrl->in(0)->in(0);
3556 }
3557 }
3558 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3559 }
3560 }
3561 }
3562 }
3563 }
3564 }
3565 }
3566 }
3568 #endif
3570 // The Compile object keeps track of failure reasons separately from the ciEnv.
3571 // This is required because there is not quite a 1-1 relation between the
3572 // ciEnv and its compilation task and the Compile object. Note that one
3573 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3574 // to backtrack and retry without subsuming loads. Other than this backtracking
3575 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3576 // by the logic in C2Compiler.
3577 void Compile::record_failure(const char* reason) {
3578 if (log() != NULL) {
3579 log()->elem("failure reason='%s' phase='compile'", reason);
3580 }
3581 if (_failure_reason == NULL) {
3582 // Record the first failure reason.
3583 _failure_reason = reason;
3584 }
3586 EventCompilerFailure event;
3587 if (event.should_commit()) {
3588 event.set_compileID(Compile::compile_id());
3589 event.set_failure(reason);
3590 event.commit();
3591 }
3593 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3594 C->print_method(PHASE_FAILURE);
3595 }
3596 _root = NULL; // flush the graph, too
3597 }
3599 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3600 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3601 _phase_name(name), _dolog(dolog)
3602 {
3603 if (dolog) {
3604 C = Compile::current();
3605 _log = C->log();
3606 } else {
3607 C = NULL;
3608 _log = NULL;
3609 }
3610 if (_log != NULL) {
3611 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3612 _log->stamp();
3613 _log->end_head();
3614 }
3615 }
3617 Compile::TracePhase::~TracePhase() {
3619 C = Compile::current();
3620 if (_dolog) {
3621 _log = C->log();
3622 } else {
3623 _log = NULL;
3624 }
3626 #ifdef ASSERT
3627 if (PrintIdealNodeCount) {
3628 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3629 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3630 }
3632 if (VerifyIdealNodeCount) {
3633 Compile::current()->print_missing_nodes();
3634 }
3635 #endif
3637 if (_log != NULL) {
3638 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3639 }
3640 }
3642 //=============================================================================
3643 // Two Constant's are equal when the type and the value are equal.
3644 bool Compile::Constant::operator==(const Constant& other) {
3645 if (type() != other.type() ) return false;
3646 if (can_be_reused() != other.can_be_reused()) return false;
3647 // For floating point values we compare the bit pattern.
3648 switch (type()) {
3649 case T_FLOAT: return (_v._value.i == other._v._value.i);
3650 case T_LONG:
3651 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3652 case T_OBJECT:
3653 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3654 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3655 case T_METADATA: return (_v._metadata == other._v._metadata);
3656 default: ShouldNotReachHere();
3657 }
3658 return false;
3659 }
3661 static int type_to_size_in_bytes(BasicType t) {
3662 switch (t) {
3663 case T_LONG: return sizeof(jlong );
3664 case T_FLOAT: return sizeof(jfloat );
3665 case T_DOUBLE: return sizeof(jdouble);
3666 case T_METADATA: return sizeof(Metadata*);
3667 // We use T_VOID as marker for jump-table entries (labels) which
3668 // need an internal word relocation.
3669 case T_VOID:
3670 case T_ADDRESS:
3671 case T_OBJECT: return sizeof(jobject);
3672 }
3674 ShouldNotReachHere();
3675 return -1;
3676 }
3678 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3679 // sort descending
3680 if (a->freq() > b->freq()) return -1;
3681 if (a->freq() < b->freq()) return 1;
3682 return 0;
3683 }
3685 void Compile::ConstantTable::calculate_offsets_and_size() {
3686 // First, sort the array by frequencies.
3687 _constants.sort(qsort_comparator);
3689 #ifdef ASSERT
3690 // Make sure all jump-table entries were sorted to the end of the
3691 // array (they have a negative frequency).
3692 bool found_void = false;
3693 for (int i = 0; i < _constants.length(); i++) {
3694 Constant con = _constants.at(i);
3695 if (con.type() == T_VOID)
3696 found_void = true; // jump-tables
3697 else
3698 assert(!found_void, "wrong sorting");
3699 }
3700 #endif
3702 int offset = 0;
3703 for (int i = 0; i < _constants.length(); i++) {
3704 Constant* con = _constants.adr_at(i);
3706 // Align offset for type.
