Thu, 14 Feb 2019 14:31:32 +0100
8087128: C2: Disallow definition split on MachCopySpill nodes
Reviewed-by: kvn
1 /*
2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "compiler/disassembler.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/block.hpp"
37 #include "opto/c2compiler.hpp"
38 #include "opto/callGenerator.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/cfgnode.hpp"
41 #include "opto/chaitin.hpp"
42 #include "opto/compile.hpp"
43 #include "opto/connode.hpp"
44 #include "opto/divnode.hpp"
45 #include "opto/escape.hpp"
46 #include "opto/idealGraphPrinter.hpp"
47 #include "opto/loopnode.hpp"
48 #include "opto/machnode.hpp"
49 #include "opto/macro.hpp"
50 #include "opto/matcher.hpp"
51 #include "opto/mathexactnode.hpp"
52 #include "opto/memnode.hpp"
53 #include "opto/mulnode.hpp"
54 #include "opto/node.hpp"
55 #include "opto/opcodes.hpp"
56 #include "opto/output.hpp"
57 #include "opto/parse.hpp"
58 #include "opto/phaseX.hpp"
59 #include "opto/rootnode.hpp"
60 #include "opto/runtime.hpp"
61 #include "opto/stringopts.hpp"
62 #include "opto/type.hpp"
63 #include "opto/vectornode.hpp"
64 #include "runtime/arguments.hpp"
65 #include "runtime/signature.hpp"
66 #include "runtime/stubRoutines.hpp"
67 #include "runtime/timer.hpp"
68 #include "trace/tracing.hpp"
69 #include "utilities/copy.hpp"
70 #if defined AD_MD_HPP
71 # include AD_MD_HPP
72 #elif defined TARGET_ARCH_MODEL_x86_32
73 # include "adfiles/ad_x86_32.hpp"
74 #elif defined TARGET_ARCH_MODEL_x86_64
75 # include "adfiles/ad_x86_64.hpp"
76 #elif defined TARGET_ARCH_MODEL_sparc
77 # include "adfiles/ad_sparc.hpp"
78 #elif defined TARGET_ARCH_MODEL_zero
79 # include "adfiles/ad_zero.hpp"
80 #elif defined TARGET_ARCH_MODEL_ppc_64
81 # include "adfiles/ad_ppc_64.hpp"
82 #endif
85 // -------------------- Compile::mach_constant_base_node -----------------------
86 // Constant table base node singleton.
87 MachConstantBaseNode* Compile::mach_constant_base_node() {
88 if (_mach_constant_base_node == NULL) {
89 _mach_constant_base_node = new (C) MachConstantBaseNode();
90 _mach_constant_base_node->add_req(C->root());
91 }
92 return _mach_constant_base_node;
93 }
96 /// Support for intrinsics.
98 // Return the index at which m must be inserted (or already exists).
99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
100 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
101 #ifdef ASSERT
102 for (int i = 1; i < _intrinsics->length(); i++) {
103 CallGenerator* cg1 = _intrinsics->at(i-1);
104 CallGenerator* cg2 = _intrinsics->at(i);
105 assert(cg1->method() != cg2->method()
106 ? cg1->method() < cg2->method()
107 : cg1->is_virtual() < cg2->is_virtual(),
108 "compiler intrinsics list must stay sorted");
109 }
110 #endif
111 // Binary search sorted list, in decreasing intervals [lo, hi].
112 int lo = 0, hi = _intrinsics->length()-1;
113 while (lo <= hi) {
114 int mid = (uint)(hi + lo) / 2;
115 ciMethod* mid_m = _intrinsics->at(mid)->method();
116 if (m < mid_m) {
117 hi = mid-1;
118 } else if (m > mid_m) {
119 lo = mid+1;
120 } else {
121 // look at minor sort key
122 bool mid_virt = _intrinsics->at(mid)->is_virtual();
123 if (is_virtual < mid_virt) {
124 hi = mid-1;
125 } else if (is_virtual > mid_virt) {
126 lo = mid+1;
127 } else {
128 return mid; // exact match
129 }
130 }
131 }
132 return lo; // inexact match
133 }
135 void Compile::register_intrinsic(CallGenerator* cg) {
136 if (_intrinsics == NULL) {
137 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
138 }
139 // This code is stolen from ciObjectFactory::insert.
140 // Really, GrowableArray should have methods for
141 // insert_at, remove_at, and binary_search.
142 int len = _intrinsics->length();
143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
144 if (index == len) {
145 _intrinsics->append(cg);
146 } else {
147 #ifdef ASSERT
148 CallGenerator* oldcg = _intrinsics->at(index);
149 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
150 #endif
151 _intrinsics->append(_intrinsics->at(len-1));
152 int pos;
153 for (pos = len-2; pos >= index; pos--) {
154 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
155 }
156 _intrinsics->at_put(index, cg);
157 }
158 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
159 }
161 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
162 assert(m->is_loaded(), "don't try this on unloaded methods");
163 if (_intrinsics != NULL) {
164 int index = intrinsic_insertion_index(m, is_virtual);
165 if (index < _intrinsics->length()
166 && _intrinsics->at(index)->method() == m
167 && _intrinsics->at(index)->is_virtual() == is_virtual) {
168 return _intrinsics->at(index);
169 }
170 }
171 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
172 if (m->intrinsic_id() != vmIntrinsics::_none &&
173 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
174 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
175 if (cg != NULL) {
176 // Save it for next time:
177 register_intrinsic(cg);
178 return cg;
179 } else {
180 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
181 }
182 }
183 return NULL;
184 }
186 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
187 // in library_call.cpp.
190 #ifndef PRODUCT
191 // statistics gathering...
193 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
194 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
196 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
197 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
198 int oflags = _intrinsic_hist_flags[id];
199 assert(flags != 0, "what happened?");
200 if (is_virtual) {
201 flags |= _intrinsic_virtual;
202 }
203 bool changed = (flags != oflags);
204 if ((flags & _intrinsic_worked) != 0) {
205 juint count = (_intrinsic_hist_count[id] += 1);
206 if (count == 1) {
207 changed = true; // first time
208 }
209 // increment the overall count also:
210 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
211 }
212 if (changed) {
213 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
214 // Something changed about the intrinsic's virtuality.
215 if ((flags & _intrinsic_virtual) != 0) {
216 // This is the first use of this intrinsic as a virtual call.
217 if (oflags != 0) {
218 // We already saw it as a non-virtual, so note both cases.
219 flags |= _intrinsic_both;
220 }
221 } else if ((oflags & _intrinsic_both) == 0) {
222 // This is the first use of this intrinsic as a non-virtual
223 flags |= _intrinsic_both;
224 }
225 }
226 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
227 }
228 // update the overall flags also:
229 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
230 return changed;
231 }
233 static char* format_flags(int flags, char* buf) {
234 buf[0] = 0;
235 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
236 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
237 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
238 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
239 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
240 if (buf[0] == 0) strcat(buf, ",");
241 assert(buf[0] == ',', "must be");
242 return &buf[1];
243 }
245 void Compile::print_intrinsic_statistics() {
246 char flagsbuf[100];
247 ttyLocker ttyl;
248 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
249 tty->print_cr("Compiler intrinsic usage:");
250 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
251 if (total == 0) total = 1; // avoid div0 in case of no successes
252 #define PRINT_STAT_LINE(name, c, f) \
253 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
254 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
255 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
256 int flags = _intrinsic_hist_flags[id];
257 juint count = _intrinsic_hist_count[id];
258 if ((flags | count) != 0) {
259 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
260 }
261 }
262 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
263 if (xtty != NULL) xtty->tail("statistics");
264 }
266 void Compile::print_statistics() {
267 { ttyLocker ttyl;
268 if (xtty != NULL) xtty->head("statistics type='opto'");
269 Parse::print_statistics();
270 PhaseCCP::print_statistics();
271 PhaseRegAlloc::print_statistics();
272 Scheduling::print_statistics();
273 PhasePeephole::print_statistics();
274 PhaseIdealLoop::print_statistics();
275 if (xtty != NULL) xtty->tail("statistics");
276 }
277 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
278 // put this under its own <statistics> element.
279 print_intrinsic_statistics();
280 }
281 }
282 #endif //PRODUCT
284 // Support for bundling info
285 Bundle* Compile::node_bundling(const Node *n) {
286 assert(valid_bundle_info(n), "oob");
287 return &_node_bundling_base[n->_idx];
288 }
290 bool Compile::valid_bundle_info(const Node *n) {
291 return (_node_bundling_limit > n->_idx);
292 }
295 void Compile::gvn_replace_by(Node* n, Node* nn) {
296 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
297 Node* use = n->last_out(i);
298 bool is_in_table = initial_gvn()->hash_delete(use);
299 uint uses_found = 0;
300 for (uint j = 0; j < use->len(); j++) {
301 if (use->in(j) == n) {
302 if (j < use->req())
303 use->set_req(j, nn);
304 else
305 use->set_prec(j, nn);
306 uses_found++;
307 }
308 }
309 if (is_in_table) {
310 // reinsert into table
311 initial_gvn()->hash_find_insert(use);
312 }
313 record_for_igvn(use);
314 i -= uses_found; // we deleted 1 or more copies of this edge
315 }
316 }
319 static inline bool not_a_node(const Node* n) {
320 if (n == NULL) return true;
321 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
322 if (*(address*)n == badAddress) return true; // kill by Node::destruct
323 return false;
324 }
326 // Identify all nodes that are reachable from below, useful.
327 // Use breadth-first pass that records state in a Unique_Node_List,
328 // recursive traversal is slower.
329 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
330 int estimated_worklist_size = live_nodes();
331 useful.map( estimated_worklist_size, NULL ); // preallocate space
333 // Initialize worklist
334 if (root() != NULL) { useful.push(root()); }
335 // If 'top' is cached, declare it useful to preserve cached node
336 if( cached_top_node() ) { useful.push(cached_top_node()); }
338 // Push all useful nodes onto the list, breadthfirst
339 for( uint next = 0; next < useful.size(); ++next ) {
340 assert( next < unique(), "Unique useful nodes < total nodes");
341 Node *n = useful.at(next);
342 uint max = n->len();
343 for( uint i = 0; i < max; ++i ) {
344 Node *m = n->in(i);
345 if (not_a_node(m)) continue;
346 useful.push(m);
347 }
348 }
349 }
351 // Update dead_node_list with any missing dead nodes using useful
352 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
353 void Compile::update_dead_node_list(Unique_Node_List &useful) {
354 uint max_idx = unique();
355 VectorSet& useful_node_set = useful.member_set();
357 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
358 // If node with index node_idx is not in useful set,
359 // mark it as dead in dead node list.
360 if (! useful_node_set.test(node_idx) ) {
361 record_dead_node(node_idx);
362 }
363 }
364 }
366 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
367 int shift = 0;
368 for (int i = 0; i < inlines->length(); i++) {
369 CallGenerator* cg = inlines->at(i);
370 CallNode* call = cg->call_node();
371 if (shift > 0) {
372 inlines->at_put(i-shift, cg);
373 }
374 if (!useful.member(call)) {
375 shift++;
376 }
377 }
378 inlines->trunc_to(inlines->length()-shift);
379 }
381 // Disconnect all useless nodes by disconnecting those at the boundary.
382 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
383 uint next = 0;
384 while (next < useful.size()) {
385 Node *n = useful.at(next++);
386 if (n->is_SafePoint()) {
387 // We're done with a parsing phase. Replaced nodes are not valid
388 // beyond that point.
389 n->as_SafePoint()->delete_replaced_nodes();
390 }
391 // Use raw traversal of out edges since this code removes out edges
392 int max = n->outcnt();
393 for (int j = 0; j < max; ++j) {
394 Node* child = n->raw_out(j);
395 if (! useful.member(child)) {
396 assert(!child->is_top() || child != top(),
397 "If top is cached in Compile object it is in useful list");
398 // Only need to remove this out-edge to the useless node
399 n->raw_del_out(j);
400 --j;
401 --max;
402 }
403 }
404 if (n->outcnt() == 1 && n->has_special_unique_user()) {
405 record_for_igvn(n->unique_out());
406 }
407 }
408 // Remove useless macro and predicate opaq nodes
409 for (int i = C->macro_count()-1; i >= 0; i--) {
410 Node* n = C->macro_node(i);
411 if (!useful.member(n)) {
412 remove_macro_node(n);
413 }
414 }
415 // Remove useless CastII nodes with range check dependency
416 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
417 Node* cast = range_check_cast_node(i);
418 if (!useful.member(cast)) {
419 remove_range_check_cast(cast);
420 }
421 }
422 // Remove useless expensive node
423 for (int i = C->expensive_count()-1; i >= 0; i--) {
424 Node* n = C->expensive_node(i);
425 if (!useful.member(n)) {
426 remove_expensive_node(n);
427 }
428 }
429 // clean up the late inline lists
430 remove_useless_late_inlines(&_string_late_inlines, useful);
431 remove_useless_late_inlines(&_boxing_late_inlines, useful);
432 remove_useless_late_inlines(&_late_inlines, useful);
433 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
434 }
436 //------------------------------frame_size_in_words-----------------------------
437 // frame_slots in units of words
438 int Compile::frame_size_in_words() const {
439 // shift is 0 in LP32 and 1 in LP64
440 const int shift = (LogBytesPerWord - LogBytesPerInt);
441 int words = _frame_slots >> shift;
442 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
443 return words;
444 }
446 // To bang the stack of this compiled method we use the stack size
447 // that the interpreter would need in case of a deoptimization. This
448 // removes the need to bang the stack in the deoptimization blob which
449 // in turn simplifies stack overflow handling.
450 int Compile::bang_size_in_bytes() const {
451 return MAX2(_interpreter_frame_size, frame_size_in_bytes());
452 }
454 // ============================================================================
455 //------------------------------CompileWrapper---------------------------------
456 class CompileWrapper : public StackObj {
457 Compile *const _compile;
458 public:
459 CompileWrapper(Compile* compile);
461 ~CompileWrapper();
462 };
464 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
465 // the Compile* pointer is stored in the current ciEnv:
466 ciEnv* env = compile->env();
467 assert(env == ciEnv::current(), "must already be a ciEnv active");
468 assert(env->compiler_data() == NULL, "compile already active?");
469 env->set_compiler_data(compile);
470 assert(compile == Compile::current(), "sanity");
472 compile->set_type_dict(NULL);
473 compile->set_type_hwm(NULL);
474 compile->set_type_last_size(0);
475 compile->set_last_tf(NULL, NULL);
476 compile->set_indexSet_arena(NULL);
477 compile->set_indexSet_free_block_list(NULL);
478 compile->init_type_arena();
479 Type::Initialize(compile);
480 _compile->set_scratch_buffer_blob(NULL);
481 _compile->begin_method();
482 }
483 CompileWrapper::~CompileWrapper() {
484 _compile->end_method();
485 if (_compile->scratch_buffer_blob() != NULL)
486 BufferBlob::free(_compile->scratch_buffer_blob());
487 _compile->env()->set_compiler_data(NULL);
488 }
491 //----------------------------print_compile_messages---------------------------
492 void Compile::print_compile_messages() {
493 #ifndef PRODUCT
494 // Check if recompiling
495 if (_subsume_loads == false && PrintOpto) {
496 // Recompiling without allowing machine instructions to subsume loads
497 tty->print_cr("*********************************************************");
498 tty->print_cr("** Bailout: Recompile without subsuming loads **");
499 tty->print_cr("*********************************************************");
500 }
501 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
502 // Recompiling without escape analysis
503 tty->print_cr("*********************************************************");
504 tty->print_cr("** Bailout: Recompile without escape analysis **");
505 tty->print_cr("*********************************************************");
506 }
507 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
508 // Recompiling without boxing elimination
509 tty->print_cr("*********************************************************");
510 tty->print_cr("** Bailout: Recompile without boxing elimination **");
511 tty->print_cr("*********************************************************");
512 }
513 if (env()->break_at_compile()) {
514 // Open the debugger when compiling this method.
515 tty->print("### Breaking when compiling: ");
516 method()->print_short_name();
517 tty->cr();
518 BREAKPOINT;
519 }
521 if( PrintOpto ) {
522 if (is_osr_compilation()) {
523 tty->print("[OSR]%3d", _compile_id);
524 } else {
525 tty->print("%3d", _compile_id);
526 }
527 }
528 #endif
529 }
532 //-----------------------init_scratch_buffer_blob------------------------------
533 // Construct a temporary BufferBlob and cache it for this compile.
