Thu, 20 Feb 2014 11:05:12 +0100
8035394: PPC64: Make usage of intrinsic dsqrt depend on processor recognition.
Reviewed-by: kvn
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_32
86 # include "adfiles/ad_ppc_32.hpp"
87 #endif
88 #ifdef TARGET_ARCH_MODEL_ppc_64
89 # include "adfiles/ad_ppc_64.hpp"
90 #endif
93 // -------------------- Compile::mach_constant_base_node -----------------------
94 // Constant table base node singleton.
95 MachConstantBaseNode* Compile::mach_constant_base_node() {
96 if (_mach_constant_base_node == NULL) {
97 _mach_constant_base_node = new (C) MachConstantBaseNode();
98 _mach_constant_base_node->add_req(C->root());
99 }
100 return _mach_constant_base_node;
101 }
104 /// Support for intrinsics.
106 // Return the index at which m must be inserted (or already exists).
107 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
108 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
109 #ifdef ASSERT
110 for (int i = 1; i < _intrinsics->length(); i++) {
111 CallGenerator* cg1 = _intrinsics->at(i-1);
112 CallGenerator* cg2 = _intrinsics->at(i);
113 assert(cg1->method() != cg2->method()
114 ? cg1->method() < cg2->method()
115 : cg1->is_virtual() < cg2->is_virtual(),
116 "compiler intrinsics list must stay sorted");
117 }
118 #endif
119 // Binary search sorted list, in decreasing intervals [lo, hi].
120 int lo = 0, hi = _intrinsics->length()-1;
121 while (lo <= hi) {
122 int mid = (uint)(hi + lo) / 2;
123 ciMethod* mid_m = _intrinsics->at(mid)->method();
124 if (m < mid_m) {
125 hi = mid-1;
126 } else if (m > mid_m) {
127 lo = mid+1;
128 } else {
129 // look at minor sort key
130 bool mid_virt = _intrinsics->at(mid)->is_virtual();
131 if (is_virtual < mid_virt) {
132 hi = mid-1;
133 } else if (is_virtual > mid_virt) {
134 lo = mid+1;
135 } else {
136 return mid; // exact match
137 }
138 }
139 }
140 return lo; // inexact match
141 }
143 void Compile::register_intrinsic(CallGenerator* cg) {
144 if (_intrinsics == NULL) {
145 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
146 }
147 // This code is stolen from ciObjectFactory::insert.
148 // Really, GrowableArray should have methods for
149 // insert_at, remove_at, and binary_search.
150 int len = _intrinsics->length();
151 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
152 if (index == len) {
153 _intrinsics->append(cg);
154 } else {
155 #ifdef ASSERT
156 CallGenerator* oldcg = _intrinsics->at(index);
157 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
158 #endif
159 _intrinsics->append(_intrinsics->at(len-1));
160 int pos;
161 for (pos = len-2; pos >= index; pos--) {
162 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
163 }
164 _intrinsics->at_put(index, cg);
165 }
166 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
167 }
169 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
170 assert(m->is_loaded(), "don't try this on unloaded methods");
171 if (_intrinsics != NULL) {
172 int index = intrinsic_insertion_index(m, is_virtual);
173 if (index < _intrinsics->length()
174 && _intrinsics->at(index)->method() == m
175 && _intrinsics->at(index)->is_virtual() == is_virtual) {
176 return _intrinsics->at(index);
177 }
178 }
179 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
180 if (m->intrinsic_id() != vmIntrinsics::_none &&
181 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
182 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
183 if (cg != NULL) {
184 // Save it for next time:
185 register_intrinsic(cg);
186 return cg;
187 } else {
188 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
189 }
190 }
191 return NULL;
192 }
194 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
195 // in library_call.cpp.
198 #ifndef PRODUCT
199 // statistics gathering...
201 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
202 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
204 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
205 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
206 int oflags = _intrinsic_hist_flags[id];
207 assert(flags != 0, "what happened?");
208 if (is_virtual) {
209 flags |= _intrinsic_virtual;
210 }
211 bool changed = (flags != oflags);
212 if ((flags & _intrinsic_worked) != 0) {
213 juint count = (_intrinsic_hist_count[id] += 1);
214 if (count == 1) {
215 changed = true; // first time
216 }
217 // increment the overall count also:
218 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
219 }
220 if (changed) {
221 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
222 // Something changed about the intrinsic's virtuality.
223 if ((flags & _intrinsic_virtual) != 0) {
224 // This is the first use of this intrinsic as a virtual call.
225 if (oflags != 0) {
226 // We already saw it as a non-virtual, so note both cases.
227 flags |= _intrinsic_both;
228 }
229 } else if ((oflags & _intrinsic_both) == 0) {
230 // This is the first use of this intrinsic as a non-virtual
231 flags |= _intrinsic_both;
232 }
233 }
234 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
235 }
236 // update the overall flags also:
237 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
238 return changed;
239 }
241 static char* format_flags(int flags, char* buf) {
242 buf[0] = 0;
243 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
244 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
245 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
246 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
247 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
248 if (buf[0] == 0) strcat(buf, ",");
249 assert(buf[0] == ',', "must be");
250 return &buf[1];
251 }
253 void Compile::print_intrinsic_statistics() {
254 char flagsbuf[100];
255 ttyLocker ttyl;
256 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
257 tty->print_cr("Compiler intrinsic usage:");
258 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
259 if (total == 0) total = 1; // avoid div0 in case of no successes
260 #define PRINT_STAT_LINE(name, c, f) \
261 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
262 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
263 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
264 int flags = _intrinsic_hist_flags[id];
265 juint count = _intrinsic_hist_count[id];
266 if ((flags | count) != 0) {
267 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
268 }
269 }
270 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
271 if (xtty != NULL) xtty->tail("statistics");
272 }
274 void Compile::print_statistics() {
275 { ttyLocker ttyl;
276 if (xtty != NULL) xtty->head("statistics type='opto'");
277 Parse::print_statistics();
278 PhaseCCP::print_statistics();
279 PhaseRegAlloc::print_statistics();
280 Scheduling::print_statistics();
281 PhasePeephole::print_statistics();
282 PhaseIdealLoop::print_statistics();
283 if (xtty != NULL) xtty->tail("statistics");
284 }
285 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
286 // put this under its own <statistics> element.
287 print_intrinsic_statistics();
288 }
289 }
290 #endif //PRODUCT
292 // Support for bundling info
293 Bundle* Compile::node_bundling(const Node *n) {
294 assert(valid_bundle_info(n), "oob");
295 return &_node_bundling_base[n->_idx];
296 }
298 bool Compile::valid_bundle_info(const Node *n) {
299 return (_node_bundling_limit > n->_idx);
300 }
303 void Compile::gvn_replace_by(Node* n, Node* nn) {
304 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
305 Node* use = n->last_out(i);
306 bool is_in_table = initial_gvn()->hash_delete(use);
307 uint uses_found = 0;
308 for (uint j = 0; j < use->len(); j++) {
309 if (use->in(j) == n) {
310 if (j < use->req())
311 use->set_req(j, nn);
312 else
313 use->set_prec(j, nn);
314 uses_found++;
315 }
316 }
317 if (is_in_table) {
318 // reinsert into table
319 initial_gvn()->hash_find_insert(use);
320 }
321 record_for_igvn(use);
322 i -= uses_found; // we deleted 1 or more copies of this edge
323 }
324 }
327 static inline bool not_a_node(const Node* n) {
328 if (n == NULL) return true;
329 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
330 if (*(address*)n == badAddress) return true; // kill by Node::destruct
331 return false;
332 }
334 // Identify all nodes that are reachable from below, useful.
335 // Use breadth-first pass that records state in a Unique_Node_List,
336 // recursive traversal is slower.
337 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
338 int estimated_worklist_size = unique();
339 useful.map( estimated_worklist_size, NULL ); // preallocate space
341 // Initialize worklist
342 if (root() != NULL) { useful.push(root()); }
343 // If 'top' is cached, declare it useful to preserve cached node
344 if( cached_top_node() ) { useful.push(cached_top_node()); }
346 // Push all useful nodes onto the list, breadthfirst
347 for( uint next = 0; next < useful.size(); ++next ) {
348 assert( next < unique(), "Unique useful nodes < total nodes");
349 Node *n = useful.at(next);
350 uint max = n->len();
351 for( uint i = 0; i < max; ++i ) {
352 Node *m = n->in(i);
353 if (not_a_node(m)) continue;
354 useful.push(m);
355 }
356 }
357 }
359 // Update dead_node_list with any missing dead nodes using useful
360 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
361 void Compile::update_dead_node_list(Unique_Node_List &useful) {
362 uint max_idx = unique();
363 VectorSet& useful_node_set = useful.member_set();
365 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
366 // If node with index node_idx is not in useful set,
367 // mark it as dead in dead node list.
368 if (! useful_node_set.test(node_idx) ) {
369 record_dead_node(node_idx);
370 }
371 }
372 }
374 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
375 int shift = 0;
376 for (int i = 0; i < inlines->length(); i++) {
377 CallGenerator* cg = inlines->at(i);
378 CallNode* call = cg->call_node();
379 if (shift > 0) {
380 inlines->at_put(i-shift, cg);
381 }
382 if (!useful.member(call)) {
383 shift++;
384 }
385 }
386 inlines->trunc_to(inlines->length()-shift);
387 }
389 // Disconnect all useless nodes by disconnecting those at the boundary.
390 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
391 uint next = 0;
392 while (next < useful.size()) {
393 Node *n = useful.at(next++);
394 // Use raw traversal of out edges since this code removes out edges
395 int max = n->outcnt();
396 for (int j = 0; j < max; ++j) {
397 Node* child = n->raw_out(j);
398 if (! useful.member(child)) {
399 assert(!child->is_top() || child != top(),
400 "If top is cached in Compile object it is in useful list");
401 // Only need to remove this out-edge to the useless node
402 n->raw_del_out(j);
403 --j;
404 --max;
405 }
406 }
407 if (n->outcnt() == 1 && n->has_special_unique_user()) {
408 record_for_igvn(n->unique_out());
409 }
410 }
411 // Remove useless macro and predicate opaq nodes
412 for (int i = C->macro_count()-1; i >= 0; i--) {
413 Node* n = C->macro_node(i);
414 if (!useful.member(n)) {
415 remove_macro_node(n);
416 }
417 }
418 // Remove useless expensive node
419 for (int i = C->expensive_count()-1; i >= 0; i--) {
420 Node* n = C->expensive_node(i);
421 if (!useful.member(n)) {
422 remove_expensive_node(n);
423 }
424 }
425 // clean up the late inline lists
426 remove_useless_late_inlines(&_string_late_inlines, useful);
427 remove_useless_late_inlines(&_boxing_late_inlines, useful);
428 remove_useless_late_inlines(&_late_inlines, useful);
429 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
430 }
432 //------------------------------frame_size_in_words-----------------------------
433 // frame_slots in units of words
434 int Compile::frame_size_in_words() const {
435 // shift is 0 in LP32 and 1 in LP64
436 const int shift = (LogBytesPerWord - LogBytesPerInt);
437 int words = _frame_slots >> shift;
438 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
439 return words;
440 }
442 // ============================================================================
443 //------------------------------CompileWrapper---------------------------------
444 class CompileWrapper : public StackObj {
445 Compile *const _compile;
446 public:
447 CompileWrapper(Compile* compile);
449 ~CompileWrapper();
450 };
452 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
453 // the Compile* pointer is stored in the current ciEnv:
454 ciEnv* env = compile->env();
455 assert(env == ciEnv::current(), "must already be a ciEnv active");
456 assert(env->compiler_data() == NULL, "compile already active?");
457 env->set_compiler_data(compile);
458 assert(compile == Compile::current(), "sanity");
460 compile->set_type_dict(NULL);
461 compile->set_type_hwm(NULL);
462 compile->set_type_last_size(0);
463 compile->set_last_tf(NULL, NULL);
464 compile->set_indexSet_arena(NULL);
465 compile->set_indexSet_free_block_list(NULL);
466 compile->init_type_arena();
467 Type::Initialize(compile);
468 _compile->set_scratch_buffer_blob(NULL);
469 _compile->begin_method();
470 }
471 CompileWrapper::~CompileWrapper() {
472 _compile->end_method();
473 if (_compile->scratch_buffer_blob() != NULL)
474 BufferBlob::free(_compile->scratch_buffer_blob());
475 _compile->env()->set_compiler_data(NULL);
476 }
479 //----------------------------print_compile_messages---------------------------
480 void Compile::print_compile_messages() {
481 #ifndef PRODUCT
482 // Check if recompiling
483 if (_subsume_loads == false && PrintOpto) {
484 // Recompiling without allowing machine instructions to subsume loads
485 tty->print_cr("*********************************************************");
486 tty->print_cr("** Bailout: Recompile without subsuming loads **");
487 tty->print_cr("*********************************************************");
488 }
489 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
490 // Recompiling without escape analysis
491 tty->print_cr("*********************************************************");
492 tty->print_cr("** Bailout: Recompile without escape analysis **");
493 tty->print_cr("*********************************************************");
494 }
495 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
496 // Recompiling without boxing elimination
497 tty->print_cr("*********************************************************");
498 tty->print_cr("** Bailout: Recompile without boxing elimination **");
499 tty->print_cr("*********************************************************");
500 }
501 if (env()->break_at_compile()) {
502 // Open the debugger when compiling this method.
503 tty->print("### Breaking when compiling: ");
504 method()->print_short_name();
505 tty->cr();
506 BREAKPOINT;
507 }
509 if( PrintOpto ) {
510 if (is_osr_compilation()) {
511 tty->print("[OSR]%3d", _compile_id);
512 } else {
513 tty->print("%3d", _compile_id);
514 }
515 }
516 #endif
517 }
520 //-----------------------init_scratch_buffer_blob------------------------------
521 // Construct a temporary BufferBlob and cache it for this compile.
522 void Compile::init_scratch_buffer_blob(int const_size) {
523 // If there is already a scratch buffer blob allocated and the
524 // constant section is big enough, use it. Otherwise free the
525 // current and allocate a new one.
526 BufferBlob* blob = scratch_buffer_blob();
527 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
528 // Use the current blob.
529 } else {
530 if (blob != NULL) {
531 BufferBlob::free(blob);
532 }
534 ResourceMark rm;
535 _scratch_const_size = const_size;
536 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
537 blob = BufferBlob::create("Compile::scratch_buffer", size);
538 // Record the buffer blob for next time.
539 set_scratch_buffer_blob(blob);
540 // Have we run out of code space?
541 if (scratch_buffer_blob() == NULL) {
542 // Let CompilerBroker disable further compilations.
543 record_failure("Not enough space for scratch buffer in CodeCache");
544 return;
545 }
546 }
548 // Initialize the relocation buffers
549 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
550 set_scratch_locs_memory(locs_buf);
551 }
554 //-----------------------scratch_emit_size-------------------------------------
555 // Helper function that computes size by emitting code
556 uint Compile::scratch_emit_size(const Node* n) {
557 // Start scratch_emit_size section.
558 set_in_scratch_emit_size(true);
560 // Emit into a trash buffer and count bytes emitted.
561 // This is a pretty expensive way to compute a size,
562 // but it works well enough if seldom used.
563 // All common fixed-size instructions are given a size
564 // method by the AD file.
565 // Note that the scratch buffer blob and locs memory are
566 // allocated at the beginning of the compile task, and
567 // may be shared by several calls to scratch_emit_size.
568 // The allocation of the scratch buffer blob is particularly
569 // expensive, since it has to grab the code cache lock.