3707 int typesize = type_to_size_in_bytes(con->type());
3708 offset = align_size_up(offset, typesize);
3709 con->set_offset(offset); // set constant's offset
3711 if (con->type() == T_VOID) {
3712 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3713 offset = offset + typesize * n->outcnt(); // expand jump-table
3714 } else {
3715 offset = offset + typesize;
3716 }
3717 }
3719 // Align size up to the next section start (which is insts; see
3720 // CodeBuffer::align_at_start).
3721 assert(_size == -1, "already set?");
3722 _size = align_size_up(offset, CodeEntryAlignment);
3723 }
3725 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3726 MacroAssembler _masm(&cb);
3727 for (int i = 0; i < _constants.length(); i++) {
3728 Constant con = _constants.at(i);
3729 address constant_addr = NULL;
3730 switch (con.type()) {
3731 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3732 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3733 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3734 case T_OBJECT: {
3735 jobject obj = con.get_jobject();
3736 int oop_index = _masm.oop_recorder()->find_index(obj);
3737 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3738 break;
3739 }
3740 case T_ADDRESS: {
3741 address addr = (address) con.get_jobject();
3742 constant_addr = _masm.address_constant(addr);
3743 break;
3744 }
3745 // We use T_VOID as marker for jump-table entries (labels) which
3746 // need an internal word relocation.
3747 case T_VOID: {
3748 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3749 // Fill the jump-table with a dummy word. The real value is
3750 // filled in later in fill_jump_table.
3751 address dummy = (address) n;
3752 constant_addr = _masm.address_constant(dummy);
3753 // Expand jump-table
3754 for (uint i = 1; i < n->outcnt(); i++) {
3755 address temp_addr = _masm.address_constant(dummy + i);
3756 assert(temp_addr, "consts section too small");
3757 }
3758 break;
3759 }
3760 case T_METADATA: {
3761 Metadata* obj = con.get_metadata();
3762 int metadata_index = _masm.oop_recorder()->find_index(obj);
3763 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3764 break;
3765 }
3766 default: ShouldNotReachHere();
3767 }
3768 assert(constant_addr, "consts section too small");
3769 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3770 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())));
3771 }
3772 }
3774 int Compile::ConstantTable::find_offset(Constant& con) const {
3775 int idx = _constants.find(con);
3776 assert(idx != -1, "constant must be in constant table");
3777 int offset = _constants.at(idx).offset();
3778 assert(offset != -1, "constant table not emitted yet?");
3779 return offset;
3780 }
3782 void Compile::ConstantTable::add(Constant& con) {
3783 if (con.can_be_reused()) {
3784 int idx = _constants.find(con);
3785 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3786 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3787 return;
3788 }
3789 }
3790 (void) _constants.append(con);
3791 }
3793 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3794 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3795 Constant con(type, value, b->_freq);
3796 add(con);
3797 return con;
3798 }
3800 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3801 Constant con(metadata);
3802 add(con);
3803 return con;
3804 }
3806 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3807 jvalue value;
3808 BasicType type = oper->type()->basic_type();
3809 switch (type) {
3810 case T_LONG: value.j = oper->constantL(); break;
3811 case T_FLOAT: value.f = oper->constantF(); break;
3812 case T_DOUBLE: value.d = oper->constantD(); break;
3813 case T_OBJECT:
3814 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3815 case T_METADATA: return add((Metadata*)oper->constant()); break;
3816 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3817 }
3818 return add(n, type, value);
3819 }
3821 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3822 jvalue value;
3823 // We can use the node pointer here to identify the right jump-table
3824 // as this method is called from Compile::Fill_buffer right before
3825 // the MachNodes are emitted and the jump-table is filled (means the
3826 // MachNode pointers do not change anymore).
3827 value.l = (jobject) n;
3828 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3829 add(con);
3830 return con;
3831 }
3833 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3834 // If called from Compile::scratch_emit_size do nothing.
3835 if (Compile::current()->in_scratch_emit_size()) return;
3837 assert(labels.is_nonempty(), "must be");
3838 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3840 // Since MachConstantNode::constant_offset() also contains
3841 // table_base_offset() we need to subtract the table_base_offset()
3842 // to get the plain offset into the constant table.
