Tue, 14 Jan 2014 17:46:48 -0800
8002074: Support for AES on SPARC
Summary: Add intrinsics/stub routines support for single-block and multi-block (as used by Cipher Block Chaining mode) AES encryption and decryption operations on the SPARC platform.
Reviewed-by: kvn, roland
Contributed-by: shrinivas.joshi@oracle.com
1 /*
2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "compiler/disassembler.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/block.hpp"
37 #include "opto/c2compiler.hpp"
38 #include "opto/callGenerator.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/cfgnode.hpp"
41 #include "opto/chaitin.hpp"
42 #include "opto/compile.hpp"
43 #include "opto/connode.hpp"
44 #include "opto/divnode.hpp"
45 #include "opto/escape.hpp"
46 #include "opto/idealGraphPrinter.hpp"
47 #include "opto/loopnode.hpp"
48 #include "opto/machnode.hpp"
49 #include "opto/macro.hpp"
50 #include "opto/matcher.hpp"
51 #include "opto/mathexactnode.hpp"
52 #include "opto/memnode.hpp"
53 #include "opto/mulnode.hpp"
54 #include "opto/node.hpp"
55 #include "opto/opcodes.hpp"
56 #include "opto/output.hpp"
57 #include "opto/parse.hpp"
58 #include "opto/phaseX.hpp"
59 #include "opto/rootnode.hpp"
60 #include "opto/runtime.hpp"
61 #include "opto/stringopts.hpp"
62 #include "opto/type.hpp"
63 #include "opto/vectornode.hpp"
64 #include "runtime/arguments.hpp"
65 #include "runtime/signature.hpp"
66 #include "runtime/stubRoutines.hpp"
67 #include "runtime/timer.hpp"
68 #include "trace/tracing.hpp"
69 #include "utilities/copy.hpp"
70 #ifdef TARGET_ARCH_MODEL_x86_32
71 # include "adfiles/ad_x86_32.hpp"
72 #endif
73 #ifdef TARGET_ARCH_MODEL_x86_64
74 # include "adfiles/ad_x86_64.hpp"
75 #endif
76 #ifdef TARGET_ARCH_MODEL_sparc
77 # include "adfiles/ad_sparc.hpp"
78 #endif
79 #ifdef TARGET_ARCH_MODEL_zero
80 # include "adfiles/ad_zero.hpp"
81 #endif
82 #ifdef TARGET_ARCH_MODEL_arm
83 # include "adfiles/ad_arm.hpp"
84 #endif
85 #ifdef TARGET_ARCH_MODEL_ppc
86 # include "adfiles/ad_ppc.hpp"
87 #endif
90 // -------------------- Compile::mach_constant_base_node -----------------------
91 // Constant table base node singleton.
92 MachConstantBaseNode* Compile::mach_constant_base_node() {
93 if (_mach_constant_base_node == NULL) {
94 _mach_constant_base_node = new (C) MachConstantBaseNode();
95 _mach_constant_base_node->add_req(C->root());
96 }
97 return _mach_constant_base_node;
98 }
101 /// Support for intrinsics.
103 // Return the index at which m must be inserted (or already exists).
104 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
105 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
106 #ifdef ASSERT
107 for (int i = 1; i < _intrinsics->length(); i++) {
108 CallGenerator* cg1 = _intrinsics->at(i-1);
109 CallGenerator* cg2 = _intrinsics->at(i);
110 assert(cg1->method() != cg2->method()
111 ? cg1->method() < cg2->method()
112 : cg1->is_virtual() < cg2->is_virtual(),
113 "compiler intrinsics list must stay sorted");
114 }
115 #endif
116 // Binary search sorted list, in decreasing intervals [lo, hi].
117 int lo = 0, hi = _intrinsics->length()-1;
118 while (lo <= hi) {
119 int mid = (uint)(hi + lo) / 2;
120 ciMethod* mid_m = _intrinsics->at(mid)->method();
121 if (m < mid_m) {
122 hi = mid-1;
123 } else if (m > mid_m) {
124 lo = mid+1;
125 } else {
126 // look at minor sort key
127 bool mid_virt = _intrinsics->at(mid)->is_virtual();
128 if (is_virtual < mid_virt) {
129 hi = mid-1;
130 } else if (is_virtual > mid_virt) {
131 lo = mid+1;
132 } else {
133 return mid; // exact match
134 }
135 }
136 }
137 return lo; // inexact match
138 }
140 void Compile::register_intrinsic(CallGenerator* cg) {
141 if (_intrinsics == NULL) {
142 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
143 }
144 // This code is stolen from ciObjectFactory::insert.
145 // Really, GrowableArray should have methods for
146 // insert_at, remove_at, and binary_search.
147 int len = _intrinsics->length();
148 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
149 if (index == len) {
150 _intrinsics->append(cg);
151 } else {
152 #ifdef ASSERT
153 CallGenerator* oldcg = _intrinsics->at(index);
154 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
155 #endif
156 _intrinsics->append(_intrinsics->at(len-1));
157 int pos;
158 for (pos = len-2; pos >= index; pos--) {
159 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
160 }
161 _intrinsics->at_put(index, cg);
162 }
163 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
164 }
166 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
167 assert(m->is_loaded(), "don't try this on unloaded methods");
168 if (_intrinsics != NULL) {
169 int index = intrinsic_insertion_index(m, is_virtual);
170 if (index < _intrinsics->length()
171 && _intrinsics->at(index)->method() == m
172 && _intrinsics->at(index)->is_virtual() == is_virtual) {
173 return _intrinsics->at(index);
174 }
175 }
176 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
177 if (m->intrinsic_id() != vmIntrinsics::_none &&
178 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
179 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
180 if (cg != NULL) {
181 // Save it for next time:
182 register_intrinsic(cg);
183 return cg;
184 } else {
185 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
186 }
187 }
188 return NULL;
189 }
191 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
192 // in library_call.cpp.
195 #ifndef PRODUCT
196 // statistics gathering...
198 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
199 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
201 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
202 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
203 int oflags = _intrinsic_hist_flags[id];
204 assert(flags != 0, "what happened?");
205 if (is_virtual) {
206 flags |= _intrinsic_virtual;
207 }
208 bool changed = (flags != oflags);
209 if ((flags & _intrinsic_worked) != 0) {
210 juint count = (_intrinsic_hist_count[id] += 1);
211 if (count == 1) {
212 changed = true; // first time
213 }
214 // increment the overall count also:
215 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
216 }
217 if (changed) {
218 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
219 // Something changed about the intrinsic's virtuality.
220 if ((flags & _intrinsic_virtual) != 0) {
221 // This is the first use of this intrinsic as a virtual call.
222 if (oflags != 0) {
223 // We already saw it as a non-virtual, so note both cases.
224 flags |= _intrinsic_both;
225 }
226 } else if ((oflags & _intrinsic_both) == 0) {
227 // This is the first use of this intrinsic as a non-virtual
228 flags |= _intrinsic_both;
229 }
230 }
231 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
232 }
233 // update the overall flags also:
234 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
235 return changed;
236 }
238 static char* format_flags(int flags, char* buf) {
239 buf[0] = 0;
240 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
241 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
242 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
243 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
244 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
245 if (buf[0] == 0) strcat(buf, ",");
246 assert(buf[0] == ',', "must be");
247 return &buf[1];
248 }
250 void Compile::print_intrinsic_statistics() {
251 char flagsbuf[100];
252 ttyLocker ttyl;
253 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
254 tty->print_cr("Compiler intrinsic usage:");
255 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
256 if (total == 0) total = 1; // avoid div0 in case of no successes
257 #define PRINT_STAT_LINE(name, c, f) \
258 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
259 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
260 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
261 int flags = _intrinsic_hist_flags[id];
262 juint count = _intrinsic_hist_count[id];
263 if ((flags | count) != 0) {
264 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
265 }
266 }
267 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
268 if (xtty != NULL) xtty->tail("statistics");
269 }
271 void Compile::print_statistics() {
272 { ttyLocker ttyl;
273 if (xtty != NULL) xtty->head("statistics type='opto'");
274 Parse::print_statistics();
275 PhaseCCP::print_statistics();
276 PhaseRegAlloc::print_statistics();
277 Scheduling::print_statistics();
278 PhasePeephole::print_statistics();
279 PhaseIdealLoop::print_statistics();
280 if (xtty != NULL) xtty->tail("statistics");
281 }
282 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
283 // put this under its own <statistics> element.
284 print_intrinsic_statistics();
285 }
286 }
287 #endif //PRODUCT
289 // Support for bundling info
290 Bundle* Compile::node_bundling(const Node *n) {
291 assert(valid_bundle_info(n), "oob");
292 return &_node_bundling_base[n->_idx];
293 }
295 bool Compile::valid_bundle_info(const Node *n) {
296 return (_node_bundling_limit > n->_idx);
297 }
300 void Compile::gvn_replace_by(Node* n, Node* nn) {
301 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
302 Node* use = n->last_out(i);
303 bool is_in_table = initial_gvn()->hash_delete(use);
304 uint uses_found = 0;
305 for (uint j = 0; j < use->len(); j++) {
306 if (use->in(j) == n) {
307 if (j < use->req())
308 use->set_req(j, nn);
309 else
310 use->set_prec(j, nn);
311 uses_found++;
312 }
313 }
314 if (is_in_table) {
315 // reinsert into table
316 initial_gvn()->hash_find_insert(use);
317 }
318 record_for_igvn(use);
319 i -= uses_found; // we deleted 1 or more copies of this edge
320 }
321 }
324 static inline bool not_a_node(const Node* n) {
325 if (n == NULL) return true;
326 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
327 if (*(address*)n == badAddress) return true; // kill by Node::destruct
328 return false;
329 }
331 // Identify all nodes that are reachable from below, useful.
332 // Use breadth-first pass that records state in a Unique_Node_List,
333 // recursive traversal is slower.
334 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
335 int estimated_worklist_size = unique();
336 useful.map( estimated_worklist_size, NULL ); // preallocate space
338 // Initialize worklist
339 if (root() != NULL) { useful.push(root()); }
340 // If 'top' is cached, declare it useful to preserve cached node
341 if( cached_top_node() ) { useful.push(cached_top_node()); }
343 // Push all useful nodes onto the list, breadthfirst
344 for( uint next = 0; next < useful.size(); ++next ) {
345 assert( next < unique(), "Unique useful nodes < total nodes");
346 Node *n = useful.at(next);
347 uint max = n->len();
348 for( uint i = 0; i < max; ++i ) {
349 Node *m = n->in(i);
350 if (not_a_node(m)) continue;
351 useful.push(m);
352 }
353 }
354 }
356 // Update dead_node_list with any missing dead nodes using useful
357 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
358 void Compile::update_dead_node_list(Unique_Node_List &useful) {
359 uint max_idx = unique();
360 VectorSet& useful_node_set = useful.member_set();
362 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
363 // If node with index node_idx is not in useful set,
364 // mark it as dead in dead node list.
365 if (! useful_node_set.test(node_idx) ) {
366 record_dead_node(node_idx);
367 }
368 }
369 }
371 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
372 int shift = 0;
373 for (int i = 0; i < inlines->length(); i++) {
374 CallGenerator* cg = inlines->at(i);
375 CallNode* call = cg->call_node();
376 if (shift > 0) {
377 inlines->at_put(i-shift, cg);
378 }
379 if (!useful.member(call)) {
380 shift++;
381 }
382 }
383 inlines->trunc_to(inlines->length()-shift);
384 }
386 // Disconnect all useless nodes by disconnecting those at the boundary.
387 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
388 uint next = 0;
389 while (next < useful.size()) {
390 Node *n = useful.at(next++);
391 // Use raw traversal of out edges since this code removes out edges
392 int max = n->outcnt();
393 for (int j = 0; j < max; ++j) {
394 Node* child = n->raw_out(j);
395 if (! useful.member(child)) {
396 assert(!child->is_top() || child != top(),
397 "If top is cached in Compile object it is in useful list");
398 // Only need to remove this out-edge to the useless node
399 n->raw_del_out(j);
400 --j;
401 --max;
402 }
403 }
404 if (n->outcnt() == 1 && n->has_special_unique_user()) {
405 record_for_igvn(n->unique_out());
406 }
407 }
408 // Remove useless macro and predicate opaq nodes
409 for (int i = C->macro_count()-1; i >= 0; i--) {
410 Node* n = C->macro_node(i);
411 if (!useful.member(n)) {
412 remove_macro_node(n);
413 }
414 }
415 // Remove useless expensive node
416 for (int i = C->expensive_count()-1; i >= 0; i--) {
417 Node* n = C->expensive_node(i);
418 if (!useful.member(n)) {
419 remove_expensive_node(n);
420 }
421 }
422 // clean up the late inline lists
423 remove_useless_late_inlines(&_string_late_inlines, useful);
424 remove_useless_late_inlines(&_boxing_late_inlines, useful);
425 remove_useless_late_inlines(&_late_inlines, useful);
426 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
427 }
429 //------------------------------frame_size_in_words-----------------------------
430 // frame_slots in units of words
431 int Compile::frame_size_in_words() const {
432 // shift is 0 in LP32 and 1 in LP64
433 const int shift = (LogBytesPerWord - LogBytesPerInt);
434 int words = _frame_slots >> shift;
435 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
436 return words;
437 }
439 // ============================================================================
440 //------------------------------CompileWrapper---------------------------------
441 class CompileWrapper : public StackObj {
442 Compile *const _compile;
443 public:
444 CompileWrapper(Compile* compile);
446 ~CompileWrapper();
447 };
449 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
450 // the Compile* pointer is stored in the current ciEnv:
451 ciEnv* env = compile->env();
452 assert(env == ciEnv::current(), "must already be a ciEnv active");
453 assert(env->compiler_data() == NULL, "compile already active?");
454 env->set_compiler_data(compile);
455 assert(compile == Compile::current(), "sanity");
457 compile->set_type_dict(NULL);
458 compile->set_type_hwm(NULL);
459 compile->set_type_last_size(0);
460 compile->set_last_tf(NULL, NULL);
461 compile->set_indexSet_arena(NULL);
462 compile->set_indexSet_free_block_list(NULL);
463 compile->init_type_arena();
464 Type::Initialize(compile);
465 _compile->set_scratch_buffer_blob(NULL);
466 _compile->begin_method();
467 }
468 CompileWrapper::~CompileWrapper() {
469 _compile->end_method();
470 if (_compile->scratch_buffer_blob() != NULL)
471 BufferBlob::free(_compile->scratch_buffer_blob());
472 _compile->env()->set_compiler_data(NULL);
473 }
476 //----------------------------print_compile_messages---------------------------
477 void Compile::print_compile_messages() {
478 #ifndef PRODUCT
479 // Check if recompiling
480 if (_subsume_loads == false && PrintOpto) {
481 // Recompiling without allowing machine instructions to subsume loads
482 tty->print_cr("*********************************************************");
483 tty->print_cr("** Bailout: Recompile without subsuming loads **");
484 tty->print_cr("*********************************************************");
485 }
486 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
487 // Recompiling without escape analysis
488 tty->print_cr("*********************************************************");
489 tty->print_cr("** Bailout: Recompile without escape analysis **");
490 tty->print_cr("*********************************************************");
491 }
492 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
493 // Recompiling without boxing elimination
494 tty->print_cr("*********************************************************");
495 tty->print_cr("** Bailout: Recompile without boxing elimination **");
496 tty->print_cr("*********************************************************");
497 }
498 if (env()->break_at_compile()) {
499 // Open the debugger when compiling this method.
500 tty->print("### Breaking when compiling: ");
501 method()->print_short_name();
502 tty->cr();
503 BREAKPOINT;
504 }
506 if( PrintOpto ) {
507 if (is_osr_compilation()) {
508 tty->print("[OSR]%3d", _compile_id);
509 } else {
510 tty->print("%3d", _compile_id);
511 }
512 }
513 #endif
514 }
517 //-----------------------init_scratch_buffer_blob------------------------------
518 // Construct a temporary BufferBlob and cache it for this compile.
