Wed, 18 Sep 2013 14:34:56 -0700
8024342: PPC64 (part 111): Support for C calling conventions that require 64-bit ints.
Summary: Some platforms, as ppc and s390x/zArch require that 32-bit ints are passed as 64-bit values to C functions. This change adds support to adapt the signature and to issue proper casts to c2-compiled stubs. The functions are used in generate_native_wrapper(). Adapt signature used by the compiler as in PhaseIdealLoop::intrinsify_fill().
Reviewed-by: kvn
1 /*
2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "code/exceptionHandlerTable.hpp"
30 #include "code/nmethod.hpp"
31 #include "compiler/compileLog.hpp"
32 #include "compiler/disassembler.hpp"
33 #include "compiler/oopMap.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/block.hpp"
36 #include "opto/c2compiler.hpp"
37 #include "opto/callGenerator.hpp"
38 #include "opto/callnode.hpp"
39 #include "opto/cfgnode.hpp"
40 #include "opto/chaitin.hpp"
41 #include "opto/compile.hpp"
42 #include "opto/connode.hpp"
43 #include "opto/divnode.hpp"
44 #include "opto/escape.hpp"
45 #include "opto/idealGraphPrinter.hpp"
46 #include "opto/loopnode.hpp"
47 #include "opto/machnode.hpp"
48 #include "opto/macro.hpp"
49 #include "opto/matcher.hpp"
50 #include "opto/memnode.hpp"
51 #include "opto/mulnode.hpp"
52 #include "opto/node.hpp"
53 #include "opto/opcodes.hpp"
54 #include "opto/output.hpp"
55 #include "opto/parse.hpp"
56 #include "opto/phaseX.hpp"
57 #include "opto/rootnode.hpp"
58 #include "opto/runtime.hpp"
59 #include "opto/stringopts.hpp"
60 #include "opto/type.hpp"
61 #include "opto/vectornode.hpp"
62 #include "runtime/arguments.hpp"
63 #include "runtime/signature.hpp"
64 #include "runtime/stubRoutines.hpp"
65 #include "runtime/timer.hpp"
66 #include "trace/tracing.hpp"
67 #include "utilities/copy.hpp"
68 #ifdef TARGET_ARCH_MODEL_x86_32
69 # include "adfiles/ad_x86_32.hpp"
70 #endif
71 #ifdef TARGET_ARCH_MODEL_x86_64
72 # include "adfiles/ad_x86_64.hpp"
73 #endif
74 #ifdef TARGET_ARCH_MODEL_sparc
75 # include "adfiles/ad_sparc.hpp"
76 #endif
77 #ifdef TARGET_ARCH_MODEL_zero
78 # include "adfiles/ad_zero.hpp"
79 #endif
80 #ifdef TARGET_ARCH_MODEL_arm
81 # include "adfiles/ad_arm.hpp"
82 #endif
83 #ifdef TARGET_ARCH_MODEL_ppc_32
84 # include "adfiles/ad_ppc_32.hpp"
85 #endif
86 #ifdef TARGET_ARCH_MODEL_ppc_64
87 # include "adfiles/ad_ppc_64.hpp"
88 #endif
91 // -------------------- Compile::mach_constant_base_node -----------------------
92 // Constant table base node singleton.
93 MachConstantBaseNode* Compile::mach_constant_base_node() {
94 if (_mach_constant_base_node == NULL) {
95 _mach_constant_base_node = new (C) MachConstantBaseNode();
96 _mach_constant_base_node->add_req(C->root());
97 }
98 return _mach_constant_base_node;
99 }
102 /// Support for intrinsics.
104 // Return the index at which m must be inserted (or already exists).
105 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
106 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
107 #ifdef ASSERT
108 for (int i = 1; i < _intrinsics->length(); i++) {
109 CallGenerator* cg1 = _intrinsics->at(i-1);
110 CallGenerator* cg2 = _intrinsics->at(i);
111 assert(cg1->method() != cg2->method()
112 ? cg1->method() < cg2->method()
113 : cg1->is_virtual() < cg2->is_virtual(),
114 "compiler intrinsics list must stay sorted");
115 }
116 #endif
117 // Binary search sorted list, in decreasing intervals [lo, hi].
118 int lo = 0, hi = _intrinsics->length()-1;
119 while (lo <= hi) {
120 int mid = (uint)(hi + lo) / 2;
121 ciMethod* mid_m = _intrinsics->at(mid)->method();
122 if (m < mid_m) {
123 hi = mid-1;
124 } else if (m > mid_m) {
125 lo = mid+1;
126 } else {
127 // look at minor sort key
128 bool mid_virt = _intrinsics->at(mid)->is_virtual();
129 if (is_virtual < mid_virt) {
130 hi = mid-1;
131 } else if (is_virtual > mid_virt) {
132 lo = mid+1;
133 } else {
134 return mid; // exact match
135 }
136 }
137 }
138 return lo; // inexact match
139 }
141 void Compile::register_intrinsic(CallGenerator* cg) {
142 if (_intrinsics == NULL) {
143 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
144 }
145 // This code is stolen from ciObjectFactory::insert.
146 // Really, GrowableArray should have methods for
147 // insert_at, remove_at, and binary_search.
148 int len = _intrinsics->length();
149 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
150 if (index == len) {
151 _intrinsics->append(cg);
152 } else {
153 #ifdef ASSERT
154 CallGenerator* oldcg = _intrinsics->at(index);
155 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
156 #endif
157 _intrinsics->append(_intrinsics->at(len-1));
158 int pos;
159 for (pos = len-2; pos >= index; pos--) {
160 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
161 }
162 _intrinsics->at_put(index, cg);
163 }
164 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
165 }
167 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
168 assert(m->is_loaded(), "don't try this on unloaded methods");
169 if (_intrinsics != NULL) {
170 int index = intrinsic_insertion_index(m, is_virtual);
171 if (index < _intrinsics->length()
172 && _intrinsics->at(index)->method() == m
173 && _intrinsics->at(index)->is_virtual() == is_virtual) {
174 return _intrinsics->at(index);
175 }
176 }
177 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
178 if (m->intrinsic_id() != vmIntrinsics::_none &&
179 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
180 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
181 if (cg != NULL) {
182 // Save it for next time:
183 register_intrinsic(cg);
184 return cg;
185 } else {
186 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
187 }
188 }
189 return NULL;
190 }
192 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
193 // in library_call.cpp.
196 #ifndef PRODUCT
197 // statistics gathering...
199 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
200 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
202 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
203 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
204 int oflags = _intrinsic_hist_flags[id];
205 assert(flags != 0, "what happened?");
206 if (is_virtual) {
207 flags |= _intrinsic_virtual;
208 }
209 bool changed = (flags != oflags);
210 if ((flags & _intrinsic_worked) != 0) {
211 juint count = (_intrinsic_hist_count[id] += 1);
212 if (count == 1) {
213 changed = true; // first time
214 }
215 // increment the overall count also:
216 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
217 }
218 if (changed) {
219 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
220 // Something changed about the intrinsic's virtuality.
221 if ((flags & _intrinsic_virtual) != 0) {
222 // This is the first use of this intrinsic as a virtual call.
223 if (oflags != 0) {
224 // We already saw it as a non-virtual, so note both cases.
225 flags |= _intrinsic_both;
226 }
227 } else if ((oflags & _intrinsic_both) == 0) {
228 // This is the first use of this intrinsic as a non-virtual
229 flags |= _intrinsic_both;
230 }
231 }
232 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
233 }
234 // update the overall flags also:
235 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
236 return changed;
237 }
239 static char* format_flags(int flags, char* buf) {
240 buf[0] = 0;
241 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
242 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
243 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
244 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
245 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
246 if (buf[0] == 0) strcat(buf, ",");
247 assert(buf[0] == ',', "must be");
248 return &buf[1];
249 }
251 void Compile::print_intrinsic_statistics() {
252 char flagsbuf[100];
253 ttyLocker ttyl;
254 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
255 tty->print_cr("Compiler intrinsic usage:");
256 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
257 if (total == 0) total = 1; // avoid div0 in case of no successes
258 #define PRINT_STAT_LINE(name, c, f) \
259 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
260 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
261 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
262 int flags = _intrinsic_hist_flags[id];
263 juint count = _intrinsic_hist_count[id];
264 if ((flags | count) != 0) {
265 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
266 }
267 }
268 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
269 if (xtty != NULL) xtty->tail("statistics");
270 }
272 void Compile::print_statistics() {
273 { ttyLocker ttyl;
274 if (xtty != NULL) xtty->head("statistics type='opto'");
275 Parse::print_statistics();
276 PhaseCCP::print_statistics();
277 PhaseRegAlloc::print_statistics();
278 Scheduling::print_statistics();
279 PhasePeephole::print_statistics();
280 PhaseIdealLoop::print_statistics();
281 if (xtty != NULL) xtty->tail("statistics");
282 }
283 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
284 // put this under its own <statistics> element.
285 print_intrinsic_statistics();
286 }
287 }
288 #endif //PRODUCT
290 // Support for bundling info
291 Bundle* Compile::node_bundling(const Node *n) {
292 assert(valid_bundle_info(n), "oob");
293 return &_node_bundling_base[n->_idx];
294 }
296 bool Compile::valid_bundle_info(const Node *n) {
297 return (_node_bundling_limit > n->_idx);
298 }
301 void Compile::gvn_replace_by(Node* n, Node* nn) {
302 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
303 Node* use = n->last_out(i);
304 bool is_in_table = initial_gvn()->hash_delete(use);
305 uint uses_found = 0;
306 for (uint j = 0; j < use->len(); j++) {
307 if (use->in(j) == n) {
308 if (j < use->req())
309 use->set_req(j, nn);
310 else
311 use->set_prec(j, nn);
312 uses_found++;
313 }
314 }
315 if (is_in_table) {
316 // reinsert into table
317 initial_gvn()->hash_find_insert(use);
318 }
319 record_for_igvn(use);
320 i -= uses_found; // we deleted 1 or more copies of this edge
321 }
322 }
325 static inline bool not_a_node(const Node* n) {
326 if (n == NULL) return true;
327 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
328 if (*(address*)n == badAddress) return true; // kill by Node::destruct
329 return false;
330 }
332 // Identify all nodes that are reachable from below, useful.