534 void Compile::init_scratch_buffer_blob(int const_size) {
535 // If there is already a scratch buffer blob allocated and the
536 // constant section is big enough, use it. Otherwise free the
537 // current and allocate a new one.
538 BufferBlob* blob = scratch_buffer_blob();
539 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
540 // Use the current blob.
541 } else {
542 if (blob != NULL) {
543 BufferBlob::free(blob);
544 }
546 ResourceMark rm;
547 _scratch_const_size = const_size;
548 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
549 blob = BufferBlob::create("Compile::scratch_buffer", size);
550 // Record the buffer blob for next time.
551 set_scratch_buffer_blob(blob);
552 // Have we run out of code space?
553 if (scratch_buffer_blob() == NULL) {
554 // Let CompilerBroker disable further compilations.
555 record_failure("Not enough space for scratch buffer in CodeCache");
556 return;
557 }
558 }
560 // Initialize the relocation buffers
561 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
562 set_scratch_locs_memory(locs_buf);
563 }
566 //-----------------------scratch_emit_size-------------------------------------
567 // Helper function that computes size by emitting code
568 uint Compile::scratch_emit_size(const Node* n) {
569 // Start scratch_emit_size section.
570 set_in_scratch_emit_size(true);
572 // Emit into a trash buffer and count bytes emitted.
573 // This is a pretty expensive way to compute a size,
574 // but it works well enough if seldom used.
575 // All common fixed-size instructions are given a size
576 // method by the AD file.
577 // Note that the scratch buffer blob and locs memory are
578 // allocated at the beginning of the compile task, and
579 // may be shared by several calls to scratch_emit_size.
580 // The allocation of the scratch buffer blob is particularly
581 // expensive, since it has to grab the code cache lock.
582 BufferBlob* blob = this->scratch_buffer_blob();
583 assert(blob != NULL, "Initialize BufferBlob at start");
584 assert(blob->size() > MAX_inst_size, "sanity");
585 relocInfo* locs_buf = scratch_locs_memory();
586 address blob_begin = blob->content_begin();
587 address blob_end = (address)locs_buf;
588 assert(blob->content_contains(blob_end), "sanity");
589 CodeBuffer buf(blob_begin, blob_end - blob_begin);
590 buf.initialize_consts_size(_scratch_const_size);
591 buf.initialize_stubs_size(MAX_stubs_size);
592 assert(locs_buf != NULL, "sanity");
593 int lsize = MAX_locs_size / 3;
594 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
595 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
596 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
598 // Do the emission.
600 Label fakeL; // Fake label for branch instructions.
601 Label* saveL = NULL;
602 uint save_bnum = 0;
603 bool is_branch = n->is_MachBranch();
604 if (is_branch) {
605 MacroAssembler masm(&buf);
606 masm.bind(fakeL);
607 n->as_MachBranch()->save_label(&saveL, &save_bnum);
608 n->as_MachBranch()->label_set(&fakeL, 0);
609 }
610 n->emit(buf, this->regalloc());
612 // Emitting into the scratch buffer should not fail
613 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason()));
615 if (is_branch) // Restore label.
616 n->as_MachBranch()->label_set(saveL, save_bnum);
618 // End scratch_emit_size section.
619 set_in_scratch_emit_size(false);
621 return buf.insts_size();
622 }
625 // ============================================================================
626 //------------------------------Compile standard-------------------------------
627 debug_only( int Compile::_debug_idx = 100000; )
629 // Compile a method. entry_bci is -1 for normal compilations and indicates
630 // the continuation bci for on stack replacement.
633 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
634 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
635 : Phase(Compiler),
636 _env(ci_env),
637 _log(ci_env->log()),
638 _compile_id(ci_env->compile_id()),
639 _save_argument_registers(false),
640 _stub_name(NULL),
641 _stub_function(NULL),
642 _stub_entry_point(NULL),
643 _method(target),
644 _entry_bci(osr_bci),
645 _initial_gvn(NULL),
646 _for_igvn(NULL),
647 _warm_calls(NULL),
648 _subsume_loads(subsume_loads),
649 _do_escape_analysis(do_escape_analysis),
650 _eliminate_boxing(eliminate_boxing),
651 _failure_reason(NULL),
652 _code_buffer("Compile::Fill_buffer"),
653 _orig_pc_slot(0),
654 _orig_pc_slot_offset_in_bytes(0),
655 _has_method_handle_invokes(false),
656 _mach_constant_base_node(NULL),
657 _node_bundling_limit(0),
658 _node_bundling_base(NULL),
659 _java_calls(0),
660 _inner_loops(0),
661 _scratch_const_size(-1),
662 _in_scratch_emit_size(false),
663 _dead_node_list(comp_arena()),
664 _dead_node_count(0),
665 #ifndef PRODUCT
666 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
667 _in_dump_cnt(0),
668 _printer(IdealGraphPrinter::printer()),
669 #endif
670 _congraph(NULL),
671 _comp_arena(mtCompiler),
672 _node_arena(mtCompiler),
673 _old_arena(mtCompiler),
674 _Compile_types(mtCompiler),
675 _replay_inline_data(NULL),
676 _late_inlines(comp_arena(), 2, 0, NULL),
677 _string_late_inlines(comp_arena(), 2, 0, NULL),
678 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
679 _late_inlines_pos(0),
680 _number_of_mh_late_inlines(0),
681 _inlining_progress(false),
682 _inlining_incrementally(false),
683 _print_inlining_list(NULL),
684 _print_inlining_idx(0),
685 _interpreter_frame_size(0),
686 _max_node_limit(MaxNodeLimit) {
687 C = this;
689 CompileWrapper cw(this);
690 #ifndef PRODUCT
691 if (TimeCompiler2) {
692 tty->print(" ");
693 target->holder()->name()->print();
694 tty->print(".");
695 target->print_short_name();
696 tty->print(" ");
697 }
698 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
699 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
700 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
701 if (!print_opto_assembly) {
702 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
703 if (print_assembly && !Disassembler::can_decode()) {
704 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
705 print_opto_assembly = true;
706 }
707 }
708 set_print_assembly(print_opto_assembly);
709 set_parsed_irreducible_loop(false);
711 if (method()->has_option("ReplayInline")) {
712 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
713 }
714 #endif
715 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
716 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
717 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
719 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
720 // Make sure the method being compiled gets its own MDO,
721 // so we can at least track the decompile_count().
722 // Need MDO to record RTM code generation state.
723 method()->ensure_method_data();
724 }
726 Init(::AliasLevel);
729 print_compile_messages();
731 _ilt = InlineTree::build_inline_tree_root();
733 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
734 assert(num_alias_types() >= AliasIdxRaw, "");
736 #define MINIMUM_NODE_HASH 1023
737 // Node list that Iterative GVN will start with
738 Unique_Node_List for_igvn(comp_arena());
739 set_for_igvn(&for_igvn);
741 // GVN that will be run immediately on new nodes
742 uint estimated_size = method()->code_size()*4+64;
743 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
744 PhaseGVN gvn(node_arena(), estimated_size);
745 set_initial_gvn(&gvn);
747 if (print_inlining() || print_intrinsics()) {
748 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
749 }
750 { // Scope for timing the parser
751 TracePhase t3("parse", &_t_parser, true);
753 // Put top into the hash table ASAP.
754 initial_gvn()->transform_no_reclaim(top());
756 // Set up tf(), start(), and find a CallGenerator.
757 CallGenerator* cg = NULL;
758 if (is_osr_compilation()) {
759 const TypeTuple *domain = StartOSRNode::osr_domain();
760 const TypeTuple *range = TypeTuple::make_range(method()->signature());
761 init_tf(TypeFunc::make(domain, range));
762 StartNode* s = new (this) StartOSRNode(root(), domain);
763 initial_gvn()->set_type_bottom(s);
764 init_start(s);
765 cg = CallGenerator::for_osr(method(), entry_bci());
766 } else {
767 // Normal case.
768 init_tf(TypeFunc::make(method()));
769 StartNode* s = new (this) StartNode(root(), tf()->domain());
770 initial_gvn()->set_type_bottom(s);
771 init_start(s);
772 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
773 // With java.lang.ref.reference.get() we must go through the
774 // intrinsic when G1 is enabled - even when get() is the root
775 // method of the compile - so that, if necessary, the value in
776 // the referent field of the reference object gets recorded by
777 // the pre-barrier code.
778 // Specifically, if G1 is enabled, the value in the referent
779 // field is recorded by the G1 SATB pre barrier. This will
780 // result in the referent being marked live and the reference
781 // object removed from the list of discovered references during
782 // reference processing.
783 cg = find_intrinsic(method(), false);
784 }
785 if (cg == NULL) {
786 float past_uses = method()->interpreter_invocation_count();
787 float expected_uses = past_uses;
788 cg = CallGenerator::for_inline(method(), expected_uses);
789 }
790 }
791 if (failing()) return;
792 if (cg == NULL) {
793 record_method_not_compilable_all_tiers("cannot parse method");
794 return;
795 }
796 JVMState* jvms = build_start_state(start(), tf());
797 if ((jvms = cg->generate(jvms)) == NULL) {
798 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
799 record_method_not_compilable("method parse failed");
800 }
801 return;
802 }
803 GraphKit kit(jvms);
805 if (!kit.stopped()) {
806 // Accept return values, and transfer control we know not where.
807 // This is done by a special, unique ReturnNode bound to root.
808 return_values(kit.jvms());
809 }
811 if (kit.has_exceptions()) {
812 // Any exceptions that escape from this call must be rethrown
813 // to whatever caller is dynamically above us on the stack.
814 // This is done by a special, unique RethrowNode bound to root.
815 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
816 }
818 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
820 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
821 inline_string_calls(true);
822 }
824 if (failing()) return;
826 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
828 // Remove clutter produced by parsing.
829 if (!failing()) {
830 ResourceMark rm;
831 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
832 }
833 }
835 // Note: Large methods are capped off in do_one_bytecode().
836 if (failing()) return;
838 // After parsing, node notes are no longer automagic.
839 // They must be propagated by register_new_node_with_optimizer(),
840 // clone(), or the like.
841 set_default_node_notes(NULL);
843 for (;;) {
844 int successes = Inline_Warm();
845 if (failing()) return;
846 if (successes == 0) break;
847 }
849 // Drain the list.
850 Finish_Warm();
851 #ifndef PRODUCT
852 if (_printer) {
853 _printer->print_inlining(this);
854 }
855 #endif
857 if (failing()) return;
858 NOT_PRODUCT( verify_graph_edges(); )
860 // Now optimize
861 Optimize();
862 if (failing()) return;
863 NOT_PRODUCT( verify_graph_edges(); )
865 #ifndef PRODUCT
866 if (PrintIdeal) {
867 ttyLocker ttyl; // keep the following output all in one block
868 // This output goes directly to the tty, not the compiler log.
869 // To enable tools to match it up with the compilation activity,
870 // be sure to tag this tty output with the compile ID.
871 if (xtty != NULL) {
872 xtty->head("ideal compile_id='%d'%s", compile_id(),
873 is_osr_compilation() ? " compile_kind='osr'" :
874 "");
875 }
876 root()->dump(9999);
877 if (xtty != NULL) {
878 xtty->tail("ideal");
879 }
880 }
881 #endif
883 NOT_PRODUCT( verify_barriers(); )
885 // Dump compilation data to replay it.
886 if (method()->has_option("DumpReplay")) {
887 env()->dump_replay_data(_compile_id);
888 }
889 if (method()->has_option("DumpInline") && (ilt() != NULL)) {
890 env()->dump_inline_data(_compile_id);
891 }
893 // Now that we know the size of all the monitors we can add a fixed slot
894 // for the original deopt pc.
896 _orig_pc_slot = fixed_slots();
897 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
898 set_fixed_slots(next_slot);
900 // Compute when to use implicit null checks. Used by matching trap based
901 // nodes and NullCheck optimization.
902 set_allowed_deopt_reasons();
904 // Now generate code
905 Code_Gen();
906 if (failing()) return;
908 // Check if we want to skip execution of all compiled code.
909 {
910 #ifndef PRODUCT
911 if (OptoNoExecute) {
912 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
913 return;
914 }
915 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
916 #endif
918 if (is_osr_compilation()) {
919 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
920 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
921 } else {
922 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
923 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
924 }
926 env()->register_method(_method, _entry_bci,
927 &_code_offsets,
928 _orig_pc_slot_offset_in_bytes,
929 code_buffer(),
930 frame_size_in_words(), _oop_map_set,
931 &_handler_table, &_inc_table,
932 compiler,
933 env()->comp_level(),
934 has_unsafe_access(),
935 SharedRuntime::is_wide_vector(max_vector_size()),
936 rtm_state()
937 );
939 if (log() != NULL) // Print code cache state into compiler log
940 log()->code_cache_state();
941 }
942 }
944 //------------------------------Compile----------------------------------------
945 // Compile a runtime stub
946 Compile::Compile( ciEnv* ci_env,
947 TypeFunc_generator generator,
948 address stub_function,
949 const char *stub_name,
950 int is_fancy_jump,
951 bool pass_tls,
952 bool save_arg_registers,
953 bool return_pc )
954 : Phase(Compiler),
955 _env(ci_env),
956 _log(ci_env->log()),
957 _compile_id(0),
958 _save_argument_registers(save_arg_registers),
959 _method(NULL),
960 _stub_name(stub_name),
961 _stub_function(stub_function),
962 _stub_entry_point(NULL),
963 _entry_bci(InvocationEntryBci),
964 _initial_gvn(NULL),
965 _for_igvn(NULL),
966 _warm_calls(NULL),
967 _orig_pc_slot(0),
968 _orig_pc_slot_offset_in_bytes(0),
969 _subsume_loads(true),
970 _do_escape_analysis(false),
971 _eliminate_boxing(false),
972 _failure_reason(NULL),
973 _code_buffer("Compile::Fill_buffer"),
974 _has_method_handle_invokes(false),
975 _mach_constant_base_node(NULL),
976 _node_bundling_limit(0),
977 _node_bundling_base(NULL),
978 _java_calls(0),
979 _inner_loops(0),
980 #ifndef PRODUCT
981 _trace_opto_output(TraceOptoOutput),
982 _in_dump_cnt(0),
983 _printer(NULL),
984 #endif
985 _comp_arena(mtCompiler),
986 _node_arena(mtCompiler),
987 _old_arena(mtCompiler),
988 _Compile_types(mtCompiler),
989 _dead_node_list(comp_arena()),
990 _dead_node_count(0),
991 _congraph(NULL),
992 _replay_inline_data(NULL),
993 _number_of_mh_late_inlines(0),
994 _inlining_progress(false),
995 _inlining_incrementally(false),
996 _print_inlining_list(NULL),
997 _print_inlining_idx(0),
998 _allowed_reasons(0),
999 _interpreter_frame_size(0),
1000 _max_node_limit(MaxNodeLimit) {
1001 C = this;
1003 #ifndef PRODUCT
1004 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
1005 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
1006 set_print_assembly(PrintFrameConverterAssembly);
1007 set_parsed_irreducible_loop(false);
1008 #endif
1009 set_has_irreducible_loop(false); // no loops
1011 CompileWrapper cw(this);
1012 Init(/*AliasLevel=*/ 0);
1013 init_tf((*generator)());
1015 {
1016 // The following is a dummy for the sake of GraphKit::gen_stub
1017 Unique_Node_List for_igvn(comp_arena());
1018 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1019 PhaseGVN gvn(Thread::current()->resource_area(),255);
1020 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1021 gvn.transform_no_reclaim(top());
1023 GraphKit kit;
1024 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1025 }
1027 NOT_PRODUCT( verify_graph_edges(); )
1028 Code_Gen();
1029 if (failing()) return;
1032 // Entry point will be accessed using compile->stub_entry_point();
1033 if (code_buffer() == NULL) {
1034 Matcher::soft_match_failure();
1035 } else {
1036 if (PrintAssembly && (WizardMode || Verbose))
1037 tty->print_cr("### Stub::%s", stub_name);
1039 if (!failing()) {
1040 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1042 // Make the NMethod
1043 // For now we mark the frame as never safe for profile stackwalking
1044 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1045 code_buffer(),
1046 CodeOffsets::frame_never_safe,
1047 // _code_offsets.value(CodeOffsets::Frame_Complete),
1048 frame_size_in_words(),
1049 _oop_map_set,
1050 save_arg_registers);
1051 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1053 _stub_entry_point = rs->entry_point();
1054 }
1055 }
1056 }
1058 //------------------------------Init-------------------------------------------
1059 // Prepare for a single compilation
1060 void Compile::Init(int aliaslevel) {
1061 _unique = 0;
1062 _regalloc = NULL;
1064 _tf = NULL; // filled in later
1065 _top = NULL; // cached later
1066 _matcher = NULL; // filled in later
1067 _cfg = NULL; // filled in later
1069 set_24_bit_selection_and_mode(Use24BitFP, false);
1071 _node_note_array = NULL;
1072 _default_node_notes = NULL;
1074 _immutable_memory = NULL; // filled in at first inquiry
1076 // Globally visible Nodes
1077 // First set TOP to NULL to give safe behavior during creation of RootNode
1078 set_cached_top_node(NULL);
1079 set_root(new (this) RootNode());
1080 // Now that you have a Root to point to, create the real TOP
1081 set_cached_top_node( new (this) ConNode(Type::TOP) );
1082 set_recent_alloc(NULL, NULL);
1084 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1085 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1086 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1087 env()->set_dependencies(new Dependencies(env()));
1089 _fixed_slots = 0;
1090 set_has_split_ifs(false);
1091 set_has_loops(has_method() && method()->has_loops()); // first approximation
1092 set_has_stringbuilder(false);
1093 set_has_boxed_value(false);
1094 _trap_can_recompile = false; // no traps emitted yet
1095 _major_progress = true; // start out assuming good things will happen
1096 set_has_unsafe_access(false);
1097 set_max_vector_size(0);
1098 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1099 set_decompile_count(0);
1101 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1102 set_num_loop_opts(LoopOptsCount);
1103 set_do_inlining(Inline);
1104 set_max_inline_size(MaxInlineSize);
1105 set_freq_inline_size(FreqInlineSize);
1106 set_do_scheduling(OptoScheduling);
1107 set_do_count_invocations(false);
1108 set_do_method_data_update(false);
1109 set_rtm_state(NoRTM); // No RTM lock eliding by default
1110 method_has_option_value("MaxNodeLimit", _max_node_limit);
1111 #if INCLUDE_RTM_OPT
1112 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1113 int rtm_state = method()->method_data()->rtm_state();
1114 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1115 // Don't generate RTM lock eliding code.