570 BufferBlob* blob = this->scratch_buffer_blob();
571 assert(blob != NULL, "Initialize BufferBlob at start");
572 assert(blob->size() > MAX_inst_size, "sanity");
573 relocInfo* locs_buf = scratch_locs_memory();
574 address blob_begin = blob->content_begin();
575 address blob_end = (address)locs_buf;
576 assert(blob->content_contains(blob_end), "sanity");
577 CodeBuffer buf(blob_begin, blob_end - blob_begin);
578 buf.initialize_consts_size(_scratch_const_size);
579 buf.initialize_stubs_size(MAX_stubs_size);
580 assert(locs_buf != NULL, "sanity");
581 int lsize = MAX_locs_size / 3;
582 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
583 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
584 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
586 // Do the emission.
588 Label fakeL; // Fake label for branch instructions.
589 Label* saveL = NULL;
590 uint save_bnum = 0;
591 bool is_branch = n->is_MachBranch();
592 if (is_branch) {
593 MacroAssembler masm(&buf);
594 masm.bind(fakeL);
595 n->as_MachBranch()->save_label(&saveL, &save_bnum);
596 n->as_MachBranch()->label_set(&fakeL, 0);
597 }
598 n->emit(buf, this->regalloc());
599 if (is_branch) // Restore label.
600 n->as_MachBranch()->label_set(saveL, save_bnum);
602 // End scratch_emit_size section.
603 set_in_scratch_emit_size(false);
605 return buf.insts_size();
606 }
609 // ============================================================================
610 //------------------------------Compile standard-------------------------------
611 debug_only( int Compile::_debug_idx = 100000; )
613 // Compile a method. entry_bci is -1 for normal compilations and indicates
614 // the continuation bci for on stack replacement.
617 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
618 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
619 : Phase(Compiler),
620 _env(ci_env),
621 _log(ci_env->log()),
622 _compile_id(ci_env->compile_id()),
623 _save_argument_registers(false),
624 _stub_name(NULL),
625 _stub_function(NULL),
626 _stub_entry_point(NULL),
627 _method(target),
628 _entry_bci(osr_bci),
629 _initial_gvn(NULL),
630 _for_igvn(NULL),
631 _warm_calls(NULL),
632 _subsume_loads(subsume_loads),
633 _do_escape_analysis(do_escape_analysis),
634 _eliminate_boxing(eliminate_boxing),
635 _failure_reason(NULL),
636 _code_buffer("Compile::Fill_buffer"),
637 _orig_pc_slot(0),
638 _orig_pc_slot_offset_in_bytes(0),
639 _has_method_handle_invokes(false),
640 _mach_constant_base_node(NULL),
641 _node_bundling_limit(0),
642 _node_bundling_base(NULL),
643 _java_calls(0),
644 _inner_loops(0),
645 _scratch_const_size(-1),
646 _in_scratch_emit_size(false),
647 _dead_node_list(comp_arena()),
648 _dead_node_count(0),
649 #ifndef PRODUCT
650 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
651 _in_dump_cnt(0),
652 _printer(IdealGraphPrinter::printer()),
653 #endif
654 _congraph(NULL),
655 _replay_inline_data(NULL),
656 _late_inlines(comp_arena(), 2, 0, NULL),
657 _string_late_inlines(comp_arena(), 2, 0, NULL),
658 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
659 _late_inlines_pos(0),
660 _number_of_mh_late_inlines(0),
661 _inlining_progress(false),
662 _inlining_incrementally(false),
663 _print_inlining_list(NULL),
664 _print_inlining_idx(0),
665 _preserve_jvm_state(0) {
666 C = this;
668 CompileWrapper cw(this);
669 #ifndef PRODUCT
670 if (TimeCompiler2) {
671 tty->print(" ");
672 target->holder()->name()->print();
673 tty->print(".");
674 target->print_short_name();
675 tty->print(" ");
676 }
677 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
678 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
679 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
680 if (!print_opto_assembly) {
681 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
682 if (print_assembly && !Disassembler::can_decode()) {
683 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
684 print_opto_assembly = true;
685 }
686 }
687 set_print_assembly(print_opto_assembly);
688 set_parsed_irreducible_loop(false);
690 if (method()->has_option("ReplayInline")) {
691 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
692 }
693 #endif
694 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
695 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
697 if (ProfileTraps) {
698 // Make sure the method being compiled gets its own MDO,
699 // so we can at least track the decompile_count().
700 method()->ensure_method_data();
701 }
703 Init(::AliasLevel);
706 print_compile_messages();
708 _ilt = InlineTree::build_inline_tree_root();
710 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
711 assert(num_alias_types() >= AliasIdxRaw, "");
713 #define MINIMUM_NODE_HASH 1023
714 // Node list that Iterative GVN will start with
715 Unique_Node_List for_igvn(comp_arena());
716 set_for_igvn(&for_igvn);
718 // GVN that will be run immediately on new nodes
719 uint estimated_size = method()->code_size()*4+64;
720 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
721 PhaseGVN gvn(node_arena(), estimated_size);
722 set_initial_gvn(&gvn);
724 if (print_inlining() || print_intrinsics()) {
725 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
726 }
727 { // Scope for timing the parser
728 TracePhase t3("parse", &_t_parser, true);
730 // Put top into the hash table ASAP.
731 initial_gvn()->transform_no_reclaim(top());
733 // Set up tf(), start(), and find a CallGenerator.
734 CallGenerator* cg = NULL;
735 if (is_osr_compilation()) {
736 const TypeTuple *domain = StartOSRNode::osr_domain();
737 const TypeTuple *range = TypeTuple::make_range(method()->signature());
738 init_tf(TypeFunc::make(domain, range));
739 StartNode* s = new (this) StartOSRNode(root(), domain);
740 initial_gvn()->set_type_bottom(s);
741 init_start(s);
742 cg = CallGenerator::for_osr(method(), entry_bci());
743 } else {
744 // Normal case.
745 init_tf(TypeFunc::make(method()));
746 StartNode* s = new (this) StartNode(root(), tf()->domain());
747 initial_gvn()->set_type_bottom(s);
748 init_start(s);
749 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
750 // With java.lang.ref.reference.get() we must go through the
751 // intrinsic when G1 is enabled - even when get() is the root
752 // method of the compile - so that, if necessary, the value in
753 // the referent field of the reference object gets recorded by
754 // the pre-barrier code.
755 // Specifically, if G1 is enabled, the value in the referent
756 // field is recorded by the G1 SATB pre barrier. This will
757 // result in the referent being marked live and the reference
758 // object removed from the list of discovered references during
759 // reference processing.
760 cg = find_intrinsic(method(), false);
761 }
762 if (cg == NULL) {
763 float past_uses = method()->interpreter_invocation_count();
764 float expected_uses = past_uses;
765 cg = CallGenerator::for_inline(method(), expected_uses);
766 }
767 }
768 if (failing()) return;
769 if (cg == NULL) {
770 record_method_not_compilable_all_tiers("cannot parse method");
771 return;
772 }
773 JVMState* jvms = build_start_state(start(), tf());
774 if ((jvms = cg->generate(jvms, NULL)) == NULL) {
775 record_method_not_compilable("method parse failed");
776 return;
777 }
778 GraphKit kit(jvms);
780 if (!kit.stopped()) {
781 // Accept return values, and transfer control we know not where.
782 // This is done by a special, unique ReturnNode bound to root.
783 return_values(kit.jvms());
784 }
786 if (kit.has_exceptions()) {
787 // Any exceptions that escape from this call must be rethrown
788 // to whatever caller is dynamically above us on the stack.
789 // This is done by a special, unique RethrowNode bound to root.
790 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
791 }
793 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
795 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
796 inline_string_calls(true);
797 }
799 if (failing()) return;
801 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
803 // Remove clutter produced by parsing.
804 if (!failing()) {
805 ResourceMark rm;
806 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
807 }
808 }
810 // Note: Large methods are capped off in do_one_bytecode().
811 if (failing()) return;
813 // After parsing, node notes are no longer automagic.
814 // They must be propagated by register_new_node_with_optimizer(),
815 // clone(), or the like.
816 set_default_node_notes(NULL);
818 for (;;) {
819 int successes = Inline_Warm();
820 if (failing()) return;
821 if (successes == 0) break;
822 }
824 // Drain the list.
825 Finish_Warm();
826 #ifndef PRODUCT
827 if (_printer) {
828 _printer->print_inlining(this);
829 }
830 #endif
832 if (failing()) return;
833 NOT_PRODUCT( verify_graph_edges(); )
835 // Now optimize
836 Optimize();
837 if (failing()) return;
838 NOT_PRODUCT( verify_graph_edges(); )
840 #ifndef PRODUCT
841 if (PrintIdeal) {
842 ttyLocker ttyl; // keep the following output all in one block
843 // This output goes directly to the tty, not the compiler log.
844 // To enable tools to match it up with the compilation activity,
845 // be sure to tag this tty output with the compile ID.
846 if (xtty != NULL) {
847 xtty->head("ideal compile_id='%d'%s", compile_id(),
848 is_osr_compilation() ? " compile_kind='osr'" :
849 "");
850 }
851 root()->dump(9999);
852 if (xtty != NULL) {
853 xtty->tail("ideal");
854 }
855 }
856 #endif
858 NOT_PRODUCT( verify_barriers(); )
860 // Dump compilation data to replay it.
861 if (method()->has_option("DumpReplay")) {
862 env()->dump_replay_data(_compile_id);
863 }
864 if (method()->has_option("DumpInline") && (ilt() != NULL)) {
865 env()->dump_inline_data(_compile_id);
866 }
868 // Now that we know the size of all the monitors we can add a fixed slot
869 // for the original deopt pc.
871 _orig_pc_slot = fixed_slots();
872 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
873 set_fixed_slots(next_slot);
875 // Compute when to use implicit null checks. Used by matching trap based
876 // nodes and NullCheck optimization.
877 set_allowed_deopt_reasons();
879 // Now generate code
880 Code_Gen();
881 if (failing()) return;
883 // Check if we want to skip execution of all compiled code.
884 {
885 #ifndef PRODUCT
886 if (OptoNoExecute) {
887 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
888 return;
889 }
890 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
891 #endif
893 if (is_osr_compilation()) {
894 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
895 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
896 } else {
897 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
898 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
899 }
901 env()->register_method(_method, _entry_bci,
902 &_code_offsets,
903 _orig_pc_slot_offset_in_bytes,
904 code_buffer(),
905 frame_size_in_words(), _oop_map_set,
906 &_handler_table, &_inc_table,
907 compiler,
908 env()->comp_level(),
909 has_unsafe_access(),
910 SharedRuntime::is_wide_vector(max_vector_size())
911 );
913 if (log() != NULL) // Print code cache state into compiler log
914 log()->code_cache_state();
915 }
916 }
918 //------------------------------Compile----------------------------------------
919 // Compile a runtime stub
920 Compile::Compile( ciEnv* ci_env,
921 TypeFunc_generator generator,
922 address stub_function,
923 const char *stub_name,
924 int is_fancy_jump,
925 bool pass_tls,
926 bool save_arg_registers,
927 bool return_pc )
928 : Phase(Compiler),
929 _env(ci_env),
930 _log(ci_env->log()),
931 _compile_id(0),
932 _save_argument_registers(save_arg_registers),
933 _method(NULL),
934 _stub_name(stub_name),
935 _stub_function(stub_function),
936 _stub_entry_point(NULL),
937 _entry_bci(InvocationEntryBci),
938 _initial_gvn(NULL),
939 _for_igvn(NULL),
940 _warm_calls(NULL),
941 _orig_pc_slot(0),
942 _orig_pc_slot_offset_in_bytes(0),
943 _subsume_loads(true),
944 _do_escape_analysis(false),
945 _eliminate_boxing(false),
946 _failure_reason(NULL),
947 _code_buffer("Compile::Fill_buffer"),
948 _has_method_handle_invokes(false),
949 _mach_constant_base_node(NULL),
950 _node_bundling_limit(0),
951 _node_bundling_base(NULL),
952 _java_calls(0),
953 _inner_loops(0),
954 #ifndef PRODUCT
955 _trace_opto_output(TraceOptoOutput),
956 _in_dump_cnt(0),
957 _printer(NULL),
958 #endif
959 _dead_node_list(comp_arena()),
960 _dead_node_count(0),
961 _congraph(NULL),
962 _replay_inline_data(NULL),
963 _number_of_mh_late_inlines(0),
964 _inlining_progress(false),
965 _inlining_incrementally(false),
966 _print_inlining_list(NULL),
967 _print_inlining_idx(0),
968 _preserve_jvm_state(0),
969 _allowed_reasons(0) {
970 C = this;
972 #ifndef PRODUCT
973 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
974 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
975 set_print_assembly(PrintFrameConverterAssembly);
976 set_parsed_irreducible_loop(false);
977 #endif
978 CompileWrapper cw(this);
979 Init(/*AliasLevel=*/ 0);
980 init_tf((*generator)());
982 {
983 // The following is a dummy for the sake of GraphKit::gen_stub
984 Unique_Node_List for_igvn(comp_arena());
985 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
986 PhaseGVN gvn(Thread::current()->resource_area(),255);
987 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
988 gvn.transform_no_reclaim(top());
990 GraphKit kit;
991 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
992 }
994 NOT_PRODUCT( verify_graph_edges(); )
995 Code_Gen();
996 if (failing()) return;
999 // Entry point will be accessed using compile->stub_entry_point();
1000 if (code_buffer() == NULL) {
1001 Matcher::soft_match_failure();
1002 } else {
1003 if (PrintAssembly && (WizardMode || Verbose))
1004 tty->print_cr("### Stub::%s", stub_name);
1006 if (!failing()) {
1007 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1009 // Make the NMethod
1010 // For now we mark the frame as never safe for profile stackwalking
1011 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1012 code_buffer(),
1013 CodeOffsets::frame_never_safe,
1014 // _code_offsets.value(CodeOffsets::Frame_Complete),
1015 frame_size_in_words(),
1016 _oop_map_set,
1017 save_arg_registers);
1018 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1020 _stub_entry_point = rs->entry_point();
1021 }
1022 }
1023 }
1025 //------------------------------Init-------------------------------------------
1026 // Prepare for a single compilation
1027 void Compile::Init(int aliaslevel) {
1028 _unique = 0;
1029 _regalloc = NULL;
1031 _tf = NULL; // filled in later
1032 _top = NULL; // cached later
1033 _matcher = NULL; // filled in later
1034 _cfg = NULL; // filled in later
1036 set_24_bit_selection_and_mode(Use24BitFP, false);
1038 _node_note_array = NULL;
1039 _default_node_notes = NULL;
1041 _immutable_memory = NULL; // filled in at first inquiry
1043 // Globally visible Nodes
1044 // First set TOP to NULL to give safe behavior during creation of RootNode
1045 set_cached_top_node(NULL);
1046 set_root(new (this) RootNode());
1047 // Now that you have a Root to point to, create the real TOP
1048 set_cached_top_node( new (this) ConNode(Type::TOP) );
1049 set_recent_alloc(NULL, NULL);
1051 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1052 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1053 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1054 env()->set_dependencies(new Dependencies(env()));
1056 _fixed_slots = 0;
1057 set_has_split_ifs(false);
1058 set_has_loops(has_method() && method()->has_loops()); // first approximation
1059 set_has_stringbuilder(false);
1060 set_has_boxed_value(false);
1061 _trap_can_recompile = false; // no traps emitted yet
1062 _major_progress = true; // start out assuming good things will happen
1063 set_has_unsafe_access(false);
1064 set_max_vector_size(0);
1065 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1066 set_decompile_count(0);
1068 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1069 set_num_loop_opts(LoopOptsCount);
1070 set_do_inlining(Inline);
1071 set_max_inline_size(MaxInlineSize);
1072 set_freq_inline_size(FreqInlineSize);
1073 set_do_scheduling(OptoScheduling);
1074 set_do_count_invocations(false);
1075 set_do_method_data_update(false);
1077 if (debug_info()->recording_non_safepoints()) {
1078 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1079 (comp_arena(), 8, 0, NULL));
1080 set_default_node_notes(Node_Notes::make(this));
1081 }
1083 // // -- Initialize types before each compile --
1084 // // Update cached type information
1085 // if( _method && _method->constants() )
1086 // Type::update_loaded_types(_method, _method->constants());
1088 // Init alias_type map.