333 // Use breadth-first pass that records state in a Unique_Node_List,
334 // recursive traversal is slower.
335 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
336 int estimated_worklist_size = unique();
337 useful.map( estimated_worklist_size, NULL ); // preallocate space
339 // Initialize worklist
340 if (root() != NULL) { useful.push(root()); }
341 // If 'top' is cached, declare it useful to preserve cached node
342 if( cached_top_node() ) { useful.push(cached_top_node()); }
344 // Push all useful nodes onto the list, breadthfirst
345 for( uint next = 0; next < useful.size(); ++next ) {
346 assert( next < unique(), "Unique useful nodes < total nodes");
347 Node *n = useful.at(next);
348 uint max = n->len();
349 for( uint i = 0; i < max; ++i ) {
350 Node *m = n->in(i);
351 if (not_a_node(m)) continue;
352 useful.push(m);
353 }
354 }
355 }
357 // Update dead_node_list with any missing dead nodes using useful
358 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
359 void Compile::update_dead_node_list(Unique_Node_List &useful) {
360 uint max_idx = unique();
361 VectorSet& useful_node_set = useful.member_set();
363 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
364 // If node with index node_idx is not in useful set,
365 // mark it as dead in dead node list.
366 if (! useful_node_set.test(node_idx) ) {
367 record_dead_node(node_idx);
368 }
369 }
370 }
372 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
373 int shift = 0;
374 for (int i = 0; i < inlines->length(); i++) {
375 CallGenerator* cg = inlines->at(i);
376 CallNode* call = cg->call_node();
377 if (shift > 0) {
378 inlines->at_put(i-shift, cg);
379 }
380 if (!useful.member(call)) {
381 shift++;
382 }
383 }
384 inlines->trunc_to(inlines->length()-shift);
385 }
387 // Disconnect all useless nodes by disconnecting those at the boundary.
388 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
389 uint next = 0;
390 while (next < useful.size()) {
391 Node *n = useful.at(next++);
392 // Use raw traversal of out edges since this code removes out edges
393 int max = n->outcnt();
394 for (int j = 0; j < max; ++j) {
395 Node* child = n->raw_out(j);
396 if (! useful.member(child)) {
397 assert(!child->is_top() || child != top(),
398 "If top is cached in Compile object it is in useful list");
399 // Only need to remove this out-edge to the useless node
400 n->raw_del_out(j);
401 --j;
402 --max;
403 }
404 }
405 if (n->outcnt() == 1 && n->has_special_unique_user()) {
406 record_for_igvn(n->unique_out());
407 }
408 }
409 // Remove useless macro and predicate opaq nodes
410 for (int i = C->macro_count()-1; i >= 0; i--) {
411 Node* n = C->macro_node(i);
412 if (!useful.member(n)) {
413 remove_macro_node(n);
414 }
415 }
416 // Remove useless expensive node
417 for (int i = C->expensive_count()-1; i >= 0; i--) {
418 Node* n = C->expensive_node(i);
419 if (!useful.member(n)) {
420 remove_expensive_node(n);
421 }
422 }
423 // clean up the late inline lists
424 remove_useless_late_inlines(&_string_late_inlines, useful);
425 remove_useless_late_inlines(&_boxing_late_inlines, useful);
426 remove_useless_late_inlines(&_late_inlines, useful);
427 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
428 }
430 //------------------------------frame_size_in_words-----------------------------
431 // frame_slots in units of words
432 int Compile::frame_size_in_words() const {
433 // shift is 0 in LP32 and 1 in LP64
434 const int shift = (LogBytesPerWord - LogBytesPerInt);
435 int words = _frame_slots >> shift;
436 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
437 return words;
438 }
440 // ============================================================================
441 //------------------------------CompileWrapper---------------------------------
442 class CompileWrapper : public StackObj {
443 Compile *const _compile;
444 public:
445 CompileWrapper(Compile* compile);
447 ~CompileWrapper();
448 };
450 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
451 // the Compile* pointer is stored in the current ciEnv:
452 ciEnv* env = compile->env();
453 assert(env == ciEnv::current(), "must already be a ciEnv active");
454 assert(env->compiler_data() == NULL, "compile already active?");
455 env->set_compiler_data(compile);
456 assert(compile == Compile::current(), "sanity");
458 compile->set_type_dict(NULL);
459 compile->set_type_hwm(NULL);
460 compile->set_type_last_size(0);
461 compile->set_last_tf(NULL, NULL);
462 compile->set_indexSet_arena(NULL);
463 compile->set_indexSet_free_block_list(NULL);
464 compile->init_type_arena();
465 Type::Initialize(compile);
466 _compile->set_scratch_buffer_blob(NULL);
467 _compile->begin_method();
468 }
469 CompileWrapper::~CompileWrapper() {
470 _compile->end_method();
471 if (_compile->scratch_buffer_blob() != NULL)
472 BufferBlob::free(_compile->scratch_buffer_blob());
473 _compile->env()->set_compiler_data(NULL);
474 }
477 //----------------------------print_compile_messages---------------------------
478 void Compile::print_compile_messages() {
479 #ifndef PRODUCT
480 // Check if recompiling
481 if (_subsume_loads == false && PrintOpto) {
482 // Recompiling without allowing machine instructions to subsume loads
483 tty->print_cr("*********************************************************");
484 tty->print_cr("** Bailout: Recompile without subsuming loads **");
485 tty->print_cr("*********************************************************");
486 }
487 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
488 // Recompiling without escape analysis
489 tty->print_cr("*********************************************************");
490 tty->print_cr("** Bailout: Recompile without escape analysis **");
491 tty->print_cr("*********************************************************");
492 }
493 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
494 // Recompiling without boxing elimination
495 tty->print_cr("*********************************************************");
496 tty->print_cr("** Bailout: Recompile without boxing elimination **");
497 tty->print_cr("*********************************************************");
498 }
499 if (env()->break_at_compile()) {
500 // Open the debugger when compiling this method.
501 tty->print("### Breaking when compiling: ");
502 method()->print_short_name();
503 tty->cr();
504 BREAKPOINT;
505 }
507 if( PrintOpto ) {
508 if (is_osr_compilation()) {
509 tty->print("[OSR]%3d", _compile_id);
510 } else {
511 tty->print("%3d", _compile_id);
512 }
513 }
514 #endif
515 }
518 //-----------------------init_scratch_buffer_blob------------------------------
519 // Construct a temporary BufferBlob and cache it for this compile.
520 void Compile::init_scratch_buffer_blob(int const_size) {
521 // If there is already a scratch buffer blob allocated and the
522 // constant section is big enough, use it. Otherwise free the
523 // current and allocate a new one.
524 BufferBlob* blob = scratch_buffer_blob();
525 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
526 // Use the current blob.
527 } else {
528 if (blob != NULL) {
529 BufferBlob::free(blob);
530 }
532 ResourceMark rm;
533 _scratch_const_size = const_size;
534 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
535 blob = BufferBlob::create("Compile::scratch_buffer", size);
536 // Record the buffer blob for next time.
537 set_scratch_buffer_blob(blob);
538 // Have we run out of code space?
539 if (scratch_buffer_blob() == NULL) {
540 // Let CompilerBroker disable further compilations.
541 record_failure("Not enough space for scratch buffer in CodeCache");
542 return;
543 }
544 }
546 // Initialize the relocation buffers
547 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
548 set_scratch_locs_memory(locs_buf);
549 }
552 //-----------------------scratch_emit_size-------------------------------------
553 // Helper function that computes size by emitting code
554 uint Compile::scratch_emit_size(const Node* n) {
555 // Start scratch_emit_size section.
556 set_in_scratch_emit_size(true);
558 // Emit into a trash buffer and count bytes emitted.
559 // This is a pretty expensive way to compute a size,
560 // but it works well enough if seldom used.
561 // All common fixed-size instructions are given a size
562 // method by the AD file.
563 // Note that the scratch buffer blob and locs memory are
564 // allocated at the beginning of the compile task, and
565 // may be shared by several calls to scratch_emit_size.
566 // The allocation of the scratch buffer blob is particularly
567 // expensive, since it has to grab the code cache lock.
568 BufferBlob* blob = this->scratch_buffer_blob();
569 assert(blob != NULL, "Initialize BufferBlob at start");
570 assert(blob->size() > MAX_inst_size, "sanity");
571 relocInfo* locs_buf = scratch_locs_memory();
572 address blob_begin = blob->content_begin();
573 address blob_end = (address)locs_buf;
574 assert(blob->content_contains(blob_end), "sanity");
575 CodeBuffer buf(blob_begin, blob_end - blob_begin);
576 buf.initialize_consts_size(_scratch_const_size);
577 buf.initialize_stubs_size(MAX_stubs_size);
578 assert(locs_buf != NULL, "sanity");
579 int lsize = MAX_locs_size / 3;
580 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
581 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
582 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
584 // Do the emission.
586 Label fakeL; // Fake label for branch instructions.
587 Label* saveL = NULL;
588 uint save_bnum = 0;
589 bool is_branch = n->is_MachBranch();
590 if (is_branch) {
591 MacroAssembler masm(&buf);
592 masm.bind(fakeL);
593 n->as_MachBranch()->save_label(&saveL, &save_bnum);
594 n->as_MachBranch()->label_set(&fakeL, 0);
595 }
596 n->emit(buf, this->regalloc());
597 if (is_branch) // Restore label.
598 n->as_MachBranch()->label_set(saveL, save_bnum);
600 // End scratch_emit_size section.