1116 set_rtm_state(NoRTM);
1117 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1118 // Generate RTM lock eliding code without abort ratio calculation code.
1119 set_rtm_state(UseRTM);
1120 } else if (UseRTMDeopt) {
1121 // Generate RTM lock eliding code and include abort ratio calculation
1122 // code if UseRTMDeopt is on.
1123 set_rtm_state(ProfileRTM);
1124 }
1125 }
1126 #endif
1127 if (debug_info()->recording_non_safepoints()) {
1128 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1129 (comp_arena(), 8, 0, NULL));
1130 set_default_node_notes(Node_Notes::make(this));
1131 }
1133 // // -- Initialize types before each compile --
1134 // // Update cached type information
1135 // if( _method && _method->constants() )
1136 // Type::update_loaded_types(_method, _method->constants());
1138 // Init alias_type map.
1139 if (!_do_escape_analysis && aliaslevel == 3)
1140 aliaslevel = 2; // No unique types without escape analysis
1141 _AliasLevel = aliaslevel;
1142 const int grow_ats = 16;
1143 _max_alias_types = grow_ats;
1144 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1145 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1146 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1147 {
1148 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1149 }
1150 // Initialize the first few types.
1151 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1152 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1153 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1154 _num_alias_types = AliasIdxRaw+1;
1155 // Zero out the alias type cache.
1156 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1157 // A NULL adr_type hits in the cache right away. Preload the right answer.
1158 probe_alias_cache(NULL)->_index = AliasIdxTop;
1160 _intrinsics = NULL;
1161 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1162 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1163 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1164 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1165 register_library_intrinsics();
1166 }
1168 //---------------------------init_start----------------------------------------
1169 // Install the StartNode on this compile object.
1170 void Compile::init_start(StartNode* s) {
1171 if (failing())
1172 return; // already failing
1173 assert(s == start(), "");
1174 }
1176 StartNode* Compile::start() const {
1177 assert(!failing(), "");
1178 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1179 Node* start = root()->fast_out(i);
1180 if( start->is_Start() )
1181 return start->as_Start();
1182 }
1183 fatal("Did not find Start node!");
1184 return NULL;
1185 }
1187 //-------------------------------immutable_memory-------------------------------------
1188 // Access immutable memory
1189 Node* Compile::immutable_memory() {
1190 if (_immutable_memory != NULL) {
1191 return _immutable_memory;
1192 }
1193 StartNode* s = start();
1194 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1195 Node *p = s->fast_out(i);
1196 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1197 _immutable_memory = p;
1198 return _immutable_memory;
1199 }
1200 }
1201 ShouldNotReachHere();
1202 return NULL;
1203 }
1205 //----------------------set_cached_top_node------------------------------------
1206 // Install the cached top node, and make sure Node::is_top works correctly.
1207 void Compile::set_cached_top_node(Node* tn) {
1208 if (tn != NULL) verify_top(tn);
1209 Node* old_top = _top;
1210 _top = tn;
1211 // Calling Node::setup_is_top allows the nodes the chance to adjust
1212 // their _out arrays.
1213 if (_top != NULL) _top->setup_is_top();
1214 if (old_top != NULL) old_top->setup_is_top();
1215 assert(_top == NULL || top()->is_top(), "");
1216 }
1218 #ifdef ASSERT
1219 uint Compile::count_live_nodes_by_graph_walk() {
1220 Unique_Node_List useful(comp_arena());
1221 // Get useful node list by walking the graph.
1222 identify_useful_nodes(useful);
1223 return useful.size();
1224 }
1226 void Compile::print_missing_nodes() {
1228 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1229 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1230 return;
1231 }
1233 // This is an expensive function. It is executed only when the user
1234 // specifies VerifyIdealNodeCount option or otherwise knows the
1235 // additional work that needs to be done to identify reachable nodes
1236 // by walking the flow graph and find the missing ones using
1237 // _dead_node_list.
1239 Unique_Node_List useful(comp_arena());
1240 // Get useful node list by walking the graph.
1241 identify_useful_nodes(useful);
1243 uint l_nodes = C->live_nodes();
1244 uint l_nodes_by_walk = useful.size();
1246 if (l_nodes != l_nodes_by_walk) {
1247 if (_log != NULL) {
1248 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1249 _log->stamp();
1250 _log->end_head();
1251 }
1252 VectorSet& useful_member_set = useful.member_set();
1253 int last_idx = l_nodes_by_walk;
1254 for (int i = 0; i < last_idx; i++) {
1255 if (useful_member_set.test(i)) {
1256 if (_dead_node_list.test(i)) {
1257 if (_log != NULL) {
1258 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1259 }
1260 if (PrintIdealNodeCount) {
1261 // Print the log message to tty
1262 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1263 useful.at(i)->dump();
1264 }
1265 }
1266 }
1267 else if (! _dead_node_list.test(i)) {
1268 if (_log != NULL) {
1269 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1270 }
1271 if (PrintIdealNodeCount) {
1272 // Print the log message to tty
1273 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1274 }
1275 }
1276 }
1277 if (_log != NULL) {
1278 _log->tail("mismatched_nodes");
1279 }
1280 }
1281 }
1282 #endif
1284 #ifndef PRODUCT
1285 void Compile::verify_top(Node* tn) const {
1286 if (tn != NULL) {
1287 assert(tn->is_Con(), "top node must be a constant");
1288 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1289 assert(tn->in(0) != NULL, "must have live top node");
1290 }
1291 }
1292 #endif
1295 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1297 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1298 guarantee(arr != NULL, "");
1299 int num_blocks = arr->length();
1300 if (grow_by < num_blocks) grow_by = num_blocks;
1301 int num_notes = grow_by * _node_notes_block_size;
1302 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1303 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1304 while (num_notes > 0) {
1305 arr->append(notes);
1306 notes += _node_notes_block_size;
1307 num_notes -= _node_notes_block_size;
1308 }
1309 assert(num_notes == 0, "exact multiple, please");
1310 }
1312 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1313 if (source == NULL || dest == NULL) return false;
1315 if (dest->is_Con())
1316 return false; // Do not push debug info onto constants.
1318 #ifdef ASSERT
1319 // Leave a bread crumb trail pointing to the original node:
1320 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1321 dest->set_debug_orig(source);
1322 }
1323 #endif
1325 if (node_note_array() == NULL)
1326 return false; // Not collecting any notes now.
1328 // This is a copy onto a pre-existing node, which may already have notes.
1329 // If both nodes have notes, do not overwrite any pre-existing notes.
1330 Node_Notes* source_notes = node_notes_at(source->_idx);
1331 if (source_notes == NULL || source_notes->is_clear()) return false;
1332 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1333 if (dest_notes == NULL || dest_notes->is_clear()) {
1334 return set_node_notes_at(dest->_idx, source_notes);
1335 }
1337 Node_Notes merged_notes = (*source_notes);
1338 // The order of operations here ensures that dest notes will win...
1339 merged_notes.update_from(dest_notes);
1340 return set_node_notes_at(dest->_idx, &merged_notes);
1341 }
1344 //--------------------------allow_range_check_smearing-------------------------
1345 // Gating condition for coalescing similar range checks.
1346 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1347 // single covering check that is at least as strong as any of them.
1348 // If the optimization succeeds, the simplified (strengthened) range check
1349 // will always succeed. If it fails, we will deopt, and then give up
1350 // on the optimization.
1351 bool Compile::allow_range_check_smearing() const {
1352 // If this method has already thrown a range-check,
1353 // assume it was because we already tried range smearing
1354 // and it failed.
1355 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1356 return !already_trapped;
1357 }
1360 //------------------------------flatten_alias_type-----------------------------
1361 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1362 int offset = tj->offset();
1363 TypePtr::PTR ptr = tj->ptr();
1365 // Known instance (scalarizable allocation) alias only with itself.
1366 bool is_known_inst = tj->isa_oopptr() != NULL &&
1367 tj->is_oopptr()->is_known_instance();
1369 // Process weird unsafe references.
1370 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1371 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1372 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1373 tj = TypeOopPtr::BOTTOM;
1374 ptr = tj->ptr();
1375 offset = tj->offset();
1376 }
1378 // Array pointers need some flattening
1379 const TypeAryPtr *ta = tj->isa_aryptr();
1380 if (ta && ta->is_stable()) {
1381 // Erase stability property for alias analysis.
1382 tj = ta = ta->cast_to_stable(false);
1383 }
1384 if( ta && is_known_inst ) {
1385 if ( offset != Type::OffsetBot &&
1386 offset > arrayOopDesc::length_offset_in_bytes() ) {
1387 offset = Type::OffsetBot; // Flatten constant access into array body only
1388 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1389 }
1390 } else if( ta && _AliasLevel >= 2 ) {
1391 // For arrays indexed by constant indices, we flatten the alias
1392 // space to include all of the array body. Only the header, klass
1393 // and array length can be accessed un-aliased.
1394 if( offset != Type::OffsetBot ) {
1395 if( ta->const_oop() ) { // MethodData* or Method*
1396 offset = Type::OffsetBot; // Flatten constant access into array body
1397 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1398 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1399 // range is OK as-is.
1400 tj = ta = TypeAryPtr::RANGE;
1401 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1402 tj = TypeInstPtr::KLASS; // all klass loads look alike
1403 ta = TypeAryPtr::RANGE; // generic ignored junk
1404 ptr = TypePtr::BotPTR;
1405 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1406 tj = TypeInstPtr::MARK;
1407 ta = TypeAryPtr::RANGE; // generic ignored junk
1408 ptr = TypePtr::BotPTR;
1409 } else { // Random constant offset into array body
1410 offset = Type::OffsetBot; // Flatten constant access into array body
1411 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1412 }
1413 }
1414 // Arrays of fixed size alias with arrays of unknown size.
1415 if (ta->size() != TypeInt::POS) {
1416 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1417 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1418 }
1419 // Arrays of known objects become arrays of unknown objects.
1420 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1421 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1422 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1423 }
1424 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1425 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1426 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1427 }
1428 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1429 // cannot be distinguished by bytecode alone.
1430 if (ta->elem() == TypeInt::BOOL) {
1431 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1432 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1433 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1434 }
1435 // During the 2nd round of IterGVN, NotNull castings are removed.
1436 // Make sure the Bottom and NotNull variants alias the same.
1437 // Also, make sure exact and non-exact variants alias the same.
1438 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1439 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1440 }
1441 }
1443 // Oop pointers need some flattening
1444 const TypeInstPtr *to = tj->isa_instptr();
1445 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1446 ciInstanceKlass *k = to->klass()->as_instance_klass();
1447 if( ptr == TypePtr::Constant ) {
1448 if (to->klass() != ciEnv::current()->Class_klass() ||
1449 offset < k->size_helper() * wordSize) {
1450 // No constant oop pointers (such as Strings); they alias with
1451 // unknown strings.
1452 assert(!is_known_inst, "not scalarizable allocation");
1453 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1454 }
1455 } else if( is_known_inst ) {
1456 tj = to; // Keep NotNull and klass_is_exact for instance type
1457 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1458 // During the 2nd round of IterGVN, NotNull castings are removed.
1459 // Make sure the Bottom and NotNull variants alias the same.
1460 // Also, make sure exact and non-exact variants alias the same.
1461 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1462 }
1463 if (to->speculative() != NULL) {
1464 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1465 }
1466 // Canonicalize the holder of this field
1467 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1468 // First handle header references such as a LoadKlassNode, even if the
1469 // object's klass is unloaded at compile time (4965979).
1470 if (!is_known_inst) { // Do it only for non-instance types
1471 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1472 }
1473 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1474 // Static fields are in the space above the normal instance
1475 // fields in the java.lang.Class instance.
1476 if (to->klass() != ciEnv::current()->Class_klass()) {
1477 to = NULL;
1478 tj = TypeOopPtr::BOTTOM;
1479 offset = tj->offset();
1480 }
1481 } else {
1482 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1483 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1484 if( is_known_inst ) {
1485 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1486 } else {
1487 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1488 }
1489 }
1490 }
1491 }
1493 // Klass pointers to object array klasses need some flattening
1494 const TypeKlassPtr *tk = tj->isa_klassptr();
1495 if( tk ) {
1496 // If we are referencing a field within a Klass, we need
1497 // to assume the worst case of an Object. Both exact and
1498 // inexact types must flatten to the same alias class so
1499 // use NotNull as the PTR.
1500 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1502 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1503 TypeKlassPtr::OBJECT->klass(),
1504 offset);
1505 }
1507 ciKlass* klass = tk->klass();
1508 if( klass->is_obj_array_klass() ) {
1509 ciKlass* k = TypeAryPtr::OOPS->klass();
1510 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1511 k = TypeInstPtr::BOTTOM->klass();
1512 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1513 }
1515 // Check for precise loads from the primary supertype array and force them
1516 // to the supertype cache alias index. Check for generic array loads from
1517 // the primary supertype array and also force them to the supertype cache
1518 // alias index. Since the same load can reach both, we need to merge
1519 // these 2 disparate memories into the same alias class. Since the
1520 // primary supertype array is read-only, there's no chance of confusion
1521 // where we bypass an array load and an array store.
1522 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1523 if (offset == Type::OffsetBot ||
1524 (offset >= primary_supers_offset &&
1525 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1526 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1527 offset = in_bytes(Klass::secondary_super_cache_offset());
1528 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1529 }
1530 }
1532 // Flatten all Raw pointers together.