1089 if (!_do_escape_analysis && aliaslevel == 3)
1090 aliaslevel = 2; // No unique types without escape analysis
1091 _AliasLevel = aliaslevel;
1092 const int grow_ats = 16;
1093 _max_alias_types = grow_ats;
1094 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1095 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1096 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1097 {
1098 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1099 }
1100 // Initialize the first few types.
1101 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1102 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1103 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1104 _num_alias_types = AliasIdxRaw+1;
1105 // Zero out the alias type cache.
1106 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1107 // A NULL adr_type hits in the cache right away. Preload the right answer.
1108 probe_alias_cache(NULL)->_index = AliasIdxTop;
1110 _intrinsics = NULL;
1111 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1112 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1113 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1114 register_library_intrinsics();
1115 }
1117 //---------------------------init_start----------------------------------------
1118 // Install the StartNode on this compile object.
1119 void Compile::init_start(StartNode* s) {
1120 if (failing())
1121 return; // already failing
1122 assert(s == start(), "");
1123 }
1125 StartNode* Compile::start() const {
1126 assert(!failing(), "");
1127 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1128 Node* start = root()->fast_out(i);
1129 if( start->is_Start() )
1130 return start->as_Start();
1131 }
1132 ShouldNotReachHere();
1133 return NULL;
1134 }
1136 //-------------------------------immutable_memory-------------------------------------
1137 // Access immutable memory
1138 Node* Compile::immutable_memory() {
1139 if (_immutable_memory != NULL) {
1140 return _immutable_memory;
1141 }
1142 StartNode* s = start();
1143 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1144 Node *p = s->fast_out(i);
1145 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1146 _immutable_memory = p;
1147 return _immutable_memory;
1148 }
1149 }
1150 ShouldNotReachHere();
1151 return NULL;
1152 }
1154 //----------------------set_cached_top_node------------------------------------
1155 // Install the cached top node, and make sure Node::is_top works correctly.
1156 void Compile::set_cached_top_node(Node* tn) {
1157 if (tn != NULL) verify_top(tn);
1158 Node* old_top = _top;
1159 _top = tn;
1160 // Calling Node::setup_is_top allows the nodes the chance to adjust
1161 // their _out arrays.
1162 if (_top != NULL) _top->setup_is_top();
1163 if (old_top != NULL) old_top->setup_is_top();
1164 assert(_top == NULL || top()->is_top(), "");
1165 }
1167 #ifdef ASSERT
1168 uint Compile::count_live_nodes_by_graph_walk() {
1169 Unique_Node_List useful(comp_arena());
1170 // Get useful node list by walking the graph.
1171 identify_useful_nodes(useful);
1172 return useful.size();
1173 }
1175 void Compile::print_missing_nodes() {
1177 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1178 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1179 return;
1180 }
1182 // This is an expensive function. It is executed only when the user
1183 // specifies VerifyIdealNodeCount option or otherwise knows the
1184 // additional work that needs to be done to identify reachable nodes
1185 // by walking the flow graph and find the missing ones using
1186 // _dead_node_list.
1188 Unique_Node_List useful(comp_arena());
1189 // Get useful node list by walking the graph.
1190 identify_useful_nodes(useful);
1192 uint l_nodes = C->live_nodes();
1193 uint l_nodes_by_walk = useful.size();
1195 if (l_nodes != l_nodes_by_walk) {
1196 if (_log != NULL) {
1197 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1198 _log->stamp();
1199 _log->end_head();
1200 }
1201 VectorSet& useful_member_set = useful.member_set();
1202 int last_idx = l_nodes_by_walk;
1203 for (int i = 0; i < last_idx; i++) {
1204 if (useful_member_set.test(i)) {
1205 if (_dead_node_list.test(i)) {
1206 if (_log != NULL) {
1207 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1208 }
1209 if (PrintIdealNodeCount) {
1210 // Print the log message to tty
1211 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1212 useful.at(i)->dump();
1213 }
1214 }
1215 }
1216 else if (! _dead_node_list.test(i)) {
1217 if (_log != NULL) {
1218 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1219 }
1220 if (PrintIdealNodeCount) {
1221 // Print the log message to tty
1222 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1223 }
1224 }
1225 }
1226 if (_log != NULL) {
1227 _log->tail("mismatched_nodes");
1228 }
1229 }
1230 }
1231 #endif
1233 #ifndef PRODUCT
1234 void Compile::verify_top(Node* tn) const {
1235 if (tn != NULL) {
1236 assert(tn->is_Con(), "top node must be a constant");
1237 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1238 assert(tn->in(0) != NULL, "must have live top node");
1239 }
1240 }
1241 #endif
1244 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1246 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1247 guarantee(arr != NULL, "");
1248 int num_blocks = arr->length();
1249 if (grow_by < num_blocks) grow_by = num_blocks;
1250 int num_notes = grow_by * _node_notes_block_size;
1251 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1252 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1253 while (num_notes > 0) {
1254 arr->append(notes);
1255 notes += _node_notes_block_size;
1256 num_notes -= _node_notes_block_size;
1257 }
1258 assert(num_notes == 0, "exact multiple, please");
1259 }
1261 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1262 if (source == NULL || dest == NULL) return false;
1264 if (dest->is_Con())
1265 return false; // Do not push debug info onto constants.
1267 #ifdef ASSERT
1268 // Leave a bread crumb trail pointing to the original node:
1269 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1270 dest->set_debug_orig(source);
1271 }
1272 #endif
1274 if (node_note_array() == NULL)
1275 return false; // Not collecting any notes now.
1277 // This is a copy onto a pre-existing node, which may already have notes.
1278 // If both nodes have notes, do not overwrite any pre-existing notes.
1279 Node_Notes* source_notes = node_notes_at(source->_idx);
1280 if (source_notes == NULL || source_notes->is_clear()) return false;
1281 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1282 if (dest_notes == NULL || dest_notes->is_clear()) {
1283 return set_node_notes_at(dest->_idx, source_notes);
1284 }
1286 Node_Notes merged_notes = (*source_notes);
1287 // The order of operations here ensures that dest notes will win...
1288 merged_notes.update_from(dest_notes);
1289 return set_node_notes_at(dest->_idx, &merged_notes);
1290 }
1293 //--------------------------allow_range_check_smearing-------------------------
1294 // Gating condition for coalescing similar range checks.
1295 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1296 // single covering check that is at least as strong as any of them.
1297 // If the optimization succeeds, the simplified (strengthened) range check
1298 // will always succeed. If it fails, we will deopt, and then give up
1299 // on the optimization.
1300 bool Compile::allow_range_check_smearing() const {
1301 // If this method has already thrown a range-check,
1302 // assume it was because we already tried range smearing
1303 // and it failed.
1304 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1305 return !already_trapped;
1306 }
1309 //------------------------------flatten_alias_type-----------------------------
1310 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1311 int offset = tj->offset();
1312 TypePtr::PTR ptr = tj->ptr();
1314 // Known instance (scalarizable allocation) alias only with itself.
1315 bool is_known_inst = tj->isa_oopptr() != NULL &&
1316 tj->is_oopptr()->is_known_instance();
1318 // Process weird unsafe references.
1319 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1320 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1321 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1322 tj = TypeOopPtr::BOTTOM;
1323 ptr = tj->ptr();
1324 offset = tj->offset();
1325 }
1327 // Array pointers need some flattening
1328 const TypeAryPtr *ta = tj->isa_aryptr();
1329 if (ta && ta->is_stable()) {
1330 // Erase stability property for alias analysis.
1331 tj = ta = ta->cast_to_stable(false);
1332 }
1333 if( ta && is_known_inst ) {
1334 if ( offset != Type::OffsetBot &&
1335 offset > arrayOopDesc::length_offset_in_bytes() ) {
1336 offset = Type::OffsetBot; // Flatten constant access into array body only
1337 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1338 }
1339 } else if( ta && _AliasLevel >= 2 ) {
1340 // For arrays indexed by constant indices, we flatten the alias
1341 // space to include all of the array body. Only the header, klass
1342 // and array length can be accessed un-aliased.
1343 if( offset != Type::OffsetBot ) {
1344 if( ta->const_oop() ) { // MethodData* or Method*
1345 offset = Type::OffsetBot; // Flatten constant access into array body
1346 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1347 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1348 // range is OK as-is.
1349 tj = ta = TypeAryPtr::RANGE;
1350 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1351 tj = TypeInstPtr::KLASS; // all klass loads look alike
1352 ta = TypeAryPtr::RANGE; // generic ignored junk
1353 ptr = TypePtr::BotPTR;
1354 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1355 tj = TypeInstPtr::MARK;
1356 ta = TypeAryPtr::RANGE; // generic ignored junk
1357 ptr = TypePtr::BotPTR;
1358 } else { // Random constant offset into array body
1359 offset = Type::OffsetBot; // Flatten constant access into array body
1360 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1361 }
1362 }
1363 // Arrays of fixed size alias with arrays of unknown size.
1364 if (ta->size() != TypeInt::POS) {
1365 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1366 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1367 }
1368 // Arrays of known objects become arrays of unknown objects.
1369 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1370 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1371 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1372 }
1373 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1374 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1375 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1376 }
1377 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1378 // cannot be distinguished by bytecode alone.
1379 if (ta->elem() == TypeInt::BOOL) {
1380 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1381 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1382 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1383 }
1384 // During the 2nd round of IterGVN, NotNull castings are removed.
1385 // Make sure the Bottom and NotNull variants alias the same.
1386 // Also, make sure exact and non-exact variants alias the same.
1387 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1388 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1389 }
1390 }
1392 // Oop pointers need some flattening
1393 const TypeInstPtr *to = tj->isa_instptr();
1394 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1395 ciInstanceKlass *k = to->klass()->as_instance_klass();
1396 if( ptr == TypePtr::Constant ) {
1397 if (to->klass() != ciEnv::current()->Class_klass() ||
1398 offset < k->size_helper() * wordSize) {
1399 // No constant oop pointers (such as Strings); they alias with
1400 // unknown strings.
1401 assert(!is_known_inst, "not scalarizable allocation");
1402 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1403 }
1404 } else if( is_known_inst ) {
1405 tj = to; // Keep NotNull and klass_is_exact for instance type
1406 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1407 // During the 2nd round of IterGVN, NotNull castings are removed.
1408 // Make sure the Bottom and NotNull variants alias the same.
1409 // Also, make sure exact and non-exact variants alias the same.
1410 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1411 }
1412 if (to->speculative() != NULL) {
1413 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1414 }
1415 // Canonicalize the holder of this field
1416 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1417 // First handle header references such as a LoadKlassNode, even if the
1418 // object's klass is unloaded at compile time (4965979).
1419 if (!is_known_inst) { // Do it only for non-instance types
1420 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1421 }
1422 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1423 // Static fields are in the space above the normal instance
1424 // fields in the java.lang.Class instance.
1425 if (to->klass() != ciEnv::current()->Class_klass()) {
1426 to = NULL;
1427 tj = TypeOopPtr::BOTTOM;
1428 offset = tj->offset();
1429 }
1430 } else {
1431 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1432 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1433 if( is_known_inst ) {
1434 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1435 } else {
1436 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1437 }
1438 }
1439 }
1440 }
1442 // Klass pointers to object array klasses need some flattening
1443 const TypeKlassPtr *tk = tj->isa_klassptr();
1444 if( tk ) {
1445 // If we are referencing a field within a Klass, we need
1446 // to assume the worst case of an Object. Both exact and
1447 // inexact types must flatten to the same alias class so
1448 // use NotNull as the PTR.
1449 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1451 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1452 TypeKlassPtr::OBJECT->klass(),
1453 offset);
1454 }
1456 ciKlass* klass = tk->klass();
1457 if( klass->is_obj_array_klass() ) {
1458 ciKlass* k = TypeAryPtr::OOPS->klass();
1459 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1460 k = TypeInstPtr::BOTTOM->klass();
1461 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1462 }
1464 // Check for precise loads from the primary supertype array and force them
1465 // to the supertype cache alias index. Check for generic array loads from
1466 // the primary supertype array and also force them to the supertype cache
1467 // alias index. Since the same load can reach both, we need to merge
1468 // these 2 disparate memories into the same alias class. Since the
1469 // primary supertype array is read-only, there's no chance of confusion
1470 // where we bypass an array load and an array store.
1471 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1472 if (offset == Type::OffsetBot ||
1473 (offset >= primary_supers_offset &&
1474 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1475 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1476 offset = in_bytes(Klass::secondary_super_cache_offset());
1477 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1478 }
1479 }
1481 // Flatten all Raw pointers together.
1482 if (tj->base() == Type::RawPtr)
1483 tj = TypeRawPtr::BOTTOM;
1485 if (tj->base() == Type::AnyPtr)
1486 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1488 // Flatten all to bottom for now
1489 switch( _AliasLevel ) {
1490 case 0:
1491 tj = TypePtr::BOTTOM;
1492 break;
1493 case 1: // Flatten to: oop, static, field or array
1494 switch (tj->base()) {
1495 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1496 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1497 case Type::AryPtr: // do not distinguish arrays at all
1498 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1499 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1500 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1501 default: ShouldNotReachHere();
1502 }
1503 break;
1504 case 2: // No collapsing at level 2; keep all splits
1505 case 3: // No collapsing at level 3; keep all splits
1506 break;
1507 default:
1508 Unimplemented();
1509 }
1511 offset = tj->offset();
1512 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1514 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1515 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1516 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1517 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1518 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1519 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1520 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1521 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1522 assert( tj->ptr() != TypePtr::TopPTR &&
1523 tj->ptr() != TypePtr::AnyNull &&
1524 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1525 // assert( tj->ptr() != TypePtr::Constant ||
1526 // tj->base() == Type::RawPtr ||
1527 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1529 return tj;
1530 }
1532 void Compile::AliasType::Init(int i, const TypePtr* at) {
1533 _index = i;
1534 _adr_type = at;
1535 _field = NULL;
1536 _element = NULL;
1537 _is_rewritable = true; // default
1538 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1539 if (atoop != NULL && atoop->is_known_instance()) {
1540 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1541 _general_index = Compile::current()->get_alias_index(gt);
1542 } else {
1543 _general_index = 0;
1544 }
1545 }
1547 //---------------------------------print_on------------------------------------
1548 #ifndef PRODUCT
1549 void Compile::AliasType::print_on(outputStream* st) {
1550 if (index() < 10)
1551 st->print("@ <%d> ", index());
1552 else st->print("@ <%d>", index());
1553 st->print(is_rewritable() ? " " : " RO");
1554 int offset = adr_type()->offset();
1555 if (offset == Type::OffsetBot)
1556 st->print(" +any");
1557 else st->print(" +%-3d", offset);
1558 st->print(" in ");
1559 adr_type()->dump_on(st);
1560 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1561 if (field() != NULL && tjp) {
1562 if (tjp->klass() != field()->holder() ||
1563 tjp->offset() != field()->offset_in_bytes()) {
1564 st->print(" != ");
1565 field()->print();
1566 st->print(" ***");
1567 }
1568 }
1569 }
1571 void print_alias_types() {
1572 Compile* C = Compile::current();
1573 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1574 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1575 C->alias_type(idx)->print_on(tty);
1576 tty->cr();
1577 }
1578 }
1579 #endif
1582 //----------------------------probe_alias_cache--------------------------------
1583 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1584 intptr_t key = (intptr_t) adr_type;
1585 key ^= key >> logAliasCacheSize;
1586 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1587 }
1590 //-----------------------------grow_alias_types--------------------------------
1591 void Compile::grow_alias_types() {
1592 const int old_ats = _max_alias_types; // how many before?