601 set_in_scratch_emit_size(false);
603 return buf.insts_size();
604 }
607 // ============================================================================
608 //------------------------------Compile standard-------------------------------
609 debug_only( int Compile::_debug_idx = 100000; )
611 // Compile a method. entry_bci is -1 for normal compilations and indicates
612 // the continuation bci for on stack replacement.
615 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
616 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
617 : Phase(Compiler),
618 _env(ci_env),
619 _log(ci_env->log()),
620 _compile_id(ci_env->compile_id()),
621 _save_argument_registers(false),
622 _stub_name(NULL),
623 _stub_function(NULL),
624 _stub_entry_point(NULL),
625 _method(target),
626 _entry_bci(osr_bci),
627 _initial_gvn(NULL),
628 _for_igvn(NULL),
629 _warm_calls(NULL),
630 _subsume_loads(subsume_loads),
631 _do_escape_analysis(do_escape_analysis),
632 _eliminate_boxing(eliminate_boxing),
633 _failure_reason(NULL),
634 _code_buffer("Compile::Fill_buffer"),
635 _orig_pc_slot(0),
636 _orig_pc_slot_offset_in_bytes(0),
637 _has_method_handle_invokes(false),
638 _mach_constant_base_node(NULL),
639 _node_bundling_limit(0),
640 _node_bundling_base(NULL),
641 _java_calls(0),
642 _inner_loops(0),
643 _scratch_const_size(-1),
644 _in_scratch_emit_size(false),
645 _dead_node_list(comp_arena()),
646 _dead_node_count(0),
647 #ifndef PRODUCT
648 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
649 _printer(IdealGraphPrinter::printer()),
650 #endif
651 _congraph(NULL),
652 _late_inlines(comp_arena(), 2, 0, NULL),
653 _string_late_inlines(comp_arena(), 2, 0, NULL),
654 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
655 _late_inlines_pos(0),
656 _number_of_mh_late_inlines(0),
657 _inlining_progress(false),
658 _inlining_incrementally(false),
659 _print_inlining_list(NULL),
660 _print_inlining(0) {
661 C = this;
663 CompileWrapper cw(this);
664 #ifndef PRODUCT
665 if (TimeCompiler2) {
666 tty->print(" ");
667 target->holder()->name()->print();
668 tty->print(".");
669 target->print_short_name();
670 tty->print(" ");
671 }
672 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
673 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
674 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
675 if (!print_opto_assembly) {
676 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
677 if (print_assembly && !Disassembler::can_decode()) {
678 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
679 print_opto_assembly = true;
680 }
681 }
682 set_print_assembly(print_opto_assembly);
683 set_parsed_irreducible_loop(false);
684 #endif
686 if (ProfileTraps) {
687 // Make sure the method being compiled gets its own MDO,
688 // so we can at least track the decompile_count().
689 method()->ensure_method_data();
690 }
692 Init(::AliasLevel);
695 print_compile_messages();
697 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
698 _ilt = InlineTree::build_inline_tree_root();
699 else
700 _ilt = NULL;
702 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
703 assert(num_alias_types() >= AliasIdxRaw, "");
705 #define MINIMUM_NODE_HASH 1023
706 // Node list that Iterative GVN will start with
707 Unique_Node_List for_igvn(comp_arena());
708 set_for_igvn(&for_igvn);
710 // GVN that will be run immediately on new nodes
711 uint estimated_size = method()->code_size()*4+64;
712 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
713 PhaseGVN gvn(node_arena(), estimated_size);
714 set_initial_gvn(&gvn);
716 if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) {
717 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
718 }
719 { // Scope for timing the parser
720 TracePhase t3("parse", &_t_parser, true);
722 // Put top into the hash table ASAP.
723 initial_gvn()->transform_no_reclaim(top());
725 // Set up tf(), start(), and find a CallGenerator.
726 CallGenerator* cg = NULL;
727 if (is_osr_compilation()) {
728 const TypeTuple *domain = StartOSRNode::osr_domain();
729 const TypeTuple *range = TypeTuple::make_range(method()->signature());
730 init_tf(TypeFunc::make(domain, range));
731 StartNode* s = new (this) StartOSRNode(root(), domain);
732 initial_gvn()->set_type_bottom(s);
733 init_start(s);
734 cg = CallGenerator::for_osr(method(), entry_bci());
735 } else {
736 // Normal case.
737 init_tf(TypeFunc::make(method()));
738 StartNode* s = new (this) StartNode(root(), tf()->domain());
739 initial_gvn()->set_type_bottom(s);
740 init_start(s);
741 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
742 // With java.lang.ref.reference.get() we must go through the
743 // intrinsic when G1 is enabled - even when get() is the root
744 // method of the compile - so that, if necessary, the value in
745 // the referent field of the reference object gets recorded by
746 // the pre-barrier code.
747 // Specifically, if G1 is enabled, the value in the referent
748 // field is recorded by the G1 SATB pre barrier. This will
749 // result in the referent being marked live and the reference
750 // object removed from the list of discovered references during
751 // reference processing.
752 cg = find_intrinsic(method(), false);
753 }
754 if (cg == NULL) {
755 float past_uses = method()->interpreter_invocation_count();
756 float expected_uses = past_uses;
757 cg = CallGenerator::for_inline(method(), expected_uses);
758 }
759 }
760 if (failing()) return;
761 if (cg == NULL) {
762 record_method_not_compilable_all_tiers("cannot parse method");
763 return;
764 }
765 JVMState* jvms = build_start_state(start(), tf());
766 if ((jvms = cg->generate(jvms)) == NULL) {
767 record_method_not_compilable("method parse failed");
768 return;
769 }
770 GraphKit kit(jvms);
772 if (!kit.stopped()) {
773 // Accept return values, and transfer control we know not where.
774 // This is done by a special, unique ReturnNode bound to root.
775 return_values(kit.jvms());
776 }
778 if (kit.has_exceptions()) {
779 // Any exceptions that escape from this call must be rethrown
780 // to whatever caller is dynamically above us on the stack.
781 // This is done by a special, unique RethrowNode bound to root.
782 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
783 }
785 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
787 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
788 inline_string_calls(true);
789 }
791 if (failing()) return;
793 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
795 // Remove clutter produced by parsing.
796 if (!failing()) {
797 ResourceMark rm;
798 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
799 }
800 }
802 // Note: Large methods are capped off in do_one_bytecode().
803 if (failing()) return;
805 // After parsing, node notes are no longer automagic.
806 // They must be propagated by register_new_node_with_optimizer(),
807 // clone(), or the like.
808 set_default_node_notes(NULL);
810 for (;;) {
811 int successes = Inline_Warm();
812 if (failing()) return;
813 if (successes == 0) break;
814 }
816 // Drain the list.
817 Finish_Warm();
818 #ifndef PRODUCT
819 if (_printer) {
820 _printer->print_inlining(this);
821 }
822 #endif
824 if (failing()) return;
825 NOT_PRODUCT( verify_graph_edges(); )
827 // Now optimize
828 Optimize();
829 if (failing()) return;
830 NOT_PRODUCT( verify_graph_edges(); )
832 #ifndef PRODUCT
833 if (PrintIdeal) {
834 ttyLocker ttyl; // keep the following output all in one block
835 // This output goes directly to the tty, not the compiler log.
836 // To enable tools to match it up with the compilation activity,
837 // be sure to tag this tty output with the compile ID.
838 if (xtty != NULL) {
839 xtty->head("ideal compile_id='%d'%s", compile_id(),
840 is_osr_compilation() ? " compile_kind='osr'" :
841 "");
842 }
843 root()->dump(9999);
844 if (xtty != NULL) {
845 xtty->tail("ideal");
846 }
847 }
848 #endif
850 // Now that we know the size of all the monitors we can add a fixed slot
851 // for the original deopt pc.
853 _orig_pc_slot = fixed_slots();
854 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
855 set_fixed_slots(next_slot);
857 // Now generate code
858 Code_Gen();
859 if (failing()) return;
861 // Check if we want to skip execution of all compiled code.