519 void Compile::init_scratch_buffer_blob(int const_size) {
520 // If there is already a scratch buffer blob allocated and the
521 // constant section is big enough, use it. Otherwise free the
522 // current and allocate a new one.
523 BufferBlob* blob = scratch_buffer_blob();
524 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
525 // Use the current blob.
526 } else {
527 if (blob != NULL) {
528 BufferBlob::free(blob);
529 }
531 ResourceMark rm;
532 _scratch_const_size = const_size;
533 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
534 blob = BufferBlob::create("Compile::scratch_buffer", size);
535 // Record the buffer blob for next time.
536 set_scratch_buffer_blob(blob);
537 // Have we run out of code space?
538 if (scratch_buffer_blob() == NULL) {
539 // Let CompilerBroker disable further compilations.
540 record_failure("Not enough space for scratch buffer in CodeCache");
541 return;
542 }
543 }
545 // Initialize the relocation buffers
546 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
547 set_scratch_locs_memory(locs_buf);
548 }
551 //-----------------------scratch_emit_size-------------------------------------
552 // Helper function that computes size by emitting code
553 uint Compile::scratch_emit_size(const Node* n) {
554 // Start scratch_emit_size section.
555 set_in_scratch_emit_size(true);
557 // Emit into a trash buffer and count bytes emitted.
558 // This is a pretty expensive way to compute a size,
559 // but it works well enough if seldom used.
560 // All common fixed-size instructions are given a size
561 // method by the AD file.
562 // Note that the scratch buffer blob and locs memory are
563 // allocated at the beginning of the compile task, and
564 // may be shared by several calls to scratch_emit_size.
565 // The allocation of the scratch buffer blob is particularly
566 // expensive, since it has to grab the code cache lock.
567 BufferBlob* blob = this->scratch_buffer_blob();
568 assert(blob != NULL, "Initialize BufferBlob at start");
569 assert(blob->size() > MAX_inst_size, "sanity");
570 relocInfo* locs_buf = scratch_locs_memory();
571 address blob_begin = blob->content_begin();
572 address blob_end = (address)locs_buf;
573 assert(blob->content_contains(blob_end), "sanity");
574 CodeBuffer buf(blob_begin, blob_end - blob_begin);
575 buf.initialize_consts_size(_scratch_const_size);
576 buf.initialize_stubs_size(MAX_stubs_size);
577 assert(locs_buf != NULL, "sanity");
578 int lsize = MAX_locs_size / 3;
579 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
580 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
581 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
583 // Do the emission.
585 Label fakeL; // Fake label for branch instructions.
586 Label* saveL = NULL;
587 uint save_bnum = 0;
588 bool is_branch = n->is_MachBranch();
589 if (is_branch) {
590 MacroAssembler masm(&buf);
591 masm.bind(fakeL);
592 n->as_MachBranch()->save_label(&saveL, &save_bnum);
593 n->as_MachBranch()->label_set(&fakeL, 0);
594 }
595 n->emit(buf, this->regalloc());
596 if (is_branch) // Restore label.
597 n->as_MachBranch()->label_set(saveL, save_bnum);
599 // End scratch_emit_size section.
600 set_in_scratch_emit_size(false);
602 return buf.insts_size();
603 }
606 // ============================================================================
607 //------------------------------Compile standard-------------------------------
608 debug_only( int Compile::_debug_idx = 100000; )
610 // Compile a method. entry_bci is -1 for normal compilations and indicates
611 // the continuation bci for on stack replacement.
614 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
615 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
616 : Phase(Compiler),
617 _env(ci_env),
618 _log(ci_env->log()),
619 _compile_id(ci_env->compile_id()),
620 _save_argument_registers(false),
621 _stub_name(NULL),
622 _stub_function(NULL),
623 _stub_entry_point(NULL),
624 _method(target),
625 _entry_bci(osr_bci),
626 _initial_gvn(NULL),
627 _for_igvn(NULL),
628 _warm_calls(NULL),
629 _subsume_loads(subsume_loads),
630 _do_escape_analysis(do_escape_analysis),
631 _eliminate_boxing(eliminate_boxing),
632 _failure_reason(NULL),
633 _code_buffer("Compile::Fill_buffer"),
634 _orig_pc_slot(0),
635 _orig_pc_slot_offset_in_bytes(0),
636 _has_method_handle_invokes(false),
637 _mach_constant_base_node(NULL),
638 _node_bundling_limit(0),
639 _node_bundling_base(NULL),
640 _java_calls(0),
641 _inner_loops(0),
642 _scratch_const_size(-1),
643 _in_scratch_emit_size(false),
644 _dead_node_list(comp_arena()),
645 _dead_node_count(0),
646 #ifndef PRODUCT
647 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
648 _printer(IdealGraphPrinter::printer()),
649 #endif
650 _congraph(NULL),
651 _replay_inline_data(NULL),
652 _late_inlines(comp_arena(), 2, 0, NULL),
653 _string_late_inlines(comp_arena(), 2, 0, NULL),
654 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
655 _late_inlines_pos(0),
656 _number_of_mh_late_inlines(0),
657 _inlining_progress(false),
658 _inlining_incrementally(false),
659 _print_inlining_list(NULL),
660 _print_inlining_idx(0),
661 _preserve_jvm_state(0) {
662 C = this;
664 CompileWrapper cw(this);
665 #ifndef PRODUCT
666 if (TimeCompiler2) {
667 tty->print(" ");
668 target->holder()->name()->print();
669 tty->print(".");
670 target->print_short_name();
671 tty->print(" ");
672 }
673 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
674 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
675 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
676 if (!print_opto_assembly) {
677 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
678 if (print_assembly && !Disassembler::can_decode()) {
679 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
680 print_opto_assembly = true;
681 }
682 }
683 set_print_assembly(print_opto_assembly);
684 set_parsed_irreducible_loop(false);
686 if (method()->has_option("ReplayInline")) {
687 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
688 }
689 #endif
690 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
691 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
693 if (ProfileTraps) {
694 // Make sure the method being compiled gets its own MDO,
695 // so we can at least track the decompile_count().
696 method()->ensure_method_data();
697 }
699 Init(::AliasLevel);
702 print_compile_messages();
704 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
705 _ilt = InlineTree::build_inline_tree_root();
706 else
707 _ilt = NULL;
709 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
710 assert(num_alias_types() >= AliasIdxRaw, "");
712 #define MINIMUM_NODE_HASH 1023
713 // Node list that Iterative GVN will start with
714 Unique_Node_List for_igvn(comp_arena());
715 set_for_igvn(&for_igvn);
717 // GVN that will be run immediately on new nodes
718 uint estimated_size = method()->code_size()*4+64;
719 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
720 PhaseGVN gvn(node_arena(), estimated_size);
721 set_initial_gvn(&gvn);
723 if (print_inlining() || print_intrinsics()) {
724 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
725 }
726 { // Scope for timing the parser
727 TracePhase t3("parse", &_t_parser, true);
729 // Put top into the hash table ASAP.
730 initial_gvn()->transform_no_reclaim(top());
732 // Set up tf(), start(), and find a CallGenerator.
733 CallGenerator* cg = NULL;
734 if (is_osr_compilation()) {
735 const TypeTuple *domain = StartOSRNode::osr_domain();
736 const TypeTuple *range = TypeTuple::make_range(method()->signature());
737 init_tf(TypeFunc::make(domain, range));
738 StartNode* s = new (this) StartOSRNode(root(), domain);
739 initial_gvn()->set_type_bottom(s);
740 init_start(s);
741 cg = CallGenerator::for_osr(method(), entry_bci());
742 } else {
743 // Normal case.
744 init_tf(TypeFunc::make(method()));
745 StartNode* s = new (this) StartNode(root(), tf()->domain());
746 initial_gvn()->set_type_bottom(s);
747 init_start(s);
748 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
749 // With java.lang.ref.reference.get() we must go through the
750 // intrinsic when G1 is enabled - even when get() is the root
751 // method of the compile - so that, if necessary, the value in
752 // the referent field of the reference object gets recorded by
753 // the pre-barrier code.
754 // Specifically, if G1 is enabled, the value in the referent
755 // field is recorded by the G1 SATB pre barrier. This will
756 // result in the referent being marked live and the reference
757 // object removed from the list of discovered references during
758 // reference processing.
759 cg = find_intrinsic(method(), false);
760 }
761 if (cg == NULL) {
762 float past_uses = method()->interpreter_invocation_count();
763 float expected_uses = past_uses;
764 cg = CallGenerator::for_inline(method(), expected_uses);
765 }
766 }
767 if (failing()) return;
768 if (cg == NULL) {
769 record_method_not_compilable_all_tiers("cannot parse method");
770 return;
771 }
772 JVMState* jvms = build_start_state(start(), tf());
773 if ((jvms = cg->generate(jvms, NULL)) == NULL) {
774 record_method_not_compilable("method parse failed");
775 return;
776 }
777 GraphKit kit(jvms);
779 if (!kit.stopped()) {
780 // Accept return values, and transfer control we know not where.
781 // This is done by a special, unique ReturnNode bound to root.
782 return_values(kit.jvms());
783 }
785 if (kit.has_exceptions()) {
786 // Any exceptions that escape from this call must be rethrown
787 // to whatever caller is dynamically above us on the stack.
788 // This is done by a special, unique RethrowNode bound to root.
789 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
790 }
792 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
794 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
795 inline_string_calls(true);
796 }
798 if (failing()) return;
800 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
802 // Remove clutter produced by parsing.
803 if (!failing()) {
804 ResourceMark rm;
805 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
806 }
807 }
809 // Note: Large methods are capped off in do_one_bytecode().
810 if (failing()) return;
812 // After parsing, node notes are no longer automagic.
813 // They must be propagated by register_new_node_with_optimizer(),
814 // clone(), or the like.
815 set_default_node_notes(NULL);
817 for (;;) {
818 int successes = Inline_Warm();
819 if (failing()) return;
820 if (successes == 0) break;
821 }
823 // Drain the list.
824 Finish_Warm();
825 #ifndef PRODUCT
826 if (_printer) {
827 _printer->print_inlining(this);
828 }
829 #endif
831 if (failing()) return;
832 NOT_PRODUCT( verify_graph_edges(); )
834 // Now optimize
835 Optimize();
836 if (failing()) return;
837 NOT_PRODUCT( verify_graph_edges(); )
839 #ifndef PRODUCT
840 if (PrintIdeal) {
841 ttyLocker ttyl; // keep the following output all in one block
842 // This output goes directly to the tty, not the compiler log.
843 // To enable tools to match it up with the compilation activity,
844 // be sure to tag this tty output with the compile ID.
845 if (xtty != NULL) {
846 xtty->head("ideal compile_id='%d'%s", compile_id(),
847 is_osr_compilation() ? " compile_kind='osr'" :
848 "");
849 }
850 root()->dump(9999);
851 if (xtty != NULL) {
852 xtty->tail("ideal");
853 }
854 }
855 #endif
857 NOT_PRODUCT( verify_barriers(); )
859 // Dump compilation data to replay it.
860 if (method()->has_option("DumpReplay")) {
861 env()->dump_replay_data(_compile_id);
862 }
863 if (method()->has_option("DumpInline") && (ilt() != NULL)) {
864 env()->dump_inline_data(_compile_id);
865 }
867 // Now that we know the size of all the monitors we can add a fixed slot
868 // for the original deopt pc.
870 _orig_pc_slot = fixed_slots();
871 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
872 set_fixed_slots(next_slot);
874 // Now generate code
875 Code_Gen();
876 if (failing()) return;
878 // Check if we want to skip execution of all compiled code.
879 {
880 #ifndef PRODUCT
881 if (OptoNoExecute) {
882 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
883 return;
884 }
885 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
886 #endif
888 if (is_osr_compilation()) {
889 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
890 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
891 } else {
892 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
893 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
894 }
896 env()->register_method(_method, _entry_bci,
897 &_code_offsets,
898 _orig_pc_slot_offset_in_bytes,
899 code_buffer(),
900 frame_size_in_words(), _oop_map_set,
901 &_handler_table, &_inc_table,
902 compiler,
903 env()->comp_level(),
904 has_unsafe_access(),
905 SharedRuntime::is_wide_vector(max_vector_size())
906 );
908 if (log() != NULL) // Print code cache state into compiler log
909 log()->code_cache_state();
910 }
911 }
913 //------------------------------Compile----------------------------------------
914 // Compile a runtime stub
915 Compile::Compile( ciEnv* ci_env,
916 TypeFunc_generator generator,
917 address stub_function,
918 const char *stub_name,
919 int is_fancy_jump,
920 bool pass_tls,
921 bool save_arg_registers,
922 bool return_pc )
923 : Phase(Compiler),
924 _env(ci_env),
925 _log(ci_env->log()),
926 _compile_id(0),
927 _save_argument_registers(save_arg_registers),
928 _method(NULL),
929 _stub_name(stub_name),
930 _stub_function(stub_function),
931 _stub_entry_point(NULL),
932 _entry_bci(InvocationEntryBci),
933 _initial_gvn(NULL),
934 _for_igvn(NULL),
935 _warm_calls(NULL),
936 _orig_pc_slot(0),
937 _orig_pc_slot_offset_in_bytes(0),
938 _subsume_loads(true),
939 _do_escape_analysis(false),
940 _eliminate_boxing(false),
941 _failure_reason(NULL),
942 _code_buffer("Compile::Fill_buffer"),
943 _has_method_handle_invokes(false),
944 _mach_constant_base_node(NULL),
945 _node_bundling_limit(0),
946 _node_bundling_base(NULL),
947 _java_calls(0),
948 _inner_loops(0),
949 #ifndef PRODUCT
950 _trace_opto_output(TraceOptoOutput),
951 _printer(NULL),
952 #endif
953 _dead_node_list(comp_arena()),
954 _dead_node_count(0),
955 _congraph(NULL),
956 _replay_inline_data(NULL),
957 _number_of_mh_late_inlines(0),
958 _inlining_progress(false),
959 _inlining_incrementally(false),
960 _print_inlining_list(NULL),
961 _print_inlining_idx(0),
962 _preserve_jvm_state(0) {
963 C = this;
965 #ifndef PRODUCT
966 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
967 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
968 set_print_assembly(PrintFrameConverterAssembly);
969 set_parsed_irreducible_loop(false);
970 #endif
971 CompileWrapper cw(this);
972 Init(/*AliasLevel=*/ 0);
973 init_tf((*generator)());
975 {
976 // The following is a dummy for the sake of GraphKit::gen_stub
977 Unique_Node_List for_igvn(comp_arena());
978 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
979 PhaseGVN gvn(Thread::current()->resource_area(),255);
980 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
981 gvn.transform_no_reclaim(top());
983 GraphKit kit;
984 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
985 }
987 NOT_PRODUCT( verify_graph_edges(); )
988 Code_Gen();
989 if (failing()) return;
992 // Entry point will be accessed using compile->stub_entry_point();
993 if (code_buffer() == NULL) {
994 Matcher::soft_match_failure();
995 } else {
996 if (PrintAssembly && (WizardMode || Verbose))
997 tty->print_cr("### Stub::%s", stub_name);
999 if (!failing()) {
1000 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1002 // Make the NMethod
1003 // For now we mark the frame as never safe for profile stackwalking
1004 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1005 code_buffer(),
1006 CodeOffsets::frame_never_safe,
1007 // _code_offsets.value(CodeOffsets::Frame_Complete),
1008 frame_size_in_words(),
1009 _oop_map_set,
1010 save_arg_registers);
1011 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1013 _stub_entry_point = rs->entry_point();
1014 }
1015 }
1016 }
1018 //------------------------------Init-------------------------------------------
1019 // Prepare for a single compilation
1020 void Compile::Init(int aliaslevel) {
1021 _unique = 0;
1022 _regalloc = NULL;
1024 _tf = NULL; // filled in later
1025 _top = NULL; // cached later
1026 _matcher = NULL; // filled in later
1027 _cfg = NULL; // filled in later
1029 set_24_bit_selection_and_mode(Use24BitFP, false);
1031 _node_note_array = NULL;
1032 _default_node_notes = NULL;
1034 _immutable_memory = NULL; // filled in at first inquiry
1036 // Globally visible Nodes
1037 // First set TOP to NULL to give safe behavior during creation of RootNode
1038 set_cached_top_node(NULL);
1039 set_root(new (this) RootNode());
1040 // Now that you have a Root to point to, create the real TOP
1041 set_cached_top_node( new (this) ConNode(Type::TOP) );
1042 set_recent_alloc(NULL, NULL);
1044 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1045 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1046 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1047 env()->set_dependencies(new Dependencies(env()));
1049 _fixed_slots = 0;
1050 set_has_split_ifs(false);
1051 set_has_loops(has_method() && method()->has_loops()); // first approximation
1052 set_has_stringbuilder(false);
1053 set_has_boxed_value(false);
1054 _trap_can_recompile = false; // no traps emitted yet
1055 _major_progress = true; // start out assuming good things will happen
1056 set_has_unsafe_access(false);
1057 set_max_vector_size(0);
1058 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1059 set_decompile_count(0);
1061 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1062 set_num_loop_opts(LoopOptsCount);
1063 set_do_inlining(Inline);
1064 set_max_inline_size(MaxInlineSize);
1065 set_freq_inline_size(FreqInlineSize);
1066 set_do_scheduling(OptoScheduling);
1067 set_do_count_invocations(false);
1068 set_do_method_data_update(false);
1070 if (debug_info()->recording_non_safepoints()) {
1071 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1072 (comp_arena(), 8, 0, NULL));
1073 set_default_node_notes(Node_Notes::make(this));
1074 }
1076 // // -- Initialize types before each compile --
1077 // // Update cached type information
1078 // if( _method && _method->constants() )
1079 // Type::update_loaded_types(_method, _method->constants());
1081 // Init alias_type map.