1593 const int new_ats = old_ats; // how many more?
1594 const int grow_ats = old_ats+new_ats; // how many now?
1595 _max_alias_types = grow_ats;
1596 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1597 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1598 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1599 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1600 }
1603 //--------------------------------find_alias_type------------------------------
1604 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1605 if (_AliasLevel == 0)
1606 return alias_type(AliasIdxBot);
1608 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1609 if (ace->_adr_type == adr_type) {
1610 return alias_type(ace->_index);
1611 }
1613 // Handle special cases.
1614 if (adr_type == NULL) return alias_type(AliasIdxTop);
1615 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1617 // Do it the slow way.
1618 const TypePtr* flat = flatten_alias_type(adr_type);
1620 #ifdef ASSERT
1621 assert(flat == flatten_alias_type(flat), "idempotent");
1622 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1623 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1624 const TypeOopPtr* foop = flat->is_oopptr();
1625 // Scalarizable allocations have exact klass always.
1626 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1627 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1628 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1629 }
1630 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1631 #endif
1633 int idx = AliasIdxTop;
1634 for (int i = 0; i < num_alias_types(); i++) {
1635 if (alias_type(i)->adr_type() == flat) {
1636 idx = i;
1637 break;
1638 }
1639 }
1641 if (idx == AliasIdxTop) {
1642 if (no_create) return NULL;
1643 // Grow the array if necessary.
1644 if (_num_alias_types == _max_alias_types) grow_alias_types();
1645 // Add a new alias type.
1646 idx = _num_alias_types++;
1647 _alias_types[idx]->Init(idx, flat);
1648 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1649 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1650 if (flat->isa_instptr()) {
1651 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1652 && flat->is_instptr()->klass() == env()->Class_klass())
1653 alias_type(idx)->set_rewritable(false);
1654 }
1655 if (flat->isa_aryptr()) {
1656 #ifdef ASSERT
1657 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1658 // (T_BYTE has the weakest alignment and size restrictions...)
1659 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1660 #endif
1661 if (flat->offset() == TypePtr::OffsetBot) {
1662 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1663 }
1664 }
1665 if (flat->isa_klassptr()) {
1666 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1667 alias_type(idx)->set_rewritable(false);
1668 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1669 alias_type(idx)->set_rewritable(false);
1670 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1671 alias_type(idx)->set_rewritable(false);
1672 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1673 alias_type(idx)->set_rewritable(false);
1674 }
1675 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1676 // but the base pointer type is not distinctive enough to identify
1677 // references into JavaThread.)
1679 // Check for final fields.
1680 const TypeInstPtr* tinst = flat->isa_instptr();
1681 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1682 ciField* field;
1683 if (tinst->const_oop() != NULL &&
1684 tinst->klass() == ciEnv::current()->Class_klass() &&
1685 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1686 // static field
1687 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1688 field = k->get_field_by_offset(tinst->offset(), true);
1689 } else {
1690 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1691 field = k->get_field_by_offset(tinst->offset(), false);
1692 }
1693 assert(field == NULL ||
1694 original_field == NULL ||
1695 (field->holder() == original_field->holder() &&
1696 field->offset() == original_field->offset() &&
1697 field->is_static() == original_field->is_static()), "wrong field?");
1698 // Set field() and is_rewritable() attributes.
1699 if (field != NULL) alias_type(idx)->set_field(field);
1700 }
1701 }
1703 // Fill the cache for next time.
1704 ace->_adr_type = adr_type;
1705 ace->_index = idx;
1706 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1708 // Might as well try to fill the cache for the flattened version, too.
1709 AliasCacheEntry* face = probe_alias_cache(flat);
1710 if (face->_adr_type == NULL) {
1711 face->_adr_type = flat;
1712 face->_index = idx;
1713 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1714 }
1716 return alias_type(idx);
1717 }
1720 Compile::AliasType* Compile::alias_type(ciField* field) {
1721 const TypeOopPtr* t;
1722 if (field->is_static())
1723 t = TypeInstPtr::make(field->holder()->java_mirror());
1724 else
1725 t = TypeOopPtr::make_from_klass_raw(field->holder());
1726 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1727 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1728 return atp;
1729 }
1732 //------------------------------have_alias_type--------------------------------
1733 bool Compile::have_alias_type(const TypePtr* adr_type) {
1734 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1735 if (ace->_adr_type == adr_type) {
1736 return true;
1737 }
1739 // Handle special cases.
1740 if (adr_type == NULL) return true;
1741 if (adr_type == TypePtr::BOTTOM) return true;
1743 return find_alias_type(adr_type, true, NULL) != NULL;
1744 }
1746 //-----------------------------must_alias--------------------------------------
1747 // True if all values of the given address type are in the given alias category.
1748 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1749 if (alias_idx == AliasIdxBot) return true; // the universal category
1750 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1751 if (alias_idx == AliasIdxTop) return false; // the empty category
1752 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1754 // the only remaining possible overlap is identity
1755 int adr_idx = get_alias_index(adr_type);
1756 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1757 assert(adr_idx == alias_idx ||
1758 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1759 && adr_type != TypeOopPtr::BOTTOM),
1760 "should not be testing for overlap with an unsafe pointer");
1761 return adr_idx == alias_idx;
1762 }
1764 //------------------------------can_alias--------------------------------------
1765 // True if any values of the given address type are in the given alias category.
1766 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1767 if (alias_idx == AliasIdxTop) return false; // the empty category
1768 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1769 if (alias_idx == AliasIdxBot) return true; // the universal category
1770 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1772 // the only remaining possible overlap is identity
1773 int adr_idx = get_alias_index(adr_type);
1774 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1775 return adr_idx == alias_idx;
1776 }
1780 //---------------------------pop_warm_call-------------------------------------
1781 WarmCallInfo* Compile::pop_warm_call() {
1782 WarmCallInfo* wci = _warm_calls;
1783 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1784 return wci;
1785 }
1787 //----------------------------Inline_Warm--------------------------------------
1788 int Compile::Inline_Warm() {
1789 // If there is room, try to inline some more warm call sites.
1790 // %%% Do a graph index compaction pass when we think we're out of space?
1791 if (!InlineWarmCalls) return 0;
1793 int calls_made_hot = 0;
1794 int room_to_grow = NodeCountInliningCutoff - unique();
1795 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1796 int amount_grown = 0;
1797 WarmCallInfo* call;
1798 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1799 int est_size = (int)call->size();
1800 if (est_size > (room_to_grow - amount_grown)) {
1801 // This one won't fit anyway. Get rid of it.
1802 call->make_cold();
1803 continue;
1804 }
1805 call->make_hot();
1806 calls_made_hot++;
1807 amount_grown += est_size;
1808 amount_to_grow -= est_size;
1809 }
1811 if (calls_made_hot > 0) set_major_progress();
1812 return calls_made_hot;
1813 }
1816 //----------------------------Finish_Warm--------------------------------------
1817 void Compile::Finish_Warm() {
1818 if (!InlineWarmCalls) return;
1819 if (failing()) return;
1820 if (warm_calls() == NULL) return;
1822 // Clean up loose ends, if we are out of space for inlining.
1823 WarmCallInfo* call;
1824 while ((call = pop_warm_call()) != NULL) {
1825 call->make_cold();
1826 }
1827 }
1829 //---------------------cleanup_loop_predicates-----------------------
1830 // Remove the opaque nodes that protect the predicates so that all unused
1831 // checks and uncommon_traps will be eliminated from the ideal graph
1832 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1833 if (predicate_count()==0) return;
1834 for (int i = predicate_count(); i > 0; i--) {
1835 Node * n = predicate_opaque1_node(i-1);
1836 assert(n->Opcode() == Op_Opaque1, "must be");
1837 igvn.replace_node(n, n->in(1));
1838 }
1839 assert(predicate_count()==0, "should be clean!");
1840 }
1842 // StringOpts and late inlining of string methods
1843 void Compile::inline_string_calls(bool parse_time) {
1844 {
1845 // remove useless nodes to make the usage analysis simpler
1846 ResourceMark rm;
1847 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1848 }
1850 {
1851 ResourceMark rm;
1852 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1853 PhaseStringOpts pso(initial_gvn(), for_igvn());
1854 print_method(PHASE_AFTER_STRINGOPTS, 3);
1855 }
1857 // now inline anything that we skipped the first time around
1858 if (!parse_time) {
1859 _late_inlines_pos = _late_inlines.length();
1860 }
1862 while (_string_late_inlines.length() > 0) {
1863 CallGenerator* cg = _string_late_inlines.pop();
1864 cg->do_late_inline();
1865 if (failing()) return;
1866 }
1867 _string_late_inlines.trunc_to(0);
1868 }
1870 // Late inlining of boxing methods
1871 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1872 if (_boxing_late_inlines.length() > 0) {
1873 assert(has_boxed_value(), "inconsistent");
1875 PhaseGVN* gvn = initial_gvn();
1876 set_inlining_incrementally(true);
1878 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1879 for_igvn()->clear();
1880 gvn->replace_with(&igvn);
1882 while (_boxing_late_inlines.length() > 0) {
1883 CallGenerator* cg = _boxing_late_inlines.pop();
1884 cg->do_late_inline();
1885 if (failing()) return;
1886 }
1887 _boxing_late_inlines.trunc_to(0);
1889 {
1890 ResourceMark rm;
1891 PhaseRemoveUseless pru(gvn, for_igvn());
1892 }
1894 igvn = PhaseIterGVN(gvn);
1895 igvn.optimize();
1897 set_inlining_progress(false);
1898 set_inlining_incrementally(false);
1899 }
1900 }
1902 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1903 assert(IncrementalInline, "incremental inlining should be on");
1904 PhaseGVN* gvn = initial_gvn();
1906 set_inlining_progress(false);
1907 for_igvn()->clear();
1908 gvn->replace_with(&igvn);
1910 int i = 0;
1912 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1913 CallGenerator* cg = _late_inlines.at(i);
1914 _late_inlines_pos = i+1;
1915 cg->do_late_inline();
1916 if (failing()) return;
1917 }
1918 int j = 0;
1919 for (; i < _late_inlines.length(); i++, j++) {
1920 _late_inlines.at_put(j, _late_inlines.at(i));
1921 }
1922 _late_inlines.trunc_to(j);
1924 {
1925 ResourceMark rm;
1926 PhaseRemoveUseless pru(gvn, for_igvn());
1927 }
1929 igvn = PhaseIterGVN(gvn);
1930 }
1932 // Perform incremental inlining until bound on number of live nodes is reached
1933 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1934 PhaseGVN* gvn = initial_gvn();
1936 set_inlining_incrementally(true);
1937 set_inlining_progress(true);
1938 uint low_live_nodes = 0;
1940 while(inlining_progress() && _late_inlines.length() > 0) {
1942 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1943 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1944 // PhaseIdealLoop is expensive so we only try it once we are
1945 // out of loop and we only try it again if the previous helped
1946 // got the number of nodes down significantly
1947 PhaseIdealLoop ideal_loop( igvn, false, true );
1948 if (failing()) return;
1949 low_live_nodes = live_nodes();
1950 _major_progress = true;
1951 }
1953 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1954 break;
1955 }
1956 }
1958 inline_incrementally_one(igvn);
1960 if (failing()) return;
1962 igvn.optimize();
1964 if (failing()) return;
1965 }
1967 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1969 if (_string_late_inlines.length() > 0) {
1970 assert(has_stringbuilder(), "inconsistent");
1971 for_igvn()->clear();
1972 initial_gvn()->replace_with(&igvn);
1974 inline_string_calls(false);
1976 if (failing()) return;
1978 {
1979 ResourceMark rm;
1980 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1981 }
1983 igvn = PhaseIterGVN(gvn);
1985 igvn.optimize();
1986 }
1988 set_inlining_incrementally(false);
1989 }
1992 //------------------------------Optimize---------------------------------------
1993 // Given a graph, optimize it.
1994 void Compile::Optimize() {
1995 TracePhase t1("optimizer", &_t_optimizer, true);
1997 #ifndef PRODUCT
1998 if (env()->break_at_compile()) {
1999 BREAKPOINT;
2000 }
2002 #endif
2004 ResourceMark rm;
2005 int loop_opts_cnt;
2007 NOT_PRODUCT( verify_graph_edges(); )
2009 print_method(PHASE_AFTER_PARSING);
2011 {
2012 // Iterative Global Value Numbering, including ideal transforms
2013 // Initialize IterGVN with types and values from parse-time GVN
2014 PhaseIterGVN igvn(initial_gvn());
2015 {
2016 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2017 igvn.optimize();
2018 }
2020 print_method(PHASE_ITER_GVN1, 2);
2022 if (failing()) return;
2024 {
2025 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2026 inline_incrementally(igvn);
2027 }
2029 print_method(PHASE_INCREMENTAL_INLINE, 2);
2031 if (failing()) return;
2033 if (eliminate_boxing()) {
2034 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2035 // Inline valueOf() methods now.
2036 inline_boxing_calls(igvn);
2038 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2040 if (failing()) return;
2041 }
2043 // Remove the speculative part of types and clean up the graph from
2044 // the extra CastPP nodes whose only purpose is to carry them. Do
2045 // that early so that optimizations are not disrupted by the extra
2046 // CastPP nodes.
2047 remove_speculative_types(igvn);
2049 // No more new expensive nodes will be added to the list from here
2050 // so keep only the actual candidates for optimizations.
2051 cleanup_expensive_nodes(igvn);
2053 // Perform escape analysis
2054 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2055 if (has_loops()) {
2056 // Cleanup graph (remove dead nodes).
2057 TracePhase t2("idealLoop", &_t_idealLoop, true);
2058 PhaseIdealLoop ideal_loop( igvn, false, true );
2059 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2060 if (failing()) return;
2061 }
2062 ConnectionGraph::do_analysis(this, &igvn);
2064 if (failing()) return;
2066 // Optimize out fields loads from scalar replaceable allocations.
2067 igvn.optimize();
2068 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2070 if (failing()) return;
2072 if (congraph() != NULL && macro_count() > 0) {
2073 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2074 PhaseMacroExpand mexp(igvn);
2075 mexp.eliminate_macro_nodes();
2076 igvn.set_delay_transform(false);
2078 igvn.optimize();
2079 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2081 if (failing()) return;
2082 }
2083 }
2085 // Loop transforms on the ideal graph. Range Check Elimination,
2086 // peeling, unrolling, etc.
2088 // Set loop opts counter
2089 loop_opts_cnt = num_loop_opts();
2090 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2091 {
2092 TracePhase t2("idealLoop", &_t_idealLoop, true);
2093 PhaseIdealLoop ideal_loop( igvn, true );
2094 loop_opts_cnt--;
2095 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2096 if (failing()) return;
2097 }
2098 // Loop opts pass if partial peeling occurred in previous pass
2099 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2100 TracePhase t3("idealLoop", &_t_idealLoop, true);
2101 PhaseIdealLoop ideal_loop( igvn, false );
2102 loop_opts_cnt--;
2103 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2104 if (failing()) return;
2105 }
2106 // Loop opts pass for loop-unrolling before CCP
2107 if(major_progress() && (loop_opts_cnt > 0)) {
2108 TracePhase t4("idealLoop", &_t_idealLoop, true);
2109 PhaseIdealLoop ideal_loop( igvn, false );
2110 loop_opts_cnt--;
2111 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2112 }
2113 if (!failing()) {
2114 // Verify that last round of loop opts produced a valid graph
2115 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2116 PhaseIdealLoop::verify(igvn);
2117 }
2118 }
2119 if (failing()) return;
2121 // Conditional Constant Propagation;
2122 PhaseCCP ccp( &igvn );
2123 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2124 {
2125 TracePhase t2("ccp", &_t_ccp, true);
2126 ccp.do_transform();
2127 }
2128 print_method(PHASE_CPP1, 2);
2130 assert( true, "Break here to ccp.dump_old2new_map()");
2132 // Iterative Global Value Numbering, including ideal transforms
2133 {
2134 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2135 igvn = ccp;
2136 igvn.optimize();
2137 }
2139 print_method(PHASE_ITER_GVN2, 2);
2141 if (failing()) return;
2143 // Loop transforms on the ideal graph. Range Check Elimination,
2144 // peeling, unrolling, etc.