862 {
863 #ifndef PRODUCT
864 if (OptoNoExecute) {
865 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
866 return;
867 }
868 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
869 #endif
871 if (is_osr_compilation()) {
872 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
873 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
874 } else {
875 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
876 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
877 }
879 env()->register_method(_method, _entry_bci,
880 &_code_offsets,
881 _orig_pc_slot_offset_in_bytes,
882 code_buffer(),
883 frame_size_in_words(), _oop_map_set,
884 &_handler_table, &_inc_table,
885 compiler,
886 env()->comp_level(),
887 has_unsafe_access(),
888 SharedRuntime::is_wide_vector(max_vector_size())
889 );
891 if (log() != NULL) // Print code cache state into compiler log
892 log()->code_cache_state();
893 }
894 }
896 //------------------------------Compile----------------------------------------
897 // Compile a runtime stub
898 Compile::Compile( ciEnv* ci_env,
899 TypeFunc_generator generator,
900 address stub_function,
901 const char *stub_name,
902 int is_fancy_jump,
903 bool pass_tls,
904 bool save_arg_registers,
905 bool return_pc )
906 : Phase(Compiler),
907 _env(ci_env),
908 _log(ci_env->log()),
909 _compile_id(0),
910 _save_argument_registers(save_arg_registers),
911 _method(NULL),
912 _stub_name(stub_name),
913 _stub_function(stub_function),
914 _stub_entry_point(NULL),
915 _entry_bci(InvocationEntryBci),
916 _initial_gvn(NULL),
917 _for_igvn(NULL),
918 _warm_calls(NULL),
919 _orig_pc_slot(0),
920 _orig_pc_slot_offset_in_bytes(0),
921 _subsume_loads(true),
922 _do_escape_analysis(false),
923 _eliminate_boxing(false),
924 _failure_reason(NULL),
925 _code_buffer("Compile::Fill_buffer"),
926 _has_method_handle_invokes(false),
927 _mach_constant_base_node(NULL),
928 _node_bundling_limit(0),
929 _node_bundling_base(NULL),
930 _java_calls(0),
931 _inner_loops(0),
932 #ifndef PRODUCT
933 _trace_opto_output(TraceOptoOutput),
934 _printer(NULL),
935 #endif
936 _dead_node_list(comp_arena()),
937 _dead_node_count(0),
938 _congraph(NULL),
939 _number_of_mh_late_inlines(0),
940 _inlining_progress(false),
941 _inlining_incrementally(false),
942 _print_inlining_list(NULL),
943 _print_inlining(0) {
944 C = this;
946 #ifndef PRODUCT
947 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
948 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
949 set_print_assembly(PrintFrameConverterAssembly);
950 set_parsed_irreducible_loop(false);
951 #endif
952 CompileWrapper cw(this);
953 Init(/*AliasLevel=*/ 0);
954 init_tf((*generator)());
956 {
957 // The following is a dummy for the sake of GraphKit::gen_stub
958 Unique_Node_List for_igvn(comp_arena());
959 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
960 PhaseGVN gvn(Thread::current()->resource_area(),255);
961 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
962 gvn.transform_no_reclaim(top());
964 GraphKit kit;
965 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
966 }
968 NOT_PRODUCT( verify_graph_edges(); )
969 Code_Gen();
970 if (failing()) return;
973 // Entry point will be accessed using compile->stub_entry_point();
974 if (code_buffer() == NULL) {
975 Matcher::soft_match_failure();
976 } else {
977 if (PrintAssembly && (WizardMode || Verbose))
978 tty->print_cr("### Stub::%s", stub_name);
980 if (!failing()) {
981 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
983 // Make the NMethod
984 // For now we mark the frame as never safe for profile stackwalking
985 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
986 code_buffer(),
987 CodeOffsets::frame_never_safe,
988 // _code_offsets.value(CodeOffsets::Frame_Complete),
989 frame_size_in_words(),
990 _oop_map_set,
991 save_arg_registers);
992 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
994 _stub_entry_point = rs->entry_point();
995 }
996 }
997 }
999 //------------------------------Init-------------------------------------------
1000 // Prepare for a single compilation
1001 void Compile::Init(int aliaslevel) {
1002 _unique = 0;
1003 _regalloc = NULL;
1005 _tf = NULL; // filled in later
1006 _top = NULL; // cached later
1007 _matcher = NULL; // filled in later
1008 _cfg = NULL; // filled in later
1010 set_24_bit_selection_and_mode(Use24BitFP, false);
1012 _node_note_array = NULL;
1013 _default_node_notes = NULL;
1015 _immutable_memory = NULL; // filled in at first inquiry
1017 // Globally visible Nodes
1018 // First set TOP to NULL to give safe behavior during creation of RootNode
1019 set_cached_top_node(NULL);
1020 set_root(new (this) RootNode());
1021 // Now that you have a Root to point to, create the real TOP
1022 set_cached_top_node( new (this) ConNode(Type::TOP) );
1023 set_recent_alloc(NULL, NULL);
1025 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1026 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1027 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1028 env()->set_dependencies(new Dependencies(env()));
1030 _fixed_slots = 0;
1031 set_has_split_ifs(false);
1032 set_has_loops(has_method() && method()->has_loops()); // first approximation
1033 set_has_stringbuilder(false);
1034 set_has_boxed_value(false);
1035 _trap_can_recompile = false; // no traps emitted yet
1036 _major_progress = true; // start out assuming good things will happen
1037 set_has_unsafe_access(false);
1038 set_max_vector_size(0);
1039 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1040 set_decompile_count(0);
1042 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1043 set_num_loop_opts(LoopOptsCount);
1044 set_do_inlining(Inline);
1045 set_max_inline_size(MaxInlineSize);
1046 set_freq_inline_size(FreqInlineSize);
1047 set_do_scheduling(OptoScheduling);
1048 set_do_count_invocations(false);
1049 set_do_method_data_update(false);
1051 if (debug_info()->recording_non_safepoints()) {
1052 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1053 (comp_arena(), 8, 0, NULL));
1054 set_default_node_notes(Node_Notes::make(this));
1055 }
1057 // // -- Initialize types before each compile --
1058 // // Update cached type information
1059 // if( _method && _method->constants() )
1060 // Type::update_loaded_types(_method, _method->constants());
1062 // Init alias_type map.
1063 if (!_do_escape_analysis && aliaslevel == 3)
1064 aliaslevel = 2; // No unique types without escape analysis
1065 _AliasLevel = aliaslevel;
1066 const int grow_ats = 16;
1067 _max_alias_types = grow_ats;
1068 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1069 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1070 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1071 {
1072 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1073 }
1074 // Initialize the first few types.
1075 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1076 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1077 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1078 _num_alias_types = AliasIdxRaw+1;
1079 // Zero out the alias type cache.
1080 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1081 // A NULL adr_type hits in the cache right away. Preload the right answer.
1082 probe_alias_cache(NULL)->_index = AliasIdxTop;
1084 _intrinsics = NULL;
1085 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1086 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1087 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1088 register_library_intrinsics();
1089 }
1091 //---------------------------init_start----------------------------------------
1092 // Install the StartNode on this compile object.
1093 void Compile::init_start(StartNode* s) {
1094 if (failing())
1095 return; // already failing
1096 assert(s == start(), "");
1097 }
1099 StartNode* Compile::start() const {
1100 assert(!failing(), "");
1101 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1102 Node* start = root()->fast_out(i);
1103 if( start->is_Start() )
1104 return start->as_Start();
1105 }
1106 ShouldNotReachHere();
1107 return NULL;
1108 }
1110 //-------------------------------immutable_memory-------------------------------------
1111 // Access immutable memory
1112 Node* Compile::immutable_memory() {
1113 if (_immutable_memory != NULL) {
1114 return _immutable_memory;
1115 }
1116 StartNode* s = start();
1117 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1118 Node *p = s->fast_out(i);
1119 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1120 _immutable_memory = p;
1121 return _immutable_memory;
1122 }
1123 }
1124 ShouldNotReachHere();
1125 return NULL;
1126 }
1128 //----------------------set_cached_top_node------------------------------------
1129 // Install the cached top node, and make sure Node::is_top works correctly.
1130 void Compile::set_cached_top_node(Node* tn) {
1131 if (tn != NULL) verify_top(tn);
1132 Node* old_top = _top;
1133 _top = tn;
1134 // Calling Node::setup_is_top allows the nodes the chance to adjust
1135 // their _out arrays.
1136 if (_top != NULL) _top->setup_is_top();
1137 if (old_top != NULL) old_top->setup_is_top();
1138 assert(_top == NULL || top()->is_top(), "");
1139 }
1141 #ifdef ASSERT
1142 uint Compile::count_live_nodes_by_graph_walk() {
1143 Unique_Node_List useful(comp_arena());
1144 // Get useful node list by walking the graph.
1145 identify_useful_nodes(useful);
1146 return useful.size();
1147 }
1149 void Compile::print_missing_nodes() {
1151 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1152 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1153 return;
1154 }
1156 // This is an expensive function. It is executed only when the user
1157 // specifies VerifyIdealNodeCount option or otherwise knows the
1158 // additional work that needs to be done to identify reachable nodes
1159 // by walking the flow graph and find the missing ones using
1160 // _dead_node_list.
1162 Unique_Node_List useful(comp_arena());
1163 // Get useful node list by walking the graph.
1164 identify_useful_nodes(useful);
1166 uint l_nodes = C->live_nodes();
1167 uint l_nodes_by_walk = useful.size();
1169 if (l_nodes != l_nodes_by_walk) {
1170 if (_log != NULL) {
1171 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1172 _log->stamp();
1173 _log->end_head();
1174 }
1175 VectorSet& useful_member_set = useful.member_set();
1176 int last_idx = l_nodes_by_walk;
1177 for (int i = 0; i < last_idx; i++) {
1178 if (useful_member_set.test(i)) {
1179 if (_dead_node_list.test(i)) {
1180 if (_log != NULL) {
1181 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1182 }
1183 if (PrintIdealNodeCount) {
1184 // Print the log message to tty
1185 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1186 useful.at(i)->dump();
1187 }
1188 }
1189 }
1190 else if (! _dead_node_list.test(i)) {
1191 if (_log != NULL) {
1192 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1193 }
1194 if (PrintIdealNodeCount) {
1195 // Print the log message to tty
1196 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1197 }
1198 }
1199 }
1200 if (_log != NULL) {
1201 _log->tail("mismatched_nodes");
1202 }
1203 }
1204 }
1205 #endif
1207 #ifndef PRODUCT
1208 void Compile::verify_top(Node* tn) const {
1209 if (tn != NULL) {
1210 assert(tn->is_Con(), "top node must be a constant");
1211 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1212 assert(tn->in(0) != NULL, "must have live top node");
1213 }
1214 }
1215 #endif
1218 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1220 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1221 guarantee(arr != NULL, "");
1222 int num_blocks = arr->length();
1223 if (grow_by < num_blocks) grow_by = num_blocks;
1224 int num_notes = grow_by * _node_notes_block_size;
1225 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1226 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1227 while (num_notes > 0) {
1228 arr->append(notes);
1229 notes += _node_notes_block_size;
1230 num_notes -= _node_notes_block_size;
1231 }
1232 assert(num_notes == 0, "exact multiple, please");
1233 }
1235 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1236 if (source == NULL || dest == NULL) return false;
1238 if (dest->is_Con())
1239 return false; // Do not push debug info onto constants.
1241 #ifdef ASSERT
1242 // Leave a bread crumb trail pointing to the original node:
1243 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1244 dest->set_debug_orig(source);
1245 }
1246 #endif
1248 if (node_note_array() == NULL)
1249 return false; // Not collecting any notes now.
1251 // This is a copy onto a pre-existing node, which may already have notes.
1252 // If both nodes have notes, do not overwrite any pre-existing notes.