1533 if (tj->base() == Type::RawPtr)
1534 tj = TypeRawPtr::BOTTOM;
1536 if (tj->base() == Type::AnyPtr)
1537 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1539 // Flatten all to bottom for now
1540 switch( _AliasLevel ) {
1541 case 0:
1542 tj = TypePtr::BOTTOM;
1543 break;
1544 case 1: // Flatten to: oop, static, field or array
1545 switch (tj->base()) {
1546 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1547 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1548 case Type::AryPtr: // do not distinguish arrays at all
1549 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1550 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1551 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1552 default: ShouldNotReachHere();
1553 }
1554 break;
1555 case 2: // No collapsing at level 2; keep all splits
1556 case 3: // No collapsing at level 3; keep all splits
1557 break;
1558 default:
1559 Unimplemented();
1560 }
1562 offset = tj->offset();
1563 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1565 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1566 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1567 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1568 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1569 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1570 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1571 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1572 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1573 assert( tj->ptr() != TypePtr::TopPTR &&
1574 tj->ptr() != TypePtr::AnyNull &&
1575 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1576 // assert( tj->ptr() != TypePtr::Constant ||
1577 // tj->base() == Type::RawPtr ||
1578 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1580 return tj;
1581 }
1583 void Compile::AliasType::Init(int i, const TypePtr* at) {
1584 _index = i;
1585 _adr_type = at;
1586 _field = NULL;
1587 _element = NULL;
1588 _is_rewritable = true; // default
1589 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1590 if (atoop != NULL && atoop->is_known_instance()) {
1591 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1592 _general_index = Compile::current()->get_alias_index(gt);
1593 } else {
1594 _general_index = 0;
1595 }
1596 }
1598 BasicType Compile::AliasType::basic_type() const {
1599 if (element() != NULL) {
1600 const Type* element = adr_type()->is_aryptr()->elem();
1601 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1602 } if (field() != NULL) {
1603 return field()->layout_type();
1604 } else {
1605 return T_ILLEGAL; // unknown
1606 }
1607 }
1609 //---------------------------------print_on------------------------------------
1610 #ifndef PRODUCT
1611 void Compile::AliasType::print_on(outputStream* st) {
1612 if (index() < 10)
1613 st->print("@ <%d> ", index());
1614 else st->print("@ <%d>", index());
1615 st->print(is_rewritable() ? " " : " RO");
1616 int offset = adr_type()->offset();
1617 if (offset == Type::OffsetBot)
1618 st->print(" +any");
1619 else st->print(" +%-3d", offset);
1620 st->print(" in ");
1621 adr_type()->dump_on(st);
1622 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1623 if (field() != NULL && tjp) {
1624 if (tjp->klass() != field()->holder() ||
1625 tjp->offset() != field()->offset_in_bytes()) {
1626 st->print(" != ");
1627 field()->print();
1628 st->print(" ***");
1629 }
1630 }
1631 }
1633 void print_alias_types() {
1634 Compile* C = Compile::current();
1635 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1636 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1637 C->alias_type(idx)->print_on(tty);
1638 tty->cr();
1639 }
1640 }
1641 #endif
1644 //----------------------------probe_alias_cache--------------------------------
1645 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1646 intptr_t key = (intptr_t) adr_type;
1647 key ^= key >> logAliasCacheSize;
1648 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1649 }
1652 //-----------------------------grow_alias_types--------------------------------
1653 void Compile::grow_alias_types() {
1654 const int old_ats = _max_alias_types; // how many before?
1655 const int new_ats = old_ats; // how many more?
1656 const int grow_ats = old_ats+new_ats; // how many now?
1657 _max_alias_types = grow_ats;
1658 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1659 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1660 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1661 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1662 }
1665 //--------------------------------find_alias_type------------------------------
1666 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1667 if (_AliasLevel == 0)
1668 return alias_type(AliasIdxBot);
1670 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1671 if (ace->_adr_type == adr_type) {
1672 return alias_type(ace->_index);
1673 }
1675 // Handle special cases.
1676 if (adr_type == NULL) return alias_type(AliasIdxTop);
1677 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1679 // Do it the slow way.
1680 const TypePtr* flat = flatten_alias_type(adr_type);
1682 #ifdef ASSERT
1683 {
1684 ResourceMark rm;
1685 assert(flat == flatten_alias_type(flat),
1686 err_msg("not idempotent: adr_type = %s; flat = %s => %s", Type::str(adr_type),
1687 Type::str(flat), Type::str(flatten_alias_type(flat))));
1688 assert(flat != TypePtr::BOTTOM,
1689 err_msg("cannot alias-analyze an untyped ptr: adr_type = %s", Type::str(adr_type)));
1690 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1691 const TypeOopPtr* foop = flat->is_oopptr();
1692 // Scalarizable allocations have exact klass always.
1693 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1694 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1695 assert(foop == flatten_alias_type(xoop),
1696 err_msg("exactness must not affect alias type: foop = %s; xoop = %s",
1697 Type::str(foop), Type::str(xoop)));
1698 }
1699 }
1700 #endif
1702 int idx = AliasIdxTop;
1703 for (int i = 0; i < num_alias_types(); i++) {
1704 if (alias_type(i)->adr_type() == flat) {
1705 idx = i;
1706 break;
1707 }
1708 }
1710 if (idx == AliasIdxTop) {
1711 if (no_create) return NULL;
1712 // Grow the array if necessary.
1713 if (_num_alias_types == _max_alias_types) grow_alias_types();
1714 // Add a new alias type.
1715 idx = _num_alias_types++;
1716 _alias_types[idx]->Init(idx, flat);
1717 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1718 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1719 if (flat->isa_instptr()) {
1720 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1721 && flat->is_instptr()->klass() == env()->Class_klass())
1722 alias_type(idx)->set_rewritable(false);
1723 }
1724 if (flat->isa_aryptr()) {
1725 #ifdef ASSERT
1726 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1727 // (T_BYTE has the weakest alignment and size restrictions...)
1728 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1729 #endif
1730 if (flat->offset() == TypePtr::OffsetBot) {
1731 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1732 }
1733 }
1734 if (flat->isa_klassptr()) {
1735 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1736 alias_type(idx)->set_rewritable(false);
1737 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1738 alias_type(idx)->set_rewritable(false);
1739 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1740 alias_type(idx)->set_rewritable(false);
1741 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1742 alias_type(idx)->set_rewritable(false);
1743 }
1744 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1745 // but the base pointer type is not distinctive enough to identify
1746 // references into JavaThread.)
1748 // Check for final fields.
1749 const TypeInstPtr* tinst = flat->isa_instptr();
1750 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1751 ciField* field;
1752 if (tinst->const_oop() != NULL &&
1753 tinst->klass() == ciEnv::current()->Class_klass() &&
1754 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1755 // static field
1756 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1757 field = k->get_field_by_offset(tinst->offset(), true);
1758 } else {
1759 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1760 field = k->get_field_by_offset(tinst->offset(), false);
1761 }
1762 assert(field == NULL ||
1763 original_field == NULL ||
1764 (field->holder() == original_field->holder() &&
1765 field->offset() == original_field->offset() &&
1766 field->is_static() == original_field->is_static()), "wrong field?");
1767 // Set field() and is_rewritable() attributes.
1768 if (field != NULL) alias_type(idx)->set_field(field);
1769 }
1770 }
1772 // Fill the cache for next time.
1773 ace->_adr_type = adr_type;
1774 ace->_index = idx;
1775 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1777 // Might as well try to fill the cache for the flattened version, too.
1778 AliasCacheEntry* face = probe_alias_cache(flat);
1779 if (face->_adr_type == NULL) {
1780 face->_adr_type = flat;
1781 face->_index = idx;
1782 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1783 }
1785 return alias_type(idx);
1786 }
1789 Compile::AliasType* Compile::alias_type(ciField* field) {
1790 const TypeOopPtr* t;
1791 if (field->is_static())
1792 t = TypeInstPtr::make(field->holder()->java_mirror());
1793 else
1794 t = TypeOopPtr::make_from_klass_raw(field->holder());
1795 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1796 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1797 return atp;
1798 }
1801 //------------------------------have_alias_type--------------------------------
1802 bool Compile::have_alias_type(const TypePtr* adr_type) {
1803 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1804 if (ace->_adr_type == adr_type) {
1805 return true;
1806 }
1808 // Handle special cases.
1809 if (adr_type == NULL) return true;
1810 if (adr_type == TypePtr::BOTTOM) return true;
1812 return find_alias_type(adr_type, true, NULL) != NULL;
1813 }
1815 //-----------------------------must_alias--------------------------------------
1816 // True if all values of the given address type are in the given alias category.
1817 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1818 if (alias_idx == AliasIdxBot) return true; // the universal category
1819 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1820 if (alias_idx == AliasIdxTop) return false; // the empty category
1821 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1823 // the only remaining possible overlap is identity
1824 int adr_idx = get_alias_index(adr_type);
1825 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1826 assert(adr_idx == alias_idx ||
1827 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1828 && adr_type != TypeOopPtr::BOTTOM),
1829 "should not be testing for overlap with an unsafe pointer");
1830 return adr_idx == alias_idx;
1831 }
1833 //------------------------------can_alias--------------------------------------
1834 // True if any values of the given address type are in the given alias category.
1835 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1836 if (alias_idx == AliasIdxTop) return false; // the empty category
1837 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1838 if (alias_idx == AliasIdxBot) return true; // the universal category
1839 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1841 // the only remaining possible overlap is identity
1842 int adr_idx = get_alias_index(adr_type);
1843 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1844 return adr_idx == alias_idx;
1845 }
1849 //---------------------------pop_warm_call-------------------------------------
1850 WarmCallInfo* Compile::pop_warm_call() {
1851 WarmCallInfo* wci = _warm_calls;
1852 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1853 return wci;
1854 }
1856 //----------------------------Inline_Warm--------------------------------------
1857 int Compile::Inline_Warm() {
1858 // If there is room, try to inline some more warm call sites.
1859 // %%% Do a graph index compaction pass when we think we're out of space?
1860 if (!InlineWarmCalls) return 0;
1862 int calls_made_hot = 0;
1863 int room_to_grow = NodeCountInliningCutoff - unique();
1864 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1865 int amount_grown = 0;
1866 WarmCallInfo* call;
1867 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1868 int est_size = (int)call->size();
1869 if (est_size > (room_to_grow - amount_grown)) {
1870 // This one won't fit anyway. Get rid of it.
1871 call->make_cold();
1872 continue;
1873 }
1874 call->make_hot();
1875 calls_made_hot++;
1876 amount_grown += est_size;
1877 amount_to_grow -= est_size;
1878 }
1880 if (calls_made_hot > 0) set_major_progress();
1881 return calls_made_hot;
1882 }
1885 //----------------------------Finish_Warm--------------------------------------
1886 void Compile::Finish_Warm() {
1887 if (!InlineWarmCalls) return;
1888 if (failing()) return;
1889 if (warm_calls() == NULL) return;
1891 // Clean up loose ends, if we are out of space for inlining.
1892 WarmCallInfo* call;
1893 while ((call = pop_warm_call()) != NULL) {
1894 call->make_cold();
1895 }
1896 }
1898 //---------------------cleanup_loop_predicates-----------------------
1899 // Remove the opaque nodes that protect the predicates so that all unused
1900 // checks and uncommon_traps will be eliminated from the ideal graph
1901 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1902 if (predicate_count()==0) return;
1903 for (int i = predicate_count(); i > 0; i--) {
1904 Node * n = predicate_opaque1_node(i-1);
1905 assert(n->Opcode() == Op_Opaque1, "must be");
1906 igvn.replace_node(n, n->in(1));
1907 }
1908 assert(predicate_count()==0, "should be clean!");
1909 }
1911 void Compile::add_range_check_cast(Node* n) {
1912 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1913 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1914 _range_check_casts->append(n);
1915 }
1917 // Remove all range check dependent CastIINodes.
1918 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1919 for (int i = range_check_cast_count(); i > 0; i--) {
1920 Node* cast = range_check_cast_node(i-1);
1921 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1922 igvn.replace_node(cast, cast->in(1));
1923 }
1924 assert(range_check_cast_count() == 0, "should be empty");
1925 }
1927 // StringOpts and late inlining of string methods
1928 void Compile::inline_string_calls(bool parse_time) {
1929 {
1930 // remove useless nodes to make the usage analysis simpler
1931 ResourceMark rm;
1932 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1933 }
1935 {
1936 ResourceMark rm;
1937 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1938 PhaseStringOpts pso(initial_gvn(), for_igvn());
1939 print_method(PHASE_AFTER_STRINGOPTS, 3);
1940 }
1942 // now inline anything that we skipped the first time around
1943 if (!parse_time) {
1944 _late_inlines_pos = _late_inlines.length();
1945 }
1947 while (_string_late_inlines.length() > 0) {
1948 CallGenerator* cg = _string_late_inlines.pop();
1949 cg->do_late_inline();
1950 if (failing()) return;
1951 }
1952 _string_late_inlines.trunc_to(0);
1953 }
1955 // Late inlining of boxing methods
1956 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1957 if (_boxing_late_inlines.length() > 0) {
1958 assert(has_boxed_value(), "inconsistent");
1960 PhaseGVN* gvn = initial_gvn();
1961 set_inlining_incrementally(true);
1963 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1964 for_igvn()->clear();
1965 gvn->replace_with(&igvn);
1967 _late_inlines_pos = _late_inlines.length();
1969 while (_boxing_late_inlines.length() > 0) {
1970 CallGenerator* cg = _boxing_late_inlines.pop();
1971 cg->do_late_inline();
1972 if (failing()) return;
1973 }
1974 _boxing_late_inlines.trunc_to(0);
1976 {
1977 ResourceMark rm;
1978 PhaseRemoveUseless pru(gvn, for_igvn());
1979 }
1981 igvn = PhaseIterGVN(gvn);
1982 igvn.optimize();
1984 set_inlining_progress(false);
1985 set_inlining_incrementally(false);
1986 }
1987 }
1989 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1990 assert(IncrementalInline, "incremental inlining should be on");
1991 PhaseGVN* gvn = initial_gvn();
1993 set_inlining_progress(false);
1994 for_igvn()->clear();
1995 gvn->replace_with(&igvn);
1997 int i = 0;
1999 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2000 CallGenerator* cg = _late_inlines.at(i);
2001 _late_inlines_pos = i+1;
2002 cg->do_late_inline();
2003 if (failing()) return;
2004 }
2005 int j = 0;
2006 for (; i < _late_inlines.length(); i++, j++) {
2007 _late_inlines.at_put(j, _late_inlines.at(i));
2008 }
2009 _late_inlines.trunc_to(j);
2011 {
2012 ResourceMark rm;
2013 PhaseRemoveUseless pru(gvn, for_igvn());
2014 }
2016 igvn = PhaseIterGVN(gvn);
2017 }
2019 // Perform incremental inlining until bound on number of live nodes is reached
2020 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2021 PhaseGVN* gvn = initial_gvn();
2023 set_inlining_incrementally(true);
2024 set_inlining_progress(true);
2025 uint low_live_nodes = 0;
2027 while(inlining_progress() && _late_inlines.length() > 0) {
2029 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2030 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2031 // PhaseIdealLoop is expensive so we only try it once we are
2032 // out of live nodes and we only try it again if the previous
2033 // helped got the number of nodes down significantly
2034 PhaseIdealLoop ideal_loop( igvn, false, true );
2035 if (failing()) return;
2036 low_live_nodes = live_nodes();
2037 _major_progress = true;
2038 }
2040 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2041 break;
2042 }
2043 }
2045 inline_incrementally_one(igvn);
2047 if (failing()) return;
2049 igvn.optimize();
2051 if (failing()) return;
2052 }
2054 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2056 if (_string_late_inlines.length() > 0) {
2057 assert(has_stringbuilder(), "inconsistent");
2058 for_igvn()->clear();
2059 initial_gvn()->replace_with(&igvn);
2061 inline_string_calls(false);
2063 if (failing()) return;
2065 {
2066 ResourceMark rm;
2067 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2068 }
2070 igvn = PhaseIterGVN(gvn);
2072 igvn.optimize();
2073 }
2075 set_inlining_incrementally(false);
2076 }
2079 //------------------------------Optimize---------------------------------------
2080 // Given a graph, optimize it.
2081 void Compile::Optimize() {
2082 TracePhase t1("optimizer", &_t_optimizer, true);
2084 #ifndef PRODUCT
2085 if (env()->break_at_compile()) {
2086 BREAKPOINT;
2087 }
2089 #endif
2091 ResourceMark rm;
2092 int loop_opts_cnt;
2094 NOT_PRODUCT( verify_graph_edges(); )
2096 print_method(PHASE_AFTER_PARSING);
2098 {
2099 // Iterative Global Value Numbering, including ideal transforms
2100 // Initialize IterGVN with types and values from parse-time GVN
2101 PhaseIterGVN igvn(initial_gvn());
2102 {
2103 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2104 igvn.optimize();
2105 }
2107 print_method(PHASE_ITER_GVN1, 2);
2109 if (failing()) return;
2111 {
2112 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2113 inline_incrementally(igvn);
2114 }
2116 print_method(PHASE_INCREMENTAL_INLINE, 2);
2118 if (failing()) return;
2120 if (eliminate_boxing()) {
2121 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2122 // Inline valueOf() methods now.
2123 inline_boxing_calls(igvn);
2125 if (AlwaysIncrementalInline) {
2126 inline_incrementally(igvn);
2127 }
2129 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2131 if (failing()) return;
2132 }
2134 // Remove the speculative part of types and clean up the graph from
2135 // the extra CastPP nodes whose only purpose is to carry them. Do
2136 // that early so that optimizations are not disrupted by the extra
2137 // CastPP nodes.