2145 if(loop_opts_cnt > 0) {
2146 debug_only( int cnt = 0; );
2147 while(major_progress() && (loop_opts_cnt > 0)) {
2148 TracePhase t2("idealLoop", &_t_idealLoop, true);
2149 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2150 PhaseIdealLoop ideal_loop( igvn, true);
2151 loop_opts_cnt--;
2152 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2153 if (failing()) return;
2154 }
2155 }
2157 {
2158 // Verify that all previous optimizations produced a valid graph
2159 // at least to this point, even if no loop optimizations were done.
2160 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2161 PhaseIdealLoop::verify(igvn);
2162 }
2164 {
2165 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2166 PhaseMacroExpand mex(igvn);
2167 if (mex.expand_macro_nodes()) {
2168 assert(failing(), "must bail out w/ explicit message");
2169 return;
2170 }
2171 }
2173 } // (End scope of igvn; run destructor if necessary for asserts.)
2175 dump_inlining();
2176 // A method with only infinite loops has no edges entering loops from root
2177 {
2178 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2179 if (final_graph_reshaping()) {
2180 assert(failing(), "must bail out w/ explicit message");
2181 return;
2182 }
2183 }
2185 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2186 }
2189 //------------------------------Code_Gen---------------------------------------
2190 // Given a graph, generate code for it
2191 void Compile::Code_Gen() {
2192 if (failing()) {
2193 return;
2194 }
2196 // Perform instruction selection. You might think we could reclaim Matcher
2197 // memory PDQ, but actually the Matcher is used in generating spill code.
2198 // Internals of the Matcher (including some VectorSets) must remain live
2199 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2200 // set a bit in reclaimed memory.
2202 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2203 // nodes. Mapping is only valid at the root of each matched subtree.
2204 NOT_PRODUCT( verify_graph_edges(); )
2206 Matcher matcher;
2207 _matcher = &matcher;
2208 {
2209 TracePhase t2("matcher", &_t_matcher, true);
2210 matcher.match();
2211 }
2212 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2213 // nodes. Mapping is only valid at the root of each matched subtree.
2214 NOT_PRODUCT( verify_graph_edges(); )
2216 // If you have too many nodes, or if matching has failed, bail out
2217 check_node_count(0, "out of nodes matching instructions");
2218 if (failing()) {
2219 return;
2220 }
2222 // Build a proper-looking CFG
2223 PhaseCFG cfg(node_arena(), root(), matcher);
2224 _cfg = &cfg;
2225 {
2226 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2227 bool success = cfg.do_global_code_motion();
2228 if (!success) {
2229 return;
2230 }
2232 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2233 NOT_PRODUCT( verify_graph_edges(); )
2234 debug_only( cfg.verify(); )
2235 }
2237 PhaseChaitin regalloc(unique(), cfg, matcher);
2238 _regalloc = ®alloc;
2239 {
2240 TracePhase t2("regalloc", &_t_registerAllocation, true);
2241 // Perform register allocation. After Chaitin, use-def chains are
2242 // no longer accurate (at spill code) and so must be ignored.
2243 // Node->LRG->reg mappings are still accurate.
2244 _regalloc->Register_Allocate();
2246 // Bail out if the allocator builds too many nodes
2247 if (failing()) {
2248 return;
2249 }
2250 }
2252 // Prior to register allocation we kept empty basic blocks in case the
2253 // the allocator needed a place to spill. After register allocation we
2254 // are not adding any new instructions. If any basic block is empty, we
2255 // can now safely remove it.
2256 {
2257 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2258 cfg.remove_empty_blocks();
2259 if (do_freq_based_layout()) {
2260 PhaseBlockLayout layout(cfg);
2261 } else {
2262 cfg.set_loop_alignment();
2263 }
2264 cfg.fixup_flow();
2265 }
2267 // Apply peephole optimizations
2268 if( OptoPeephole ) {
2269 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2270 PhasePeephole peep( _regalloc, cfg);
2271 peep.do_transform();
2272 }
2274 // Do late expand if CPU requires this.
2275 if (Matcher::require_postalloc_expand) {
2276 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2277 cfg.postalloc_expand(_regalloc);
2278 }
2280 // Convert Nodes to instruction bits in a buffer
2281 {
2282 // %%%% workspace merge brought two timers together for one job
2283 TracePhase t2a("output", &_t_output, true);
2284 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2285 Output();
2286 }
2288 print_method(PHASE_FINAL_CODE);
2290 // He's dead, Jim.
2291 _cfg = (PhaseCFG*)0xdeadbeef;
2292 _regalloc = (PhaseChaitin*)0xdeadbeef;
2293 }
2296 //------------------------------dump_asm---------------------------------------
2297 // Dump formatted assembly
2298 #ifndef PRODUCT
2299 void Compile::dump_asm(int *pcs, uint pc_limit) {
2300 bool cut_short = false;
2301 tty->print_cr("#");
2302 tty->print("# "); _tf->dump(); tty->cr();
2303 tty->print_cr("#");
2305 // For all blocks
2306 int pc = 0x0; // Program counter
2307 char starts_bundle = ' ';
2308 _regalloc->dump_frame();
2310 Node *n = NULL;
2311 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2312 if (VMThread::should_terminate()) {
2313 cut_short = true;
2314 break;
2315 }
2316 Block* block = _cfg->get_block(i);
2317 if (block->is_connector() && !Verbose) {
2318 continue;
2319 }
2320 n = block->head();
2321 if (pcs && n->_idx < pc_limit) {
2322 tty->print("%3.3x ", pcs[n->_idx]);
2323 } else {
2324 tty->print(" ");
2325 }
2326 block->dump_head(_cfg);
2327 if (block->is_connector()) {
2328 tty->print_cr(" # Empty connector block");
2329 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2330 tty->print_cr(" # Block is sole successor of call");
2331 }
2333 // For all instructions
2334 Node *delay = NULL;
2335 for (uint j = 0; j < block->number_of_nodes(); j++) {
2336 if (VMThread::should_terminate()) {
2337 cut_short = true;
2338 break;
2339 }
2340 n = block->get_node(j);
2341 if (valid_bundle_info(n)) {
2342 Bundle* bundle = node_bundling(n);
2343 if (bundle->used_in_unconditional_delay()) {
2344 delay = n;
2345 continue;
2346 }
2347 if (bundle->starts_bundle()) {
2348 starts_bundle = '+';
2349 }
2350 }
2352 if (WizardMode) {
2353 n->dump();
2354 }
2356 if( !n->is_Region() && // Dont print in the Assembly
2357 !n->is_Phi() && // a few noisely useless nodes
2358 !n->is_Proj() &&
2359 !n->is_MachTemp() &&
2360 !n->is_SafePointScalarObject() &&
2361 !n->is_Catch() && // Would be nice to print exception table targets
2362 !n->is_MergeMem() && // Not very interesting
2363 !n->is_top() && // Debug info table constants
2364 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2365 ) {
2366 if (pcs && n->_idx < pc_limit)
2367 tty->print("%3.3x", pcs[n->_idx]);
2368 else
2369 tty->print(" ");
2370 tty->print(" %c ", starts_bundle);
2371 starts_bundle = ' ';
2372 tty->print("\t");
2373 n->format(_regalloc, tty);
2374 tty->cr();
2375 }
2377 // If we have an instruction with a delay slot, and have seen a delay,
2378 // then back up and print it
2379 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2380 assert(delay != NULL, "no unconditional delay instruction");
2381 if (WizardMode) delay->dump();
2383 if (node_bundling(delay)->starts_bundle())
2384 starts_bundle = '+';
2385 if (pcs && n->_idx < pc_limit)
2386 tty->print("%3.3x", pcs[n->_idx]);
2387 else
2388 tty->print(" ");
2389 tty->print(" %c ", starts_bundle);
2390 starts_bundle = ' ';
2391 tty->print("\t");
2392 delay->format(_regalloc, tty);
2393 tty->print_cr("");
2394 delay = NULL;
2395 }
2397 // Dump the exception table as well
2398 if( n->is_Catch() && (Verbose || WizardMode) ) {
2399 // Print the exception table for this offset
2400 _handler_table.print_subtable_for(pc);
2401 }
2402 }
2404 if (pcs && n->_idx < pc_limit)
2405 tty->print_cr("%3.3x", pcs[n->_idx]);
2406 else
2407 tty->print_cr("");
2409 assert(cut_short || delay == NULL, "no unconditional delay branch");
2411 } // End of per-block dump
2412 tty->print_cr("");
2414 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2415 }
2416 #endif
2418 //------------------------------Final_Reshape_Counts---------------------------
2419 // This class defines counters to help identify when a method
2420 // may/must be executed using hardware with only 24-bit precision.
2421 struct Final_Reshape_Counts : public StackObj {
2422 int _call_count; // count non-inlined 'common' calls
2423 int _float_count; // count float ops requiring 24-bit precision
2424 int _double_count; // count double ops requiring more precision
2425 int _java_call_count; // count non-inlined 'java' calls
2426 int _inner_loop_count; // count loops which need alignment
2427 VectorSet _visited; // Visitation flags
2428 Node_List _tests; // Set of IfNodes & PCTableNodes
2430 Final_Reshape_Counts() :
2431 _call_count(0), _float_count(0), _double_count(0),
2432 _java_call_count(0), _inner_loop_count(0),
2433 _visited( Thread::current()->resource_area() ) { }
2435 void inc_call_count () { _call_count ++; }
2436 void inc_float_count () { _float_count ++; }
2437 void inc_double_count() { _double_count++; }
2438 void inc_java_call_count() { _java_call_count++; }
2439 void inc_inner_loop_count() { _inner_loop_count++; }
2441 int get_call_count () const { return _call_count ; }
2442 int get_float_count () const { return _float_count ; }
2443 int get_double_count() const { return _double_count; }
2444 int get_java_call_count() const { return _java_call_count; }
2445 int get_inner_loop_count() const { return _inner_loop_count; }
2446 };
2448 #ifdef ASSERT
2449 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2450 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2451 // Make sure the offset goes inside the instance layout.
2452 return k->contains_field_offset(tp->offset());
2453 // Note that OffsetBot and OffsetTop are very negative.
2454 }
2455 #endif
2457 // Eliminate trivially redundant StoreCMs and accumulate their
2458 // precedence edges.
2459 void Compile::eliminate_redundant_card_marks(Node* n) {
2460 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2461 if (n->in(MemNode::Address)->outcnt() > 1) {
2462 // There are multiple users of the same address so it might be
2463 // possible to eliminate some of the StoreCMs
2464 Node* mem = n->in(MemNode::Memory);
2465 Node* adr = n->in(MemNode::Address);
2466 Node* val = n->in(MemNode::ValueIn);
2467 Node* prev = n;
2468 bool done = false;
2469 // Walk the chain of StoreCMs eliminating ones that match. As
2470 // long as it's a chain of single users then the optimization is
2471 // safe. Eliminating partially redundant StoreCMs would require
2472 // cloning copies down the other paths.
2473 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2474 if (adr == mem->in(MemNode::Address) &&
2475 val == mem->in(MemNode::ValueIn)) {
2476 // redundant StoreCM
2477 if (mem->req() > MemNode::OopStore) {
2478 // Hasn't been processed by this code yet.
2479 n->add_prec(mem->in(MemNode::OopStore));
2480 } else {
2481 // Already converted to precedence edge
2482 for (uint i = mem->req(); i < mem->len(); i++) {
2483 // Accumulate any precedence edges
2484 if (mem->in(i) != NULL) {
2485 n->add_prec(mem->in(i));
2486 }
2487 }
2488 // Everything above this point has been processed.
2489 done = true;
2490 }
2491 // Eliminate the previous StoreCM
2492 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2493 assert(mem->outcnt() == 0, "should be dead");
2494 mem->disconnect_inputs(NULL, this);
2495 } else {
2496 prev = mem;
2497 }
2498 mem = prev->in(MemNode::Memory);
2499 }
2500 }
2501 }
2503 //------------------------------final_graph_reshaping_impl----------------------
2504 // Implement items 1-5 from final_graph_reshaping below.
2505 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2507 if ( n->outcnt() == 0 ) return; // dead node
2508 uint nop = n->Opcode();
2510 // Check for 2-input instruction with "last use" on right input.
2511 // Swap to left input. Implements item (2).
2512 if( n->req() == 3 && // two-input instruction
2513 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2514 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2515 n->in(2)->outcnt() == 1 &&// right use IS a last use
2516 !n->in(2)->is_Con() ) { // right use is not a constant
2517 // Check for commutative opcode
2518 switch( nop ) {
2519 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2520 case Op_MaxI: case Op_MinI:
2521 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2522 case Op_AndL: case Op_XorL: case Op_OrL:
2523 case Op_AndI: case Op_XorI: case Op_OrI: {
2524 // Move "last use" input to left by swapping inputs
2525 n->swap_edges(1, 2);
2526 break;
2527 }
2528 default:
2529 break;
2530 }
2531 }
2533 #ifdef ASSERT
2534 if( n->is_Mem() ) {
2535 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2536 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2537 // oop will be recorded in oop map if load crosses safepoint
2538 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2539 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2540 "raw memory operations should have control edge");
2541 }
2542 #endif
2543 // Count FPU ops and common calls, implements item (3)
2544 switch( nop ) {
2545 // Count all float operations that may use FPU
2546 case Op_AddF:
2547 case Op_SubF:
2548 case Op_MulF:
2549 case Op_DivF:
2550 case Op_NegF:
2551 case Op_ModF:
2552 case Op_ConvI2F:
2553 case Op_ConF:
2554 case Op_CmpF:
2555 case Op_CmpF3:
2556 // case Op_ConvL2F: // longs are split into 32-bit halves
2557 frc.inc_float_count();
2558 break;
2560 case Op_ConvF2D:
2561 case Op_ConvD2F:
2562 frc.inc_float_count();
2563 frc.inc_double_count();
2564 break;
2566 // Count all double operations that may use FPU
2567 case Op_AddD:
2568 case Op_SubD:
2569 case Op_MulD:
2570 case Op_DivD:
2571 case Op_NegD:
2572 case Op_ModD:
2573 case Op_ConvI2D:
2574 case Op_ConvD2I:
2575 // case Op_ConvL2D: // handled by leaf call
2576 // case Op_ConvD2L: // handled by leaf call
2577 case Op_ConD:
2578 case Op_CmpD:
2579 case Op_CmpD3:
2580 frc.inc_double_count();
2581 break;
2582 case Op_Opaque1: // Remove Opaque Nodes before matching
2583 case Op_Opaque2: // Remove Opaque Nodes before matching
2584 n->subsume_by(n->in(1), this);
2585 break;
2586 case Op_CallStaticJava:
2587 case Op_CallJava:
2588 case Op_CallDynamicJava:
2589 frc.inc_java_call_count(); // Count java call site;
2590 case Op_CallRuntime:
2591 case Op_CallLeaf:
2592 case Op_CallLeafNoFP: {
2593 assert( n->is_Call(), "" );
2594 CallNode *call = n->as_Call();
2595 // Count call sites where the FP mode bit would have to be flipped.
2596 // Do not count uncommon runtime calls:
2597 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2598 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2599 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2600 frc.inc_call_count(); // Count the call site
2601 } else { // See if uncommon argument is shared
2602 Node *n = call->in(TypeFunc::Parms);
2603 int nop = n->Opcode();
2604 // Clone shared simple arguments to uncommon calls, item (1).