1082 if (!_do_escape_analysis && aliaslevel == 3)
1083 aliaslevel = 2; // No unique types without escape analysis
1084 _AliasLevel = aliaslevel;
1085 const int grow_ats = 16;
1086 _max_alias_types = grow_ats;
1087 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1088 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1089 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1090 {
1091 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1092 }
1093 // Initialize the first few types.
1094 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1095 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1096 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1097 _num_alias_types = AliasIdxRaw+1;
1098 // Zero out the alias type cache.
1099 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1100 // A NULL adr_type hits in the cache right away. Preload the right answer.
1101 probe_alias_cache(NULL)->_index = AliasIdxTop;
1103 _intrinsics = NULL;
1104 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1105 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1106 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1107 register_library_intrinsics();
1108 }
1110 //---------------------------init_start----------------------------------------
1111 // Install the StartNode on this compile object.
1112 void Compile::init_start(StartNode* s) {
1113 if (failing())
1114 return; // already failing
1115 assert(s == start(), "");
1116 }
1118 StartNode* Compile::start() const {
1119 assert(!failing(), "");
1120 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1121 Node* start = root()->fast_out(i);
1122 if( start->is_Start() )
1123 return start->as_Start();
1124 }
1125 ShouldNotReachHere();
1126 return NULL;
1127 }
1129 //-------------------------------immutable_memory-------------------------------------
1130 // Access immutable memory
1131 Node* Compile::immutable_memory() {
1132 if (_immutable_memory != NULL) {
1133 return _immutable_memory;
1134 }
1135 StartNode* s = start();
1136 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1137 Node *p = s->fast_out(i);
1138 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1139 _immutable_memory = p;
1140 return _immutable_memory;
1141 }
1142 }
1143 ShouldNotReachHere();
1144 return NULL;
1145 }
1147 //----------------------set_cached_top_node------------------------------------
1148 // Install the cached top node, and make sure Node::is_top works correctly.
1149 void Compile::set_cached_top_node(Node* tn) {
1150 if (tn != NULL) verify_top(tn);
1151 Node* old_top = _top;
1152 _top = tn;
1153 // Calling Node::setup_is_top allows the nodes the chance to adjust
1154 // their _out arrays.
1155 if (_top != NULL) _top->setup_is_top();
1156 if (old_top != NULL) old_top->setup_is_top();
1157 assert(_top == NULL || top()->is_top(), "");
1158 }
1160 #ifdef ASSERT
1161 uint Compile::count_live_nodes_by_graph_walk() {
1162 Unique_Node_List useful(comp_arena());
1163 // Get useful node list by walking the graph.
1164 identify_useful_nodes(useful);
1165 return useful.size();
1166 }
1168 void Compile::print_missing_nodes() {
1170 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1171 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1172 return;
1173 }
1175 // This is an expensive function. It is executed only when the user
1176 // specifies VerifyIdealNodeCount option or otherwise knows the
1177 // additional work that needs to be done to identify reachable nodes
1178 // by walking the flow graph and find the missing ones using
1179 // _dead_node_list.
1181 Unique_Node_List useful(comp_arena());
1182 // Get useful node list by walking the graph.
1183 identify_useful_nodes(useful);
1185 uint l_nodes = C->live_nodes();
1186 uint l_nodes_by_walk = useful.size();
1188 if (l_nodes != l_nodes_by_walk) {
1189 if (_log != NULL) {
1190 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1191 _log->stamp();
1192 _log->end_head();
1193 }
1194 VectorSet& useful_member_set = useful.member_set();
1195 int last_idx = l_nodes_by_walk;
1196 for (int i = 0; i < last_idx; i++) {
1197 if (useful_member_set.test(i)) {
1198 if (_dead_node_list.test(i)) {
1199 if (_log != NULL) {
1200 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1201 }
1202 if (PrintIdealNodeCount) {
1203 // Print the log message to tty
1204 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1205 useful.at(i)->dump();
1206 }
1207 }
1208 }
1209 else if (! _dead_node_list.test(i)) {
1210 if (_log != NULL) {
1211 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1212 }
1213 if (PrintIdealNodeCount) {
1214 // Print the log message to tty
1215 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1216 }
1217 }
1218 }
1219 if (_log != NULL) {
1220 _log->tail("mismatched_nodes");
1221 }
1222 }
1223 }
1224 #endif
1226 #ifndef PRODUCT
1227 void Compile::verify_top(Node* tn) const {
1228 if (tn != NULL) {
1229 assert(tn->is_Con(), "top node must be a constant");
1230 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1231 assert(tn->in(0) != NULL, "must have live top node");
1232 }
1233 }
1234 #endif
1237 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1239 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1240 guarantee(arr != NULL, "");
1241 int num_blocks = arr->length();
1242 if (grow_by < num_blocks) grow_by = num_blocks;
1243 int num_notes = grow_by * _node_notes_block_size;
1244 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1245 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1246 while (num_notes > 0) {
1247 arr->append(notes);
1248 notes += _node_notes_block_size;
1249 num_notes -= _node_notes_block_size;
1250 }
1251 assert(num_notes == 0, "exact multiple, please");
1252 }
1254 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1255 if (source == NULL || dest == NULL) return false;
1257 if (dest->is_Con())
1258 return false; // Do not push debug info onto constants.
1260 #ifdef ASSERT
1261 // Leave a bread crumb trail pointing to the original node:
1262 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1263 dest->set_debug_orig(source);
1264 }
1265 #endif
1267 if (node_note_array() == NULL)
1268 return false; // Not collecting any notes now.
1270 // This is a copy onto a pre-existing node, which may already have notes.
1271 // If both nodes have notes, do not overwrite any pre-existing notes.
1272 Node_Notes* source_notes = node_notes_at(source->_idx);
1273 if (source_notes == NULL || source_notes->is_clear()) return false;
1274 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1275 if (dest_notes == NULL || dest_notes->is_clear()) {
1276 return set_node_notes_at(dest->_idx, source_notes);
1277 }
1279 Node_Notes merged_notes = (*source_notes);
1280 // The order of operations here ensures that dest notes will win...
1281 merged_notes.update_from(dest_notes);
1282 return set_node_notes_at(dest->_idx, &merged_notes);
1283 }
1286 //--------------------------allow_range_check_smearing-------------------------
1287 // Gating condition for coalescing similar range checks.
1288 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1289 // single covering check that is at least as strong as any of them.
1290 // If the optimization succeeds, the simplified (strengthened) range check
1291 // will always succeed. If it fails, we will deopt, and then give up
1292 // on the optimization.
1293 bool Compile::allow_range_check_smearing() const {
1294 // If this method has already thrown a range-check,
1295 // assume it was because we already tried range smearing
1296 // and it failed.
1297 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1298 return !already_trapped;
1299 }
1302 //------------------------------flatten_alias_type-----------------------------
1303 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1304 int offset = tj->offset();
1305 TypePtr::PTR ptr = tj->ptr();
1307 // Known instance (scalarizable allocation) alias only with itself.
1308 bool is_known_inst = tj->isa_oopptr() != NULL &&
1309 tj->is_oopptr()->is_known_instance();
1311 // Process weird unsafe references.
1312 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1313 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1314 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1315 tj = TypeOopPtr::BOTTOM;
1316 ptr = tj->ptr();
1317 offset = tj->offset();
1318 }
1320 // Array pointers need some flattening
1321 const TypeAryPtr *ta = tj->isa_aryptr();
1322 if (ta && ta->is_stable()) {
1323 // Erase stability property for alias analysis.
1324 tj = ta = ta->cast_to_stable(false);
1325 }
1326 if( ta && is_known_inst ) {
1327 if ( offset != Type::OffsetBot &&
1328 offset > arrayOopDesc::length_offset_in_bytes() ) {
1329 offset = Type::OffsetBot; // Flatten constant access into array body only
1330 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1331 }
1332 } else if( ta && _AliasLevel >= 2 ) {
1333 // For arrays indexed by constant indices, we flatten the alias
1334 // space to include all of the array body. Only the header, klass
1335 // and array length can be accessed un-aliased.
1336 if( offset != Type::OffsetBot ) {
1337 if( ta->const_oop() ) { // MethodData* or Method*
1338 offset = Type::OffsetBot; // Flatten constant access into array body
1339 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1340 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1341 // range is OK as-is.
1342 tj = ta = TypeAryPtr::RANGE;
1343 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1344 tj = TypeInstPtr::KLASS; // all klass loads look alike
1345 ta = TypeAryPtr::RANGE; // generic ignored junk
1346 ptr = TypePtr::BotPTR;
1347 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1348 tj = TypeInstPtr::MARK;
1349 ta = TypeAryPtr::RANGE; // generic ignored junk
1350 ptr = TypePtr::BotPTR;
1351 } else { // Random constant offset into array body
1352 offset = Type::OffsetBot; // Flatten constant access into array body
1353 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1354 }
1355 }
1356 // Arrays of fixed size alias with arrays of unknown size.
1357 if (ta->size() != TypeInt::POS) {
1358 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1359 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1360 }
1361 // Arrays of known objects become arrays of unknown objects.
1362 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1363 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1364 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1365 }
1366 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1367 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1368 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1369 }
1370 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1371 // cannot be distinguished by bytecode alone.
1372 if (ta->elem() == TypeInt::BOOL) {
1373 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1374 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1375 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1376 }
1377 // During the 2nd round of IterGVN, NotNull castings are removed.
1378 // Make sure the Bottom and NotNull variants alias the same.
1379 // Also, make sure exact and non-exact variants alias the same.
1380 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1381 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1382 }
1383 }
1385 // Oop pointers need some flattening
1386 const TypeInstPtr *to = tj->isa_instptr();
1387 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1388 ciInstanceKlass *k = to->klass()->as_instance_klass();
1389 if( ptr == TypePtr::Constant ) {
1390 if (to->klass() != ciEnv::current()->Class_klass() ||
1391 offset < k->size_helper() * wordSize) {
1392 // No constant oop pointers (such as Strings); they alias with
1393 // unknown strings.
1394 assert(!is_known_inst, "not scalarizable allocation");
1395 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1396 }
1397 } else if( is_known_inst ) {
1398 tj = to; // Keep NotNull and klass_is_exact for instance type
1399 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1400 // During the 2nd round of IterGVN, NotNull castings are removed.
1401 // Make sure the Bottom and NotNull variants alias the same.
1402 // Also, make sure exact and non-exact variants alias the same.
1403 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1404 }
1405 if (to->speculative() != NULL) {
1406 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1407 }
1408 // Canonicalize the holder of this field
1409 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1410 // First handle header references such as a LoadKlassNode, even if the
1411 // object's klass is unloaded at compile time (4965979).
1412 if (!is_known_inst) { // Do it only for non-instance types
1413 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1414 }
1415 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1416 // Static fields are in the space above the normal instance
1417 // fields in the java.lang.Class instance.
1418 if (to->klass() != ciEnv::current()->Class_klass()) {
1419 to = NULL;
1420 tj = TypeOopPtr::BOTTOM;
1421 offset = tj->offset();
1422 }
1423 } else {
1424 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1425 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1426 if( is_known_inst ) {
1427 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1428 } else {
1429 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1430 }
1431 }
1432 }
1433 }
1435 // Klass pointers to object array klasses need some flattening
1436 const TypeKlassPtr *tk = tj->isa_klassptr();
1437 if( tk ) {
1438 // If we are referencing a field within a Klass, we need
1439 // to assume the worst case of an Object. Both exact and
1440 // inexact types must flatten to the same alias class so
1441 // use NotNull as the PTR.
1442 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1444 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1445 TypeKlassPtr::OBJECT->klass(),
1446 offset);
1447 }
1449 ciKlass* klass = tk->klass();
1450 if( klass->is_obj_array_klass() ) {
1451 ciKlass* k = TypeAryPtr::OOPS->klass();
1452 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1453 k = TypeInstPtr::BOTTOM->klass();
1454 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1455 }
1457 // Check for precise loads from the primary supertype array and force them
1458 // to the supertype cache alias index. Check for generic array loads from
1459 // the primary supertype array and also force them to the supertype cache
1460 // alias index. Since the same load can reach both, we need to merge
1461 // these 2 disparate memories into the same alias class. Since the
1462 // primary supertype array is read-only, there's no chance of confusion
1463 // where we bypass an array load and an array store.
1464 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1465 if (offset == Type::OffsetBot ||
1466 (offset >= primary_supers_offset &&
1467 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1468 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1469 offset = in_bytes(Klass::secondary_super_cache_offset());
1470 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1471 }
1472 }
1474 // Flatten all Raw pointers together.