2138 remove_speculative_types(igvn);
2140 // No more new expensive nodes will be added to the list from here
2141 // so keep only the actual candidates for optimizations.
2142 cleanup_expensive_nodes(igvn);
2144 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2145 NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);)
2146 initial_gvn()->replace_with(&igvn);
2147 for_igvn()->clear();
2148 Unique_Node_List new_worklist(C->comp_arena());
2149 {
2150 ResourceMark rm;
2151 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2152 }
2153 set_for_igvn(&new_worklist);
2154 igvn = PhaseIterGVN(initial_gvn());
2155 igvn.optimize();
2156 }
2158 // Perform escape analysis
2159 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2160 if (has_loops()) {
2161 // Cleanup graph (remove dead nodes).
2162 TracePhase t2("idealLoop", &_t_idealLoop, true);
2163 PhaseIdealLoop ideal_loop( igvn, false, true );
2164 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2165 if (failing()) return;
2166 }
2167 ConnectionGraph::do_analysis(this, &igvn);
2169 if (failing()) return;
2171 // Optimize out fields loads from scalar replaceable allocations.
2172 igvn.optimize();
2173 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2175 if (failing()) return;
2177 if (congraph() != NULL && macro_count() > 0) {
2178 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2179 PhaseMacroExpand mexp(igvn);
2180 mexp.eliminate_macro_nodes();
2181 igvn.set_delay_transform(false);
2183 igvn.optimize();
2184 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2186 if (failing()) return;
2187 }
2188 }
2190 // Loop transforms on the ideal graph. Range Check Elimination,
2191 // peeling, unrolling, etc.
2193 // Set loop opts counter
2194 loop_opts_cnt = num_loop_opts();
2195 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2196 {
2197 TracePhase t2("idealLoop", &_t_idealLoop, true);
2198 PhaseIdealLoop ideal_loop( igvn, true );
2199 loop_opts_cnt--;
2200 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2201 if (failing()) return;
2202 }
2203 // Loop opts pass if partial peeling occurred in previous pass
2204 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2205 TracePhase t3("idealLoop", &_t_idealLoop, true);
2206 PhaseIdealLoop ideal_loop( igvn, false );
2207 loop_opts_cnt--;
2208 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2209 if (failing()) return;
2210 }
2211 // Loop opts pass for loop-unrolling before CCP
2212 if(major_progress() && (loop_opts_cnt > 0)) {
2213 TracePhase t4("idealLoop", &_t_idealLoop, true);
2214 PhaseIdealLoop ideal_loop( igvn, false );
2215 loop_opts_cnt--;
2216 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2217 }
2218 if (!failing()) {
2219 // Verify that last round of loop opts produced a valid graph
2220 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2221 PhaseIdealLoop::verify(igvn);
2222 }
2223 }
2224 if (failing()) return;
2226 // Conditional Constant Propagation;
2227 PhaseCCP ccp( &igvn );
2228 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2229 {
2230 TracePhase t2("ccp", &_t_ccp, true);
2231 ccp.do_transform();
2232 }
2233 print_method(PHASE_CPP1, 2);
2235 assert( true, "Break here to ccp.dump_old2new_map()");
2237 // Iterative Global Value Numbering, including ideal transforms
2238 {
2239 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2240 igvn = ccp;
2241 igvn.optimize();
2242 }
2244 print_method(PHASE_ITER_GVN2, 2);
2246 if (failing()) return;
2248 // Loop transforms on the ideal graph. Range Check Elimination,
2249 // peeling, unrolling, etc.
2250 if(loop_opts_cnt > 0) {
2251 debug_only( int cnt = 0; );
2252 while(major_progress() && (loop_opts_cnt > 0)) {
2253 TracePhase t2("idealLoop", &_t_idealLoop, true);
2254 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2255 PhaseIdealLoop ideal_loop( igvn, true);
2256 loop_opts_cnt--;
2257 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2258 if (failing()) return;
2259 }
2260 }
2262 {
2263 // Verify that all previous optimizations produced a valid graph
2264 // at least to this point, even if no loop optimizations were done.
2265 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2266 PhaseIdealLoop::verify(igvn);
2267 }
2269 if (range_check_cast_count() > 0) {
2270 // No more loop optimizations. Remove all range check dependent CastIINodes.
2271 C->remove_range_check_casts(igvn);
2272 igvn.optimize();
2273 }
2275 {
2276 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2277 PhaseMacroExpand mex(igvn);
2278 if (mex.expand_macro_nodes()) {
2279 assert(failing(), "must bail out w/ explicit message");
2280 return;
2281 }
2282 }
2284 } // (End scope of igvn; run destructor if necessary for asserts.)
2286 dump_inlining();
2287 // A method with only infinite loops has no edges entering loops from root
2288 {
2289 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2290 if (final_graph_reshaping()) {
2291 assert(failing(), "must bail out w/ explicit message");
2292 return;
2293 }
2294 }
2296 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2297 }
2300 //------------------------------Code_Gen---------------------------------------
2301 // Given a graph, generate code for it
2302 void Compile::Code_Gen() {
2303 if (failing()) {
2304 return;
2305 }
2307 // Perform instruction selection. You might think we could reclaim Matcher
2308 // memory PDQ, but actually the Matcher is used in generating spill code.
2309 // Internals of the Matcher (including some VectorSets) must remain live
2310 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2311 // set a bit in reclaimed memory.
2313 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2314 // nodes. Mapping is only valid at the root of each matched subtree.
2315 NOT_PRODUCT( verify_graph_edges(); )
2317 Matcher matcher;
2318 _matcher = &matcher;
2319 {
2320 TracePhase t2("matcher", &_t_matcher, true);
2321 matcher.match();
2322 }
2323 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2324 // nodes. Mapping is only valid at the root of each matched subtree.
2325 NOT_PRODUCT( verify_graph_edges(); )
2327 // If you have too many nodes, or if matching has failed, bail out
2328 check_node_count(0, "out of nodes matching instructions");
2329 if (failing()) {
2330 return;
2331 }
2333 // Build a proper-looking CFG
2334 PhaseCFG cfg(node_arena(), root(), matcher);
2335 _cfg = &cfg;
2336 {
2337 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2338 bool success = cfg.do_global_code_motion();
2339 if (!success) {
2340 return;
2341 }
2343 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2344 NOT_PRODUCT( verify_graph_edges(); )
2345 debug_only( cfg.verify(); )
2346 }
2348 PhaseChaitin regalloc(unique(), cfg, matcher);
2349 _regalloc = ®alloc;
2350 {
2351 TracePhase t2("regalloc", &_t_registerAllocation, true);
2352 // Perform register allocation. After Chaitin, use-def chains are
2353 // no longer accurate (at spill code) and so must be ignored.
2354 // Node->LRG->reg mappings are still accurate.
2355 _regalloc->Register_Allocate();
2357 // Bail out if the allocator builds too many nodes
2358 if (failing()) {
2359 return;
2360 }
2361 }
2363 // Prior to register allocation we kept empty basic blocks in case the
2364 // the allocator needed a place to spill. After register allocation we
2365 // are not adding any new instructions. If any basic block is empty, we
2366 // can now safely remove it.
2367 {
2368 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2369 cfg.remove_empty_blocks();
2370 if (do_freq_based_layout()) {
2371 PhaseBlockLayout layout(cfg);
2372 } else {
2373 cfg.set_loop_alignment();
2374 }
2375 cfg.fixup_flow();
2376 }
2378 // Apply peephole optimizations
2379 if( OptoPeephole ) {
2380 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2381 PhasePeephole peep( _regalloc, cfg);
2382 peep.do_transform();
2383 }
2385 // Do late expand if CPU requires this.
2386 if (Matcher::require_postalloc_expand) {
2387 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2388 cfg.postalloc_expand(_regalloc);
2389 }
2391 // Convert Nodes to instruction bits in a buffer
2392 {
2393 // %%%% workspace merge brought two timers together for one job
2394 TracePhase t2a("output", &_t_output, true);
2395 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2396 Output();
2397 }
2399 print_method(PHASE_FINAL_CODE);
2401 // He's dead, Jim.
2402 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2403 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2404 }
2407 //------------------------------dump_asm---------------------------------------
2408 // Dump formatted assembly
2409 #ifndef PRODUCT
2410 void Compile::dump_asm(int *pcs, uint pc_limit) {
2411 bool cut_short = false;
2412 tty->print_cr("#");
2413 tty->print("# "); _tf->dump(); tty->cr();
2414 tty->print_cr("#");
2416 // For all blocks
2417 int pc = 0x0; // Program counter
2418 char starts_bundle = ' ';
2419 _regalloc->dump_frame();
2421 Node *n = NULL;
2422 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2423 if (VMThread::should_terminate()) {
2424 cut_short = true;
2425 break;
2426 }
2427 Block* block = _cfg->get_block(i);
2428 if (block->is_connector() && !Verbose) {
2429 continue;
2430 }
2431 n = block->head();
2432 if (pcs && n->_idx < pc_limit) {
2433 tty->print("%3.3x ", pcs[n->_idx]);
2434 } else {
2435 tty->print(" ");
2436 }
2437 block->dump_head(_cfg);
2438 if (block->is_connector()) {
2439 tty->print_cr(" # Empty connector block");
2440 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2441 tty->print_cr(" # Block is sole successor of call");
2442 }
2444 // For all instructions
2445 Node *delay = NULL;
2446 for (uint j = 0; j < block->number_of_nodes(); j++) {
2447 if (VMThread::should_terminate()) {
2448 cut_short = true;
2449 break;
2450 }
2451 n = block->get_node(j);
2452 if (valid_bundle_info(n)) {
2453 Bundle* bundle = node_bundling(n);
2454 if (bundle->used_in_unconditional_delay()) {
2455 delay = n;
2456 continue;
2457 }
2458 if (bundle->starts_bundle()) {
2459 starts_bundle = '+';
2460 }
2461 }
2463 if (WizardMode) {
2464 n->dump();
2465 }
2467 if( !n->is_Region() && // Dont print in the Assembly
2468 !n->is_Phi() && // a few noisely useless nodes
2469 !n->is_Proj() &&
2470 !n->is_MachTemp() &&
2471 !n->is_SafePointScalarObject() &&
2472 !n->is_Catch() && // Would be nice to print exception table targets
2473 !n->is_MergeMem() && // Not very interesting
2474 !n->is_top() && // Debug info table constants
2475 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2476 ) {
2477 if (pcs && n->_idx < pc_limit)
2478 tty->print("%3.3x", pcs[n->_idx]);
2479 else
2480 tty->print(" ");
2481 tty->print(" %c ", starts_bundle);
2482 starts_bundle = ' ';
2483 tty->print("\t");
2484 n->format(_regalloc, tty);
2485 tty->cr();
2486 }
2488 // If we have an instruction with a delay slot, and have seen a delay,
2489 // then back up and print it
2490 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2491 assert(delay != NULL, "no unconditional delay instruction");
2492 if (WizardMode) delay->dump();
2494 if (node_bundling(delay)->starts_bundle())
2495 starts_bundle = '+';
2496 if (pcs && n->_idx < pc_limit)
2497 tty->print("%3.3x", pcs[n->_idx]);
2498 else
2499 tty->print(" ");
2500 tty->print(" %c ", starts_bundle);
2501 starts_bundle = ' ';
2502 tty->print("\t");
2503 delay->format(_regalloc, tty);
2504 tty->cr();
2505 delay = NULL;
2506 }
2508 // Dump the exception table as well
2509 if( n->is_Catch() && (Verbose || WizardMode) ) {
2510 // Print the exception table for this offset
2511 _handler_table.print_subtable_for(pc);
2512 }
2513 }
2515 if (pcs && n->_idx < pc_limit)
2516 tty->print_cr("%3.3x", pcs[n->_idx]);
2517 else
2518 tty->cr();
2520 assert(cut_short || delay == NULL, "no unconditional delay branch");
2522 } // End of per-block dump
2523 tty->cr();
2525 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2526 }
2527 #endif
2529 //------------------------------Final_Reshape_Counts---------------------------
2530 // This class defines counters to help identify when a method
2531 // may/must be executed using hardware with only 24-bit precision.
2532 struct Final_Reshape_Counts : public StackObj {
2533 int _call_count; // count non-inlined 'common' calls
2534 int _float_count; // count float ops requiring 24-bit precision
2535 int _double_count; // count double ops requiring more precision
2536 int _java_call_count; // count non-inlined 'java' calls
2537 int _inner_loop_count; // count loops which need alignment
2538 VectorSet _visited; // Visitation flags
2539 Node_List _tests; // Set of IfNodes & PCTableNodes
2541 Final_Reshape_Counts() :
2542 _call_count(0), _float_count(0), _double_count(0),
2543 _java_call_count(0), _inner_loop_count(0),
2544 _visited( Thread::current()->resource_area() ) { }
2546 void inc_call_count () { _call_count ++; }
2547 void inc_float_count () { _float_count ++; }
2548 void inc_double_count() { _double_count++; }
2549 void inc_java_call_count() { _java_call_count++; }
2550 void inc_inner_loop_count() { _inner_loop_count++; }
2552 int get_call_count () const { return _call_count ; }
2553 int get_float_count () const { return _float_count ; }
2554 int get_double_count() const { return _double_count; }
2555 int get_java_call_count() const { return _java_call_count; }
2556 int get_inner_loop_count() const { return _inner_loop_count; }
2557 };
2559 #ifdef ASSERT
2560 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2561 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2562 // Make sure the offset goes inside the instance layout.
2563 return k->contains_field_offset(tp->offset());
2564 // Note that OffsetBot and OffsetTop are very negative.
2565 }
2566 #endif
2568 // Eliminate trivially redundant StoreCMs and accumulate their
2569 // precedence edges.
2570 void Compile::eliminate_redundant_card_marks(Node* n) {
2571 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2572 if (n->in(MemNode::Address)->outcnt() > 1) {
2573 // There are multiple users of the same address so it might be
2574 // possible to eliminate some of the StoreCMs
2575 Node* mem = n->in(MemNode::Memory);
2576 Node* adr = n->in(MemNode::Address);
2577 Node* val = n->in(MemNode::ValueIn);
2578 Node* prev = n;
2579 bool done = false;
2580 // Walk the chain of StoreCMs eliminating ones that match. As
2581 // long as it's a chain of single users then the optimization is
2582 // safe. Eliminating partially redundant StoreCMs would require
2583 // cloning copies down the other paths.
2584 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2585 if (adr == mem->in(MemNode::Address) &&
2586 val == mem->in(MemNode::ValueIn)) {
2587 // redundant StoreCM
2588 if (mem->req() > MemNode::OopStore) {
2589 // Hasn't been processed by this code yet.
2590 n->add_prec(mem->in(MemNode::OopStore));
2591 } else {
2592 // Already converted to precedence edge
2593 for (uint i = mem->req(); i < mem->len(); i++) {
2594 // Accumulate any precedence edges
2595 if (mem->in(i) != NULL) {
2596 n->add_prec(mem->in(i));
2597 }
2598 }
2599 // Everything above this point has been processed.
2600 done = true;
2601 }
2602 // Eliminate the previous StoreCM
2603 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2604 assert(mem->outcnt() == 0, "should be dead");
2605 mem->disconnect_inputs(NULL, this);
2606 } else {
2607 prev = mem;
2608 }
2609 mem = prev->in(MemNode::Memory);
2610 }
2611 }
2612 }
2614 //------------------------------final_graph_reshaping_impl----------------------
2615 // Implement items 1-5 from final_graph_reshaping below.
2616 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2618 if ( n->outcnt() == 0 ) return; // dead node
2619 uint nop = n->Opcode();
2621 // Check for 2-input instruction with "last use" on right input.
2622 // Swap to left input. Implements item (2).