2605 if( n->outcnt() > 1 &&
2606 !n->is_Proj() &&
2607 nop != Op_CreateEx &&
2608 nop != Op_CheckCastPP &&
2609 nop != Op_DecodeN &&
2610 nop != Op_DecodeNKlass &&
2611 !n->is_Mem() ) {
2612 Node *x = n->clone();
2613 call->set_req( TypeFunc::Parms, x );
2614 }
2615 }
2616 break;
2617 }
2619 case Op_StoreD:
2620 case Op_LoadD:
2621 case Op_LoadD_unaligned:
2622 frc.inc_double_count();
2623 goto handle_mem;
2624 case Op_StoreF:
2625 case Op_LoadF:
2626 frc.inc_float_count();
2627 goto handle_mem;
2629 case Op_StoreCM:
2630 {
2631 // Convert OopStore dependence into precedence edge
2632 Node* prec = n->in(MemNode::OopStore);
2633 n->del_req(MemNode::OopStore);
2634 n->add_prec(prec);
2635 eliminate_redundant_card_marks(n);
2636 }
2638 // fall through
2640 case Op_StoreB:
2641 case Op_StoreC:
2642 case Op_StorePConditional:
2643 case Op_StoreI:
2644 case Op_StoreL:
2645 case Op_StoreIConditional:
2646 case Op_StoreLConditional:
2647 case Op_CompareAndSwapI:
2648 case Op_CompareAndSwapL:
2649 case Op_CompareAndSwapP:
2650 case Op_CompareAndSwapN:
2651 case Op_GetAndAddI:
2652 case Op_GetAndAddL:
2653 case Op_GetAndSetI:
2654 case Op_GetAndSetL:
2655 case Op_GetAndSetP:
2656 case Op_GetAndSetN:
2657 case Op_StoreP:
2658 case Op_StoreN:
2659 case Op_StoreNKlass:
2660 case Op_LoadB:
2661 case Op_LoadUB:
2662 case Op_LoadUS:
2663 case Op_LoadI:
2664 case Op_LoadKlass:
2665 case Op_LoadNKlass:
2666 case Op_LoadL:
2667 case Op_LoadL_unaligned:
2668 case Op_LoadPLocked:
2669 case Op_LoadP:
2670 case Op_LoadN:
2671 case Op_LoadRange:
2672 case Op_LoadS: {
2673 handle_mem:
2674 #ifdef ASSERT
2675 if( VerifyOptoOopOffsets ) {
2676 assert( n->is_Mem(), "" );
2677 MemNode *mem = (MemNode*)n;
2678 // Check to see if address types have grounded out somehow.
2679 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2680 assert( !tp || oop_offset_is_sane(tp), "" );
2681 }
2682 #endif
2683 break;
2684 }
2686 case Op_AddP: { // Assert sane base pointers
2687 Node *addp = n->in(AddPNode::Address);
2688 assert( !addp->is_AddP() ||
2689 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2690 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2691 "Base pointers must match" );
2692 #ifdef _LP64
2693 if ((UseCompressedOops || UseCompressedClassPointers) &&
2694 addp->Opcode() == Op_ConP &&
2695 addp == n->in(AddPNode::Base) &&
2696 n->in(AddPNode::Offset)->is_Con()) {
2697 // Use addressing with narrow klass to load with offset on x86.
2698 // On sparc loading 32-bits constant and decoding it have less
2699 // instructions (4) then load 64-bits constant (7).
2700 // Do this transformation here since IGVN will convert ConN back to ConP.
2701 const Type* t = addp->bottom_type();
2702 if (t->isa_oopptr() || t->isa_klassptr()) {
2703 Node* nn = NULL;
2705 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2707 // Look for existing ConN node of the same exact type.
2708 Node* r = root();
2709 uint cnt = r->outcnt();
2710 for (uint i = 0; i < cnt; i++) {
2711 Node* m = r->raw_out(i);
2712 if (m!= NULL && m->Opcode() == op &&
2713 m->bottom_type()->make_ptr() == t) {
2714 nn = m;
2715 break;
2716 }
2717 }
2718 if (nn != NULL) {
2719 // Decode a narrow oop to match address
2720 // [R12 + narrow_oop_reg<<3 + offset]
2721 if (t->isa_oopptr()) {
2722 nn = new (this) DecodeNNode(nn, t);
2723 } else {
2724 nn = new (this) DecodeNKlassNode(nn, t);
2725 }
2726 n->set_req(AddPNode::Base, nn);
2727 n->set_req(AddPNode::Address, nn);
2728 if (addp->outcnt() == 0) {
2729 addp->disconnect_inputs(NULL, this);
2730 }
2731 }
2732 }
2733 }
2734 #endif
2735 break;
2736 }
2738 #ifdef _LP64
2739 case Op_CastPP:
2740 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2741 Node* in1 = n->in(1);
2742 const Type* t = n->bottom_type();
2743 Node* new_in1 = in1->clone();
2744 new_in1->as_DecodeN()->set_type(t);
2746 if (!Matcher::narrow_oop_use_complex_address()) {
2747 //
2748 // x86, ARM and friends can handle 2 adds in addressing mode
2749 // and Matcher can fold a DecodeN node into address by using
2750 // a narrow oop directly and do implicit NULL check in address:
2751 //
2752 // [R12 + narrow_oop_reg<<3 + offset]
2753 // NullCheck narrow_oop_reg
2754 //
2755 // On other platforms (Sparc) we have to keep new DecodeN node and
2756 // use it to do implicit NULL check in address:
2757 //
2758 // decode_not_null narrow_oop_reg, base_reg
2759 // [base_reg + offset]
2760 // NullCheck base_reg
2761 //
2762 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2763 // to keep the information to which NULL check the new DecodeN node
2764 // corresponds to use it as value in implicit_null_check().
2765 //
2766 new_in1->set_req(0, n->in(0));
2767 }
2769 n->subsume_by(new_in1, this);
2770 if (in1->outcnt() == 0) {
2771 in1->disconnect_inputs(NULL, this);
2772 }
2773 }
2774 break;
2776 case Op_CmpP:
2777 // Do this transformation here to preserve CmpPNode::sub() and
2778 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2779 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2780 Node* in1 = n->in(1);
2781 Node* in2 = n->in(2);
2782 if (!in1->is_DecodeNarrowPtr()) {
2783 in2 = in1;
2784 in1 = n->in(2);
2785 }
2786 assert(in1->is_DecodeNarrowPtr(), "sanity");
2788 Node* new_in2 = NULL;
2789 if (in2->is_DecodeNarrowPtr()) {
2790 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2791 new_in2 = in2->in(1);
2792 } else if (in2->Opcode() == Op_ConP) {
2793 const Type* t = in2->bottom_type();
2794 if (t == TypePtr::NULL_PTR) {
2795 assert(in1->is_DecodeN(), "compare klass to null?");
2796 // Don't convert CmpP null check into CmpN if compressed
2797 // oops implicit null check is not generated.
2798 // This will allow to generate normal oop implicit null check.
2799 if (Matcher::gen_narrow_oop_implicit_null_checks())
2800 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2801 //
2802 // This transformation together with CastPP transformation above
2803 // will generated code for implicit NULL checks for compressed oops.
2804 //
2805 // The original code after Optimize()
2806 //
2807 // LoadN memory, narrow_oop_reg
2808 // decode narrow_oop_reg, base_reg
2809 // CmpP base_reg, NULL
2810 // CastPP base_reg // NotNull
2811 // Load [base_reg + offset], val_reg
2812 //
2813 // after these transformations will be
2814 //
2815 // LoadN memory, narrow_oop_reg
2816 // CmpN narrow_oop_reg, NULL
2817 // decode_not_null narrow_oop_reg, base_reg
2818 // Load [base_reg + offset], val_reg
2819 //
2820 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2821 // since narrow oops can be used in debug info now (see the code in
2822 // final_graph_reshaping_walk()).
2823 //
2824 // At the end the code will be matched to
2825 // on x86:
2826 //
2827 // Load_narrow_oop memory, narrow_oop_reg
2828 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2829 // NullCheck narrow_oop_reg
2830 //
2831 // and on sparc:
2832 //
2833 // Load_narrow_oop memory, narrow_oop_reg
2834 // decode_not_null narrow_oop_reg, base_reg
2835 // Load [base_reg + offset], val_reg
2836 // NullCheck base_reg
2837 //
2838 } else if (t->isa_oopptr()) {
2839 new_in2 = ConNode::make(this, t->make_narrowoop());
2840 } else if (t->isa_klassptr()) {
2841 new_in2 = ConNode::make(this, t->make_narrowklass());
2842 }
2843 }
2844 if (new_in2 != NULL) {
2845 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2846 n->subsume_by(cmpN, this);
2847 if (in1->outcnt() == 0) {
2848 in1->disconnect_inputs(NULL, this);
2849 }
2850 if (in2->outcnt() == 0) {
2851 in2->disconnect_inputs(NULL, this);
2852 }
2853 }
2854 }
2855 break;
2857 case Op_DecodeN:
2858 case Op_DecodeNKlass:
2859 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2860 // DecodeN could be pinned when it can't be fold into
2861 // an address expression, see the code for Op_CastPP above.
2862 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2863 break;
2865 case Op_EncodeP:
2866 case Op_EncodePKlass: {
2867 Node* in1 = n->in(1);
2868 if (in1->is_DecodeNarrowPtr()) {
2869 n->subsume_by(in1->in(1), this);
2870 } else if (in1->Opcode() == Op_ConP) {
2871 const Type* t = in1->bottom_type();
2872 if (t == TypePtr::NULL_PTR) {
2873 assert(t->isa_oopptr(), "null klass?");
2874 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2875 } else if (t->isa_oopptr()) {
2876 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2877 } else if (t->isa_klassptr()) {
2878 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2879 }
2880 }
2881 if (in1->outcnt() == 0) {
2882 in1->disconnect_inputs(NULL, this);
2883 }
2884 break;
2885 }
2887 case Op_Proj: {
2888 if (OptimizeStringConcat) {
2889 ProjNode* p = n->as_Proj();
2890 if (p->_is_io_use) {
2891 // Separate projections were used for the exception path which
2892 // are normally removed by a late inline. If it wasn't inlined
2893 // then they will hang around and should just be replaced with
2894 // the original one.
2895 Node* proj = NULL;
2896 // Replace with just one
2897 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2898 Node *use = i.get();
2899 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2900 proj = use;
2901 break;
2902 }
2903 }
2904 assert(proj != NULL, "must be found");
2905 p->subsume_by(proj, this);
2906 }
2907 }
2908 break;
2909 }
2911 case Op_Phi:
2912 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
2913 // The EncodeP optimization may create Phi with the same edges
2914 // for all paths. It is not handled well by Register Allocator.
2915 Node* unique_in = n->in(1);
2916 assert(unique_in != NULL, "");
2917 uint cnt = n->req();
2918 for (uint i = 2; i < cnt; i++) {
2919 Node* m = n->in(i);
2920 assert(m != NULL, "");
2921 if (unique_in != m)
2922 unique_in = NULL;
2923 }
2924 if (unique_in != NULL) {
2925 n->subsume_by(unique_in, this);
2926 }
2927 }
2928 break;
2930 #endif
2932 case Op_ModI:
2933 if (UseDivMod) {
2934 // Check if a%b and a/b both exist
2935 Node* d = n->find_similar(Op_DivI);
2936 if (d) {
2937 // Replace them with a fused divmod if supported
2938 if (Matcher::has_match_rule(Op_DivModI)) {
2939 DivModINode* divmod = DivModINode::make(this, n);
2940 d->subsume_by(divmod->div_proj(), this);
2941 n->subsume_by(divmod->mod_proj(), this);
2942 } else {
2943 // replace a%b with a-((a/b)*b)
2944 Node* mult = new (this) MulINode(d, d->in(2));
2945 Node* sub = new (this) SubINode(d->in(1), mult);
2946 n->subsume_by(sub, this);
2947 }
2948 }
2949 }
2950 break;
2952 case Op_ModL:
2953 if (UseDivMod) {
2954 // Check if a%b and a/b both exist
2955 Node* d = n->find_similar(Op_DivL);
2956 if (d) {
2957 // Replace them with a fused divmod if supported
2958 if (Matcher::has_match_rule(Op_DivModL)) {
2959 DivModLNode* divmod = DivModLNode::make(this, n);
2960 d->subsume_by(divmod->div_proj(), this);
2961 n->subsume_by(divmod->mod_proj(), this);
2962 } else {
2963 // replace a%b with a-((a/b)*b)
2964 Node* mult = new (this) MulLNode(d, d->in(2));
2965 Node* sub = new (this) SubLNode(d->in(1), mult);
2966 n->subsume_by(sub, this);
2967 }
2968 }
2969 }
2970 break;
2972 case Op_LoadVector:
2973 case Op_StoreVector:
2974 break;
2976 case Op_PackB:
2977 case Op_PackS:
2978 case Op_PackI:
2979 case Op_PackF:
2980 case Op_PackL:
2981 case Op_PackD:
2982 if (n->req()-1 > 2) {
2983 // Replace many operand PackNodes with a binary tree for matching
2984 PackNode* p = (PackNode*) n;
2985 Node* btp = p->binary_tree_pack(this, 1, n->req());
2986 n->subsume_by(btp, this);
2987 }
2988 break;
2989 case Op_Loop:
2990 case Op_CountedLoop:
2991 if (n->as_Loop()->is_inner_loop()) {
2992 frc.inc_inner_loop_count();
2993 }
2994 break;
2995 case Op_LShiftI:
2996 case Op_RShiftI:
2997 case Op_URShiftI:
2998 case Op_LShiftL:
2999 case Op_RShiftL:
3000 case Op_URShiftL:
3001 if (Matcher::need_masked_shift_count) {
3002 // The cpu's shift instructions don't restrict the count to the
3003 // lower 5/6 bits. We need to do the masking ourselves.
3004 Node* in2 = n->in(2);
3005 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3006 const TypeInt* t = in2->find_int_type();
3007 if (t != NULL && t->is_con()) {
3008 juint shift = t->get_con();
3009 if (shift > mask) { // Unsigned cmp
3010 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
3011 }
3012 } else {
3013 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3014 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3015 n->set_req(2, shift);
3016 }
3017 }
3018 if (in2->outcnt() == 0) { // Remove dead node
3019 in2->disconnect_inputs(NULL, this);
3020 }
3021 }
3022 break;
3023 case Op_MemBarStoreStore:
3024 case Op_MemBarRelease:
3025 // Break the link with AllocateNode: it is no longer useful and
3026 // confuses register allocation.
3027 if (n->req() > MemBarNode::Precedent) {
3028 n->set_req(MemBarNode::Precedent, top());
3029 }
3030 break;
3031 // Must set a control edge on all nodes that produce a FlagsProj
3032 // so they can't escape the block that consumes the flags.
3033 // Must also set the non throwing branch as the control
3034 // for all nodes that depends on the result. Unless the node
3035 // already have a control that isn't the control of the
3036 // flag producer
3037 case Op_FlagsProj:
3038 {
3039 MathExactNode* math = (MathExactNode*) n->in(0);
3040 Node* ctrl = math->control_node();
3041 Node* non_throwing = math->non_throwing_branch();
3042 math->set_req(0, ctrl);
3044 Node* result = math->result_node();
3045 if (result != NULL) {
3046 for (DUIterator_Fast jmax, j = result->fast_outs(jmax); j < jmax; j++) {
3047 Node* out = result->fast_out(j);
3048 // Phi nodes shouldn't be moved. They would only match below if they
3049 // had the same control as the MathExactNode. The only time that
3050 // would happen is if the Phi is also an input to the MathExact
3051 //
3052 // Cmp nodes shouldn't have control set at all.