1253 Node_Notes* source_notes = node_notes_at(source->_idx);
1254 if (source_notes == NULL || source_notes->is_clear()) return false;
1255 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1256 if (dest_notes == NULL || dest_notes->is_clear()) {
1257 return set_node_notes_at(dest->_idx, source_notes);
1258 }
1260 Node_Notes merged_notes = (*source_notes);
1261 // The order of operations here ensures that dest notes will win...
1262 merged_notes.update_from(dest_notes);
1263 return set_node_notes_at(dest->_idx, &merged_notes);
1264 }
1267 //--------------------------allow_range_check_smearing-------------------------
1268 // Gating condition for coalescing similar range checks.
1269 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1270 // single covering check that is at least as strong as any of them.
1271 // If the optimization succeeds, the simplified (strengthened) range check
1272 // will always succeed. If it fails, we will deopt, and then give up
1273 // on the optimization.
1274 bool Compile::allow_range_check_smearing() const {
1275 // If this method has already thrown a range-check,
1276 // assume it was because we already tried range smearing
1277 // and it failed.
1278 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1279 return !already_trapped;
1280 }
1283 //------------------------------flatten_alias_type-----------------------------
1284 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1285 int offset = tj->offset();
1286 TypePtr::PTR ptr = tj->ptr();
1288 // Known instance (scalarizable allocation) alias only with itself.
1289 bool is_known_inst = tj->isa_oopptr() != NULL &&
1290 tj->is_oopptr()->is_known_instance();
1292 // Process weird unsafe references.
1293 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1294 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1295 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1296 tj = TypeOopPtr::BOTTOM;
1297 ptr = tj->ptr();
1298 offset = tj->offset();
1299 }
1301 // Array pointers need some flattening
1302 const TypeAryPtr *ta = tj->isa_aryptr();
1303 if( ta && is_known_inst ) {
1304 if ( offset != Type::OffsetBot &&
1305 offset > arrayOopDesc::length_offset_in_bytes() ) {
1306 offset = Type::OffsetBot; // Flatten constant access into array body only
1307 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1308 }
1309 } else if( ta && _AliasLevel >= 2 ) {
1310 // For arrays indexed by constant indices, we flatten the alias
1311 // space to include all of the array body. Only the header, klass
1312 // and array length can be accessed un-aliased.
1313 if( offset != Type::OffsetBot ) {
1314 if( ta->const_oop() ) { // MethodData* or Method*
1315 offset = Type::OffsetBot; // Flatten constant access into array body
1316 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1317 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1318 // range is OK as-is.
1319 tj = ta = TypeAryPtr::RANGE;
1320 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1321 tj = TypeInstPtr::KLASS; // all klass loads look alike
1322 ta = TypeAryPtr::RANGE; // generic ignored junk
1323 ptr = TypePtr::BotPTR;
1324 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1325 tj = TypeInstPtr::MARK;
1326 ta = TypeAryPtr::RANGE; // generic ignored junk
1327 ptr = TypePtr::BotPTR;
1328 } else { // Random constant offset into array body
1329 offset = Type::OffsetBot; // Flatten constant access into array body
1330 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1331 }
1332 }
1333 // Arrays of fixed size alias with arrays of unknown size.
1334 if (ta->size() != TypeInt::POS) {
1335 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1336 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1337 }
1338 // Arrays of known objects become arrays of unknown objects.
1339 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1340 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1341 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1342 }
1343 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1344 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1345 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1346 }
1347 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1348 // cannot be distinguished by bytecode alone.
1349 if (ta->elem() == TypeInt::BOOL) {
1350 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1351 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1352 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1353 }
1354 // During the 2nd round of IterGVN, NotNull castings are removed.
1355 // Make sure the Bottom and NotNull variants alias the same.
1356 // Also, make sure exact and non-exact variants alias the same.
1357 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1358 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1359 }
1360 }
1362 // Oop pointers need some flattening
1363 const TypeInstPtr *to = tj->isa_instptr();
1364 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1365 ciInstanceKlass *k = to->klass()->as_instance_klass();
1366 if( ptr == TypePtr::Constant ) {
1367 if (to->klass() != ciEnv::current()->Class_klass() ||
1368 offset < k->size_helper() * wordSize) {
1369 // No constant oop pointers (such as Strings); they alias with
1370 // unknown strings.
1371 assert(!is_known_inst, "not scalarizable allocation");
1372 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1373 }
1374 } else if( is_known_inst ) {
1375 tj = to; // Keep NotNull and klass_is_exact for instance type
1376 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1377 // During the 2nd round of IterGVN, NotNull castings are removed.
1378 // Make sure the Bottom and NotNull variants alias the same.
1379 // Also, make sure exact and non-exact variants alias the same.
1380 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1381 }
1382 // Canonicalize the holder of this field
1383 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1384 // First handle header references such as a LoadKlassNode, even if the
1385 // object's klass is unloaded at compile time (4965979).
1386 if (!is_known_inst) { // Do it only for non-instance types
1387 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1388 }
1389 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1390 // Static fields are in the space above the normal instance
1391 // fields in the java.lang.Class instance.
1392 if (to->klass() != ciEnv::current()->Class_klass()) {
1393 to = NULL;
1394 tj = TypeOopPtr::BOTTOM;
1395 offset = tj->offset();
1396 }
1397 } else {
1398 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1399 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1400 if( is_known_inst ) {
1401 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1402 } else {
1403 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1404 }
1405 }
1406 }
1407 }
1409 // Klass pointers to object array klasses need some flattening
1410 const TypeKlassPtr *tk = tj->isa_klassptr();
1411 if( tk ) {
1412 // If we are referencing a field within a Klass, we need
1413 // to assume the worst case of an Object. Both exact and
1414 // inexact types must flatten to the same alias class so
1415 // use NotNull as the PTR.
1416 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1418 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1419 TypeKlassPtr::OBJECT->klass(),
1420 offset);
1421 }
1423 ciKlass* klass = tk->klass();
1424 if( klass->is_obj_array_klass() ) {
1425 ciKlass* k = TypeAryPtr::OOPS->klass();
1426 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1427 k = TypeInstPtr::BOTTOM->klass();
1428 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1429 }
1431 // Check for precise loads from the primary supertype array and force them
1432 // to the supertype cache alias index. Check for generic array loads from
1433 // the primary supertype array and also force them to the supertype cache
1434 // alias index. Since the same load can reach both, we need to merge
1435 // these 2 disparate memories into the same alias class. Since the
1436 // primary supertype array is read-only, there's no chance of confusion
1437 // where we bypass an array load and an array store.
1438 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1439 if (offset == Type::OffsetBot ||
1440 (offset >= primary_supers_offset &&
1441 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1442 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1443 offset = in_bytes(Klass::secondary_super_cache_offset());
1444 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1445 }
1446 }
1448 // Flatten all Raw pointers together.
1449 if (tj->base() == Type::RawPtr)
1450 tj = TypeRawPtr::BOTTOM;
1452 if (tj->base() == Type::AnyPtr)
1453 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1455 // Flatten all to bottom for now
1456 switch( _AliasLevel ) {
1457 case 0:
1458 tj = TypePtr::BOTTOM;
1459 break;
1460 case 1: // Flatten to: oop, static, field or array
1461 switch (tj->base()) {
1462 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1463 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1464 case Type::AryPtr: // do not distinguish arrays at all
1465 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1466 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1467 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1468 default: ShouldNotReachHere();
1469 }
1470 break;
1471 case 2: // No collapsing at level 2; keep all splits
1472 case 3: // No collapsing at level 3; keep all splits
1473 break;
1474 default:
1475 Unimplemented();
1476 }
1478 offset = tj->offset();
1479 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1481 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1482 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1483 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1484 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1485 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1486 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1487 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1488 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1489 assert( tj->ptr() != TypePtr::TopPTR &&
1490 tj->ptr() != TypePtr::AnyNull &&
1491 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1492 // assert( tj->ptr() != TypePtr::Constant ||
1493 // tj->base() == Type::RawPtr ||
1494 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1496 return tj;
1497 }
1499 void Compile::AliasType::Init(int i, const TypePtr* at) {
1500 _index = i;
1501 _adr_type = at;
1502 _field = NULL;
1503 _is_rewritable = true; // default
1504 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1505 if (atoop != NULL && atoop->is_known_instance()) {
1506 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1507 _general_index = Compile::current()->get_alias_index(gt);
1508 } else {
1509 _general_index = 0;
1510 }
1511 }
1513 //---------------------------------print_on------------------------------------
1514 #ifndef PRODUCT
1515 void Compile::AliasType::print_on(outputStream* st) {
1516 if (index() < 10)
1517 st->print("@ <%d> ", index());
1518 else st->print("@ <%d>", index());
1519 st->print(is_rewritable() ? " " : " RO");
1520 int offset = adr_type()->offset();
1521 if (offset == Type::OffsetBot)
1522 st->print(" +any");
1523 else st->print(" +%-3d", offset);
1524 st->print(" in ");
1525 adr_type()->dump_on(st);
1526 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1527 if (field() != NULL && tjp) {
1528 if (tjp->klass() != field()->holder() ||
1529 tjp->offset() != field()->offset_in_bytes()) {
1530 st->print(" != ");
1531 field()->print();
1532 st->print(" ***");
1533 }
1534 }
1535 }
1537 void print_alias_types() {
1538 Compile* C = Compile::current();
1539 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1540 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1541 C->alias_type(idx)->print_on(tty);
1542 tty->cr();
1543 }
1544 }
1545 #endif
1548 //----------------------------probe_alias_cache--------------------------------
1549 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1550 intptr_t key = (intptr_t) adr_type;
1551 key ^= key >> logAliasCacheSize;
1552 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1553 }
1556 //-----------------------------grow_alias_types--------------------------------
1557 void Compile::grow_alias_types() {
1558 const int old_ats = _max_alias_types; // how many before?
1559 const int new_ats = old_ats; // how many more?
1560 const int grow_ats = old_ats+new_ats; // how many now?