2623 if( n->req() == 3 && // two-input instruction
2624 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2625 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2626 n->in(2)->outcnt() == 1 &&// right use IS a last use
2627 !n->in(2)->is_Con() ) { // right use is not a constant
2628 // Check for commutative opcode
2629 switch( nop ) {
2630 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2631 case Op_MaxI: case Op_MinI:
2632 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2633 case Op_AndL: case Op_XorL: case Op_OrL:
2634 case Op_AndI: case Op_XorI: case Op_OrI: {
2635 // Move "last use" input to left by swapping inputs
2636 n->swap_edges(1, 2);
2637 break;
2638 }
2639 default:
2640 break;
2641 }
2642 }
2644 #ifdef ASSERT
2645 if( n->is_Mem() ) {
2646 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2647 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2648 // oop will be recorded in oop map if load crosses safepoint
2649 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2650 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2651 "raw memory operations should have control edge");
2652 }
2653 #endif
2654 // Count FPU ops and common calls, implements item (3)
2655 switch( nop ) {
2656 // Count all float operations that may use FPU
2657 case Op_AddF:
2658 case Op_SubF:
2659 case Op_MulF:
2660 case Op_DivF:
2661 case Op_NegF:
2662 case Op_ModF:
2663 case Op_ConvI2F:
2664 case Op_ConF:
2665 case Op_CmpF:
2666 case Op_CmpF3:
2667 // case Op_ConvL2F: // longs are split into 32-bit halves
2668 frc.inc_float_count();
2669 break;
2671 case Op_ConvF2D:
2672 case Op_ConvD2F:
2673 frc.inc_float_count();
2674 frc.inc_double_count();
2675 break;
2677 // Count all double operations that may use FPU
2678 case Op_AddD:
2679 case Op_SubD:
2680 case Op_MulD:
2681 case Op_DivD:
2682 case Op_NegD:
2683 case Op_ModD:
2684 case Op_ConvI2D:
2685 case Op_ConvD2I:
2686 // case Op_ConvL2D: // handled by leaf call
2687 // case Op_ConvD2L: // handled by leaf call
2688 case Op_ConD:
2689 case Op_CmpD:
2690 case Op_CmpD3:
2691 frc.inc_double_count();
2692 break;
2693 case Op_Opaque1: // Remove Opaque Nodes before matching
2694 case Op_Opaque2: // Remove Opaque Nodes before matching
2695 case Op_Opaque3:
2696 n->subsume_by(n->in(1), this);
2697 break;
2698 case Op_CallStaticJava:
2699 case Op_CallJava:
2700 case Op_CallDynamicJava:
2701 frc.inc_java_call_count(); // Count java call site;
2702 case Op_CallRuntime:
2703 case Op_CallLeaf:
2704 case Op_CallLeafNoFP: {
2705 assert( n->is_Call(), "" );
2706 CallNode *call = n->as_Call();
2707 // Count call sites where the FP mode bit would have to be flipped.
2708 // Do not count uncommon runtime calls:
2709 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2710 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2711 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2712 frc.inc_call_count(); // Count the call site
2713 } else { // See if uncommon argument is shared
2714 Node *n = call->in(TypeFunc::Parms);
2715 int nop = n->Opcode();
2716 // Clone shared simple arguments to uncommon calls, item (1).
2717 if( n->outcnt() > 1 &&
2718 !n->is_Proj() &&
2719 nop != Op_CreateEx &&
2720 nop != Op_CheckCastPP &&
2721 nop != Op_DecodeN &&
2722 nop != Op_DecodeNKlass &&
2723 !n->is_Mem() ) {
2724 Node *x = n->clone();
2725 call->set_req( TypeFunc::Parms, x );
2726 }
2727 }
2728 break;
2729 }
2731 case Op_StoreD:
2732 case Op_LoadD:
2733 case Op_LoadD_unaligned:
2734 frc.inc_double_count();
2735 goto handle_mem;
2736 case Op_StoreF:
2737 case Op_LoadF:
2738 frc.inc_float_count();
2739 goto handle_mem;
2741 case Op_StoreCM:
2742 {
2743 // Convert OopStore dependence into precedence edge
2744 Node* prec = n->in(MemNode::OopStore);
2745 n->del_req(MemNode::OopStore);
2746 n->add_prec(prec);
2747 eliminate_redundant_card_marks(n);
2748 }
2750 // fall through
2752 case Op_StoreB:
2753 case Op_StoreC:
2754 case Op_StorePConditional:
2755 case Op_StoreI:
2756 case Op_StoreL:
2757 case Op_StoreIConditional:
2758 case Op_StoreLConditional:
2759 case Op_CompareAndSwapI:
2760 case Op_CompareAndSwapL:
2761 case Op_CompareAndSwapP:
2762 case Op_CompareAndSwapN:
2763 case Op_GetAndAddI:
2764 case Op_GetAndAddL:
2765 case Op_GetAndSetI:
2766 case Op_GetAndSetL:
2767 case Op_GetAndSetP:
2768 case Op_GetAndSetN:
2769 case Op_StoreP:
2770 case Op_StoreN:
2771 case Op_StoreNKlass:
2772 case Op_LoadB:
2773 case Op_LoadUB:
2774 case Op_LoadUS:
2775 case Op_LoadI:
2776 case Op_LoadKlass:
2777 case Op_LoadNKlass:
2778 case Op_LoadL:
2779 case Op_LoadL_unaligned:
2780 case Op_LoadPLocked:
2781 case Op_LoadP:
2782 case Op_LoadN:
2783 case Op_LoadRange:
2784 case Op_LoadS: {
2785 handle_mem:
2786 #ifdef ASSERT
2787 if( VerifyOptoOopOffsets ) {
2788 assert( n->is_Mem(), "" );
2789 MemNode *mem = (MemNode*)n;
2790 // Check to see if address types have grounded out somehow.
2791 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2792 assert( !tp || oop_offset_is_sane(tp), "" );
2793 }
2794 #endif
2795 break;
2796 }
2798 case Op_AddP: { // Assert sane base pointers
2799 Node *addp = n->in(AddPNode::Address);
2800 assert( !addp->is_AddP() ||
2801 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2802 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2803 "Base pointers must match" );
2804 #ifdef _LP64
2805 if ((UseCompressedOops || UseCompressedClassPointers) &&
2806 addp->Opcode() == Op_ConP &&
2807 addp == n->in(AddPNode::Base) &&
2808 n->in(AddPNode::Offset)->is_Con()) {
2809 // Use addressing with narrow klass to load with offset on x86.
2810 // On sparc loading 32-bits constant and decoding it have less
2811 // instructions (4) then load 64-bits constant (7).
2812 // Do this transformation here since IGVN will convert ConN back to ConP.
2813 const Type* t = addp->bottom_type();
2814 if (t->isa_oopptr() || t->isa_klassptr()) {
2815 Node* nn = NULL;
2817 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2819 // Look for existing ConN node of the same exact type.
2820 Node* r = root();
2821 uint cnt = r->outcnt();
2822 for (uint i = 0; i < cnt; i++) {
2823 Node* m = r->raw_out(i);
2824 if (m!= NULL && m->Opcode() == op &&
2825 m->bottom_type()->make_ptr() == t) {
2826 nn = m;
2827 break;
2828 }
2829 }
2830 if (nn != NULL) {
2831 // Decode a narrow oop to match address
2832 // [R12 + narrow_oop_reg<<3 + offset]
2833 if (t->isa_oopptr()) {
2834 nn = new (this) DecodeNNode(nn, t);
2835 } else {
2836 nn = new (this) DecodeNKlassNode(nn, t);
2837 }
2838 n->set_req(AddPNode::Base, nn);
2839 n->set_req(AddPNode::Address, nn);
2840 if (addp->outcnt() == 0) {
2841 addp->disconnect_inputs(NULL, this);
2842 }
2843 }
2844 }
2845 }
2846 #endif
2847 break;
2848 }
2850 #ifdef _LP64
2851 case Op_CastPP:
2852 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2853 Node* in1 = n->in(1);
2854 const Type* t = n->bottom_type();
2855 Node* new_in1 = in1->clone();
2856 new_in1->as_DecodeN()->set_type(t);
2858 if (!Matcher::narrow_oop_use_complex_address()) {
2859 //
2860 // x86, ARM and friends can handle 2 adds in addressing mode
2861 // and Matcher can fold a DecodeN node into address by using
2862 // a narrow oop directly and do implicit NULL check in address:
2863 //
2864 // [R12 + narrow_oop_reg<<3 + offset]
2865 // NullCheck narrow_oop_reg
2866 //
2867 // On other platforms (Sparc) we have to keep new DecodeN node and
2868 // use it to do implicit NULL check in address:
2869 //
2870 // decode_not_null narrow_oop_reg, base_reg
2871 // [base_reg + offset]
2872 // NullCheck base_reg
2873 //
2874 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2875 // to keep the information to which NULL check the new DecodeN node
2876 // corresponds to use it as value in implicit_null_check().
2877 //
2878 new_in1->set_req(0, n->in(0));
2879 }
2881 n->subsume_by(new_in1, this);
2882 if (in1->outcnt() == 0) {
2883 in1->disconnect_inputs(NULL, this);
2884 }
2885 }
2886 break;
2888 case Op_CmpP:
2889 // Do this transformation here to preserve CmpPNode::sub() and
2890 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2891 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2892 Node* in1 = n->in(1);
2893 Node* in2 = n->in(2);
2894 if (!in1->is_DecodeNarrowPtr()) {
2895 in2 = in1;
2896 in1 = n->in(2);
2897 }
2898 assert(in1->is_DecodeNarrowPtr(), "sanity");
2900 Node* new_in2 = NULL;
2901 if (in2->is_DecodeNarrowPtr()) {
2902 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2903 new_in2 = in2->in(1);
2904 } else if (in2->Opcode() == Op_ConP) {
2905 const Type* t = in2->bottom_type();
2906 if (t == TypePtr::NULL_PTR) {
2907 assert(in1->is_DecodeN(), "compare klass to null?");
2908 // Don't convert CmpP null check into CmpN if compressed
2909 // oops implicit null check is not generated.
2910 // This will allow to generate normal oop implicit null check.
2911 if (Matcher::gen_narrow_oop_implicit_null_checks())
2912 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2913 //
2914 // This transformation together with CastPP transformation above
2915 // will generated code for implicit NULL checks for compressed oops.
2916 //
2917 // The original code after Optimize()
2918 //
2919 // LoadN memory, narrow_oop_reg
2920 // decode narrow_oop_reg, base_reg
2921 // CmpP base_reg, NULL
2922 // CastPP base_reg // NotNull
2923 // Load [base_reg + offset], val_reg
2924 //
2925 // after these transformations will be
2926 //
2927 // LoadN memory, narrow_oop_reg
2928 // CmpN narrow_oop_reg, NULL
2929 // decode_not_null narrow_oop_reg, base_reg
2930 // Load [base_reg + offset], val_reg
2931 //
2932 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2933 // since narrow oops can be used in debug info now (see the code in
2934 // final_graph_reshaping_walk()).
2935 //
2936 // At the end the code will be matched to
2937 // on x86:
2938 //
2939 // Load_narrow_oop memory, narrow_oop_reg
2940 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2941 // NullCheck narrow_oop_reg
2942 //
2943 // and on sparc:
2944 //
2945 // Load_narrow_oop memory, narrow_oop_reg
2946 // decode_not_null narrow_oop_reg, base_reg
2947 // Load [base_reg + offset], val_reg
2948 // NullCheck base_reg
2949 //
2950 } else if (t->isa_oopptr()) {
2951 new_in2 = ConNode::make(this, t->make_narrowoop());
2952 } else if (t->isa_klassptr()) {
2953 new_in2 = ConNode::make(this, t->make_narrowklass());
2954 }
2955 }
2956 if (new_in2 != NULL) {
2957 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2958 n->subsume_by(cmpN, this);
2959 if (in1->outcnt() == 0) {
2960 in1->disconnect_inputs(NULL, this);
2961 }
2962 if (in2->outcnt() == 0) {
2963 in2->disconnect_inputs(NULL, this);
2964 }
2965 }
2966 }
2967 break;
2969 case Op_DecodeN:
2970 case Op_DecodeNKlass:
2971 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2972 // DecodeN could be pinned when it can't be fold into
2973 // an address expression, see the code for Op_CastPP above.
2974 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2975 break;
2977 case Op_EncodeP:
2978 case Op_EncodePKlass: {
2979 Node* in1 = n->in(1);
2980 if (in1->is_DecodeNarrowPtr()) {
2981 n->subsume_by(in1->in(1), this);
2982 } else if (in1->Opcode() == Op_ConP) {
2983 const Type* t = in1->bottom_type();
2984 if (t == TypePtr::NULL_PTR) {
2985 assert(t->isa_oopptr(), "null klass?");
2986 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2987 } else if (t->isa_oopptr()) {
2988 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2989 } else if (t->isa_klassptr()) {
2990 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2991 }
2992 }
2993 if (in1->outcnt() == 0) {
2994 in1->disconnect_inputs(NULL, this);
2995 }
2996 break;
2997 }
2999 case Op_Proj: {
3000 if (OptimizeStringConcat) {
3001 ProjNode* p = n->as_Proj();
3002 if (p->_is_io_use) {
3003 // Separate projections were used for the exception path which
3004 // are normally removed by a late inline. If it wasn't inlined
3005 // then they will hang around and should just be replaced with
3006 // the original one.
3007 Node* proj = NULL;
3008 // Replace with just one
3009 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3010 Node *use = i.get();
3011 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3012 proj = use;
3013 break;
3014 }
3015 }
3016 assert(proj != NULL, "must be found");
3017 p->subsume_by(proj, this);
3018 }
3019 }
3020 break;
3021 }
3023 case Op_Phi:
3024 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3025 // The EncodeP optimization may create Phi with the same edges
3026 // for all paths. It is not handled well by Register Allocator.
3027 Node* unique_in = n->in(1);
3028 assert(unique_in != NULL, "");
3029 uint cnt = n->req();
3030 for (uint i = 2; i < cnt; i++) {
3031 Node* m = n->in(i);
3032 assert(m != NULL, "");
3033 if (unique_in != m)
3034 unique_in = NULL;
3035 }
3036 if (unique_in != NULL) {
3037 n->subsume_by(unique_in, this);
3038 }
3039 }
3040 break;
3042 #endif
3044 #ifdef ASSERT
3045 case Op_CastII:
3046 // Verify that all range check dependent CastII nodes were removed.
3047 if (n->isa_CastII()->has_range_check()) {
3048 n->dump(3);
3049 assert(false, "Range check dependent CastII node was not removed");
3050 }
3051 break;
3052 #endif
3054 case Op_ModI:
3055 if (UseDivMod) {
3056 // Check if a%b and a/b both exist
3057 Node* d = n->find_similar(Op_DivI);
3058 if (d) {
3059 // Replace them with a fused divmod if supported
3060 if (Matcher::has_match_rule(Op_DivModI)) {
3061 DivModINode* divmod = DivModINode::make(this, n);
3062 d->subsume_by(divmod->div_proj(), this);
3063 n->subsume_by(divmod->mod_proj(), this);
3064 } else {
3065 // replace a%b with a-((a/b)*b)
3066 Node* mult = new (this) MulINode(d, d->in(2));
3067 Node* sub = new (this) SubINode(d->in(1), mult);
3068 n->subsume_by(sub, this);
3069 }
3070 }
3071 }
3072 break;
3074 case Op_ModL:
3075 if (UseDivMod) {
3076 // Check if a%b and a/b both exist
3077 Node* d = n->find_similar(Op_DivL);
3078 if (d) {
3079 // Replace them with a fused divmod if supported
3080 if (Matcher::has_match_rule(Op_DivModL)) {
3081 DivModLNode* divmod = DivModLNode::make(this, n);
3082 d->subsume_by(divmod->div_proj(), this);
3083 n->subsume_by(divmod->mod_proj(), this);
3084 } else {
3085 // replace a%b with a-((a/b)*b)
3086 Node* mult = new (this) MulLNode(d, d->in(2));
3087 Node* sub = new (this) SubLNode(d->in(1), mult);
3088 n->subsume_by(sub, this);
3089 }
3090 }
3091 }
3092 break;
3094 case Op_LoadVector:
3095 case Op_StoreVector:
3096 break;
3098 case Op_PackB:
3099 case Op_PackS:
3100 case Op_PackI:
3101 case Op_PackF:
3102 case Op_PackL:
3103 case Op_PackD:
3104 if (n->req()-1 > 2) {
3105 // Replace many operand PackNodes with a binary tree for matching
3106 PackNode* p = (PackNode*) n;
3107 Node* btp = p->binary_tree_pack(this, 1, n->req());
3108 n->subsume_by(btp, this);
3109 }
3110 break;
3111 case Op_Loop:
3112 case Op_CountedLoop:
3113 if (n->as_Loop()->is_inner_loop()) {
3114 frc.inc_inner_loop_count();
3115 }
3116 break;
3117 case Op_LShiftI:
3118 case Op_RShiftI:
3119 case Op_URShiftI:
3120 case Op_LShiftL:
3121 case Op_RShiftL:
3122 case Op_URShiftL:
3123 if (Matcher::need_masked_shift_count) {
3124 // The cpu's shift instructions don't restrict the count to the
3125 // lower 5/6 bits. We need to do the masking ourselves.