3053 if (out->is_Phi() ||
3054 out->is_Cmp()) {
3055 continue;
3056 }
3058 if (out->in(0) == NULL) {
3059 out->set_req(0, non_throwing);
3060 } else if (out->in(0) == ctrl) {
3061 out->set_req(0, non_throwing);
3062 }
3063 }
3064 }
3065 }
3066 break;
3067 default:
3068 assert( !n->is_Call(), "" );
3069 assert( !n->is_Mem(), "" );
3070 break;
3071 }
3073 // Collect CFG split points
3074 if (n->is_MultiBranch())
3075 frc._tests.push(n);
3076 }
3078 //------------------------------final_graph_reshaping_walk---------------------
3079 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3080 // requires that the walk visits a node's inputs before visiting the node.
3081 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3082 ResourceArea *area = Thread::current()->resource_area();
3083 Unique_Node_List sfpt(area);
3085 frc._visited.set(root->_idx); // first, mark node as visited
3086 uint cnt = root->req();
3087 Node *n = root;
3088 uint i = 0;
3089 while (true) {
3090 if (i < cnt) {
3091 // Place all non-visited non-null inputs onto stack
3092 Node* m = n->in(i);
3093 ++i;
3094 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3095 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
3096 sfpt.push(m);
3097 cnt = m->req();
3098 nstack.push(n, i); // put on stack parent and next input's index
3099 n = m;
3100 i = 0;
3101 }
3102 } else {
3103 // Now do post-visit work
3104 final_graph_reshaping_impl( n, frc );
3105 if (nstack.is_empty())
3106 break; // finished
3107 n = nstack.node(); // Get node from stack
3108 cnt = n->req();
3109 i = nstack.index();
3110 nstack.pop(); // Shift to the next node on stack
3111 }
3112 }
3114 // Skip next transformation if compressed oops are not used.
3115 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3116 (!UseCompressedOops && !UseCompressedClassPointers))
3117 return;
3119 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3120 // It could be done for an uncommon traps or any safepoints/calls
3121 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3122 while (sfpt.size() > 0) {
3123 n = sfpt.pop();
3124 JVMState *jvms = n->as_SafePoint()->jvms();
3125 assert(jvms != NULL, "sanity");
3126 int start = jvms->debug_start();
3127 int end = n->req();
3128 bool is_uncommon = (n->is_CallStaticJava() &&
3129 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3130 for (int j = start; j < end; j++) {
3131 Node* in = n->in(j);
3132 if (in->is_DecodeNarrowPtr()) {
3133 bool safe_to_skip = true;
3134 if (!is_uncommon ) {
3135 // Is it safe to skip?
3136 for (uint i = 0; i < in->outcnt(); i++) {
3137 Node* u = in->raw_out(i);
3138 if (!u->is_SafePoint() ||
3139 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3140 safe_to_skip = false;
3141 }
3142 }
3143 }
3144 if (safe_to_skip) {
3145 n->set_req(j, in->in(1));
3146 }
3147 if (in->outcnt() == 0) {
3148 in->disconnect_inputs(NULL, this);
3149 }
3150 }
3151 }
3152 }
3153 }
3155 //------------------------------final_graph_reshaping--------------------------
3156 // Final Graph Reshaping.
3157 //
3158 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3159 // and not commoned up and forced early. Must come after regular
3160 // optimizations to avoid GVN undoing the cloning. Clone constant
3161 // inputs to Loop Phis; these will be split by the allocator anyways.
3162 // Remove Opaque nodes.
3163 // (2) Move last-uses by commutative operations to the left input to encourage
3164 // Intel update-in-place two-address operations and better register usage
3165 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3166 // calls canonicalizing them back.
3167 // (3) Count the number of double-precision FP ops, single-precision FP ops
3168 // and call sites. On Intel, we can get correct rounding either by
3169 // forcing singles to memory (requires extra stores and loads after each
3170 // FP bytecode) or we can set a rounding mode bit (requires setting and
3171 // clearing the mode bit around call sites). The mode bit is only used
3172 // if the relative frequency of single FP ops to calls is low enough.
3173 // This is a key transform for SPEC mpeg_audio.
3174 // (4) Detect infinite loops; blobs of code reachable from above but not
3175 // below. Several of the Code_Gen algorithms fail on such code shapes,
3176 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3177 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3178 // Detection is by looking for IfNodes where only 1 projection is
3179 // reachable from below or CatchNodes missing some targets.
3180 // (5) Assert for insane oop offsets in debug mode.
3182 bool Compile::final_graph_reshaping() {
3183 // an infinite loop may have been eliminated by the optimizer,
3184 // in which case the graph will be empty.
3185 if (root()->req() == 1) {
3186 record_method_not_compilable("trivial infinite loop");
3187 return true;
3188 }
3190 // Expensive nodes have their control input set to prevent the GVN
3191 // from freely commoning them. There's no GVN beyond this point so
3192 // no need to keep the control input. We want the expensive nodes to
3193 // be freely moved to the least frequent code path by gcm.
3194 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3195 for (int i = 0; i < expensive_count(); i++) {
3196 _expensive_nodes->at(i)->set_req(0, NULL);
3197 }
3199 Final_Reshape_Counts frc;
3201 // Visit everybody reachable!
3202 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3203 Node_Stack nstack(unique() >> 1);
3204 final_graph_reshaping_walk(nstack, root(), frc);
3206 // Check for unreachable (from below) code (i.e., infinite loops).
3207 for( uint i = 0; i < frc._tests.size(); i++ ) {
3208 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3209 // Get number of CFG targets.
3210 // Note that PCTables include exception targets after calls.
3211 uint required_outcnt = n->required_outcnt();
3212 if (n->outcnt() != required_outcnt) {
3213 // Check for a few special cases. Rethrow Nodes never take the
3214 // 'fall-thru' path, so expected kids is 1 less.
3215 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3216 if (n->in(0)->in(0)->is_Call()) {
3217 CallNode *call = n->in(0)->in(0)->as_Call();
3218 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3219 required_outcnt--; // Rethrow always has 1 less kid
3220 } else if (call->req() > TypeFunc::Parms &&
3221 call->is_CallDynamicJava()) {
3222 // Check for null receiver. In such case, the optimizer has
3223 // detected that the virtual call will always result in a null
3224 // pointer exception. The fall-through projection of this CatchNode
3225 // will not be populated.
3226 Node *arg0 = call->in(TypeFunc::Parms);
3227 if (arg0->is_Type() &&
3228 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3229 required_outcnt--;
3230 }
3231 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3232 call->req() > TypeFunc::Parms+1 &&
3233 call->is_CallStaticJava()) {
3234 // Check for negative array length. In such case, the optimizer has
3235 // detected that the allocation attempt will always result in an
3236 // exception. There is no fall-through projection of this CatchNode .
3237 Node *arg1 = call->in(TypeFunc::Parms+1);
3238 if (arg1->is_Type() &&
3239 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3240 required_outcnt--;
3241 }
3242 }
3243 }
3244 }
3245 // Recheck with a better notion of 'required_outcnt'
3246 if (n->outcnt() != required_outcnt) {
3247 record_method_not_compilable("malformed control flow");
3248 return true; // Not all targets reachable!
3249 }
3250 }
3251 // Check that I actually visited all kids. Unreached kids
3252 // must be infinite loops.
3253 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3254 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3255 record_method_not_compilable("infinite loop");
3256 return true; // Found unvisited kid; must be unreach
3257 }
3258 }
3260 // If original bytecodes contained a mixture of floats and doubles
3261 // check if the optimizer has made it homogenous, item (3).
3262 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3263 frc.get_float_count() > 32 &&
3264 frc.get_double_count() == 0 &&
3265 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3266 set_24_bit_selection_and_mode( false, true );
3267 }
3269 set_java_calls(frc.get_java_call_count());
3270 set_inner_loops(frc.get_inner_loop_count());
3272 // No infinite loops, no reason to bail out.
3273 return false;
3274 }
3276 //-----------------------------too_many_traps----------------------------------
3277 // Report if there are too many traps at the current method and bci.
3278 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3279 bool Compile::too_many_traps(ciMethod* method,
3280 int bci,
3281 Deoptimization::DeoptReason reason) {
3282 ciMethodData* md = method->method_data();
3283 if (md->is_empty()) {
3284 // Assume the trap has not occurred, or that it occurred only
3285 // because of a transient condition during start-up in the interpreter.
3286 return false;
3287 }
3288 if (md->has_trap_at(bci, reason) != 0) {
3289 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3290 // Also, if there are multiple reasons, or if there is no per-BCI record,
3291 // assume the worst.
3292 if (log())
3293 log()->elem("observe trap='%s' count='%d'",
3294 Deoptimization::trap_reason_name(reason),
3295 md->trap_count(reason));
3296 return true;
3297 } else {
3298 // Ignore method/bci and see if there have been too many globally.
3299 return too_many_traps(reason, md);
3300 }
3301 }
3303 // Less-accurate variant which does not require a method and bci.
3304 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3305 ciMethodData* logmd) {
3306 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
3307 // Too many traps globally.
3308 // Note that we use cumulative trap_count, not just md->trap_count.
3309 if (log()) {
3310 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3311 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3312 Deoptimization::trap_reason_name(reason),
3313 mcount, trap_count(reason));
3314 }
3315 return true;
3316 } else {
3317 // The coast is clear.
3318 return false;
3319 }
3320 }
3322 //--------------------------too_many_recompiles--------------------------------
3323 // Report if there are too many recompiles at the current method and bci.
3324 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3325 // Is not eager to return true, since this will cause the compiler to use
3326 // Action_none for a trap point, to avoid too many recompilations.
3327 bool Compile::too_many_recompiles(ciMethod* method,
3328 int bci,
3329 Deoptimization::DeoptReason reason) {
3330 ciMethodData* md = method->method_data();
3331 if (md->is_empty()) {
3332 // Assume the trap has not occurred, or that it occurred only
3333 // because of a transient condition during start-up in the interpreter.
3334 return false;
3335 }
3336 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3337 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3338 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3339 Deoptimization::DeoptReason per_bc_reason
3340 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3341 if ((per_bc_reason == Deoptimization::Reason_none
3342 || md->has_trap_at(bci, reason) != 0)
3343 // The trap frequency measure we care about is the recompile count:
3344 && md->trap_recompiled_at(bci)
3345 && md->overflow_recompile_count() >= bc_cutoff) {
3346 // Do not emit a trap here if it has already caused recompilations.
3347 // Also, if there are multiple reasons, or if there is no per-BCI record,
3348 // assume the worst.
3349 if (log())
3350 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3351 Deoptimization::trap_reason_name(reason),
3352 md->trap_count(reason),
3353 md->overflow_recompile_count());
3354 return true;
3355 } else if (trap_count(reason) != 0
3356 && decompile_count() >= m_cutoff) {
3357 // Too many recompiles globally, and we have seen this sort of trap.
3358 // Use cumulative decompile_count, not just md->decompile_count.
3359 if (log())
3360 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3361 Deoptimization::trap_reason_name(reason),
3362 md->trap_count(reason), trap_count(reason),
3363 md->decompile_count(), decompile_count());
3364 return true;
3365 } else {
3366 // The coast is clear.
3367 return false;
3368 }
3369 }
3371 // Compute when not to trap. Used by matching trap based nodes and
3372 // NullCheck optimization.
3373 void Compile::set_allowed_deopt_reasons() {
3374 _allowed_reasons = 0;
3375 if (is_method_compilation()) {
3376 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3377 assert(rs < BitsPerInt, "recode bit map");
3378 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3379 _allowed_reasons |= nth_bit(rs);
3380 }
3381 }
3382 }
3383 }
3385 #ifndef PRODUCT
3386 //------------------------------verify_graph_edges---------------------------
3387 // Walk the Graph and verify that there is a one-to-one correspondence
3388 // between Use-Def edges and Def-Use edges in the graph.
3389 void Compile::verify_graph_edges(bool no_dead_code) {
3390 if (VerifyGraphEdges) {
3391 ResourceArea *area = Thread::current()->resource_area();
3392 Unique_Node_List visited(area);
3393 // Call recursive graph walk to check edges
3394 _root->verify_edges(visited);
3395 if (no_dead_code) {
3396 // Now make sure that no visited node is used by an unvisited node.
3397 bool dead_nodes = 0;
3398 Unique_Node_List checked(area);
3399 while (visited.size() > 0) {
3400 Node* n = visited.pop();
3401 checked.push(n);
3402 for (uint i = 0; i < n->outcnt(); i++) {
3403 Node* use = n->raw_out(i);
3404 if (checked.member(use)) continue; // already checked
3405 if (visited.member(use)) continue; // already in the graph
3406 if (use->is_Con()) continue; // a dead ConNode is OK
3407 // At this point, we have found a dead node which is DU-reachable.
3408 if (dead_nodes++ == 0)
3409 tty->print_cr("*** Dead nodes reachable via DU edges:");
3410 use->dump(2);
3411 tty->print_cr("---");
3412 checked.push(use); // No repeats; pretend it is now checked.
3413 }
3414 }
3415 assert(dead_nodes == 0, "using nodes must be reachable from root");
3416 }
3417 }
3418 }
3420 // Verify GC barriers consistency
3421 // Currently supported:
3422 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3423 void Compile::verify_barriers() {
3424 if (UseG1GC) {
3425 // Verify G1 pre-barriers
3426 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3428 ResourceArea *area = Thread::current()->resource_area();
3429 Unique_Node_List visited(area);
3430 Node_List worklist(area);
3431 // We're going to walk control flow backwards starting from the Root
3432 worklist.push(_root);
3433 while (worklist.size() > 0) {
3434 Node* x = worklist.pop();
3435 if (x == NULL || x == top()) continue;
3436 if (visited.member(x)) {
3437 continue;
3438 } else {
3439 visited.push(x);
3440 }
3442 if (x->is_Region()) {
3443 for (uint i = 1; i < x->req(); i++) {
3444 worklist.push(x->in(i));
3445 }
3446 } else {
3447 worklist.push(x->in(0));
3448 // We are looking for the pattern:
3449 // /->ThreadLocal
3450 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3451 // \->ConI(0)
3452 // We want to verify that the If and the LoadB have the same control
3453 // See GraphKit::g1_write_barrier_pre()
3454 if (x->is_If()) {
3455 IfNode *iff = x->as_If();
3456 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3457 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3458 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3459 && cmp->in(1)->is_Load()) {
3460 LoadNode* load = cmp->in(1)->as_Load();
3461 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3462 && load->in(2)->in(3)->is_Con()
3463 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3465 Node* if_ctrl = iff->in(0);
3466 Node* load_ctrl = load->in(0);
3468 if (if_ctrl != load_ctrl) {
3469 // Skip possible CProj->NeverBranch in infinite loops
3470 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3471 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3472 if_ctrl = if_ctrl->in(0)->in(0);
3473 }
3474 }
3475 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3476 }
3477 }
3478 }
3479 }
3480 }
3481 }
3482 }
3483 }
3485 #endif
3487 // The Compile object keeps track of failure reasons separately from the ciEnv.
3488 // This is required because there is not quite a 1-1 relation between the
3489 // ciEnv and its compilation task and the Compile object. Note that one
3490 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3491 // to backtrack and retry without subsuming loads. Other than this backtracking
3492 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3493 // by the logic in C2Compiler.
3494 void Compile::record_failure(const char* reason) {
3495 if (log() != NULL) {
3496 log()->elem("failure reason='%s' phase='compile'", reason);
3497 }
3498 if (_failure_reason == NULL) {
3499 // Record the first failure reason.