1561 _max_alias_types = grow_ats;
1562 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1563 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1564 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1565 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1566 }
1569 //--------------------------------find_alias_type------------------------------
1570 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1571 if (_AliasLevel == 0)
1572 return alias_type(AliasIdxBot);
1574 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1575 if (ace->_adr_type == adr_type) {
1576 return alias_type(ace->_index);
1577 }
1579 // Handle special cases.
1580 if (adr_type == NULL) return alias_type(AliasIdxTop);
1581 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1583 // Do it the slow way.
1584 const TypePtr* flat = flatten_alias_type(adr_type);
1586 #ifdef ASSERT
1587 assert(flat == flatten_alias_type(flat), "idempotent");
1588 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1589 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1590 const TypeOopPtr* foop = flat->is_oopptr();
1591 // Scalarizable allocations have exact klass always.
1592 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1593 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1594 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1595 }
1596 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1597 #endif
1599 int idx = AliasIdxTop;
1600 for (int i = 0; i < num_alias_types(); i++) {
1601 if (alias_type(i)->adr_type() == flat) {
1602 idx = i;
1603 break;
1604 }
1605 }
1607 if (idx == AliasIdxTop) {
1608 if (no_create) return NULL;
1609 // Grow the array if necessary.
1610 if (_num_alias_types == _max_alias_types) grow_alias_types();
1611 // Add a new alias type.
1612 idx = _num_alias_types++;
1613 _alias_types[idx]->Init(idx, flat);
1614 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1615 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1616 if (flat->isa_instptr()) {
1617 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1618 && flat->is_instptr()->klass() == env()->Class_klass())
1619 alias_type(idx)->set_rewritable(false);
1620 }
1621 if (flat->isa_klassptr()) {
1622 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1623 alias_type(idx)->set_rewritable(false);
1624 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1625 alias_type(idx)->set_rewritable(false);
1626 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1627 alias_type(idx)->set_rewritable(false);
1628 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1629 alias_type(idx)->set_rewritable(false);
1630 }
1631 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1632 // but the base pointer type is not distinctive enough to identify
1633 // references into JavaThread.)
1635 // Check for final fields.
1636 const TypeInstPtr* tinst = flat->isa_instptr();
1637 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1638 ciField* field;
1639 if (tinst->const_oop() != NULL &&
1640 tinst->klass() == ciEnv::current()->Class_klass() &&
1641 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1642 // static field
1643 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1644 field = k->get_field_by_offset(tinst->offset(), true);
1645 } else {
1646 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1647 field = k->get_field_by_offset(tinst->offset(), false);
1648 }
1649 assert(field == NULL ||
1650 original_field == NULL ||
1651 (field->holder() == original_field->holder() &&
1652 field->offset() == original_field->offset() &&
1653 field->is_static() == original_field->is_static()), "wrong field?");
1654 // Set field() and is_rewritable() attributes.
1655 if (field != NULL) alias_type(idx)->set_field(field);
1656 }
1657 }
1659 // Fill the cache for next time.
1660 ace->_adr_type = adr_type;
1661 ace->_index = idx;
1662 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1664 // Might as well try to fill the cache for the flattened version, too.
1665 AliasCacheEntry* face = probe_alias_cache(flat);
1666 if (face->_adr_type == NULL) {
1667 face->_adr_type = flat;
1668 face->_index = idx;
1669 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1670 }
1672 return alias_type(idx);
1673 }
1676 Compile::AliasType* Compile::alias_type(ciField* field) {
1677 const TypeOopPtr* t;
1678 if (field->is_static())
1679 t = TypeInstPtr::make(field->holder()->java_mirror());
1680 else
1681 t = TypeOopPtr::make_from_klass_raw(field->holder());
1682 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1683 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1684 return atp;
1685 }
1688 //------------------------------have_alias_type--------------------------------
1689 bool Compile::have_alias_type(const TypePtr* adr_type) {
1690 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1691 if (ace->_adr_type == adr_type) {
1692 return true;
1693 }
1695 // Handle special cases.
1696 if (adr_type == NULL) return true;
1697 if (adr_type == TypePtr::BOTTOM) return true;
1699 return find_alias_type(adr_type, true, NULL) != NULL;
1700 }
1702 //-----------------------------must_alias--------------------------------------
1703 // True if all values of the given address type are in the given alias category.
1704 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1705 if (alias_idx == AliasIdxBot) return true; // the universal category
1706 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1707 if (alias_idx == AliasIdxTop) return false; // the empty category
1708 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1710 // the only remaining possible overlap is identity
1711 int adr_idx = get_alias_index(adr_type);
1712 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1713 assert(adr_idx == alias_idx ||
1714 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1715 && adr_type != TypeOopPtr::BOTTOM),
1716 "should not be testing for overlap with an unsafe pointer");
1717 return adr_idx == alias_idx;
1718 }
1720 //------------------------------can_alias--------------------------------------
1721 // True if any values of the given address type are in the given alias category.
1722 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1723 if (alias_idx == AliasIdxTop) return false; // the empty category
1724 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1725 if (alias_idx == AliasIdxBot) return true; // the universal category
1726 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1728 // the only remaining possible overlap is identity
1729 int adr_idx = get_alias_index(adr_type);
1730 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1731 return adr_idx == alias_idx;
1732 }
1736 //---------------------------pop_warm_call-------------------------------------
1737 WarmCallInfo* Compile::pop_warm_call() {
1738 WarmCallInfo* wci = _warm_calls;
1739 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1740 return wci;
1741 }
1743 //----------------------------Inline_Warm--------------------------------------
1744 int Compile::Inline_Warm() {
1745 // If there is room, try to inline some more warm call sites.
1746 // %%% Do a graph index compaction pass when we think we're out of space?
1747 if (!InlineWarmCalls) return 0;
1749 int calls_made_hot = 0;
1750 int room_to_grow = NodeCountInliningCutoff - unique();
1751 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1752 int amount_grown = 0;
1753 WarmCallInfo* call;
1754 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1755 int est_size = (int)call->size();
1756 if (est_size > (room_to_grow - amount_grown)) {
1757 // This one won't fit anyway. Get rid of it.
1758 call->make_cold();
1759 continue;
1760 }
1761 call->make_hot();
1762 calls_made_hot++;
1763 amount_grown += est_size;
1764 amount_to_grow -= est_size;
1765 }
1767 if (calls_made_hot > 0) set_major_progress();
1768 return calls_made_hot;
1769 }
1772 //----------------------------Finish_Warm--------------------------------------
1773 void Compile::Finish_Warm() {
1774 if (!InlineWarmCalls) return;
1775 if (failing()) return;
1776 if (warm_calls() == NULL) return;
1778 // Clean up loose ends, if we are out of space for inlining.
1779 WarmCallInfo* call;
1780 while ((call = pop_warm_call()) != NULL) {
1781 call->make_cold();
1782 }
1783 }
1785 //---------------------cleanup_loop_predicates-----------------------
1786 // Remove the opaque nodes that protect the predicates so that all unused
1787 // checks and uncommon_traps will be eliminated from the ideal graph
1788 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1789 if (predicate_count()==0) return;
1790 for (int i = predicate_count(); i > 0; i--) {
1791 Node * n = predicate_opaque1_node(i-1);
1792 assert(n->Opcode() == Op_Opaque1, "must be");
1793 igvn.replace_node(n, n->in(1));
1794 }
1795 assert(predicate_count()==0, "should be clean!");
1796 }
1798 // StringOpts and late inlining of string methods
1799 void Compile::inline_string_calls(bool parse_time) {
1800 {
1801 // remove useless nodes to make the usage analysis simpler
1802 ResourceMark rm;
1803 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1804 }
1806 {
1807 ResourceMark rm;
1808 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1809 PhaseStringOpts pso(initial_gvn(), for_igvn());
1810 print_method(PHASE_AFTER_STRINGOPTS, 3);
1811 }
1813 // now inline anything that we skipped the first time around
1814 if (!parse_time) {
1815 _late_inlines_pos = _late_inlines.length();
1816 }
1818 while (_string_late_inlines.length() > 0) {
1819 CallGenerator* cg = _string_late_inlines.pop();
1820 cg->do_late_inline();
1821 if (failing()) return;
1822 }
1823 _string_late_inlines.trunc_to(0);
1824 }
1826 // Late inlining of boxing methods
1827 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1828 if (_boxing_late_inlines.length() > 0) {
1829 assert(has_boxed_value(), "inconsistent");
1831 PhaseGVN* gvn = initial_gvn();
1832 set_inlining_incrementally(true);
1834 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1835 for_igvn()->clear();
1836 gvn->replace_with(&igvn);
1838 while (_boxing_late_inlines.length() > 0) {
1839 CallGenerator* cg = _boxing_late_inlines.pop();
1840 cg->do_late_inline();
1841 if (failing()) return;
1842 }
1843 _boxing_late_inlines.trunc_to(0);
1845 {
1846 ResourceMark rm;
1847 PhaseRemoveUseless pru(gvn, for_igvn());
1848 }
1850 igvn = PhaseIterGVN(gvn);
1851 igvn.optimize();
1853 set_inlining_progress(false);
1854 set_inlining_incrementally(false);
1855 }
1856 }
1858 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1859 assert(IncrementalInline, "incremental inlining should be on");
1860 PhaseGVN* gvn = initial_gvn();
1862 set_inlining_progress(false);
1863 for_igvn()->clear();
1864 gvn->replace_with(&igvn);
1866 int i = 0;
1868 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1869 CallGenerator* cg = _late_inlines.at(i);
1870 _late_inlines_pos = i+1;
1871 cg->do_late_inline();
1872 if (failing()) return;
1873 }
1874 int j = 0;
1875 for (; i < _late_inlines.length(); i++, j++) {
1876 _late_inlines.at_put(j, _late_inlines.at(i));
1877 }
1878 _late_inlines.trunc_to(j);
1880 {
1881 ResourceMark rm;
1882 PhaseRemoveUseless pru(gvn, for_igvn());
1883 }
1885 igvn = PhaseIterGVN(gvn);
1886 }
1888 // Perform incremental inlining until bound on number of live nodes is reached
1889 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1890 PhaseGVN* gvn = initial_gvn();
1892 set_inlining_incrementally(true);
1893 set_inlining_progress(true);
1894 uint low_live_nodes = 0;
1896 while(inlining_progress() && _late_inlines.length() > 0) {
1898 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1899 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1900 // PhaseIdealLoop is expensive so we only try it once we are
1901 // out of loop and we only try it again if the previous helped
1902 // got the number of nodes down significantly
1903 PhaseIdealLoop ideal_loop( igvn, false, true );
1904 if (failing()) return;
1905 low_live_nodes = live_nodes();
1906 _major_progress = true;
1907 }
1909 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1910 break;
1911 }
1912 }
1914 inline_incrementally_one(igvn);
1916 if (failing()) return;
1918 igvn.optimize();
1920 if (failing()) return;
1921 }
1923 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1925 if (_string_late_inlines.length() > 0) {
1926 assert(has_stringbuilder(), "inconsistent");
1927 for_igvn()->clear();
1928 initial_gvn()->replace_with(&igvn);
1930 inline_string_calls(false);
1932 if (failing()) return;
1934 {
1935 ResourceMark rm;
1936 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1937 }
1939 igvn = PhaseIterGVN(gvn);
1941 igvn.optimize();
1942 }
1944 set_inlining_incrementally(false);
1945 }
1948 //------------------------------Optimize---------------------------------------
1949 // Given a graph, optimize it.