1475 if (tj->base() == Type::RawPtr)
1476 tj = TypeRawPtr::BOTTOM;
1478 if (tj->base() == Type::AnyPtr)
1479 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1481 // Flatten all to bottom for now
1482 switch( _AliasLevel ) {
1483 case 0:
1484 tj = TypePtr::BOTTOM;
1485 break;
1486 case 1: // Flatten to: oop, static, field or array
1487 switch (tj->base()) {
1488 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1489 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1490 case Type::AryPtr: // do not distinguish arrays at all
1491 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1492 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1493 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1494 default: ShouldNotReachHere();
1495 }
1496 break;
1497 case 2: // No collapsing at level 2; keep all splits
1498 case 3: // No collapsing at level 3; keep all splits
1499 break;
1500 default:
1501 Unimplemented();
1502 }
1504 offset = tj->offset();
1505 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1507 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1508 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1509 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1510 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1511 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1512 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1513 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1514 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1515 assert( tj->ptr() != TypePtr::TopPTR &&
1516 tj->ptr() != TypePtr::AnyNull &&
1517 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1518 // assert( tj->ptr() != TypePtr::Constant ||
1519 // tj->base() == Type::RawPtr ||
1520 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1522 return tj;
1523 }
1525 void Compile::AliasType::Init(int i, const TypePtr* at) {
1526 _index = i;
1527 _adr_type = at;
1528 _field = NULL;
1529 _element = NULL;
1530 _is_rewritable = true; // default
1531 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1532 if (atoop != NULL && atoop->is_known_instance()) {
1533 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1534 _general_index = Compile::current()->get_alias_index(gt);
1535 } else {
1536 _general_index = 0;
1537 }
1538 }
1540 //---------------------------------print_on------------------------------------
1541 #ifndef PRODUCT
1542 void Compile::AliasType::print_on(outputStream* st) {
1543 if (index() < 10)
1544 st->print("@ <%d> ", index());
1545 else st->print("@ <%d>", index());
1546 st->print(is_rewritable() ? " " : " RO");
1547 int offset = adr_type()->offset();
1548 if (offset == Type::OffsetBot)
1549 st->print(" +any");
1550 else st->print(" +%-3d", offset);
1551 st->print(" in ");
1552 adr_type()->dump_on(st);
1553 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1554 if (field() != NULL && tjp) {
1555 if (tjp->klass() != field()->holder() ||
1556 tjp->offset() != field()->offset_in_bytes()) {
1557 st->print(" != ");
1558 field()->print();
1559 st->print(" ***");
1560 }
1561 }
1562 }
1564 void print_alias_types() {
1565 Compile* C = Compile::current();
1566 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1567 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1568 C->alias_type(idx)->print_on(tty);
1569 tty->cr();
1570 }
1571 }
1572 #endif
1575 //----------------------------probe_alias_cache--------------------------------
1576 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1577 intptr_t key = (intptr_t) adr_type;
1578 key ^= key >> logAliasCacheSize;
1579 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1580 }
1583 //-----------------------------grow_alias_types--------------------------------
1584 void Compile::grow_alias_types() {
1585 const int old_ats = _max_alias_types; // how many before?
1586 const int new_ats = old_ats; // how many more?
1587 const int grow_ats = old_ats+new_ats; // how many now?
1588 _max_alias_types = grow_ats;
1589 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1590 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1591 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1592 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1593 }
1596 //--------------------------------find_alias_type------------------------------
1597 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1598 if (_AliasLevel == 0)
1599 return alias_type(AliasIdxBot);
1601 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1602 if (ace->_adr_type == adr_type) {
1603 return alias_type(ace->_index);
1604 }
1606 // Handle special cases.
1607 if (adr_type == NULL) return alias_type(AliasIdxTop);
1608 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1610 // Do it the slow way.
1611 const TypePtr* flat = flatten_alias_type(adr_type);
1613 #ifdef ASSERT
1614 assert(flat == flatten_alias_type(flat), "idempotent");
1615 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1616 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1617 const TypeOopPtr* foop = flat->is_oopptr();
1618 // Scalarizable allocations have exact klass always.
1619 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1620 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1621 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1622 }
1623 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1624 #endif
1626 int idx = AliasIdxTop;
1627 for (int i = 0; i < num_alias_types(); i++) {
1628 if (alias_type(i)->adr_type() == flat) {
1629 idx = i;
1630 break;
1631 }
1632 }
1634 if (idx == AliasIdxTop) {
1635 if (no_create) return NULL;
1636 // Grow the array if necessary.
1637 if (_num_alias_types == _max_alias_types) grow_alias_types();
1638 // Add a new alias type.
1639 idx = _num_alias_types++;
1640 _alias_types[idx]->Init(idx, flat);
1641 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1642 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1643 if (flat->isa_instptr()) {
1644 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1645 && flat->is_instptr()->klass() == env()->Class_klass())
1646 alias_type(idx)->set_rewritable(false);
1647 }
1648 if (flat->isa_aryptr()) {
1649 #ifdef ASSERT
1650 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1651 // (T_BYTE has the weakest alignment and size restrictions...)
1652 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1653 #endif
1654 if (flat->offset() == TypePtr::OffsetBot) {
1655 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1656 }
1657 }
1658 if (flat->isa_klassptr()) {
1659 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1660 alias_type(idx)->set_rewritable(false);
1661 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1662 alias_type(idx)->set_rewritable(false);
1663 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1664 alias_type(idx)->set_rewritable(false);
1665 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1666 alias_type(idx)->set_rewritable(false);
1667 }
1668 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1669 // but the base pointer type is not distinctive enough to identify
1670 // references into JavaThread.)
1672 // Check for final fields.
1673 const TypeInstPtr* tinst = flat->isa_instptr();
1674 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1675 ciField* field;
1676 if (tinst->const_oop() != NULL &&
1677 tinst->klass() == ciEnv::current()->Class_klass() &&
1678 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1679 // static field
1680 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1681 field = k->get_field_by_offset(tinst->offset(), true);
1682 } else {
1683 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1684 field = k->get_field_by_offset(tinst->offset(), false);
1685 }
1686 assert(field == NULL ||
1687 original_field == NULL ||
1688 (field->holder() == original_field->holder() &&
1689 field->offset() == original_field->offset() &&
1690 field->is_static() == original_field->is_static()), "wrong field?");
1691 // Set field() and is_rewritable() attributes.
1692 if (field != NULL) alias_type(idx)->set_field(field);
1693 }
1694 }
1696 // Fill the cache for next time.
1697 ace->_adr_type = adr_type;
1698 ace->_index = idx;
1699 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1701 // Might as well try to fill the cache for the flattened version, too.
1702 AliasCacheEntry* face = probe_alias_cache(flat);
1703 if (face->_adr_type == NULL) {
1704 face->_adr_type = flat;
1705 face->_index = idx;
1706 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1707 }
1709 return alias_type(idx);
1710 }
1713 Compile::AliasType* Compile::alias_type(ciField* field) {
1714 const TypeOopPtr* t;
1715 if (field->is_static())
1716 t = TypeInstPtr::make(field->holder()->java_mirror());
1717 else
1718 t = TypeOopPtr::make_from_klass_raw(field->holder());
1719 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1720 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1721 return atp;
1722 }
1725 //------------------------------have_alias_type--------------------------------
1726 bool Compile::have_alias_type(const TypePtr* adr_type) {
1727 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1728 if (ace->_adr_type == adr_type) {
1729 return true;
1730 }
1732 // Handle special cases.
1733 if (adr_type == NULL) return true;
1734 if (adr_type == TypePtr::BOTTOM) return true;
1736 return find_alias_type(adr_type, true, NULL) != NULL;
1737 }
1739 //-----------------------------must_alias--------------------------------------
1740 // True if all values of the given address type are in the given alias category.
1741 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1742 if (alias_idx == AliasIdxBot) return true; // the universal category
1743 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1744 if (alias_idx == AliasIdxTop) return false; // the empty category
1745 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1747 // the only remaining possible overlap is identity
1748 int adr_idx = get_alias_index(adr_type);
1749 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1750 assert(adr_idx == alias_idx ||
1751 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1752 && adr_type != TypeOopPtr::BOTTOM),
1753 "should not be testing for overlap with an unsafe pointer");
1754 return adr_idx == alias_idx;
1755 }
1757 //------------------------------can_alias--------------------------------------
1758 // True if any values of the given address type are in the given alias category.
1759 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1760 if (alias_idx == AliasIdxTop) return false; // the empty category
1761 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1762 if (alias_idx == AliasIdxBot) return true; // the universal category
1763 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1765 // the only remaining possible overlap is identity
1766 int adr_idx = get_alias_index(adr_type);
1767 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1768 return adr_idx == alias_idx;
1769 }
1773 //---------------------------pop_warm_call-------------------------------------
1774 WarmCallInfo* Compile::pop_warm_call() {
1775 WarmCallInfo* wci = _warm_calls;
1776 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1777 return wci;
1778 }
1780 //----------------------------Inline_Warm--------------------------------------
1781 int Compile::Inline_Warm() {
1782 // If there is room, try to inline some more warm call sites.
1783 // %%% Do a graph index compaction pass when we think we're out of space?
1784 if (!InlineWarmCalls) return 0;
1786 int calls_made_hot = 0;
1787 int room_to_grow = NodeCountInliningCutoff - unique();
1788 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1789 int amount_grown = 0;
1790 WarmCallInfo* call;
1791 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1792 int est_size = (int)call->size();
1793 if (est_size > (room_to_grow - amount_grown)) {
1794 // This one won't fit anyway. Get rid of it.
1795 call->make_cold();
1796 continue;
1797 }
1798 call->make_hot();
1799 calls_made_hot++;
1800 amount_grown += est_size;
1801 amount_to_grow -= est_size;
1802 }
1804 if (calls_made_hot > 0) set_major_progress();
1805 return calls_made_hot;
1806 }
1809 //----------------------------Finish_Warm--------------------------------------
1810 void Compile::Finish_Warm() {
1811 if (!InlineWarmCalls) return;
1812 if (failing()) return;
1813 if (warm_calls() == NULL) return;
1815 // Clean up loose ends, if we are out of space for inlining.
1816 WarmCallInfo* call;
1817 while ((call = pop_warm_call()) != NULL) {
1818 call->make_cold();
1819 }
1820 }
1822 //---------------------cleanup_loop_predicates-----------------------
1823 // Remove the opaque nodes that protect the predicates so that all unused
1824 // checks and uncommon_traps will be eliminated from the ideal graph
1825 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1826 if (predicate_count()==0) return;
1827 for (int i = predicate_count(); i > 0; i--) {
1828 Node * n = predicate_opaque1_node(i-1);
1829 assert(n->Opcode() == Op_Opaque1, "must be");
1830 igvn.replace_node(n, n->in(1));
1831 }
1832 assert(predicate_count()==0, "should be clean!");
1833 }
1835 // StringOpts and late inlining of string methods
1836 void Compile::inline_string_calls(bool parse_time) {
1837 {
1838 // remove useless nodes to make the usage analysis simpler
1839 ResourceMark rm;
1840 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1841 }
1843 {
1844 ResourceMark rm;
1845 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1846 PhaseStringOpts pso(initial_gvn(), for_igvn());
1847 print_method(PHASE_AFTER_STRINGOPTS, 3);
1848 }
1850 // now inline anything that we skipped the first time around
1851 if (!parse_time) {
1852 _late_inlines_pos = _late_inlines.length();
1853 }
1855 while (_string_late_inlines.length() > 0) {
1856 CallGenerator* cg = _string_late_inlines.pop();
1857 cg->do_late_inline();
1858 if (failing()) return;
1859 }
1860 _string_late_inlines.trunc_to(0);
1861 }
1863 // Late inlining of boxing methods
1864 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1865 if (_boxing_late_inlines.length() > 0) {
1866 assert(has_boxed_value(), "inconsistent");
1868 PhaseGVN* gvn = initial_gvn();
1869 set_inlining_incrementally(true);
1871 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1872 for_igvn()->clear();
1873 gvn->replace_with(&igvn);
1875 while (_boxing_late_inlines.length() > 0) {
1876 CallGenerator* cg = _boxing_late_inlines.pop();
1877 cg->do_late_inline();
1878 if (failing()) return;
1879 }
1880 _boxing_late_inlines.trunc_to(0);
1882 {
1883 ResourceMark rm;
1884 PhaseRemoveUseless pru(gvn, for_igvn());
1885 }
1887 igvn = PhaseIterGVN(gvn);
1888 igvn.optimize();
1890 set_inlining_progress(false);
1891 set_inlining_incrementally(false);
1892 }
1893 }
1895 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1896 assert(IncrementalInline, "incremental inlining should be on");
1897 PhaseGVN* gvn = initial_gvn();
1899 set_inlining_progress(false);
1900 for_igvn()->clear();
1901 gvn->replace_with(&igvn);
1903 int i = 0;
1905 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1906 CallGenerator* cg = _late_inlines.at(i);
1907 _late_inlines_pos = i+1;
1908 cg->do_late_inline();
1909 if (failing()) return;
1910 }
1911 int j = 0;
1912 for (; i < _late_inlines.length(); i++, j++) {
1913 _late_inlines.at_put(j, _late_inlines.at(i));
1914 }
1915 _late_inlines.trunc_to(j);
1917 {
1918 ResourceMark rm;
1919 PhaseRemoveUseless pru(gvn, for_igvn());
1920 }
1922 igvn = PhaseIterGVN(gvn);
1923 }
1925 // Perform incremental inlining until bound on number of live nodes is reached
1926 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1927 PhaseGVN* gvn = initial_gvn();
1929 set_inlining_incrementally(true);
1930 set_inlining_progress(true);
1931 uint low_live_nodes = 0;
1933 while(inlining_progress() && _late_inlines.length() > 0) {
1935 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1936 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1937 // PhaseIdealLoop is expensive so we only try it once we are
1938 // out of loop and we only try it again if the previous helped
1939 // got the number of nodes down significantly
1940 PhaseIdealLoop ideal_loop( igvn, false, true );
1941 if (failing()) return;
1942 low_live_nodes = live_nodes();
1943 _major_progress = true;
1944 }
1946 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1947 break;
1948 }
1949 }
1951 inline_incrementally_one(igvn);
1953 if (failing()) return;
1955 igvn.optimize();
1957 if (failing()) return;
1958 }
1960 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1962 if (_string_late_inlines.length() > 0) {
1963 assert(has_stringbuilder(), "inconsistent");
1964 for_igvn()->clear();
1965 initial_gvn()->replace_with(&igvn);
1967 inline_string_calls(false);
1969 if (failing()) return;
1971 {
1972 ResourceMark rm;
1973 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1974 }
1976 igvn = PhaseIterGVN(gvn);
1978 igvn.optimize();
1979 }
1981 set_inlining_incrementally(false);
1982 }
1985 //------------------------------Optimize---------------------------------------
1986 // Given a graph, optimize it.
1987 void Compile::Optimize() {
1988 TracePhase t1("optimizer", &_t_optimizer, true);
1990 #ifndef PRODUCT
1991 if (env()->break_at_compile()) {
1992 BREAKPOINT;
1993 }
1995 #endif
1997 ResourceMark rm;
1998 int loop_opts_cnt;
2000 NOT_PRODUCT( verify_graph_edges(); )
2002 print_method(PHASE_AFTER_PARSING);
2004 {
2005 // Iterative Global Value Numbering, including ideal transforms
2006 // Initialize IterGVN with types and values from parse-time GVN
2007 PhaseIterGVN igvn(initial_gvn());
2008 {
2009 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2010 igvn.optimize();
2011 }
2013 print_method(PHASE_ITER_GVN1, 2);
2015 if (failing()) return;
2017 {
2018 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2019 inline_incrementally(igvn);
2020 }
2022 print_method(PHASE_INCREMENTAL_INLINE, 2);
2024 if (failing()) return;
2026 if (eliminate_boxing()) {
2027 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2028 // Inline valueOf() methods now.
2029 inline_boxing_calls(igvn);
2031 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2033 if (failing()) return;
2034 }
2036 // Remove the speculative part of types and clean up the graph from
2037 // the extra CastPP nodes whose only purpose is to carry them. Do
2038 // that early so that optimizations are not disrupted by the extra
2039 // CastPP nodes.
2040 remove_speculative_types(igvn);
2042 // No more new expensive nodes will be added to the list from here
2043 // so keep only the actual candidates for optimizations.
2044 cleanup_expensive_nodes(igvn);
2046 // Perform escape analysis
2047 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2048 if (has_loops()) {
2049 // Cleanup graph (remove dead nodes).
2050 TracePhase t2("idealLoop", &_t_idealLoop, true);
2051 PhaseIdealLoop ideal_loop( igvn, false, true );
2052 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2053 if (failing()) return;
2054 }
2055 ConnectionGraph::do_analysis(this, &igvn);
2057 if (failing()) return;
2059 // Optimize out fields loads from scalar replaceable allocations.
2060 igvn.optimize();
2061 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2063 if (failing()) return;
2065 if (congraph() != NULL && macro_count() > 0) {
2066 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2067 PhaseMacroExpand mexp(igvn);
2068 mexp.eliminate_macro_nodes();
2069 igvn.set_delay_transform(false);
2071 igvn.optimize();
2072 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2074 if (failing()) return;
2075 }
2076 }
2078 // Loop transforms on the ideal graph. Range Check Elimination,
2079 // peeling, unrolling, etc.