3126 Node* in2 = n->in(2);
3127 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3128 const TypeInt* t = in2->find_int_type();
3129 if (t != NULL && t->is_con()) {
3130 juint shift = t->get_con();
3131 if (shift > mask) { // Unsigned cmp
3132 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
3133 }
3134 } else {
3135 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3136 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3137 n->set_req(2, shift);
3138 }
3139 }
3140 if (in2->outcnt() == 0) { // Remove dead node
3141 in2->disconnect_inputs(NULL, this);
3142 }
3143 }
3144 break;
3145 case Op_MemBarStoreStore:
3146 case Op_MemBarRelease:
3147 // Break the link with AllocateNode: it is no longer useful and
3148 // confuses register allocation.
3149 if (n->req() > MemBarNode::Precedent) {
3150 n->set_req(MemBarNode::Precedent, top());
3151 }
3152 break;
3153 default:
3154 assert( !n->is_Call(), "" );
3155 assert( !n->is_Mem(), "" );
3156 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3157 break;
3158 }
3160 // Collect CFG split points
3161 if (n->is_MultiBranch())
3162 frc._tests.push(n);
3163 }
3165 //------------------------------final_graph_reshaping_walk---------------------
3166 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3167 // requires that the walk visits a node's inputs before visiting the node.
3168 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3169 ResourceArea *area = Thread::current()->resource_area();
3170 Unique_Node_List sfpt(area);
3172 frc._visited.set(root->_idx); // first, mark node as visited
3173 uint cnt = root->req();
3174 Node *n = root;
3175 uint i = 0;
3176 while (true) {
3177 if (i < cnt) {
3178 // Place all non-visited non-null inputs onto stack
3179 Node* m = n->in(i);
3180 ++i;
3181 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3182 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3183 // compute worst case interpreter size in case of a deoptimization
3184 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3186 sfpt.push(m);
3187 }
3188 cnt = m->req();
3189 nstack.push(n, i); // put on stack parent and next input's index
3190 n = m;
3191 i = 0;
3192 }
3193 } else {
3194 // Now do post-visit work
3195 final_graph_reshaping_impl( n, frc );
3196 if (nstack.is_empty())
3197 break; // finished
3198 n = nstack.node(); // Get node from stack
3199 cnt = n->req();
3200 i = nstack.index();
3201 nstack.pop(); // Shift to the next node on stack
3202 }
3203 }
3205 // Skip next transformation if compressed oops are not used.
3206 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3207 (!UseCompressedOops && !UseCompressedClassPointers))
3208 return;
3210 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3211 // It could be done for an uncommon traps or any safepoints/calls
3212 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3213 while (sfpt.size() > 0) {
3214 n = sfpt.pop();
3215 JVMState *jvms = n->as_SafePoint()->jvms();
3216 assert(jvms != NULL, "sanity");
3217 int start = jvms->debug_start();
3218 int end = n->req();
3219 bool is_uncommon = (n->is_CallStaticJava() &&
3220 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3221 for (int j = start; j < end; j++) {
3222 Node* in = n->in(j);
3223 if (in->is_DecodeNarrowPtr()) {
3224 bool safe_to_skip = true;
3225 if (!is_uncommon ) {
3226 // Is it safe to skip?
3227 for (uint i = 0; i < in->outcnt(); i++) {
3228 Node* u = in->raw_out(i);
3229 if (!u->is_SafePoint() ||
3230 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3231 safe_to_skip = false;
3232 }
3233 }
3234 }
3235 if (safe_to_skip) {
3236 n->set_req(j, in->in(1));
3237 }
3238 if (in->outcnt() == 0) {
3239 in->disconnect_inputs(NULL, this);
3240 }
3241 }
3242 }
3243 }
3244 }
3246 //------------------------------final_graph_reshaping--------------------------
3247 // Final Graph Reshaping.
3248 //
3249 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3250 // and not commoned up and forced early. Must come after regular
3251 // optimizations to avoid GVN undoing the cloning. Clone constant
3252 // inputs to Loop Phis; these will be split by the allocator anyways.
3253 // Remove Opaque nodes.
3254 // (2) Move last-uses by commutative operations to the left input to encourage
3255 // Intel update-in-place two-address operations and better register usage
3256 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3257 // calls canonicalizing them back.
3258 // (3) Count the number of double-precision FP ops, single-precision FP ops
3259 // and call sites. On Intel, we can get correct rounding either by
3260 // forcing singles to memory (requires extra stores and loads after each
3261 // FP bytecode) or we can set a rounding mode bit (requires setting and
3262 // clearing the mode bit around call sites). The mode bit is only used
3263 // if the relative frequency of single FP ops to calls is low enough.
3264 // This is a key transform for SPEC mpeg_audio.
3265 // (4) Detect infinite loops; blobs of code reachable from above but not
3266 // below. Several of the Code_Gen algorithms fail on such code shapes,
3267 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3268 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3269 // Detection is by looking for IfNodes where only 1 projection is
3270 // reachable from below or CatchNodes missing some targets.
3271 // (5) Assert for insane oop offsets in debug mode.
3273 bool Compile::final_graph_reshaping() {
3274 // an infinite loop may have been eliminated by the optimizer,
3275 // in which case the graph will be empty.
3276 if (root()->req() == 1) {
3277 record_method_not_compilable("trivial infinite loop");
3278 return true;
3279 }
3281 // Expensive nodes have their control input set to prevent the GVN
3282 // from freely commoning them. There's no GVN beyond this point so
3283 // no need to keep the control input. We want the expensive nodes to
3284 // be freely moved to the least frequent code path by gcm.
3285 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3286 for (int i = 0; i < expensive_count(); i++) {
3287 _expensive_nodes->at(i)->set_req(0, NULL);
3288 }
3290 Final_Reshape_Counts frc;
3292 // Visit everybody reachable!
3293 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3294 Node_Stack nstack(live_nodes() >> 1);
3295 final_graph_reshaping_walk(nstack, root(), frc);
3297 // Check for unreachable (from below) code (i.e., infinite loops).
3298 for( uint i = 0; i < frc._tests.size(); i++ ) {
3299 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3300 // Get number of CFG targets.
3301 // Note that PCTables include exception targets after calls.
3302 uint required_outcnt = n->required_outcnt();
3303 if (n->outcnt() != required_outcnt) {
3304 // Check for a few special cases. Rethrow Nodes never take the
3305 // 'fall-thru' path, so expected kids is 1 less.
3306 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3307 if (n->in(0)->in(0)->is_Call()) {
3308 CallNode *call = n->in(0)->in(0)->as_Call();
3309 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3310 required_outcnt--; // Rethrow always has 1 less kid
3311 } else if (call->req() > TypeFunc::Parms &&
3312 call->is_CallDynamicJava()) {
3313 // Check for null receiver. In such case, the optimizer has
3314 // detected that the virtual call will always result in a null
3315 // pointer exception. The fall-through projection of this CatchNode
3316 // will not be populated.
3317 Node *arg0 = call->in(TypeFunc::Parms);
3318 if (arg0->is_Type() &&
3319 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3320 required_outcnt--;
3321 }
3322 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3323 call->req() > TypeFunc::Parms+1 &&
3324 call->is_CallStaticJava()) {
3325 // Check for negative array length. In such case, the optimizer has
3326 // detected that the allocation attempt will always result in an
3327 // exception. There is no fall-through projection of this CatchNode .
3328 Node *arg1 = call->in(TypeFunc::Parms+1);
3329 if (arg1->is_Type() &&
3330 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3331 required_outcnt--;
3332 }
3333 }
3334 }
3335 }
3336 // Recheck with a better notion of 'required_outcnt'
3337 if (n->outcnt() != required_outcnt) {
3338 record_method_not_compilable("malformed control flow");
3339 return true; // Not all targets reachable!
3340 }
3341 }
3342 // Check that I actually visited all kids. Unreached kids
3343 // must be infinite loops.
3344 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3345 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3346 record_method_not_compilable("infinite loop");
3347 return true; // Found unvisited kid; must be unreach
3348 }
3349 }
3351 // If original bytecodes contained a mixture of floats and doubles
3352 // check if the optimizer has made it homogenous, item (3).
3353 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3354 frc.get_float_count() > 32 &&
3355 frc.get_double_count() == 0 &&
3356 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3357 set_24_bit_selection_and_mode( false, true );
3358 }
3360 set_java_calls(frc.get_java_call_count());
3361 set_inner_loops(frc.get_inner_loop_count());
3363 // No infinite loops, no reason to bail out.
3364 return false;
3365 }
3367 //-----------------------------too_many_traps----------------------------------
3368 // Report if there are too many traps at the current method and bci.
3369 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3370 bool Compile::too_many_traps(ciMethod* method,
3371 int bci,
3372 Deoptimization::DeoptReason reason) {
3373 ciMethodData* md = method->method_data();
3374 if (md->is_empty()) {
3375 // Assume the trap has not occurred, or that it occurred only
3376 // because of a transient condition during start-up in the interpreter.
3377 return false;
3378 }
3379 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3380 if (md->has_trap_at(bci, m, reason) != 0) {
3381 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3382 // Also, if there are multiple reasons, or if there is no per-BCI record,
3383 // assume the worst.
3384 if (log())
3385 log()->elem("observe trap='%s' count='%d'",
3386 Deoptimization::trap_reason_name(reason),
3387 md->trap_count(reason));
3388 return true;
3389 } else {
3390 // Ignore method/bci and see if there have been too many globally.
3391 return too_many_traps(reason, md);
3392 }
3393 }
3395 // Less-accurate variant which does not require a method and bci.
3396 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3397 ciMethodData* logmd) {
3398 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3399 // Too many traps globally.
3400 // Note that we use cumulative trap_count, not just md->trap_count.
3401 if (log()) {
3402 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3403 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3404 Deoptimization::trap_reason_name(reason),
3405 mcount, trap_count(reason));
3406 }
3407 return true;
3408 } else {
3409 // The coast is clear.
3410 return false;
3411 }
3412 }
3414 //--------------------------too_many_recompiles--------------------------------
3415 // Report if there are too many recompiles at the current method and bci.
3416 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3417 // Is not eager to return true, since this will cause the compiler to use
3418 // Action_none for a trap point, to avoid too many recompilations.
3419 bool Compile::too_many_recompiles(ciMethod* method,
3420 int bci,
3421 Deoptimization::DeoptReason reason) {
3422 ciMethodData* md = method->method_data();
3423 if (md->is_empty()) {
3424 // Assume the trap has not occurred, or that it occurred only
3425 // because of a transient condition during start-up in the interpreter.
3426 return false;
3427 }
3428 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3429 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3430 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3431 Deoptimization::DeoptReason per_bc_reason
3432 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3433 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3434 if ((per_bc_reason == Deoptimization::Reason_none
3435 || md->has_trap_at(bci, m, reason) != 0)
3436 // The trap frequency measure we care about is the recompile count:
3437 && md->trap_recompiled_at(bci, m)
3438 && md->overflow_recompile_count() >= bc_cutoff) {
3439 // Do not emit a trap here if it has already caused recompilations.
3440 // Also, if there are multiple reasons, or if there is no per-BCI record,
3441 // assume the worst.
3442 if (log())
3443 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3444 Deoptimization::trap_reason_name(reason),
3445 md->trap_count(reason),
3446 md->overflow_recompile_count());
3447 return true;
3448 } else if (trap_count(reason) != 0
3449 && decompile_count() >= m_cutoff) {
3450 // Too many recompiles globally, and we have seen this sort of trap.
3451 // Use cumulative decompile_count, not just md->decompile_count.
3452 if (log())
3453 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3454 Deoptimization::trap_reason_name(reason),
3455 md->trap_count(reason), trap_count(reason),
3456 md->decompile_count(), decompile_count());
3457 return true;
3458 } else {
3459 // The coast is clear.
3460 return false;
3461 }
3462 }
3464 // Compute when not to trap. Used by matching trap based nodes and
3465 // NullCheck optimization.
3466 void Compile::set_allowed_deopt_reasons() {
3467 _allowed_reasons = 0;
3468 if (is_method_compilation()) {
3469 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3470 assert(rs < BitsPerInt, "recode bit map");
3471 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3472 _allowed_reasons |= nth_bit(rs);
3473 }
3474 }
3475 }
3476 }
3478 #ifndef PRODUCT
3479 //------------------------------verify_graph_edges---------------------------
3480 // Walk the Graph and verify that there is a one-to-one correspondence
3481 // between Use-Def edges and Def-Use edges in the graph.
3482 void Compile::verify_graph_edges(bool no_dead_code) {
3483 if (VerifyGraphEdges) {
3484 ResourceArea *area = Thread::current()->resource_area();
3485 Unique_Node_List visited(area);
3486 // Call recursive graph walk to check edges
3487 _root->verify_edges(visited);
3488 if (no_dead_code) {
3489 // Now make sure that no visited node is used by an unvisited node.
3490 bool dead_nodes = false;
3491 Unique_Node_List checked(area);
3492 while (visited.size() > 0) {
3493 Node* n = visited.pop();
3494 checked.push(n);
3495 for (uint i = 0; i < n->outcnt(); i++) {
3496 Node* use = n->raw_out(i);
3497 if (checked.member(use)) continue; // already checked
3498 if (visited.member(use)) continue; // already in the graph
3499 if (use->is_Con()) continue; // a dead ConNode is OK
3500 // At this point, we have found a dead node which is DU-reachable.
3501 if (!dead_nodes) {
3502 tty->print_cr("*** Dead nodes reachable via DU edges:");
3503 dead_nodes = true;
3504 }
3505 use->dump(2);
3506 tty->print_cr("---");
3507 checked.push(use); // No repeats; pretend it is now checked.
3508 }
3509 }
3510 assert(!dead_nodes, "using nodes must be reachable from root");
3511 }
3512 }
3513 }
3515 // Verify GC barriers consistency
3516 // Currently supported:
3517 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3518 void Compile::verify_barriers() {
3519 if (UseG1GC) {
3520 // Verify G1 pre-barriers
3521 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3523 ResourceArea *area = Thread::current()->resource_area();
3524 Unique_Node_List visited(area);
3525 Node_List worklist(area);
3526 // We're going to walk control flow backwards starting from the Root
3527 worklist.push(_root);
3528 while (worklist.size() > 0) {
3529 Node* x = worklist.pop();
3530 if (x == NULL || x == top()) continue;
3531 if (visited.member(x)) {
3532 continue;
3533 } else {
3534 visited.push(x);
3535 }
3537 if (x->is_Region()) {
3538 for (uint i = 1; i < x->req(); i++) {
3539 worklist.push(x->in(i));
3540 }
3541 } else {
3542 worklist.push(x->in(0));
3543 // We are looking for the pattern:
3544 // /->ThreadLocal
3545 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3546 // \->ConI(0)
3547 // We want to verify that the If and the LoadB have the same control
3548 // See GraphKit::g1_write_barrier_pre()
3549 if (x->is_If()) {
3550 IfNode *iff = x->as_If();
3551 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3552 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3553 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3554 && cmp->in(1)->is_Load()) {
3555 LoadNode* load = cmp->in(1)->as_Load();
3556 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3557 && load->in(2)->in(3)->is_Con()
3558 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3560 Node* if_ctrl = iff->in(0);
3561 Node* load_ctrl = load->in(0);
3563 if (if_ctrl != load_ctrl) {
3564 // Skip possible CProj->NeverBranch in infinite loops
3565 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3566 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3567 if_ctrl = if_ctrl->in(0)->in(0);
3568 }
3569 }
3570 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3571 }
3572 }
3573 }
3574 }
3575 }
3576 }
3577 }
3578 }
3580 #endif
3582 // The Compile object keeps track of failure reasons separately from the ciEnv.
3583 // This is required because there is not quite a 1-1 relation between the
3584 // ciEnv and its compilation task and the Compile object. Note that one
3585 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3586 // to backtrack and retry without subsuming loads. Other than this backtracking
3587 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3588 // by the logic in C2Compiler.
3589 void Compile::record_failure(const char* reason) {
3590 if (log() != NULL) {
3591 log()->elem("failure reason='%s' phase='compile'", reason);
3592 }
3593 if (_failure_reason == NULL) {
3594 // Record the first failure reason.