1950 void Compile::Optimize() {
1951 TracePhase t1("optimizer", &_t_optimizer, true);
1953 #ifndef PRODUCT
1954 if (env()->break_at_compile()) {
1955 BREAKPOINT;
1956 }
1958 #endif
1960 ResourceMark rm;
1961 int loop_opts_cnt;
1963 NOT_PRODUCT( verify_graph_edges(); )
1965 print_method(PHASE_AFTER_PARSING);
1967 {
1968 // Iterative Global Value Numbering, including ideal transforms
1969 // Initialize IterGVN with types and values from parse-time GVN
1970 PhaseIterGVN igvn(initial_gvn());
1971 {
1972 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1973 igvn.optimize();
1974 }
1976 print_method(PHASE_ITER_GVN1, 2);
1978 if (failing()) return;
1980 {
1981 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
1982 inline_incrementally(igvn);
1983 }
1985 print_method(PHASE_INCREMENTAL_INLINE, 2);
1987 if (failing()) return;
1989 if (eliminate_boxing()) {
1990 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
1991 // Inline valueOf() methods now.
1992 inline_boxing_calls(igvn);
1994 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
1996 if (failing()) return;
1997 }
1999 // No more new expensive nodes will be added to the list from here
2000 // so keep only the actual candidates for optimizations.
2001 cleanup_expensive_nodes(igvn);
2003 // Perform escape analysis
2004 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2005 if (has_loops()) {
2006 // Cleanup graph (remove dead nodes).
2007 TracePhase t2("idealLoop", &_t_idealLoop, true);
2008 PhaseIdealLoop ideal_loop( igvn, false, true );
2009 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2010 if (failing()) return;
2011 }
2012 ConnectionGraph::do_analysis(this, &igvn);
2014 if (failing()) return;
2016 // Optimize out fields loads from scalar replaceable allocations.
2017 igvn.optimize();
2018 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2020 if (failing()) return;
2022 if (congraph() != NULL && macro_count() > 0) {
2023 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2024 PhaseMacroExpand mexp(igvn);
2025 mexp.eliminate_macro_nodes();
2026 igvn.set_delay_transform(false);
2028 igvn.optimize();
2029 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2031 if (failing()) return;
2032 }
2033 }
2035 // Loop transforms on the ideal graph. Range Check Elimination,
2036 // peeling, unrolling, etc.
2038 // Set loop opts counter
2039 loop_opts_cnt = num_loop_opts();
2040 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2041 {
2042 TracePhase t2("idealLoop", &_t_idealLoop, true);
2043 PhaseIdealLoop ideal_loop( igvn, true );
2044 loop_opts_cnt--;
2045 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2046 if (failing()) return;
2047 }
2048 // Loop opts pass if partial peeling occurred in previous pass
2049 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2050 TracePhase t3("idealLoop", &_t_idealLoop, true);
2051 PhaseIdealLoop ideal_loop( igvn, false );
2052 loop_opts_cnt--;
2053 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2054 if (failing()) return;
2055 }
2056 // Loop opts pass for loop-unrolling before CCP
2057 if(major_progress() && (loop_opts_cnt > 0)) {
2058 TracePhase t4("idealLoop", &_t_idealLoop, true);
2059 PhaseIdealLoop ideal_loop( igvn, false );
2060 loop_opts_cnt--;
2061 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2062 }
2063 if (!failing()) {
2064 // Verify that last round of loop opts produced a valid graph
2065 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2066 PhaseIdealLoop::verify(igvn);
2067 }
2068 }
2069 if (failing()) return;
2071 // Conditional Constant Propagation;
2072 PhaseCCP ccp( &igvn );
2073 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2074 {
2075 TracePhase t2("ccp", &_t_ccp, true);
2076 ccp.do_transform();
2077 }
2078 print_method(PHASE_CPP1, 2);
2080 assert( true, "Break here to ccp.dump_old2new_map()");
2082 // Iterative Global Value Numbering, including ideal transforms
2083 {
2084 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2085 igvn = ccp;
2086 igvn.optimize();
2087 }
2089 print_method(PHASE_ITER_GVN2, 2);
2091 if (failing()) return;
2093 // Loop transforms on the ideal graph. Range Check Elimination,
2094 // peeling, unrolling, etc.
2095 if(loop_opts_cnt > 0) {
2096 debug_only( int cnt = 0; );
2097 while(major_progress() && (loop_opts_cnt > 0)) {
2098 TracePhase t2("idealLoop", &_t_idealLoop, true);
2099 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2100 PhaseIdealLoop ideal_loop( igvn, true);
2101 loop_opts_cnt--;
2102 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2103 if (failing()) return;
2104 }
2105 }
2107 {
2108 // Verify that all previous optimizations produced a valid graph
2109 // at least to this point, even if no loop optimizations were done.
2110 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2111 PhaseIdealLoop::verify(igvn);
2112 }
2114 {
2115 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2116 PhaseMacroExpand mex(igvn);
2117 if (mex.expand_macro_nodes()) {
2118 assert(failing(), "must bail out w/ explicit message");
2119 return;
2120 }
2121 }
2123 } // (End scope of igvn; run destructor if necessary for asserts.)
2125 dump_inlining();
2126 // A method with only infinite loops has no edges entering loops from root
2127 {
2128 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2129 if (final_graph_reshaping()) {
2130 assert(failing(), "must bail out w/ explicit message");
2131 return;
2132 }
2133 }
2135 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2136 }
2139 //------------------------------Code_Gen---------------------------------------
2140 // Given a graph, generate code for it
2141 void Compile::Code_Gen() {
2142 if (failing()) {
2143 return;
2144 }
2146 // Perform instruction selection. You might think we could reclaim Matcher
2147 // memory PDQ, but actually the Matcher is used in generating spill code.
2148 // Internals of the Matcher (including some VectorSets) must remain live
2149 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2150 // set a bit in reclaimed memory.
2152 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2153 // nodes. Mapping is only valid at the root of each matched subtree.
2154 NOT_PRODUCT( verify_graph_edges(); )
2156 Matcher matcher;
2157 _matcher = &matcher;
2158 {
2159 TracePhase t2("matcher", &_t_matcher, true);
2160 matcher.match();
2161 }
2162 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2163 // nodes. Mapping is only valid at the root of each matched subtree.
2164 NOT_PRODUCT( verify_graph_edges(); )
2166 // If you have too many nodes, or if matching has failed, bail out
2167 check_node_count(0, "out of nodes matching instructions");
2168 if (failing()) {
2169 return;
2170 }
2172 // Build a proper-looking CFG
2173 PhaseCFG cfg(node_arena(), root(), matcher);
2174 _cfg = &cfg;
2175 {
2176 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2177 bool success = cfg.do_global_code_motion();
2178 if (!success) {
2179 return;
2180 }
2182 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2183 NOT_PRODUCT( verify_graph_edges(); )
2184 debug_only( cfg.verify(); )
2185 }
2187 PhaseChaitin regalloc(unique(), cfg, matcher);
2188 _regalloc = ®alloc;
2189 {
2190 TracePhase t2("regalloc", &_t_registerAllocation, true);
2191 // Perform register allocation. After Chaitin, use-def chains are
2192 // no longer accurate (at spill code) and so must be ignored.
2193 // Node->LRG->reg mappings are still accurate.
2194 _regalloc->Register_Allocate();
2196 // Bail out if the allocator builds too many nodes
2197 if (failing()) {
2198 return;
2199 }
2200 }
2202 // Prior to register allocation we kept empty basic blocks in case the
2203 // the allocator needed a place to spill. After register allocation we
2204 // are not adding any new instructions. If any basic block is empty, we
2205 // can now safely remove it.
2206 {
2207 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2208 cfg.remove_empty_blocks();
2209 if (do_freq_based_layout()) {
2210 PhaseBlockLayout layout(cfg);
2211 } else {
2212 cfg.set_loop_alignment();
2213 }
2214 cfg.fixup_flow();
2215 }
2217 // Apply peephole optimizations
2218 if( OptoPeephole ) {
2219 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2220 PhasePeephole peep( _regalloc, cfg);
2221 peep.do_transform();
2222 }
2224 // Convert Nodes to instruction bits in a buffer
2225 {
2226 // %%%% workspace merge brought two timers together for one job
2227 TracePhase t2a("output", &_t_output, true);
2228 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2229 Output();
2230 }
2232 print_method(PHASE_FINAL_CODE);
2234 // He's dead, Jim.