2081 // Set loop opts counter
2082 loop_opts_cnt = num_loop_opts();
2083 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2084 {
2085 TracePhase t2("idealLoop", &_t_idealLoop, true);
2086 PhaseIdealLoop ideal_loop( igvn, true );
2087 loop_opts_cnt--;
2088 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2089 if (failing()) return;
2090 }
2091 // Loop opts pass if partial peeling occurred in previous pass
2092 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2093 TracePhase t3("idealLoop", &_t_idealLoop, true);
2094 PhaseIdealLoop ideal_loop( igvn, false );
2095 loop_opts_cnt--;
2096 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2097 if (failing()) return;
2098 }
2099 // Loop opts pass for loop-unrolling before CCP
2100 if(major_progress() && (loop_opts_cnt > 0)) {
2101 TracePhase t4("idealLoop", &_t_idealLoop, true);
2102 PhaseIdealLoop ideal_loop( igvn, false );
2103 loop_opts_cnt--;
2104 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2105 }
2106 if (!failing()) {
2107 // Verify that last round of loop opts produced a valid graph
2108 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2109 PhaseIdealLoop::verify(igvn);
2110 }
2111 }
2112 if (failing()) return;
2114 // Conditional Constant Propagation;
2115 PhaseCCP ccp( &igvn );
2116 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2117 {
2118 TracePhase t2("ccp", &_t_ccp, true);
2119 ccp.do_transform();
2120 }
2121 print_method(PHASE_CPP1, 2);
2123 assert( true, "Break here to ccp.dump_old2new_map()");
2125 // Iterative Global Value Numbering, including ideal transforms
2126 {
2127 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2128 igvn = ccp;
2129 igvn.optimize();
2130 }
2132 print_method(PHASE_ITER_GVN2, 2);
2134 if (failing()) return;
2136 // Loop transforms on the ideal graph. Range Check Elimination,
2137 // peeling, unrolling, etc.
2138 if(loop_opts_cnt > 0) {
2139 debug_only( int cnt = 0; );
2140 while(major_progress() && (loop_opts_cnt > 0)) {
2141 TracePhase t2("idealLoop", &_t_idealLoop, true);
2142 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2143 PhaseIdealLoop ideal_loop( igvn, true);
2144 loop_opts_cnt--;
2145 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2146 if (failing()) return;
2147 }
2148 }
2150 {
2151 // Verify that all previous optimizations produced a valid graph
2152 // at least to this point, even if no loop optimizations were done.
2153 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2154 PhaseIdealLoop::verify(igvn);
2155 }
2157 {
2158 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2159 PhaseMacroExpand mex(igvn);
2160 if (mex.expand_macro_nodes()) {
2161 assert(failing(), "must bail out w/ explicit message");
2162 return;
2163 }
2164 }
2166 } // (End scope of igvn; run destructor if necessary for asserts.)
2168 dump_inlining();
2169 // A method with only infinite loops has no edges entering loops from root
2170 {
2171 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2172 if (final_graph_reshaping()) {
2173 assert(failing(), "must bail out w/ explicit message");
2174 return;
2175 }
2176 }
2178 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2179 }
2182 //------------------------------Code_Gen---------------------------------------
2183 // Given a graph, generate code for it
2184 void Compile::Code_Gen() {
2185 if (failing()) {
2186 return;
2187 }
2189 // Perform instruction selection. You might think we could reclaim Matcher
2190 // memory PDQ, but actually the Matcher is used in generating spill code.
2191 // Internals of the Matcher (including some VectorSets) must remain live
2192 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2193 // set a bit in reclaimed memory.
2195 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2196 // nodes. Mapping is only valid at the root of each matched subtree.
2197 NOT_PRODUCT( verify_graph_edges(); )
2199 Matcher matcher;
2200 _matcher = &matcher;
2201 {
2202 TracePhase t2("matcher", &_t_matcher, true);
2203 matcher.match();
2204 }
2205 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2206 // nodes. Mapping is only valid at the root of each matched subtree.
2207 NOT_PRODUCT( verify_graph_edges(); )
2209 // If you have too many nodes, or if matching has failed, bail out
2210 check_node_count(0, "out of nodes matching instructions");
2211 if (failing()) {
2212 return;
2213 }
2215 // Build a proper-looking CFG
2216 PhaseCFG cfg(node_arena(), root(), matcher);
2217 _cfg = &cfg;
2218 {
2219 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2220 bool success = cfg.do_global_code_motion();
2221 if (!success) {
2222 return;
2223 }
2225 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2226 NOT_PRODUCT( verify_graph_edges(); )
2227 debug_only( cfg.verify(); )
2228 }
2230 PhaseChaitin regalloc(unique(), cfg, matcher);
2231 _regalloc = ®alloc;
2232 {
2233 TracePhase t2("regalloc", &_t_registerAllocation, true);
2234 // Perform register allocation. After Chaitin, use-def chains are
2235 // no longer accurate (at spill code) and so must be ignored.
2236 // Node->LRG->reg mappings are still accurate.
2237 _regalloc->Register_Allocate();
2239 // Bail out if the allocator builds too many nodes
2240 if (failing()) {
2241 return;
2242 }
2243 }
2245 // Prior to register allocation we kept empty basic blocks in case the
2246 // the allocator needed a place to spill. After register allocation we
2247 // are not adding any new instructions. If any basic block is empty, we
2248 // can now safely remove it.
2249 {
2250 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2251 cfg.remove_empty_blocks();
2252 if (do_freq_based_layout()) {
2253 PhaseBlockLayout layout(cfg);
2254 } else {
2255 cfg.set_loop_alignment();
2256 }
2257 cfg.fixup_flow();
2258 }
2260 // Apply peephole optimizations
2261 if( OptoPeephole ) {
2262 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2263 PhasePeephole peep( _regalloc, cfg);
2264 peep.do_transform();
2265 }
2267 // Convert Nodes to instruction bits in a buffer
2268 {
2269 // %%%% workspace merge brought two timers together for one job
2270 TracePhase t2a("output", &_t_output, true);
2271 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2272 Output();
2273 }
2275 print_method(PHASE_FINAL_CODE);
2277 // He's dead, Jim.
2278 _cfg = (PhaseCFG*)0xdeadbeef;
2279 _regalloc = (PhaseChaitin*)0xdeadbeef;
2280 }
2283 //------------------------------dump_asm---------------------------------------
2284 // Dump formatted assembly
2285 #ifndef PRODUCT
2286 void Compile::dump_asm(int *pcs, uint pc_limit) {
2287 bool cut_short = false;
2288 tty->print_cr("#");
2289 tty->print("# "); _tf->dump(); tty->cr();
2290 tty->print_cr("#");
2292 // For all blocks
2293 int pc = 0x0; // Program counter
2294 char starts_bundle = ' ';
2295 _regalloc->dump_frame();
2297 Node *n = NULL;
2298 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2299 if (VMThread::should_terminate()) {
2300 cut_short = true;
2301 break;
2302 }
2303 Block* block = _cfg->get_block(i);
2304 if (block->is_connector() && !Verbose) {
2305 continue;
2306 }
2307 n = block->head();
2308 if (pcs && n->_idx < pc_limit) {
2309 tty->print("%3.3x ", pcs[n->_idx]);
2310 } else {
2311 tty->print(" ");
2312 }
2313 block->dump_head(_cfg);
2314 if (block->is_connector()) {
2315 tty->print_cr(" # Empty connector block");
2316 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2317 tty->print_cr(" # Block is sole successor of call");
2318 }
2320 // For all instructions
2321 Node *delay = NULL;
2322 for (uint j = 0; j < block->number_of_nodes(); j++) {
2323 if (VMThread::should_terminate()) {
2324 cut_short = true;
2325 break;
2326 }
2327 n = block->get_node(j);
2328 if (valid_bundle_info(n)) {
2329 Bundle* bundle = node_bundling(n);
2330 if (bundle->used_in_unconditional_delay()) {
2331 delay = n;
2332 continue;
2333 }
2334 if (bundle->starts_bundle()) {
2335 starts_bundle = '+';
2336 }
2337 }
2339 if (WizardMode) {
2340 n->dump();
2341 }
2343 if( !n->is_Region() && // Dont print in the Assembly
2344 !n->is_Phi() && // a few noisely useless nodes
2345 !n->is_Proj() &&
2346 !n->is_MachTemp() &&
2347 !n->is_SafePointScalarObject() &&
2348 !n->is_Catch() && // Would be nice to print exception table targets
2349 !n->is_MergeMem() && // Not very interesting
2350 !n->is_top() && // Debug info table constants
2351 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2352 ) {
2353 if (pcs && n->_idx < pc_limit)
2354 tty->print("%3.3x", pcs[n->_idx]);
2355 else
2356 tty->print(" ");
2357 tty->print(" %c ", starts_bundle);
2358 starts_bundle = ' ';
2359 tty->print("\t");
2360 n->format(_regalloc, tty);
2361 tty->cr();
2362 }
2364 // If we have an instruction with a delay slot, and have seen a delay,
2365 // then back up and print it
2366 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2367 assert(delay != NULL, "no unconditional delay instruction");
2368 if (WizardMode) delay->dump();
2370 if (node_bundling(delay)->starts_bundle())
2371 starts_bundle = '+';
2372 if (pcs && n->_idx < pc_limit)
2373 tty->print("%3.3x", pcs[n->_idx]);
2374 else
2375 tty->print(" ");
2376 tty->print(" %c ", starts_bundle);
2377 starts_bundle = ' ';
2378 tty->print("\t");
2379 delay->format(_regalloc, tty);
2380 tty->print_cr("");
2381 delay = NULL;
2382 }
2384 // Dump the exception table as well
2385 if( n->is_Catch() && (Verbose || WizardMode) ) {
2386 // Print the exception table for this offset
2387 _handler_table.print_subtable_for(pc);
2388 }
2389 }
2391 if (pcs && n->_idx < pc_limit)
2392 tty->print_cr("%3.3x", pcs[n->_idx]);
2393 else
2394 tty->print_cr("");
2396 assert(cut_short || delay == NULL, "no unconditional delay branch");
2398 } // End of per-block dump
2399 tty->print_cr("");
2401 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2402 }
2403 #endif
2405 //------------------------------Final_Reshape_Counts---------------------------
2406 // This class defines counters to help identify when a method
2407 // may/must be executed using hardware with only 24-bit precision.
2408 struct Final_Reshape_Counts : public StackObj {
2409 int _call_count; // count non-inlined 'common' calls
2410 int _float_count; // count float ops requiring 24-bit precision
2411 int _double_count; // count double ops requiring more precision
2412 int _java_call_count; // count non-inlined 'java' calls
2413 int _inner_loop_count; // count loops which need alignment
2414 VectorSet _visited; // Visitation flags
2415 Node_List _tests; // Set of IfNodes & PCTableNodes
2417 Final_Reshape_Counts() :
2418 _call_count(0), _float_count(0), _double_count(0),
2419 _java_call_count(0), _inner_loop_count(0),
2420 _visited( Thread::current()->resource_area() ) { }
2422 void inc_call_count () { _call_count ++; }
2423 void inc_float_count () { _float_count ++; }
2424 void inc_double_count() { _double_count++; }
2425 void inc_java_call_count() { _java_call_count++; }
2426 void inc_inner_loop_count() { _inner_loop_count++; }
2428 int get_call_count () const { return _call_count ; }
2429 int get_float_count () const { return _float_count ; }
2430 int get_double_count() const { return _double_count; }
2431 int get_java_call_count() const { return _java_call_count; }
2432 int get_inner_loop_count() const { return _inner_loop_count; }
2433 };
2435 #ifdef ASSERT
2436 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2437 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2438 // Make sure the offset goes inside the instance layout.
2439 return k->contains_field_offset(tp->offset());
2440 // Note that OffsetBot and OffsetTop are very negative.
2441 }
2442 #endif
2444 // Eliminate trivially redundant StoreCMs and accumulate their
2445 // precedence edges.
2446 void Compile::eliminate_redundant_card_marks(Node* n) {
2447 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2448 if (n->in(MemNode::Address)->outcnt() > 1) {
2449 // There are multiple users of the same address so it might be
2450 // possible to eliminate some of the StoreCMs
2451 Node* mem = n->in(MemNode::Memory);
2452 Node* adr = n->in(MemNode::Address);
2453 Node* val = n->in(MemNode::ValueIn);
2454 Node* prev = n;
2455 bool done = false;
2456 // Walk the chain of StoreCMs eliminating ones that match. As
2457 // long as it's a chain of single users then the optimization is
2458 // safe. Eliminating partially redundant StoreCMs would require
2459 // cloning copies down the other paths.
2460 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2461 if (adr == mem->in(MemNode::Address) &&
2462 val == mem->in(MemNode::ValueIn)) {
2463 // redundant StoreCM
2464 if (mem->req() > MemNode::OopStore) {
2465 // Hasn't been processed by this code yet.
2466 n->add_prec(mem->in(MemNode::OopStore));
2467 } else {
2468 // Already converted to precedence edge
2469 for (uint i = mem->req(); i < mem->len(); i++) {
2470 // Accumulate any precedence edges
2471 if (mem->in(i) != NULL) {
2472 n->add_prec(mem->in(i));
2473 }
2474 }
2475 // Everything above this point has been processed.
2476 done = true;
2477 }
2478 // Eliminate the previous StoreCM
2479 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2480 assert(mem->outcnt() == 0, "should be dead");
2481 mem->disconnect_inputs(NULL, this);
2482 } else {
2483 prev = mem;
2484 }
2485 mem = prev->in(MemNode::Memory);
2486 }
2487 }
2488 }
2490 //------------------------------final_graph_reshaping_impl----------------------
2491 // Implement items 1-5 from final_graph_reshaping below.
2492 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2494 if ( n->outcnt() == 0 ) return; // dead node
2495 uint nop = n->Opcode();
2497 // Check for 2-input instruction with "last use" on right input.
2498 // Swap to left input. Implements item (2).
2499 if( n->req() == 3 && // two-input instruction
2500 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2501 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2502 n->in(2)->outcnt() == 1 &&// right use IS a last use
2503 !n->in(2)->is_Con() ) { // right use is not a constant
2504 // Check for commutative opcode
2505 switch( nop ) {
2506 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2507 case Op_MaxI: case Op_MinI:
2508 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2509 case Op_AndL: case Op_XorL: case Op_OrL:
2510 case Op_AndI: case Op_XorI: case Op_OrI: {
2511 // Move "last use" input to left by swapping inputs
2512 n->swap_edges(1, 2);
2513 break;
2514 }
2515 default:
2516 break;
2517 }
2518 }
2520 #ifdef ASSERT
2521 if( n->is_Mem() ) {
2522 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2523 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2524 // oop will be recorded in oop map if load crosses safepoint
2525 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2526 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2527 "raw memory operations should have control edge");
2528 }
2529 #endif
2530 // Count FPU ops and common calls, implements item (3)
2531 switch( nop ) {
2532 // Count all float operations that may use FPU
2533 case Op_AddF:
2534 case Op_SubF:
2535 case Op_MulF:
2536 case Op_DivF:
2537 case Op_NegF:
2538 case Op_ModF:
2539 case Op_ConvI2F:
2540 case Op_ConF:
2541 case Op_CmpF:
2542 case Op_CmpF3:
2543 // case Op_ConvL2F: // longs are split into 32-bit halves
2544 frc.inc_float_count();
2545 break;
2547 case Op_ConvF2D:
2548 case Op_ConvD2F:
2549 frc.inc_float_count();
2550 frc.inc_double_count();
2551 break;
2553 // Count all double operations that may use FPU
2554 case Op_AddD:
2555 case Op_SubD:
2556 case Op_MulD:
2557 case Op_DivD:
2558 case Op_NegD:
2559 case Op_ModD:
2560 case Op_ConvI2D:
2561 case Op_ConvD2I:
2562 // case Op_ConvL2D: // handled by leaf call
2563 // case Op_ConvD2L: // handled by leaf call
2564 case Op_ConD:
2565 case Op_CmpD:
2566 case Op_CmpD3:
2567 frc.inc_double_count();
2568 break;
2569 case Op_Opaque1: // Remove Opaque Nodes before matching
2570 case Op_Opaque2: // Remove Opaque Nodes before matching
2571 n->subsume_by(n->in(1), this);
2572 break;
2573 case Op_CallStaticJava:
2574 case Op_CallJava:
2575 case Op_CallDynamicJava:
2576 frc.inc_java_call_count(); // Count java call site;
2577 case Op_CallRuntime:
2578 case Op_CallLeaf:
2579 case Op_CallLeafNoFP: {
2580 assert( n->is_Call(), "" );
2581 CallNode *call = n->as_Call();
2582 // Count call sites where the FP mode bit would have to be flipped.