3500 _failure_reason = reason;
3501 }
3503 EventCompilerFailure event;
3504 if (event.should_commit()) {
3505 event.set_compileID(Compile::compile_id());
3506 event.set_failure(reason);
3507 event.commit();
3508 }
3510 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3511 C->print_method(PHASE_FAILURE);
3512 }
3513 _root = NULL; // flush the graph, too
3514 }
3516 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3517 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3518 _phase_name(name), _dolog(dolog)
3519 {
3520 if (dolog) {
3521 C = Compile::current();
3522 _log = C->log();
3523 } else {
3524 C = NULL;
3525 _log = NULL;
3526 }
3527 if (_log != NULL) {
3528 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3529 _log->stamp();
3530 _log->end_head();
3531 }
3532 }
3534 Compile::TracePhase::~TracePhase() {
3536 C = Compile::current();
3537 if (_dolog) {
3538 _log = C->log();
3539 } else {
3540 _log = NULL;
3541 }
3543 #ifdef ASSERT
3544 if (PrintIdealNodeCount) {
3545 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3546 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3547 }
3549 if (VerifyIdealNodeCount) {
3550 Compile::current()->print_missing_nodes();
3551 }
3552 #endif
3554 if (_log != NULL) {
3555 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3556 }
3557 }
3559 //=============================================================================
3560 // Two Constant's are equal when the type and the value are equal.
3561 bool Compile::Constant::operator==(const Constant& other) {
3562 if (type() != other.type() ) return false;
3563 if (can_be_reused() != other.can_be_reused()) return false;
3564 // For floating point values we compare the bit pattern.
3565 switch (type()) {
3566 case T_FLOAT: return (_v._value.i == other._v._value.i);
3567 case T_LONG:
3568 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3569 case T_OBJECT:
3570 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3571 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3572 case T_METADATA: return (_v._metadata == other._v._metadata);
3573 default: ShouldNotReachHere();
3574 }
3575 return false;
3576 }
3578 static int type_to_size_in_bytes(BasicType t) {
3579 switch (t) {
3580 case T_LONG: return sizeof(jlong );
3581 case T_FLOAT: return sizeof(jfloat );
3582 case T_DOUBLE: return sizeof(jdouble);
3583 case T_METADATA: return sizeof(Metadata*);
3584 // We use T_VOID as marker for jump-table entries (labels) which
3585 // need an internal word relocation.
3586 case T_VOID:
3587 case T_ADDRESS:
3588 case T_OBJECT: return sizeof(jobject);
3589 }
3591 ShouldNotReachHere();
3592 return -1;
3593 }
3595 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3596 // sort descending
3597 if (a->freq() > b->freq()) return -1;
3598 if (a->freq() < b->freq()) return 1;
3599 return 0;
3600 }
3602 void Compile::ConstantTable::calculate_offsets_and_size() {
3603 // First, sort the array by frequencies.
3604 _constants.sort(qsort_comparator);
3606 #ifdef ASSERT
3607 // Make sure all jump-table entries were sorted to the end of the
3608 // array (they have a negative frequency).
3609 bool found_void = false;
3610 for (int i = 0; i < _constants.length(); i++) {
3611 Constant con = _constants.at(i);
3612 if (con.type() == T_VOID)
3613 found_void = true; // jump-tables
3614 else
3615 assert(!found_void, "wrong sorting");
3616 }
3617 #endif
3619 int offset = 0;
3620 for (int i = 0; i < _constants.length(); i++) {
3621 Constant* con = _constants.adr_at(i);
3623 // Align offset for type.
3624 int typesize = type_to_size_in_bytes(con->type());
3625 offset = align_size_up(offset, typesize);
3626 con->set_offset(offset); // set constant's offset
3628 if (con->type() == T_VOID) {
3629 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3630 offset = offset + typesize * n->outcnt(); // expand jump-table
3631 } else {
3632 offset = offset + typesize;
3633 }
3634 }
3636 // Align size up to the next section start (which is insts; see
3637 // CodeBuffer::align_at_start).
3638 assert(_size == -1, "already set?");
3639 _size = align_size_up(offset, CodeEntryAlignment);
3640 }
3642 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3643 MacroAssembler _masm(&cb);
3644 for (int i = 0; i < _constants.length(); i++) {
3645 Constant con = _constants.at(i);
3646 address constant_addr;
3647 switch (con.type()) {
3648 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3649 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3650 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3651 case T_OBJECT: {
3652 jobject obj = con.get_jobject();
3653 int oop_index = _masm.oop_recorder()->find_index(obj);
3654 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3655 break;
3656 }
3657 case T_ADDRESS: {
3658 address addr = (address) con.get_jobject();
3659 constant_addr = _masm.address_constant(addr);
3660 break;
3661 }
3662 // We use T_VOID as marker for jump-table entries (labels) which
3663 // need an internal word relocation.
3664 case T_VOID: {
3665 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3666 // Fill the jump-table with a dummy word. The real value is
3667 // filled in later in fill_jump_table.
3668 address dummy = (address) n;
3669 constant_addr = _masm.address_constant(dummy);
3670 // Expand jump-table
3671 for (uint i = 1; i < n->outcnt(); i++) {
3672 address temp_addr = _masm.address_constant(dummy + i);
3673 assert(temp_addr, "consts section too small");
3674 }
3675 break;
3676 }
3677 case T_METADATA: {
3678 Metadata* obj = con.get_metadata();
3679 int metadata_index = _masm.oop_recorder()->find_index(obj);
3680 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3681 break;
3682 }
3683 default: ShouldNotReachHere();
3684 }
3685 assert(constant_addr, "consts section too small");
3686 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3687 }
3688 }
3690 int Compile::ConstantTable::find_offset(Constant& con) const {
3691 int idx = _constants.find(con);
3692 assert(idx != -1, "constant must be in constant table");
3693 int offset = _constants.at(idx).offset();
3694 assert(offset != -1, "constant table not emitted yet?");
3695 return offset;
3696 }
3698 void Compile::ConstantTable::add(Constant& con) {
3699 if (con.can_be_reused()) {
3700 int idx = _constants.find(con);
3701 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3702 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3703 return;
3704 }
3705 }
3706 (void) _constants.append(con);
3707 }
3709 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3710 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3711 Constant con(type, value, b->_freq);
3712 add(con);
3713 return con;
3714 }
3716 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3717 Constant con(metadata);
3718 add(con);
3719 return con;
3720 }
3722 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3723 jvalue value;
3724 BasicType type = oper->type()->basic_type();
3725 switch (type) {
3726 case T_LONG: value.j = oper->constantL(); break;
3727 case T_FLOAT: value.f = oper->constantF(); break;
3728 case T_DOUBLE: value.d = oper->constantD(); break;
3729 case T_OBJECT:
3730 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3731 case T_METADATA: return add((Metadata*)oper->constant()); break;
3732 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3733 }
3734 return add(n, type, value);
3735 }
3737 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3738 jvalue value;
3739 // We can use the node pointer here to identify the right jump-table
3740 // as this method is called from Compile::Fill_buffer right before
3741 // the MachNodes are emitted and the jump-table is filled (means the
3742 // MachNode pointers do not change anymore).
3743 value.l = (jobject) n;
3744 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3745 add(con);
3746 return con;
3747 }
3749 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3750 // If called from Compile::scratch_emit_size do nothing.
3751 if (Compile::current()->in_scratch_emit_size()) return;
3753 assert(labels.is_nonempty(), "must be");
3754 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3756 // Since MachConstantNode::constant_offset() also contains
3757 // table_base_offset() we need to subtract the table_base_offset()
3758 // to get the plain offset into the constant table.
3759 int offset = n->constant_offset() - table_base_offset();
3761 MacroAssembler _masm(&cb);
3762 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3764 for (uint i = 0; i < n->outcnt(); i++) {
3765 address* constant_addr = &jump_table_base[i];
3766 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)));
3767 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3768 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3769 }
3770 }
3772 void Compile::dump_inlining() {
3773 if (print_inlining() || print_intrinsics()) {
3774 // Print inlining message for candidates that we couldn't inline
3775 // for lack of space or non constant receiver
3776 for (int i = 0; i < _late_inlines.length(); i++) {
3777 CallGenerator* cg = _late_inlines.at(i);
3778 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3779 }
3780 Unique_Node_List useful;
3781 useful.push(root());
3782 for (uint next = 0; next < useful.size(); ++next) {
3783 Node* n = useful.at(next);
3784 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3785 CallNode* call = n->as_Call();
3786 CallGenerator* cg = call->generator();
3787 cg->print_inlining_late("receiver not constant");
3788 }
3789 uint max = n->len();
3790 for ( uint i = 0; i < max; ++i ) {
3791 Node *m = n->in(i);
3792 if ( m == NULL ) continue;
3793 useful.push(m);
3794 }
3795 }
3796 for (int i = 0; i < _print_inlining_list->length(); i++) {
3797 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string());
3798 }
3799 }
3800 }
3802 // Dump inlining replay data to the stream.
3803 // Don't change thread state and acquire any locks.
3804 void Compile::dump_inline_data(outputStream* out) {
3805 InlineTree* inl_tree = ilt();
3806 if (inl_tree != NULL) {
3807 out->print(" inline %d", inl_tree->count());
3808 inl_tree->dump_replay_data(out);
3809 }
3810 }
3812 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3813 if (n1->Opcode() < n2->Opcode()) return -1;
3814 else if (n1->Opcode() > n2->Opcode()) return 1;
3816 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()));
3817 for (uint i = 1; i < n1->req(); i++) {
3818 if (n1->in(i) < n2->in(i)) return -1;
3819 else if (n1->in(i) > n2->in(i)) return 1;
3820 }
3822 return 0;
3823 }
3825 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3826 Node* n1 = *n1p;
3827 Node* n2 = *n2p;
3829 return cmp_expensive_nodes(n1, n2);
3830 }
3832 void Compile::sort_expensive_nodes() {
3833 if (!expensive_nodes_sorted()) {
3834 _expensive_nodes->sort(cmp_expensive_nodes);
3835 }
3836 }
3838 bool Compile::expensive_nodes_sorted() const {
3839 for (int i = 1; i < _expensive_nodes->length(); i++) {
3840 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3841 return false;
3842 }
3843 }
3844 return true;
3845 }
3847 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3848 if (_expensive_nodes->length() == 0) {
3849 return false;
3850 }
3852 assert(OptimizeExpensiveOps, "optimization off?");
3854 // Take this opportunity to remove dead nodes from the list
3855 int j = 0;
3856 for (int i = 0; i < _expensive_nodes->length(); i++) {
3857 Node* n = _expensive_nodes->at(i);
3858 if (!n->is_unreachable(igvn)) {
3859 assert(n->is_expensive(), "should be expensive");
3860 _expensive_nodes->at_put(j, n);
3861 j++;
3862 }
3863 }
3864 _expensive_nodes->trunc_to(j);
3866 // Then sort the list so that similar nodes are next to each other
3867 // and check for at least two nodes of identical kind with same data
3868 // inputs.
3869 sort_expensive_nodes();
3871 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3872 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3873 return true;
3874 }
3875 }
3877 return false;
3878 }
3880 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3881 if (_expensive_nodes->length() == 0) {
3882 return;
3883 }
3885 assert(OptimizeExpensiveOps, "optimization off?");
3887 // Sort to bring similar nodes next to each other and clear the
3888 // control input of nodes for which there's only a single copy.
3889 sort_expensive_nodes();
3891 int j = 0;
3892 int identical = 0;
3893 int i = 0;
3894 for (; i < _expensive_nodes->length()-1; i++) {
3895 assert(j <= i, "can't write beyond current index");
3896 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3897 identical++;
3898 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3899 continue;
3900 }
3901 if (identical > 0) {
3902 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3903 identical = 0;
3904 } else {
3905 Node* n = _expensive_nodes->at(i);
3906 igvn.hash_delete(n);
3907 n->set_req(0, NULL);
3908 igvn.hash_insert(n);
3909 }
3910 }
3911 if (identical > 0) {
3912 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3913 } else if (_expensive_nodes->length() >= 1) {
3914 Node* n = _expensive_nodes->at(i);
3915 igvn.hash_delete(n);
3916 n->set_req(0, NULL);
3917 igvn.hash_insert(n);
3918 }
3919 _expensive_nodes->trunc_to(j);
3920 }
3922 void Compile::add_expensive_node(Node * n) {
3923 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3924 assert(n->is_expensive(), "expensive nodes with non-null control here only");
3925 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
3926 if (OptimizeExpensiveOps) {
3927 _expensive_nodes->append(n);
3928 } else {
3929 // Clear control input and let IGVN optimize expensive nodes if
3930 // OptimizeExpensiveOps is off.
3931 n->set_req(0, NULL);
3932 }
3933 }
3935 /**
3936 * Remove the speculative part of types and clean up the graph
3937 */
3938 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
3939 if (UseTypeSpeculation) {
3940 Unique_Node_List worklist;
3941 worklist.push(root());
3942 int modified = 0;
3943 // Go over all type nodes that carry a speculative type, drop the
3944 // speculative part of the type and enqueue the node for an igvn
3945 // which may optimize it out.
3946 for (uint next = 0; next < worklist.size(); ++next) {
3947 Node *n = worklist.at(next);
3948 if (n->is_Type()) {
3949 TypeNode* tn = n->as_Type();
3950 const Type* t = tn->type();
3951 const Type* t_no_spec = t->remove_speculative();
3952 if (t_no_spec != t) {
3953 bool in_hash = igvn.hash_delete(n);
3954 assert(in_hash, "node should be in igvn hash table");
3955 tn->set_type(t_no_spec);
3956 igvn.hash_insert(n);
3957 igvn._worklist.push(n); // give it a chance to go away
3958 modified++;
3959 }
3960 }
3961 uint max = n->len();
3962 for( uint i = 0; i < max; ++i ) {
3963 Node *m = n->in(i);
3964 if (not_a_node(m)) continue;
3965 worklist.push(m);
3966 }
3967 }
3968 // Drop the speculative part of all types in the igvn's type table
3969 igvn.remove_speculative_types();
3970 if (modified > 0) {
3971 igvn.optimize();
3972 }
3973 #ifdef ASSERT
3974 // Verify that after the IGVN is over no speculative type has resurfaced
3975 worklist.clear();
3976 worklist.push(root());
3977 for (uint next = 0; next < worklist.size(); ++next) {
3978 Node *n = worklist.at(next);
3979 const Type* t = igvn.type(n);
3980 assert(t == t->remove_speculative(), "no more speculative types");
3981 if (n->is_Type()) {
3982 t = n->as_Type()->type();
3983 assert(t == t->remove_speculative(), "no more speculative types");
3984 }
3985 uint max = n->len();
3986 for( uint i = 0; i < max; ++i ) {
3987 Node *m = n->in(i);
3988 if (not_a_node(m)) continue;
3989 worklist.push(m);
3990 }
3991 }
3992 igvn.check_no_speculative_types();
3993 #endif
3994 }
3995 }
3997 // Auxiliary method to support randomized stressing/fuzzing.
3998 //
3999 // This method can be called the arbitrary number of times, with current count
4000 // as the argument. The logic allows selecting a single candidate from the
4001 // running list of candidates as follows:
4002 // int count = 0;
4003 // Cand* selected = null;
4004 // while(cand = cand->next()) {
4005 // if (randomized_select(++count)) {
4006 // selected = cand;
4007 // }
4008 // }
4009 //
4010 // Including count equalizes the chances any candidate is "selected".
4011 // This is useful when we don't have the complete list of candidates to choose
4012 // from uniformly. In this case, we need to adjust the randomicity of the
4013 // selection, or else we will end up biasing the selection towards the latter
4014 // candidates.
4015 //
4016 // Quick back-envelope calculation shows that for the list of n candidates
4017 // the equal probability for the candidate to persist as "best" can be
4018 // achieved by replacing it with "next" k-th candidate with the probability
4019 // of 1/k. It can be easily shown that by the end of the run, the
4020 // probability for any candidate is converged to 1/n, thus giving the
4021 // uniform distribution among all the candidates.
4022 //
4023 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4024 #define RANDOMIZED_DOMAIN_POW 29
4025 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4026 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4027 bool Compile::randomized_select(int count) {
4028 assert(count > 0, "only positive");
4029 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4030 }