3843 int offset = n->constant_offset() - table_base_offset();
3845 MacroAssembler _masm(&cb);
3846 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3848 for (uint i = 0; i < n->outcnt(); i++) {
3849 address* constant_addr = &jump_table_base[i];
3850 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)));
3851 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3852 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3853 }
3854 }
3856 void Compile::dump_inlining() {
3857 if (print_inlining() || print_intrinsics()) {
3858 // Print inlining message for candidates that we couldn't inline
3859 // for lack of space or non constant receiver
3860 for (int i = 0; i < _late_inlines.length(); i++) {
3861 CallGenerator* cg = _late_inlines.at(i);
3862 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3863 }
3864 Unique_Node_List useful;
3865 useful.push(root());
3866 for (uint next = 0; next < useful.size(); ++next) {
3867 Node* n = useful.at(next);
3868 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3869 CallNode* call = n->as_Call();
3870 CallGenerator* cg = call->generator();
3871 cg->print_inlining_late("receiver not constant");
3872 }
3873 uint max = n->len();
3874 for ( uint i = 0; i < max; ++i ) {
3875 Node *m = n->in(i);
3876 if ( m == NULL ) continue;
3877 useful.push(m);
3878 }
3879 }
3880 for (int i = 0; i < _print_inlining_list->length(); i++) {
3881 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
3882 }
3883 }
3884 }
3886 // Dump inlining replay data to the stream.
3887 // Don't change thread state and acquire any locks.
3888 void Compile::dump_inline_data(outputStream* out) {
3889 InlineTree* inl_tree = ilt();
3890 if (inl_tree != NULL) {
3891 out->print(" inline %d", inl_tree->count());
3892 inl_tree->dump_replay_data(out);
3893 }
3894 }
3896 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3897 if (n1->Opcode() < n2->Opcode()) return -1;
3898 else if (n1->Opcode() > n2->Opcode()) return 1;
3900 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()));
3901 for (uint i = 1; i < n1->req(); i++) {
3902 if (n1->in(i) < n2->in(i)) return -1;
3903 else if (n1->in(i) > n2->in(i)) return 1;
3904 }
3906 return 0;
3907 }
3909 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3910 Node* n1 = *n1p;
3911 Node* n2 = *n2p;
3913 return cmp_expensive_nodes(n1, n2);
3914 }
3916 void Compile::sort_expensive_nodes() {
3917 if (!expensive_nodes_sorted()) {
3918 _expensive_nodes->sort(cmp_expensive_nodes);
3919 }
3920 }
3922 bool Compile::expensive_nodes_sorted() const {
3923 for (int i = 1; i < _expensive_nodes->length(); i++) {
3924 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3925 return false;
3926 }
3927 }
3928 return true;
3929 }
3931 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3932 if (_expensive_nodes->length() == 0) {
3933 return false;
3934 }
3936 assert(OptimizeExpensiveOps, "optimization off?");
3938 // Take this opportunity to remove dead nodes from the list
3939 int j = 0;
3940 for (int i = 0; i < _expensive_nodes->length(); i++) {
3941 Node* n = _expensive_nodes->at(i);
3942 if (!n->is_unreachable(igvn)) {
3943 assert(n->is_expensive(), "should be expensive");
3944 _expensive_nodes->at_put(j, n);
3945 j++;
3946 }
3947 }
3948 _expensive_nodes->trunc_to(j);
3950 // Then sort the list so that similar nodes are next to each other
3951 // and check for at least two nodes of identical kind with same data
3952 // inputs.
3953 sort_expensive_nodes();
3955 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3956 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3957 return true;
3958 }
3959 }
3961 return false;
3962 }
3964 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3965 if (_expensive_nodes->length() == 0) {
3966 return;
3967 }
3969 assert(OptimizeExpensiveOps, "optimization off?");
3971 // Sort to bring similar nodes next to each other and clear the
3972 // control input of nodes for which there's only a single copy.
3973 sort_expensive_nodes();
3975 int j = 0;
3976 int identical = 0;
3977 int i = 0;
3978 for (; i < _expensive_nodes->length()-1; i++) {
3979 assert(j <= i, "can't write beyond current index");
3980 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3981 identical++;
3982 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3983 continue;
3984 }
3985 if (identical > 0) {
3986 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3987 identical = 0;
3988 } else {
3989 Node* n = _expensive_nodes->at(i);
3990 igvn.hash_delete(n);
3991 n->set_req(0, NULL);
3992 igvn.hash_insert(n);
3993 }
3994 }
3995 if (identical > 0) {
3996 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3997 } else if (_expensive_nodes->length() >= 1) {
3998 Node* n = _expensive_nodes->at(i);
3999 igvn.hash_delete(n);
4000 n->set_req(0, NULL);
4001 igvn.hash_insert(n);
4002 }
4003 _expensive_nodes->trunc_to(j);
4004 }
4006 void Compile::add_expensive_node(Node * n) {
4007 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4008 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4009 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4010 if (OptimizeExpensiveOps) {
4011 _expensive_nodes->append(n);
4012 } else {
4013 // Clear control input and let IGVN optimize expensive nodes if
4014 // OptimizeExpensiveOps is off.