2583 // Do not count uncommon runtime calls:
2584 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2585 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2586 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2587 frc.inc_call_count(); // Count the call site
2588 } else { // See if uncommon argument is shared
2589 Node *n = call->in(TypeFunc::Parms);
2590 int nop = n->Opcode();
2591 // Clone shared simple arguments to uncommon calls, item (1).
2592 if( n->outcnt() > 1 &&
2593 !n->is_Proj() &&
2594 nop != Op_CreateEx &&
2595 nop != Op_CheckCastPP &&
2596 nop != Op_DecodeN &&
2597 nop != Op_DecodeNKlass &&
2598 !n->is_Mem() ) {
2599 Node *x = n->clone();
2600 call->set_req( TypeFunc::Parms, x );
2601 }
2602 }
2603 break;
2604 }
2606 case Op_StoreD:
2607 case Op_LoadD:
2608 case Op_LoadD_unaligned:
2609 frc.inc_double_count();
2610 goto handle_mem;
2611 case Op_StoreF:
2612 case Op_LoadF:
2613 frc.inc_float_count();
2614 goto handle_mem;
2616 case Op_StoreCM:
2617 {
2618 // Convert OopStore dependence into precedence edge
2619 Node* prec = n->in(MemNode::OopStore);
2620 n->del_req(MemNode::OopStore);
2621 n->add_prec(prec);
2622 eliminate_redundant_card_marks(n);
2623 }
2625 // fall through
2627 case Op_StoreB:
2628 case Op_StoreC:
2629 case Op_StorePConditional:
2630 case Op_StoreI:
2631 case Op_StoreL:
2632 case Op_StoreIConditional:
2633 case Op_StoreLConditional:
2634 case Op_CompareAndSwapI:
2635 case Op_CompareAndSwapL:
2636 case Op_CompareAndSwapP:
2637 case Op_CompareAndSwapN:
2638 case Op_GetAndAddI:
2639 case Op_GetAndAddL:
2640 case Op_GetAndSetI:
2641 case Op_GetAndSetL:
2642 case Op_GetAndSetP:
2643 case Op_GetAndSetN:
2644 case Op_StoreP:
2645 case Op_StoreN:
2646 case Op_StoreNKlass:
2647 case Op_LoadB:
2648 case Op_LoadUB:
2649 case Op_LoadUS:
2650 case Op_LoadI:
2651 case Op_LoadKlass:
2652 case Op_LoadNKlass:
2653 case Op_LoadL:
2654 case Op_LoadL_unaligned:
2655 case Op_LoadPLocked:
2656 case Op_LoadP:
2657 case Op_LoadN:
2658 case Op_LoadRange:
2659 case Op_LoadS: {
2660 handle_mem:
2661 #ifdef ASSERT
2662 if( VerifyOptoOopOffsets ) {
2663 assert( n->is_Mem(), "" );
2664 MemNode *mem = (MemNode*)n;
2665 // Check to see if address types have grounded out somehow.
2666 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2667 assert( !tp || oop_offset_is_sane(tp), "" );
2668 }
2669 #endif
2670 break;
2671 }
2673 case Op_AddP: { // Assert sane base pointers
2674 Node *addp = n->in(AddPNode::Address);
2675 assert( !addp->is_AddP() ||
2676 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2677 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2678 "Base pointers must match" );
2679 #ifdef _LP64
2680 if ((UseCompressedOops || UseCompressedClassPointers) &&
2681 addp->Opcode() == Op_ConP &&
2682 addp == n->in(AddPNode::Base) &&
2683 n->in(AddPNode::Offset)->is_Con()) {
2684 // Use addressing with narrow klass to load with offset on x86.
2685 // On sparc loading 32-bits constant and decoding it have less
2686 // instructions (4) then load 64-bits constant (7).
2687 // Do this transformation here since IGVN will convert ConN back to ConP.
2688 const Type* t = addp->bottom_type();
2689 if (t->isa_oopptr() || t->isa_klassptr()) {
2690 Node* nn = NULL;
2692 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2694 // Look for existing ConN node of the same exact type.
2695 Node* r = root();
2696 uint cnt = r->outcnt();
2697 for (uint i = 0; i < cnt; i++) {
2698 Node* m = r->raw_out(i);
2699 if (m!= NULL && m->Opcode() == op &&
2700 m->bottom_type()->make_ptr() == t) {
2701 nn = m;
2702 break;
2703 }
2704 }
2705 if (nn != NULL) {
2706 // Decode a narrow oop to match address
2707 // [R12 + narrow_oop_reg<<3 + offset]
2708 if (t->isa_oopptr()) {
2709 nn = new (this) DecodeNNode(nn, t);
2710 } else {
2711 nn = new (this) DecodeNKlassNode(nn, t);
2712 }
2713 n->set_req(AddPNode::Base, nn);
2714 n->set_req(AddPNode::Address, nn);
2715 if (addp->outcnt() == 0) {
2716 addp->disconnect_inputs(NULL, this);
2717 }
2718 }
2719 }
2720 }
2721 #endif
2722 break;
2723 }
2725 #ifdef _LP64
2726 case Op_CastPP:
2727 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2728 Node* in1 = n->in(1);
2729 const Type* t = n->bottom_type();
2730 Node* new_in1 = in1->clone();
2731 new_in1->as_DecodeN()->set_type(t);
2733 if (!Matcher::narrow_oop_use_complex_address()) {
2734 //
2735 // x86, ARM and friends can handle 2 adds in addressing mode
2736 // and Matcher can fold a DecodeN node into address by using
2737 // a narrow oop directly and do implicit NULL check in address:
2738 //
2739 // [R12 + narrow_oop_reg<<3 + offset]
2740 // NullCheck narrow_oop_reg
2741 //
2742 // On other platforms (Sparc) we have to keep new DecodeN node and
2743 // use it to do implicit NULL check in address:
2744 //
2745 // decode_not_null narrow_oop_reg, base_reg
2746 // [base_reg + offset]
2747 // NullCheck base_reg
2748 //
2749 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2750 // to keep the information to which NULL check the new DecodeN node
2751 // corresponds to use it as value in implicit_null_check().
2752 //
2753 new_in1->set_req(0, n->in(0));
2754 }
2756 n->subsume_by(new_in1, this);
2757 if (in1->outcnt() == 0) {
2758 in1->disconnect_inputs(NULL, this);
2759 }
2760 }
2761 break;
2763 case Op_CmpP:
2764 // Do this transformation here to preserve CmpPNode::sub() and
2765 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2766 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2767 Node* in1 = n->in(1);
2768 Node* in2 = n->in(2);
2769 if (!in1->is_DecodeNarrowPtr()) {
2770 in2 = in1;
2771 in1 = n->in(2);
2772 }
2773 assert(in1->is_DecodeNarrowPtr(), "sanity");
2775 Node* new_in2 = NULL;
2776 if (in2->is_DecodeNarrowPtr()) {
2777 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2778 new_in2 = in2->in(1);
2779 } else if (in2->Opcode() == Op_ConP) {
2780 const Type* t = in2->bottom_type();
2781 if (t == TypePtr::NULL_PTR) {
2782 assert(in1->is_DecodeN(), "compare klass to null?");
2783 // Don't convert CmpP null check into CmpN if compressed
2784 // oops implicit null check is not generated.
2785 // This will allow to generate normal oop implicit null check.
2786 if (Matcher::gen_narrow_oop_implicit_null_checks())
2787 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2788 //
2789 // This transformation together with CastPP transformation above
2790 // will generated code for implicit NULL checks for compressed oops.
2791 //
2792 // The original code after Optimize()
2793 //
2794 // LoadN memory, narrow_oop_reg
2795 // decode narrow_oop_reg, base_reg
2796 // CmpP base_reg, NULL
2797 // CastPP base_reg // NotNull
2798 // Load [base_reg + offset], val_reg
2799 //
2800 // after these transformations will be
2801 //
2802 // LoadN memory, narrow_oop_reg
2803 // CmpN narrow_oop_reg, NULL
2804 // decode_not_null narrow_oop_reg, base_reg
2805 // Load [base_reg + offset], val_reg
2806 //
2807 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2808 // since narrow oops can be used in debug info now (see the code in
2809 // final_graph_reshaping_walk()).
2810 //
2811 // At the end the code will be matched to
2812 // on x86:
2813 //
2814 // Load_narrow_oop memory, narrow_oop_reg
2815 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2816 // NullCheck narrow_oop_reg
2817 //
2818 // and on sparc:
2819 //
2820 // Load_narrow_oop memory, narrow_oop_reg
2821 // decode_not_null narrow_oop_reg, base_reg
2822 // Load [base_reg + offset], val_reg
2823 // NullCheck base_reg
2824 //
2825 } else if (t->isa_oopptr()) {
2826 new_in2 = ConNode::make(this, t->make_narrowoop());
2827 } else if (t->isa_klassptr()) {
2828 new_in2 = ConNode::make(this, t->make_narrowklass());
2829 }
2830 }
2831 if (new_in2 != NULL) {
2832 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2833 n->subsume_by(cmpN, this);
2834 if (in1->outcnt() == 0) {
2835 in1->disconnect_inputs(NULL, this);
2836 }
2837 if (in2->outcnt() == 0) {
2838 in2->disconnect_inputs(NULL, this);
2839 }
2840 }
2841 }
2842 break;
2844 case Op_DecodeN:
2845 case Op_DecodeNKlass:
2846 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2847 // DecodeN could be pinned when it can't be fold into
2848 // an address expression, see the code for Op_CastPP above.
2849 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2850 break;
2852 case Op_EncodeP:
2853 case Op_EncodePKlass: {
2854 Node* in1 = n->in(1);
2855 if (in1->is_DecodeNarrowPtr()) {
2856 n->subsume_by(in1->in(1), this);
2857 } else if (in1->Opcode() == Op_ConP) {
2858 const Type* t = in1->bottom_type();
2859 if (t == TypePtr::NULL_PTR) {
2860 assert(t->isa_oopptr(), "null klass?");
2861 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2862 } else if (t->isa_oopptr()) {
2863 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2864 } else if (t->isa_klassptr()) {
2865 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2866 }
2867 }
2868 if (in1->outcnt() == 0) {
2869 in1->disconnect_inputs(NULL, this);
2870 }
2871 break;
2872 }
2874 case Op_Proj: {
2875 if (OptimizeStringConcat) {
2876 ProjNode* p = n->as_Proj();
2877 if (p->_is_io_use) {
2878 // Separate projections were used for the exception path which
2879 // are normally removed by a late inline. If it wasn't inlined
2880 // then they will hang around and should just be replaced with
2881 // the original one.
2882 Node* proj = NULL;
2883 // Replace with just one
2884 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2885 Node *use = i.get();
2886 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2887 proj = use;
2888 break;
2889 }
2890 }
2891 assert(proj != NULL, "must be found");
2892 p->subsume_by(proj, this);
2893 }
2894 }
2895 break;
2896 }
2898 case Op_Phi:
2899 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
2900 // The EncodeP optimization may create Phi with the same edges
2901 // for all paths. It is not handled well by Register Allocator.
2902 Node* unique_in = n->in(1);
2903 assert(unique_in != NULL, "");
2904 uint cnt = n->req();
2905 for (uint i = 2; i < cnt; i++) {
2906 Node* m = n->in(i);
2907 assert(m != NULL, "");
2908 if (unique_in != m)
2909 unique_in = NULL;
2910 }
2911 if (unique_in != NULL) {
2912 n->subsume_by(unique_in, this);
2913 }
2914 }
2915 break;
2917 #endif
2919 case Op_ModI:
2920 if (UseDivMod) {
2921 // Check if a%b and a/b both exist
2922 Node* d = n->find_similar(Op_DivI);
2923 if (d) {
2924 // Replace them with a fused divmod if supported
2925 if (Matcher::has_match_rule(Op_DivModI)) {
2926 DivModINode* divmod = DivModINode::make(this, n);
2927 d->subsume_by(divmod->div_proj(), this);
2928 n->subsume_by(divmod->mod_proj(), this);
2929 } else {
2930 // replace a%b with a-((a/b)*b)
2931 Node* mult = new (this) MulINode(d, d->in(2));
2932 Node* sub = new (this) SubINode(d->in(1), mult);
2933 n->subsume_by(sub, this);
2934 }
2935 }
2936 }
2937 break;
2939 case Op_ModL:
2940 if (UseDivMod) {
2941 // Check if a%b and a/b both exist
2942 Node* d = n->find_similar(Op_DivL);
2943 if (d) {
2944 // Replace them with a fused divmod if supported
2945 if (Matcher::has_match_rule(Op_DivModL)) {
2946 DivModLNode* divmod = DivModLNode::make(this, n);
2947 d->subsume_by(divmod->div_proj(), this);
2948 n->subsume_by(divmod->mod_proj(), this);
2949 } else {
2950 // replace a%b with a-((a/b)*b)
2951 Node* mult = new (this) MulLNode(d, d->in(2));
2952 Node* sub = new (this) SubLNode(d->in(1), mult);
2953 n->subsume_by(sub, this);
2954 }
2955 }
2956 }
2957 break;
2959 case Op_LoadVector:
2960 case Op_StoreVector:
2961 break;
2963 case Op_PackB:
2964 case Op_PackS:
2965 case Op_PackI:
2966 case Op_PackF:
2967 case Op_PackL:
2968 case Op_PackD:
2969 if (n->req()-1 > 2) {
2970 // Replace many operand PackNodes with a binary tree for matching
2971 PackNode* p = (PackNode*) n;
2972 Node* btp = p->binary_tree_pack(this, 1, n->req());
2973 n->subsume_by(btp, this);
2974 }
2975 break;
2976 case Op_Loop:
2977 case Op_CountedLoop:
2978 if (n->as_Loop()->is_inner_loop()) {
2979 frc.inc_inner_loop_count();
2980 }
2981 break;
2982 case Op_LShiftI:
2983 case Op_RShiftI:
2984 case Op_URShiftI:
2985 case Op_LShiftL:
2986 case Op_RShiftL:
2987 case Op_URShiftL:
2988 if (Matcher::need_masked_shift_count) {
2989 // The cpu's shift instructions don't restrict the count to the
2990 // lower 5/6 bits. We need to do the masking ourselves.
2991 Node* in2 = n->in(2);
2992 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2993 const TypeInt* t = in2->find_int_type();
2994 if (t != NULL && t->is_con()) {
2995 juint shift = t->get_con();
2996 if (shift > mask) { // Unsigned cmp
2997 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
2998 }
2999 } else {
3000 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3001 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3002 n->set_req(2, shift);
3003 }
3004 }
3005 if (in2->outcnt() == 0) { // Remove dead node
3006 in2->disconnect_inputs(NULL, this);
3007 }
3008 }
3009 break;
3010 case Op_MemBarStoreStore:
3011 case Op_MemBarRelease:
3012 // Break the link with AllocateNode: it is no longer useful and
3013 // confuses register allocation.