3595 _failure_reason = reason;
3596 }
3598 EventCompilerFailure event;
3599 if (event.should_commit()) {
3600 event.set_compileID(Compile::compile_id());
3601 event.set_failure(reason);
3602 event.commit();
3603 }
3605 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3606 C->print_method(PHASE_FAILURE);
3607 }
3608 _root = NULL; // flush the graph, too
3609 }
3611 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3612 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3613 _phase_name(name), _dolog(dolog)
3614 {
3615 if (dolog) {
3616 C = Compile::current();
3617 _log = C->log();
3618 } else {
3619 C = NULL;
3620 _log = NULL;
3621 }
3622 if (_log != NULL) {
3623 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3624 _log->stamp();
3625 _log->end_head();
3626 }
3627 }
3629 Compile::TracePhase::~TracePhase() {
3631 C = Compile::current();
3632 if (_dolog) {
3633 _log = C->log();
3634 } else {
3635 _log = NULL;
3636 }
3638 #ifdef ASSERT
3639 if (PrintIdealNodeCount) {
3640 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3641 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3642 }
3644 if (VerifyIdealNodeCount) {
3645 Compile::current()->print_missing_nodes();
3646 }
3647 #endif
3649 if (_log != NULL) {
3650 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3651 }
3652 }
3654 //=============================================================================
3655 // Two Constant's are equal when the type and the value are equal.
3656 bool Compile::Constant::operator==(const Constant& other) {
3657 if (type() != other.type() ) return false;
3658 if (can_be_reused() != other.can_be_reused()) return false;
3659 // For floating point values we compare the bit pattern.
3660 switch (type()) {
3661 case T_FLOAT: return (_v._value.i == other._v._value.i);
3662 case T_LONG:
3663 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3664 case T_OBJECT:
3665 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3666 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3667 case T_METADATA: return (_v._metadata == other._v._metadata);
3668 default: ShouldNotReachHere();
3669 }
3670 return false;
3671 }
3673 static int type_to_size_in_bytes(BasicType t) {
3674 switch (t) {
3675 case T_LONG: return sizeof(jlong );
3676 case T_FLOAT: return sizeof(jfloat );
3677 case T_DOUBLE: return sizeof(jdouble);
3678 case T_METADATA: return sizeof(Metadata*);
3679 // We use T_VOID as marker for jump-table entries (labels) which
3680 // need an internal word relocation.
3681 case T_VOID:
3682 case T_ADDRESS:
3683 case T_OBJECT: return sizeof(jobject);
3684 }
3686 ShouldNotReachHere();
3687 return -1;
3688 }
3690 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3691 // sort descending
3692 if (a->freq() > b->freq()) return -1;
3693 if (a->freq() < b->freq()) return 1;
3694 return 0;
3695 }
3697 void Compile::ConstantTable::calculate_offsets_and_size() {
3698 // First, sort the array by frequencies.
3699 _constants.sort(qsort_comparator);
3701 #ifdef ASSERT
3702 // Make sure all jump-table entries were sorted to the end of the
3703 // array (they have a negative frequency).
3704 bool found_void = false;
3705 for (int i = 0; i < _constants.length(); i++) {
3706 Constant con = _constants.at(i);
3707 if (con.type() == T_VOID)
3708 found_void = true; // jump-tables
3709 else
3710 assert(!found_void, "wrong sorting");
3711 }
3712 #endif
3714 int offset = 0;
3715 for (int i = 0; i < _constants.length(); i++) {
3716 Constant* con = _constants.adr_at(i);
3718 // Align offset for type.
3719 int typesize = type_to_size_in_bytes(con->type());
3720 offset = align_size_up(offset, typesize);
3721 con->set_offset(offset); // set constant's offset
3723 if (con->type() == T_VOID) {
3724 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3725 offset = offset + typesize * n->outcnt(); // expand jump-table
3726 } else {
3727 offset = offset + typesize;
3728 }
3729 }
3731 // Align size up to the next section start (which is insts; see
3732 // CodeBuffer::align_at_start).
3733 assert(_size == -1, "already set?");
3734 _size = align_size_up(offset, CodeEntryAlignment);
3735 }
3737 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3738 MacroAssembler _masm(&cb);
3739 for (int i = 0; i < _constants.length(); i++) {
3740 Constant con = _constants.at(i);
3741 address constant_addr = NULL;
3742 switch (con.type()) {
3743 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3744 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3745 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3746 case T_OBJECT: {
3747 jobject obj = con.get_jobject();
3748 int oop_index = _masm.oop_recorder()->find_index(obj);
3749 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3750 break;
3751 }
3752 case T_ADDRESS: {
3753 address addr = (address) con.get_jobject();
3754 constant_addr = _masm.address_constant(addr);
3755 break;
3756 }
3757 // We use T_VOID as marker for jump-table entries (labels) which
3758 // need an internal word relocation.
3759 case T_VOID: {
3760 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3761 // Fill the jump-table with a dummy word. The real value is
3762 // filled in later in fill_jump_table.
3763 address dummy = (address) n;
3764 constant_addr = _masm.address_constant(dummy);
3765 // Expand jump-table
3766 for (uint i = 1; i < n->outcnt(); i++) {
3767 address temp_addr = _masm.address_constant(dummy + i);
3768 assert(temp_addr, "consts section too small");
3769 }
3770 break;
3771 }
3772 case T_METADATA: {
3773 Metadata* obj = con.get_metadata();
3774 int metadata_index = _masm.oop_recorder()->find_index(obj);
3775 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3776 break;
3777 }
3778 default: ShouldNotReachHere();
3779 }
3780 assert(constant_addr, "consts section too small");
3781 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3782 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())));
3783 }
3784 }
3786 int Compile::ConstantTable::find_offset(Constant& con) const {
3787 int idx = _constants.find(con);
3788 assert(idx != -1, "constant must be in constant table");
3789 int offset = _constants.at(idx).offset();
3790 assert(offset != -1, "constant table not emitted yet?");
3791 return offset;
3792 }
3794 void Compile::ConstantTable::add(Constant& con) {
3795 if (con.can_be_reused()) {
3796 int idx = _constants.find(con);
3797 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3798 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3799 return;
3800 }
3801 }
3802 (void) _constants.append(con);
3803 }
3805 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3806 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3807 Constant con(type, value, b->_freq);
3808 add(con);
3809 return con;
3810 }
3812 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3813 Constant con(metadata);
3814 add(con);
3815 return con;
3816 }
3818 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3819 jvalue value;
3820 BasicType type = oper->type()->basic_type();
3821 switch (type) {
3822 case T_LONG: value.j = oper->constantL(); break;
3823 case T_FLOAT: value.f = oper->constantF(); break;
3824 case T_DOUBLE: value.d = oper->constantD(); break;
3825 case T_OBJECT:
3826 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3827 case T_METADATA: return add((Metadata*)oper->constant()); break;
3828 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3829 }
3830 return add(n, type, value);
3831 }
3833 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3834 jvalue value;
3835 // We can use the node pointer here to identify the right jump-table
3836 // as this method is called from Compile::Fill_buffer right before
3837 // the MachNodes are emitted and the jump-table is filled (means the
3838 // MachNode pointers do not change anymore).
3839 value.l = (jobject) n;
3840 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3841 add(con);
3842 return con;
3843 }
3845 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3846 // If called from Compile::scratch_emit_size do nothing.
3847 if (Compile::current()->in_scratch_emit_size()) return;
3849 assert(labels.is_nonempty(), "must be");
3850 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3852 // Since MachConstantNode::constant_offset() also contains
3853 // table_base_offset() we need to subtract the table_base_offset()
3854 // to get the plain offset into the constant table.
3855 int offset = n->constant_offset() - table_base_offset();
3857 MacroAssembler _masm(&cb);
3858 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3860 for (uint i = 0; i < n->outcnt(); i++) {
3861 address* constant_addr = &jump_table_base[i];
3862 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i)));
3863 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3864 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3865 }
3866 }
3868 void Compile::dump_inlining() {
3869 if (print_inlining() || print_intrinsics()) {
3870 // Print inlining message for candidates that we couldn't inline
3871 // for lack of space or non constant receiver
3872 for (int i = 0; i < _late_inlines.length(); i++) {
3873 CallGenerator* cg = _late_inlines.at(i);
3874 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3875 }
3876 Unique_Node_List useful;
3877 useful.push(root());
3878 for (uint next = 0; next < useful.size(); ++next) {
3879 Node* n = useful.at(next);
3880 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3881 CallNode* call = n->as_Call();
3882 CallGenerator* cg = call->generator();
3883 cg->print_inlining_late("receiver not constant");
3884 }
3885 uint max = n->len();
3886 for ( uint i = 0; i < max; ++i ) {
3887 Node *m = n->in(i);
3888 if ( m == NULL ) continue;
3889 useful.push(m);
3890 }
3891 }
3892 for (int i = 0; i < _print_inlining_list->length(); i++) {
3893 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
3894 }
3895 }
3896 }
3898 // Dump inlining replay data to the stream.
3899 // Don't change thread state and acquire any locks.
3900 void Compile::dump_inline_data(outputStream* out) {
3901 InlineTree* inl_tree = ilt();
3902 if (inl_tree != NULL) {
3903 out->print(" inline %d", inl_tree->count());
3904 inl_tree->dump_replay_data(out);
3905 }
3906 }
3908 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3909 if (n1->Opcode() < n2->Opcode()) return -1;
3910 else if (n1->Opcode() > n2->Opcode()) return 1;
3912 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()));
3913 for (uint i = 1; i < n1->req(); i++) {
3914 if (n1->in(i) < n2->in(i)) return -1;
3915 else if (n1->in(i) > n2->in(i)) return 1;
3916 }
3918 return 0;
3919 }
3921 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3922 Node* n1 = *n1p;
3923 Node* n2 = *n2p;
3925 return cmp_expensive_nodes(n1, n2);
3926 }
3928 void Compile::sort_expensive_nodes() {
3929 if (!expensive_nodes_sorted()) {
3930 _expensive_nodes->sort(cmp_expensive_nodes);
3931 }
3932 }
3934 bool Compile::expensive_nodes_sorted() const {
3935 for (int i = 1; i < _expensive_nodes->length(); i++) {
3936 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3937 return false;
3938 }
3939 }
3940 return true;
3941 }
3943 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3944 if (_expensive_nodes->length() == 0) {
3945 return false;
3946 }
3948 assert(OptimizeExpensiveOps, "optimization off?");
3950 // Take this opportunity to remove dead nodes from the list
3951 int j = 0;
3952 for (int i = 0; i < _expensive_nodes->length(); i++) {
3953 Node* n = _expensive_nodes->at(i);
3954 if (!n->is_unreachable(igvn)) {
3955 assert(n->is_expensive(), "should be expensive");
3956 _expensive_nodes->at_put(j, n);
3957 j++;
3958 }
3959 }
3960 _expensive_nodes->trunc_to(j);
3962 // Then sort the list so that similar nodes are next to each other
3963 // and check for at least two nodes of identical kind with same data
3964 // inputs.
3965 sort_expensive_nodes();
3967 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3968 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3969 return true;
3970 }
3971 }
3973 return false;
3974 }
3976 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3977 if (_expensive_nodes->length() == 0) {
3978 return;
3979 }
3981 assert(OptimizeExpensiveOps, "optimization off?");
3983 // Sort to bring similar nodes next to each other and clear the
3984 // control input of nodes for which there's only a single copy.
3985 sort_expensive_nodes();
3987 int j = 0;
3988 int identical = 0;
3989 int i = 0;
3990 for (; i < _expensive_nodes->length()-1; i++) {
3991 assert(j <= i, "can't write beyond current index");
3992 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3993 identical++;
3994 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3995 continue;
3996 }
3997 if (identical > 0) {
3998 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3999 identical = 0;
4000 } else {
4001 Node* n = _expensive_nodes->at(i);
4002 igvn.hash_delete(n);
4003 n->set_req(0, NULL);
4004 igvn.hash_insert(n);
4005 }
4006 }
4007 if (identical > 0) {
4008 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4009 } else if (_expensive_nodes->length() >= 1) {
4010 Node* n = _expensive_nodes->at(i);
4011 igvn.hash_delete(n);
4012 n->set_req(0, NULL);
4013 igvn.hash_insert(n);
4014 }
4015 _expensive_nodes->trunc_to(j);
4016 }
4018 void Compile::add_expensive_node(Node * n) {
4019 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4020 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4021 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4022 if (OptimizeExpensiveOps) {
4023 _expensive_nodes->append(n);
4024 } else {
4025 // Clear control input and let IGVN optimize expensive nodes if
4026 // OptimizeExpensiveOps is off.
4027 n->set_req(0, NULL);
4028 }
4029 }
4031 /**
4032 * Remove the speculative part of types and clean up the graph
4033 */
4034 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4035 if (UseTypeSpeculation) {
4036 Unique_Node_List worklist;
4037 worklist.push(root());
4038 int modified = 0;
4039 // Go over all type nodes that carry a speculative type, drop the
4040 // speculative part of the type and enqueue the node for an igvn
4041 // which may optimize it out.
4042 for (uint next = 0; next < worklist.size(); ++next) {
4043 Node *n = worklist.at(next);
4044 if (n->is_Type()) {
4045 TypeNode* tn = n->as_Type();
4046 const Type* t = tn->type();
4047 const Type* t_no_spec = t->remove_speculative();
4048 if (t_no_spec != t) {
4049 bool in_hash = igvn.hash_delete(n);
4050 assert(in_hash, "node should be in igvn hash table");
4051 tn->set_type(t_no_spec);
4052 igvn.hash_insert(n);
4053 igvn._worklist.push(n); // give it a chance to go away
4054 modified++;
4055 }
4056 }
4057 uint max = n->len();
4058 for( uint i = 0; i < max; ++i ) {
4059 Node *m = n->in(i);
4060 if (not_a_node(m)) continue;
4061 worklist.push(m);
4062 }
4063 }
4064 // Drop the speculative part of all types in the igvn's type table
4065 igvn.remove_speculative_types();
4066 if (modified > 0) {
4067 igvn.optimize();
4068 }
4069 #ifdef ASSERT
4070 // Verify that after the IGVN is over no speculative type has resurfaced
4071 worklist.clear();
4072 worklist.push(root());
4073 for (uint next = 0; next < worklist.size(); ++next) {
4074 Node *n = worklist.at(next);
4075 const Type* t = igvn.type_or_null(n);
4076 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4077 if (n->is_Type()) {
4078 t = n->as_Type()->type();
4079 assert(t == t->remove_speculative(), "no more speculative types");
4080 }
4081 uint max = n->len();
4082 for( uint i = 0; i < max; ++i ) {
4083 Node *m = n->in(i);
4084 if (not_a_node(m)) continue;
4085 worklist.push(m);
4086 }
4087 }
4088 igvn.check_no_speculative_types();
4089 #endif
4090 }
4091 }
4093 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4094 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4095 if (ctrl != NULL) {
4096 // Express control dependency by a CastII node with a narrow type.
4097 value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */);
4098 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4099 // node from floating above the range check during loop optimizations. Otherwise, the
4100 // ConvI2L node may be eliminated independently of the range check, causing the data path
4101 // to become TOP while the control path is still there (although it's unreachable).
4102 value->set_req(0, ctrl);
4103 // Save CastII node to remove it after loop optimizations.
4104 phase->C->add_range_check_cast(value);
4105 value = phase->transform(value);
4106 }
4107 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4108 return phase->transform(new (phase->C) ConvI2LNode(value, ltype));
4109 }
4111 // Auxiliary method to support randomized stressing/fuzzing.
4112 //
4113 // This method can be called the arbitrary number of times, with current count
4114 // as the argument. The logic allows selecting a single candidate from the
4115 // running list of candidates as follows:
4116 // int count = 0;
4117 // Cand* selected = null;
4118 // while(cand = cand->next()) {
4119 // if (randomized_select(++count)) {
4120 // selected = cand;
4121 // }
4122 // }
4123 //
4124 // Including count equalizes the chances any candidate is "selected".
4125 // This is useful when we don't have the complete list of candidates to choose
4126 // from uniformly. In this case, we need to adjust the randomicity of the
4127 // selection, or else we will end up biasing the selection towards the latter
4128 // candidates.
4129 //
4130 // Quick back-envelope calculation shows that for the list of n candidates
4131 // the equal probability for the candidate to persist as "best" can be
4132 // achieved by replacing it with "next" k-th candidate with the probability
4133 // of 1/k. It can be easily shown that by the end of the run, the
4134 // probability for any candidate is converged to 1/n, thus giving the
4135 // uniform distribution among all the candidates.
4136 //
4137 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4138 #define RANDOMIZED_DOMAIN_POW 29
4139 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4140 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4141 bool Compile::randomized_select(int count) {
4142 assert(count > 0, "only positive");
4143 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4144 }