4015 n->set_req(0, NULL);
4016 }
4017 }
4019 /**
4020 * Remove the speculative part of types and clean up the graph
4021 */
4022 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4023 if (UseTypeSpeculation) {
4024 Unique_Node_List worklist;
4025 worklist.push(root());
4026 int modified = 0;
4027 // Go over all type nodes that carry a speculative type, drop the
4028 // speculative part of the type and enqueue the node for an igvn
4029 // which may optimize it out.
4030 for (uint next = 0; next < worklist.size(); ++next) {
4031 Node *n = worklist.at(next);
4032 if (n->is_Type()) {
4033 TypeNode* tn = n->as_Type();
4034 const Type* t = tn->type();
4035 const Type* t_no_spec = t->remove_speculative();
4036 if (t_no_spec != t) {
4037 bool in_hash = igvn.hash_delete(n);
4038 assert(in_hash, "node should be in igvn hash table");
4039 tn->set_type(t_no_spec);
4040 igvn.hash_insert(n);
4041 igvn._worklist.push(n); // give it a chance to go away
4042 modified++;
4043 }
4044 }
4045 uint max = n->len();
4046 for( uint i = 0; i < max; ++i ) {
4047 Node *m = n->in(i);
4048 if (not_a_node(m)) continue;
4049 worklist.push(m);
4050 }
4051 }
4052 // Drop the speculative part of all types in the igvn's type table
4053 igvn.remove_speculative_types();
4054 if (modified > 0) {
4055 igvn.optimize();
4056 }
4057 #ifdef ASSERT
4058 // Verify that after the IGVN is over no speculative type has resurfaced
4059 worklist.clear();
4060 worklist.push(root());
4061 for (uint next = 0; next < worklist.size(); ++next) {
4062 Node *n = worklist.at(next);
4063 const Type* t = igvn.type_or_null(n);
4064 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4065 if (n->is_Type()) {
4066 t = n->as_Type()->type();
4067 assert(t == t->remove_speculative(), "no more speculative types");
4068 }
4069 uint max = n->len();
4070 for( uint i = 0; i < max; ++i ) {
4071 Node *m = n->in(i);
4072 if (not_a_node(m)) continue;
4073 worklist.push(m);
4074 }
4075 }
4076 igvn.check_no_speculative_types();
4077 #endif
4078 }
4079 }
4081 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4082 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4083 if (ctrl != NULL) {
4084 // Express control dependency by a CastII node with a narrow type.
4085 value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */);
4086 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4087 // node from floating above the range check during loop optimizations. Otherwise, the
4088 // ConvI2L node may be eliminated independently of the range check, causing the data path
4089 // to become TOP while the control path is still there (although it's unreachable).
4090 value->set_req(0, ctrl);
4091 // Save CastII node to remove it after loop optimizations.
4092 phase->C->add_range_check_cast(value);
4093 value = phase->transform(value);
4094 }
4095 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4096 return phase->transform(new (phase->C) ConvI2LNode(value, ltype));
4097 }
4099 // Auxiliary method to support randomized stressing/fuzzing.
4100 //
4101 // This method can be called the arbitrary number of times, with current count
4102 // as the argument. The logic allows selecting a single candidate from the
4103 // running list of candidates as follows:
4104 // int count = 0;
4105 // Cand* selected = null;
4106 // while(cand = cand->next()) {
4107 // if (randomized_select(++count)) {
4108 // selected = cand;
4109 // }
4110 // }
4111 //
4112 // Including count equalizes the chances any candidate is "selected".
4113 // This is useful when we don't have the complete list of candidates to choose
4114 // from uniformly. In this case, we need to adjust the randomicity of the
4115 // selection, or else we will end up biasing the selection towards the latter
4116 // candidates.
4117 //
4118 // Quick back-envelope calculation shows that for the list of n candidates
4119 // the equal probability for the candidate to persist as "best" can be
4120 // achieved by replacing it with "next" k-th candidate with the probability
4121 // of 1/k. It can be easily shown that by the end of the run, the
4122 // probability for any candidate is converged to 1/n, thus giving the
4123 // uniform distribution among all the candidates.
4124 //
4125 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4126 #define RANDOMIZED_DOMAIN_POW 29
4127 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4128 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4129 bool Compile::randomized_select(int count) {
4130 assert(count > 0, "only positive");
4131 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4132 }