3014 if (n->req() > MemBarNode::Precedent) {
3015 n->set_req(MemBarNode::Precedent, top());
3016 }
3017 break;
3018 // Must set a control edge on all nodes that produce a FlagsProj
3019 // so they can't escape the block that consumes the flags.
3020 // Must also set the non throwing branch as the control
3021 // for all nodes that depends on the result. Unless the node
3022 // already have a control that isn't the control of the
3023 // flag producer
3024 case Op_FlagsProj:
3025 {
3026 MathExactNode* math = (MathExactNode*) n->in(0);
3027 Node* ctrl = math->control_node();
3028 Node* non_throwing = math->non_throwing_branch();
3029 math->set_req(0, ctrl);
3031 Node* result = math->result_node();
3032 if (result != NULL) {
3033 for (DUIterator_Fast jmax, j = result->fast_outs(jmax); j < jmax; j++) {
3034 Node* out = result->fast_out(j);
3035 // Phi nodes shouldn't be moved. They would only match below if they
3036 // had the same control as the MathExactNode. The only time that
3037 // would happen is if the Phi is also an input to the MathExact
3038 //
3039 // Cmp nodes shouldn't have control set at all.
3040 if (out->is_Phi() ||
3041 out->is_Cmp()) {
3042 continue;
3043 }
3045 if (out->in(0) == NULL) {
3046 out->set_req(0, non_throwing);
3047 } else if (out->in(0) == ctrl) {
3048 out->set_req(0, non_throwing);
3049 }
3050 }
3051 }
3052 }
3053 break;
3054 default:
3055 assert( !n->is_Call(), "" );
3056 assert( !n->is_Mem(), "" );
3057 break;
3058 }
3060 // Collect CFG split points
3061 if (n->is_MultiBranch())
3062 frc._tests.push(n);
3063 }
3065 //------------------------------final_graph_reshaping_walk---------------------
3066 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3067 // requires that the walk visits a node's inputs before visiting the node.
3068 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3069 ResourceArea *area = Thread::current()->resource_area();
3070 Unique_Node_List sfpt(area);
3072 frc._visited.set(root->_idx); // first, mark node as visited
3073 uint cnt = root->req();
3074 Node *n = root;
3075 uint i = 0;
3076 while (true) {
3077 if (i < cnt) {
3078 // Place all non-visited non-null inputs onto stack
3079 Node* m = n->in(i);
3080 ++i;
3081 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3082 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
3083 sfpt.push(m);
3084 cnt = m->req();
3085 nstack.push(n, i); // put on stack parent and next input's index
3086 n = m;
3087 i = 0;
3088 }
3089 } else {
3090 // Now do post-visit work
3091 final_graph_reshaping_impl( n, frc );
3092 if (nstack.is_empty())
3093 break; // finished
3094 n = nstack.node(); // Get node from stack
3095 cnt = n->req();
3096 i = nstack.index();
3097 nstack.pop(); // Shift to the next node on stack
3098 }
3099 }
3101 // Skip next transformation if compressed oops are not used.
3102 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3103 (!UseCompressedOops && !UseCompressedClassPointers))
3104 return;
3106 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3107 // It could be done for an uncommon traps or any safepoints/calls
3108 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3109 while (sfpt.size() > 0) {
3110 n = sfpt.pop();
3111 JVMState *jvms = n->as_SafePoint()->jvms();
3112 assert(jvms != NULL, "sanity");
3113 int start = jvms->debug_start();
3114 int end = n->req();
3115 bool is_uncommon = (n->is_CallStaticJava() &&
3116 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3117 for (int j = start; j < end; j++) {
3118 Node* in = n->in(j);
3119 if (in->is_DecodeNarrowPtr()) {
3120 bool safe_to_skip = true;
3121 if (!is_uncommon ) {
3122 // Is it safe to skip?
3123 for (uint i = 0; i < in->outcnt(); i++) {
3124 Node* u = in->raw_out(i);
3125 if (!u->is_SafePoint() ||
3126 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3127 safe_to_skip = false;
3128 }
3129 }
3130 }
3131 if (safe_to_skip) {
3132 n->set_req(j, in->in(1));
3133 }
3134 if (in->outcnt() == 0) {
3135 in->disconnect_inputs(NULL, this);
3136 }
3137 }
3138 }
3139 }
3140 }
3142 //------------------------------final_graph_reshaping--------------------------
3143 // Final Graph Reshaping.
3144 //
3145 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3146 // and not commoned up and forced early. Must come after regular
3147 // optimizations to avoid GVN undoing the cloning. Clone constant
3148 // inputs to Loop Phis; these will be split by the allocator anyways.
3149 // Remove Opaque nodes.
3150 // (2) Move last-uses by commutative operations to the left input to encourage
3151 // Intel update-in-place two-address operations and better register usage
3152 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3153 // calls canonicalizing them back.
3154 // (3) Count the number of double-precision FP ops, single-precision FP ops
3155 // and call sites. On Intel, we can get correct rounding either by
3156 // forcing singles to memory (requires extra stores and loads after each
3157 // FP bytecode) or we can set a rounding mode bit (requires setting and
3158 // clearing the mode bit around call sites). The mode bit is only used
3159 // if the relative frequency of single FP ops to calls is low enough.
3160 // This is a key transform for SPEC mpeg_audio.
3161 // (4) Detect infinite loops; blobs of code reachable from above but not
3162 // below. Several of the Code_Gen algorithms fail on such code shapes,
3163 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3164 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3165 // Detection is by looking for IfNodes where only 1 projection is
3166 // reachable from below or CatchNodes missing some targets.
3167 // (5) Assert for insane oop offsets in debug mode.
3169 bool Compile::final_graph_reshaping() {
3170 // an infinite loop may have been eliminated by the optimizer,
3171 // in which case the graph will be empty.
3172 if (root()->req() == 1) {
3173 record_method_not_compilable("trivial infinite loop");
3174 return true;
3175 }
3177 // Expensive nodes have their control input set to prevent the GVN
3178 // from freely commoning them. There's no GVN beyond this point so
3179 // no need to keep the control input. We want the expensive nodes to
3180 // be freely moved to the least frequent code path by gcm.
3181 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3182 for (int i = 0; i < expensive_count(); i++) {
3183 _expensive_nodes->at(i)->set_req(0, NULL);
3184 }
3186 Final_Reshape_Counts frc;
3188 // Visit everybody reachable!
3189 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3190 Node_Stack nstack(unique() >> 1);
3191 final_graph_reshaping_walk(nstack, root(), frc);
3193 // Check for unreachable (from below) code (i.e., infinite loops).
3194 for( uint i = 0; i < frc._tests.size(); i++ ) {
3195 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3196 // Get number of CFG targets.
3197 // Note that PCTables include exception targets after calls.
3198 uint required_outcnt = n->required_outcnt();
3199 if (n->outcnt() != required_outcnt) {
3200 // Check for a few special cases. Rethrow Nodes never take the
3201 // 'fall-thru' path, so expected kids is 1 less.
3202 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3203 if (n->in(0)->in(0)->is_Call()) {
3204 CallNode *call = n->in(0)->in(0)->as_Call();
3205 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3206 required_outcnt--; // Rethrow always has 1 less kid
3207 } else if (call->req() > TypeFunc::Parms &&
3208 call->is_CallDynamicJava()) {
3209 // Check for null receiver. In such case, the optimizer has
3210 // detected that the virtual call will always result in a null
3211 // pointer exception. The fall-through projection of this CatchNode
3212 // will not be populated.
3213 Node *arg0 = call->in(TypeFunc::Parms);
3214 if (arg0->is_Type() &&
3215 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3216 required_outcnt--;
3217 }
3218 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3219 call->req() > TypeFunc::Parms+1 &&
3220 call->is_CallStaticJava()) {
3221 // Check for negative array length. In such case, the optimizer has
3222 // detected that the allocation attempt will always result in an
3223 // exception. There is no fall-through projection of this CatchNode .
3224 Node *arg1 = call->in(TypeFunc::Parms+1);
3225 if (arg1->is_Type() &&
3226 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3227 required_outcnt--;
3228 }
3229 }
3230 }
3231 }
3232 // Recheck with a better notion of 'required_outcnt'
3233 if (n->outcnt() != required_outcnt) {
3234 record_method_not_compilable("malformed control flow");
3235 return true; // Not all targets reachable!
3236 }
3237 }
3238 // Check that I actually visited all kids. Unreached kids
3239 // must be infinite loops.
3240 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3241 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3242 record_method_not_compilable("infinite loop");
3243 return true; // Found unvisited kid; must be unreach
3244 }
3245 }
3247 // If original bytecodes contained a mixture of floats and doubles
3248 // check if the optimizer has made it homogenous, item (3).
3249 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3250 frc.get_float_count() > 32 &&
3251 frc.get_double_count() == 0 &&
3252 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3253 set_24_bit_selection_and_mode( false, true );
3254 }
3256 set_java_calls(frc.get_java_call_count());
3257 set_inner_loops(frc.get_inner_loop_count());
3259 // No infinite loops, no reason to bail out.
3260 return false;
3261 }
3263 //-----------------------------too_many_traps----------------------------------
3264 // Report if there are too many traps at the current method and bci.
3265 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3266 bool Compile::too_many_traps(ciMethod* method,
3267 int bci,
3268 Deoptimization::DeoptReason reason) {
3269 ciMethodData* md = method->method_data();
3270 if (md->is_empty()) {
3271 // Assume the trap has not occurred, or that it occurred only
3272 // because of a transient condition during start-up in the interpreter.
3273 return false;
3274 }
3275 if (md->has_trap_at(bci, reason) != 0) {
3276 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3277 // Also, if there are multiple reasons, or if there is no per-BCI record,
3278 // assume the worst.
3279 if (log())
3280 log()->elem("observe trap='%s' count='%d'",
3281 Deoptimization::trap_reason_name(reason),
3282 md->trap_count(reason));
3283 return true;
3284 } else {
3285 // Ignore method/bci and see if there have been too many globally.
3286 return too_many_traps(reason, md);
3287 }
3288 }
3290 // Less-accurate variant which does not require a method and bci.
3291 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3292 ciMethodData* logmd) {
3293 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
3294 // Too many traps globally.
3295 // Note that we use cumulative trap_count, not just md->trap_count.
3296 if (log()) {
3297 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3298 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3299 Deoptimization::trap_reason_name(reason),
3300 mcount, trap_count(reason));
3301 }
3302 return true;
3303 } else {
3304 // The coast is clear.
3305 return false;
3306 }
3307 }
3309 //--------------------------too_many_recompiles--------------------------------
3310 // Report if there are too many recompiles at the current method and bci.
3311 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3312 // Is not eager to return true, since this will cause the compiler to use
3313 // Action_none for a trap point, to avoid too many recompilations.
3314 bool Compile::too_many_recompiles(ciMethod* method,
3315 int bci,
3316 Deoptimization::DeoptReason reason) {
3317 ciMethodData* md = method->method_data();
3318 if (md->is_empty()) {
3319 // Assume the trap has not occurred, or that it occurred only
3320 // because of a transient condition during start-up in the interpreter.
3321 return false;
3322 }
3323 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3324 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3325 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3326 Deoptimization::DeoptReason per_bc_reason
3327 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3328 if ((per_bc_reason == Deoptimization::Reason_none
3329 || md->has_trap_at(bci, reason) != 0)
3330 // The trap frequency measure we care about is the recompile count:
3331 && md->trap_recompiled_at(bci)
3332 && md->overflow_recompile_count() >= bc_cutoff) {
3333 // Do not emit a trap here if it has already caused recompilations.
3334 // Also, if there are multiple reasons, or if there is no per-BCI record,
3335 // assume the worst.
3336 if (log())
3337 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3338 Deoptimization::trap_reason_name(reason),
3339 md->trap_count(reason),
3340 md->overflow_recompile_count());
3341 return true;
3342 } else if (trap_count(reason) != 0
3343 && decompile_count() >= m_cutoff) {
3344 // Too many recompiles globally, and we have seen this sort of trap.
3345 // Use cumulative decompile_count, not just md->decompile_count.
3346 if (log())
3347 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3348 Deoptimization::trap_reason_name(reason),
3349 md->trap_count(reason), trap_count(reason),
3350 md->decompile_count(), decompile_count());
3351 return true;
3352 } else {
3353 // The coast is clear.
3354 return false;
3355 }
3356 }
3359 #ifndef PRODUCT
3360 //------------------------------verify_graph_edges---------------------------
3361 // Walk the Graph and verify that there is a one-to-one correspondence
3362 // between Use-Def edges and Def-Use edges in the graph.
3363 void Compile::verify_graph_edges(bool no_dead_code) {
3364 if (VerifyGraphEdges) {
3365 ResourceArea *area = Thread::current()->resource_area();
3366 Unique_Node_List visited(area);
3367 // Call recursive graph walk to check edges
3368 _root->verify_edges(visited);
3369 if (no_dead_code) {
3370 // Now make sure that no visited node is used by an unvisited node.
3371 bool dead_nodes = 0;
3372 Unique_Node_List checked(area);
3373 while (visited.size() > 0) {
3374 Node* n = visited.pop();
3375 checked.push(n);
3376 for (uint i = 0; i < n->outcnt(); i++) {
3377 Node* use = n->raw_out(i);
3378 if (checked.member(use)) continue; // already checked
3379 if (visited.member(use)) continue; // already in the graph
3380 if (use->is_Con()) continue; // a dead ConNode is OK
3381 // At this point, we have found a dead node which is DU-reachable.
3382 if (dead_nodes++ == 0)
3383 tty->print_cr("*** Dead nodes reachable via DU edges:");
3384 use->dump(2);
3385 tty->print_cr("---");
3386 checked.push(use); // No repeats; pretend it is now checked.
3387 }
3388 }
3389 assert(dead_nodes == 0, "using nodes must be reachable from root");
3390 }
3391 }
3392 }
3394 // Verify GC barriers consistency
3395 // Currently supported:
3396 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3397 void Compile::verify_barriers() {
3398 if (UseG1GC) {
3399 // Verify G1 pre-barriers
3400 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3402 ResourceArea *area = Thread::current()->resource_area();
3403 Unique_Node_List visited(area);
3404 Node_List worklist(area);
3405 // We're going to walk control flow backwards starting from the Root
3406 worklist.push(_root);
3407 while (worklist.size() > 0) {
3408 Node* x = worklist.pop();
3409 if (x == NULL || x == top()) continue;
3410 if (visited.member(x)) {
3411 continue;
3412 } else {
3413 visited.push(x);
3414 }
3416 if (x->is_Region()) {
3417 for (uint i = 1; i < x->req(); i++) {
3418 worklist.push(x->in(i));
3419 }
3420 } else {
3421 worklist.push(x->in(0));
3422 // We are looking for the pattern:
3423 // /->ThreadLocal
3424 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3425 // \->ConI(0)
3426 // We want to verify that the If and the LoadB have the same control
3427 // See GraphKit::g1_write_barrier_pre()
3428 if (x->is_If()) {
3429 IfNode *iff = x->as_If();
3430 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3431 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3432 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3433 && cmp->in(1)->is_Load()) {
3434 LoadNode* load = cmp->in(1)->as_Load();
3435 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3436 && load->in(2)->in(3)->is_Con()
3437 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3439 Node* if_ctrl = iff->in(0);
3440 Node* load_ctrl = load->in(0);
3442 if (if_ctrl != load_ctrl) {
3443 // Skip possible CProj->NeverBranch in infinite loops
3444 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3445 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3446 if_ctrl = if_ctrl->in(0)->in(0);
3447 }
3448 }
3449 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3450 }
3451 }
3452 }
3453 }
3454 }
3455 }
3456 }
3457 }
3459 #endif
3461 // The Compile object keeps track of failure reasons separately from the ciEnv.
3462 // This is required because there is not quite a 1-1 relation between the
3463 // ciEnv and its compilation task and the Compile object. Note that one
3464 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3465 // to backtrack and retry without subsuming loads. Other than this backtracking
3466 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3467 // by the logic in C2Compiler.
3468 void Compile::record_failure(const char* reason) {
3469 if (log() != NULL) {
3470 log()->elem("failure reason='%s' phase='compile'", reason);
3471 }
3472 if (_failure_reason == NULL) {
3473 // Record the first failure reason.
3474 _failure_reason = reason;
3475 }
3477 EventCompilerFailure event;
3478 if (event.should_commit()) {
3479 event.set_compileID(Compile::compile_id());
3480 event.set_failure(reason);
3481 event.commit();
3482 }
3484 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3485 C->print_method(PHASE_FAILURE);
3486 }
3487 _root = NULL; // flush the graph, too
3488 }
3490 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3491 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3492 _phase_name(name), _dolog(dolog)
3493 {
3494 if (dolog) {
3495 C = Compile::current();
3496 _log = C->log();
3497 } else {
3498 C = NULL;
3499 _log = NULL;
3500 }
3501 if (_log != NULL) {
3502 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3503 _log->stamp();
3504 _log->end_head();
3505 }
3506 }
3508 Compile::TracePhase::~TracePhase() {
3510 C = Compile::current();
3511 if (_dolog) {
3512 _log = C->log();
3513 } else {
3514 _log = NULL;
3515 }
3517 #ifdef ASSERT
3518 if (PrintIdealNodeCount) {
3519 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3520 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3521 }
3523 if (VerifyIdealNodeCount) {
3524 Compile::current()->print_missing_nodes();
3525 }
3526 #endif
3528 if (_log != NULL) {
3529 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3530 }
3531 }
3533 //=============================================================================
3534 // Two Constant's are equal when the type and the value are equal.
3535 bool Compile::Constant::operator==(const Constant& other) {
3536 if (type() != other.type() ) return false;
3537 if (can_be_reused() != other.can_be_reused()) return false;
3538 // For floating point values we compare the bit pattern.
3539 switch (type()) {
3540 case T_FLOAT: return (_v._value.i == other._v._value.i);
3541 case T_LONG:
3542 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3543 case T_OBJECT:
3544 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3545 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3546 case T_METADATA: return (_v._metadata == other._v._metadata);
3547 default: ShouldNotReachHere();
3548 }
3549 return false;
3550 }
3552 static int type_to_size_in_bytes(BasicType t) {
3553 switch (t) {
3554 case T_LONG: return sizeof(jlong );
3555 case T_FLOAT: return sizeof(jfloat );
3556 case T_DOUBLE: return sizeof(jdouble);
3557 case T_METADATA: return sizeof(Metadata*);
3558 // We use T_VOID as marker for jump-table entries (labels) which
3559 // need an internal word relocation.
3560 case T_VOID:
3561 case T_ADDRESS:
3562 case T_OBJECT: return sizeof(jobject);
3563 }
3565 ShouldNotReachHere();
3566 return -1;
3567 }
3569 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3570 // sort descending
3571 if (a->freq() > b->freq()) return -1;
3572 if (a->freq() < b->freq()) return 1;
3573 return 0;
3574 }
3576 void Compile::ConstantTable::calculate_offsets_and_size() {
3577 // First, sort the array by frequencies.
3578 _constants.sort(qsort_comparator);
3580 #ifdef ASSERT
3581 // Make sure all jump-table entries were sorted to the end of the
3582 // array (they have a negative frequency).
3583 bool found_void = false;
3584 for (int i = 0; i < _constants.length(); i++) {
3585 Constant con = _constants.at(i);
3586 if (con.type() == T_VOID)
3587 found_void = true; // jump-tables
3588 else
3589 assert(!found_void, "wrong sorting");
3590 }
3591 #endif
3593 int offset = 0;
3594 for (int i = 0; i < _constants.length(); i++) {
3595 Constant* con = _constants.adr_at(i);
3597 // Align offset for type.
3598 int typesize = type_to_size_in_bytes(con->type());
3599 offset = align_size_up(offset, typesize);
3600 con->set_offset(offset); // set constant's offset
3602 if (con->type() == T_VOID) {
3603 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3604 offset = offset + typesize * n->outcnt(); // expand jump-table
3605 } else {
3606 offset = offset + typesize;
3607 }
3608 }
3610 // Align size up to the next section start (which is insts; see
3611 // CodeBuffer::align_at_start).
3612 assert(_size == -1, "already set?");
3613 _size = align_size_up(offset, CodeEntryAlignment);
3614 }
3616 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3617 MacroAssembler _masm(&cb);
3618 for (int i = 0; i < _constants.length(); i++) {
3619 Constant con = _constants.at(i);
3620 address constant_addr;
3621 switch (con.type()) {
3622 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3623 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3624 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3625 case T_OBJECT: {
3626 jobject obj = con.get_jobject();
3627 int oop_index = _masm.oop_recorder()->find_index(obj);
3628 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3629 break;
3630 }
3631 case T_ADDRESS: {
3632 address addr = (address) con.get_jobject();
3633 constant_addr = _masm.address_constant(addr);
3634 break;
3635 }
3636 // We use T_VOID as marker for jump-table entries (labels) which
3637 // need an internal word relocation.
3638 case T_VOID: {
3639 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3640 // Fill the jump-table with a dummy word. The real value is
3641 // filled in later in fill_jump_table.
3642 address dummy = (address) n;
3643 constant_addr = _masm.address_constant(dummy);
3644 // Expand jump-table
3645 for (uint i = 1; i < n->outcnt(); i++) {
3646 address temp_addr = _masm.address_constant(dummy + i);
3647 assert(temp_addr, "consts section too small");
3648 }
3649 break;
3650 }
3651 case T_METADATA: {
3652 Metadata* obj = con.get_metadata();
3653 int metadata_index = _masm.oop_recorder()->find_index(obj);
3654 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3655 break;
3656 }
3657 default: ShouldNotReachHere();
3658 }
3659 assert(constant_addr, "consts section too small");
3660 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3661 }
3662 }
3664 int Compile::ConstantTable::find_offset(Constant& con) const {
3665 int idx = _constants.find(con);
3666 assert(idx != -1, "constant must be in constant table");
3667 int offset = _constants.at(idx).offset();
3668 assert(offset != -1, "constant table not emitted yet?");
3669 return offset;
3670 }
3672 void Compile::ConstantTable::add(Constant& con) {
3673 if (con.can_be_reused()) {
3674 int idx = _constants.find(con);
3675 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3676 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3677 return;
3678 }
3679 }
3680 (void) _constants.append(con);
3681 }
3683 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3684 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3685 Constant con(type, value, b->_freq);
3686 add(con);
3687 return con;
3688 }
3690 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3691 Constant con(metadata);
3692 add(con);
3693 return con;
3694 }
3696 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3697 jvalue value;
3698 BasicType type = oper->type()->basic_type();
3699 switch (type) {
3700 case T_LONG: value.j = oper->constantL(); break;
3701 case T_FLOAT: value.f = oper->constantF(); break;
3702 case T_DOUBLE: value.d = oper->constantD(); break;
3703 case T_OBJECT:
3704 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3705 case T_METADATA: return add((Metadata*)oper->constant()); break;
3706 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3707 }
3708 return add(n, type, value);
3709 }
3711 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3712 jvalue value;
3713 // We can use the node pointer here to identify the right jump-table
3714 // as this method is called from Compile::Fill_buffer right before
3715 // the MachNodes are emitted and the jump-table is filled (means the
3716 // MachNode pointers do not change anymore).
3717 value.l = (jobject) n;
3718 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3719 add(con);
3720 return con;
3721 }
3723 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3724 // If called from Compile::scratch_emit_size do nothing.
3725 if (Compile::current()->in_scratch_emit_size()) return;
3727 assert(labels.is_nonempty(), "must be");
3728 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3730 // Since MachConstantNode::constant_offset() also contains
3731 // table_base_offset() we need to subtract the table_base_offset()
3732 // to get the plain offset into the constant table.
3733 int offset = n->constant_offset() - table_base_offset();
3735 MacroAssembler _masm(&cb);
3736 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3738 for (uint i = 0; i < n->outcnt(); i++) {
3739 address* constant_addr = &jump_table_base[i];
3740 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i)));
3741 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3742 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3743 }
3744 }
3746 void Compile::dump_inlining() {
3747 if (print_inlining() || print_intrinsics()) {
3748 // Print inlining message for candidates that we couldn't inline
3749 // for lack of space or non constant receiver
3750 for (int i = 0; i < _late_inlines.length(); i++) {
3751 CallGenerator* cg = _late_inlines.at(i);
3752 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3753 }
3754 Unique_Node_List useful;
3755 useful.push(root());
3756 for (uint next = 0; next < useful.size(); ++next) {
3757 Node* n = useful.at(next);
3758 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3759 CallNode* call = n->as_Call();
3760 CallGenerator* cg = call->generator();
3761 cg->print_inlining_late("receiver not constant");
3762 }
3763 uint max = n->len();
3764 for ( uint i = 0; i < max; ++i ) {
3765 Node *m = n->in(i);
3766 if ( m == NULL ) continue;
3767 useful.push(m);
3768 }
3769 }
3770 for (int i = 0; i < _print_inlining_list->length(); i++) {
3771 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string());
3772 }
3773 }
3774 }
3776 // Dump inlining replay data to the stream.
3777 // Don't change thread state and acquire any locks.
3778 void Compile::dump_inline_data(outputStream* out) {
3779 InlineTree* inl_tree = ilt();
3780 if (inl_tree != NULL) {
3781 out->print(" inline %d", inl_tree->count());
3782 inl_tree->dump_replay_data(out);
3783 }
3784 }
3786 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3787 if (n1->Opcode() < n2->Opcode()) return -1;
3788 else if (n1->Opcode() > n2->Opcode()) return 1;
3790 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()));
3791 for (uint i = 1; i < n1->req(); i++) {
3792 if (n1->in(i) < n2->in(i)) return -1;
3793 else if (n1->in(i) > n2->in(i)) return 1;
3794 }
3796 return 0;
3797 }
3799 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3800 Node* n1 = *n1p;
3801 Node* n2 = *n2p;
3803 return cmp_expensive_nodes(n1, n2);
3804 }
3806 void Compile::sort_expensive_nodes() {
3807 if (!expensive_nodes_sorted()) {
3808 _expensive_nodes->sort(cmp_expensive_nodes);
3809 }
3810 }
3812 bool Compile::expensive_nodes_sorted() const {
3813 for (int i = 1; i < _expensive_nodes->length(); i++) {
3814 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3815 return false;
3816 }
3817 }
3818 return true;
3819 }
3821 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3822 if (_expensive_nodes->length() == 0) {
3823 return false;
3824 }
3826 assert(OptimizeExpensiveOps, "optimization off?");
3828 // Take this opportunity to remove dead nodes from the list
3829 int j = 0;
3830 for (int i = 0; i < _expensive_nodes->length(); i++) {
3831 Node* n = _expensive_nodes->at(i);
3832 if (!n->is_unreachable(igvn)) {
3833 assert(n->is_expensive(), "should be expensive");
3834 _expensive_nodes->at_put(j, n);
3835 j++;
3836 }
3837 }
3838 _expensive_nodes->trunc_to(j);
3840 // Then sort the list so that similar nodes are next to each other
3841 // and check for at least two nodes of identical kind with same data
3842 // inputs.
3843 sort_expensive_nodes();
3845 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3846 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3847 return true;
3848 }
3849 }
3851 return false;
3852 }
3854 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3855 if (_expensive_nodes->length() == 0) {
3856 return;
3857 }
3859 assert(OptimizeExpensiveOps, "optimization off?");
3861 // Sort to bring similar nodes next to each other and clear the
3862 // control input of nodes for which there's only a single copy.
3863 sort_expensive_nodes();
3865 int j = 0;
3866 int identical = 0;
3867 int i = 0;
3868 for (; i < _expensive_nodes->length()-1; i++) {
3869 assert(j <= i, "can't write beyond current index");
3870 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3871 identical++;
3872 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3873 continue;
3874 }
3875 if (identical > 0) {
3876 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3877 identical = 0;
3878 } else {
3879 Node* n = _expensive_nodes->at(i);
3880 igvn.hash_delete(n);
3881 n->set_req(0, NULL);
3882 igvn.hash_insert(n);
3883 }
3884 }
3885 if (identical > 0) {
3886 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3887 } else if (_expensive_nodes->length() >= 1) {
3888 Node* n = _expensive_nodes->at(i);
3889 igvn.hash_delete(n);
3890 n->set_req(0, NULL);
3891 igvn.hash_insert(n);
3892 }
3893 _expensive_nodes->trunc_to(j);
3894 }
3896 void Compile::add_expensive_node(Node * n) {
3897 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3898 assert(n->is_expensive(), "expensive nodes with non-null control here only");
3899 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
3900 if (OptimizeExpensiveOps) {
3901 _expensive_nodes->append(n);
3902 } else {
3903 // Clear control input and let IGVN optimize expensive nodes if
3904 // OptimizeExpensiveOps is off.
3905 n->set_req(0, NULL);
3906 }
3907 }
3909 /**
3910 * Remove the speculative part of types and clean up the graph
3911 */
3912 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
3913 if (UseTypeSpeculation) {
3914 Unique_Node_List worklist;
3915 worklist.push(root());
3916 int modified = 0;
3917 // Go over all type nodes that carry a speculative type, drop the
3918 // speculative part of the type and enqueue the node for an igvn
3919 // which may optimize it out.
3920 for (uint next = 0; next < worklist.size(); ++next) {
3921 Node *n = worklist.at(next);
3922 if (n->is_Type() && n->as_Type()->type()->isa_oopptr() != NULL &&
3923 n->as_Type()->type()->is_oopptr()->speculative() != NULL) {
3924 TypeNode* tn = n->as_Type();
3925 const TypeOopPtr* t = tn->type()->is_oopptr();
3926 bool in_hash = igvn.hash_delete(n);
3927 assert(in_hash, "node should be in igvn hash table");
3928 tn->set_type(t->remove_speculative());
3929 igvn.hash_insert(n);
3930 igvn._worklist.push(n); // give it a chance to go away
3931 modified++;
3932 }
3933 uint max = n->len();
3934 for( uint i = 0; i < max; ++i ) {
3935 Node *m = n->in(i);
3936 if (not_a_node(m)) continue;
3937 worklist.push(m);
3938 }
3939 }
3940 // Drop the speculative part of all types in the igvn's type table
3941 igvn.remove_speculative_types();
3942 if (modified > 0) {
3943 igvn.optimize();
3944 }
3945 }
3946 }
3948 // Auxiliary method to support randomized stressing/fuzzing.
3949 //
3950 // This method can be called the arbitrary number of times, with current count
3951 // as the argument. The logic allows selecting a single candidate from the
3952 // running list of candidates as follows:
3953 // int count = 0;
3954 // Cand* selected = null;
3955 // while(cand = cand->next()) {
3956 // if (randomized_select(++count)) {
3957 // selected = cand;
3958 // }
3959 // }
3960 //
3961 // Including count equalizes the chances any candidate is "selected".
3962 // This is useful when we don't have the complete list of candidates to choose
3963 // from uniformly. In this case, we need to adjust the randomicity of the
3964 // selection, or else we will end up biasing the selection towards the latter
3965 // candidates.
3966 //
3967 // Quick back-envelope calculation shows that for the list of n candidates
3968 // the equal probability for the candidate to persist as "best" can be
3969 // achieved by replacing it with "next" k-th candidate with the probability
3970 // of 1/k. It can be easily shown that by the end of the run, the
3971 // probability for any candidate is converged to 1/n, thus giving the
3972 // uniform distribution among all the candidates.
3973 //
3974 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
3975 #define RANDOMIZED_DOMAIN_POW 29
3976 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
3977 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
3978 bool Compile::randomized_select(int count) {
3979 assert(count > 0, "only positive");
3980 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
3981 }