2235 _cfg = (PhaseCFG*)0xdeadbeef;
2236 _regalloc = (PhaseChaitin*)0xdeadbeef;
2237 }
2240 //------------------------------dump_asm---------------------------------------
2241 // Dump formatted assembly
2242 #ifndef PRODUCT
2243 void Compile::dump_asm(int *pcs, uint pc_limit) {
2244 bool cut_short = false;
2245 tty->print_cr("#");
2246 tty->print("# "); _tf->dump(); tty->cr();
2247 tty->print_cr("#");
2249 // For all blocks
2250 int pc = 0x0; // Program counter
2251 char starts_bundle = ' ';
2252 _regalloc->dump_frame();
2254 Node *n = NULL;
2255 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2256 if (VMThread::should_terminate()) {
2257 cut_short = true;
2258 break;
2259 }
2260 Block* block = _cfg->get_block(i);
2261 if (block->is_connector() && !Verbose) {
2262 continue;
2263 }
2264 n = block->_nodes[0];
2265 if (pcs && n->_idx < pc_limit) {
2266 tty->print("%3.3x ", pcs[n->_idx]);
2267 } else {
2268 tty->print(" ");
2269 }
2270 block->dump_head(_cfg);
2271 if (block->is_connector()) {
2272 tty->print_cr(" # Empty connector block");
2273 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2274 tty->print_cr(" # Block is sole successor of call");
2275 }
2277 // For all instructions
2278 Node *delay = NULL;
2279 for (uint j = 0; j < block->_nodes.size(); j++) {
2280 if (VMThread::should_terminate()) {
2281 cut_short = true;
2282 break;
2283 }
2284 n = block->_nodes[j];
2285 if (valid_bundle_info(n)) {
2286 Bundle* bundle = node_bundling(n);
2287 if (bundle->used_in_unconditional_delay()) {
2288 delay = n;
2289 continue;
2290 }
2291 if (bundle->starts_bundle()) {
2292 starts_bundle = '+';
2293 }
2294 }
2296 if (WizardMode) {
2297 n->dump();
2298 }
2300 if( !n->is_Region() && // Dont print in the Assembly
2301 !n->is_Phi() && // a few noisely useless nodes
2302 !n->is_Proj() &&
2303 !n->is_MachTemp() &&
2304 !n->is_SafePointScalarObject() &&
2305 !n->is_Catch() && // Would be nice to print exception table targets
2306 !n->is_MergeMem() && // Not very interesting
2307 !n->is_top() && // Debug info table constants
2308 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2309 ) {
2310 if (pcs && n->_idx < pc_limit)
2311 tty->print("%3.3x", pcs[n->_idx]);
2312 else
2313 tty->print(" ");
2314 tty->print(" %c ", starts_bundle);
2315 starts_bundle = ' ';
2316 tty->print("\t");
2317 n->format(_regalloc, tty);
2318 tty->cr();
2319 }
2321 // If we have an instruction with a delay slot, and have seen a delay,
2322 // then back up and print it
2323 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2324 assert(delay != NULL, "no unconditional delay instruction");
2325 if (WizardMode) delay->dump();
2327 if (node_bundling(delay)->starts_bundle())
2328 starts_bundle = '+';
2329 if (pcs && n->_idx < pc_limit)
2330 tty->print("%3.3x", pcs[n->_idx]);
2331 else
2332 tty->print(" ");
2333 tty->print(" %c ", starts_bundle);
2334 starts_bundle = ' ';
2335 tty->print("\t");
2336 delay->format(_regalloc, tty);
2337 tty->print_cr("");
2338 delay = NULL;
2339 }
2341 // Dump the exception table as well
2342 if( n->is_Catch() && (Verbose || WizardMode) ) {
2343 // Print the exception table for this offset
2344 _handler_table.print_subtable_for(pc);
2345 }
2346 }
2348 if (pcs && n->_idx < pc_limit)
2349 tty->print_cr("%3.3x", pcs[n->_idx]);
2350 else
2351 tty->print_cr("");
2353 assert(cut_short || delay == NULL, "no unconditional delay branch");
2355 } // End of per-block dump
2356 tty->print_cr("");
2358 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2359 }
2360 #endif
2362 //------------------------------Final_Reshape_Counts---------------------------
2363 // This class defines counters to help identify when a method
2364 // may/must be executed using hardware with only 24-bit precision.
2365 struct Final_Reshape_Counts : public StackObj {
2366 int _call_count; // count non-inlined 'common' calls
2367 int _float_count; // count float ops requiring 24-bit precision
2368 int _double_count; // count double ops requiring more precision
2369 int _java_call_count; // count non-inlined 'java' calls
2370 int _inner_loop_count; // count loops which need alignment
2371 VectorSet _visited; // Visitation flags
2372 Node_List _tests; // Set of IfNodes & PCTableNodes
2374 Final_Reshape_Counts() :
2375 _call_count(0), _float_count(0), _double_count(0),
2376 _java_call_count(0), _inner_loop_count(0),
2377 _visited( Thread::current()->resource_area() ) { }
2379 void inc_call_count () { _call_count ++; }
2380 void inc_float_count () { _float_count ++; }
2381 void inc_double_count() { _double_count++; }
2382 void inc_java_call_count() { _java_call_count++; }
2383 void inc_inner_loop_count() { _inner_loop_count++; }
2385 int get_call_count () const { return _call_count ; }
2386 int get_float_count () const { return _float_count ; }
2387 int get_double_count() const { return _double_count; }
2388 int get_java_call_count() const { return _java_call_count; }
2389 int get_inner_loop_count() const { return _inner_loop_count; }
2390 };
2392 #ifdef ASSERT
2393 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2394 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2395 // Make sure the offset goes inside the instance layout.
2396 return k->contains_field_offset(tp->offset());
2397 // Note that OffsetBot and OffsetTop are very negative.
2398 }
2399 #endif
2401 // Eliminate trivially redundant StoreCMs and accumulate their
2402 // precedence edges.
2403 void Compile::eliminate_redundant_card_marks(Node* n) {
2404 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2405 if (n->in(MemNode::Address)->outcnt() > 1) {
2406 // There are multiple users of the same address so it might be
2407 // possible to eliminate some of the StoreCMs
2408 Node* mem = n->in(MemNode::Memory);
2409 Node* adr = n->in(MemNode::Address);
2410 Node* val = n->in(MemNode::ValueIn);
2411 Node* prev = n;
2412 bool done = false;
2413 // Walk the chain of StoreCMs eliminating ones that match. As
2414 // long as it's a chain of single users then the optimization is
2415 // safe. Eliminating partially redundant StoreCMs would require
2416 // cloning copies down the other paths.
2417 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2418 if (adr == mem->in(MemNode::Address) &&
2419 val == mem->in(MemNode::ValueIn)) {
2420 // redundant StoreCM
2421 if (mem->req() > MemNode::OopStore) {
2422 // Hasn't been processed by this code yet.
2423 n->add_prec(mem->in(MemNode::OopStore));
2424 } else {
2425 // Already converted to precedence edge
2426 for (uint i = mem->req(); i < mem->len(); i++) {
2427 // Accumulate any precedence edges
2428 if (mem->in(i) != NULL) {
2429 n->add_prec(mem->in(i));
2430 }
2431 }
2432 // Everything above this point has been processed.
2433 done = true;
2434 }
2435 // Eliminate the previous StoreCM
2436 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2437 assert(mem->outcnt() == 0, "should be dead");
2438 mem->disconnect_inputs(NULL, this);
2439 } else {
2440 prev = mem;
2441 }
2442 mem = prev->in(MemNode::Memory);
2443 }
2444 }
2445 }
2447 //------------------------------final_graph_reshaping_impl----------------------
2448 // Implement items 1-5 from final_graph_reshaping below.
2449 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2451 if ( n->outcnt() == 0 ) return; // dead node
2452 uint nop = n->Opcode();
2454 // Check for 2-input instruction with "last use" on right input.
2455 // Swap to left input. Implements item (2).
2456 if( n->req() == 3 && // two-input instruction
2457 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2458 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2459 n->in(2)->outcnt() == 1 &&// right use IS a last use
2460 !n->in(2)->is_Con() ) { // right use is not a constant
2461 // Check for commutative opcode
2462 switch( nop ) {
2463 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2464 case Op_MaxI: case Op_MinI:
2465 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2466 case Op_AndL: case Op_XorL: case Op_OrL:
2467 case Op_AndI: case Op_XorI: case Op_OrI: {
2468 // Move "last use" input to left by swapping inputs
2469 n->swap_edges(1, 2);
2470 break;
2471 }
2472 default:
2473 break;
2474 }
2475 }
2477 #ifdef ASSERT
2478 if( n->is_Mem() ) {
2479 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2480 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2481 // oop will be recorded in oop map if load crosses safepoint
2482 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2483 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2484 "raw memory operations should have control edge");
2485 }
2486 #endif
2487 // Count FPU ops and common calls, implements item (3)
2488 switch( nop ) {
2489 // Count all float operations that may use FPU
2490 case Op_AddF:
2491 case Op_SubF:
2492 case Op_MulF:
2493 case Op_DivF:
2494 case Op_NegF:
2495 case Op_ModF:
2496 case Op_ConvI2F:
2497 case Op_ConF:
2498 case Op_CmpF:
2499 case Op_CmpF3:
2500 // case Op_ConvL2F: // longs are split into 32-bit halves
2501 frc.inc_float_count();
2502 break;
2504 case Op_ConvF2D:
2505 case Op_ConvD2F:
2506 frc.inc_float_count();
2507 frc.inc_double_count();
2508 break;
2510 // Count all double operations that may use FPU
2511 case Op_AddD:
2512 case Op_SubD:
2513 case Op_MulD:
2514 case Op_DivD:
2515 case Op_NegD:
2516 case Op_ModD:
2517 case Op_ConvI2D:
2518 case Op_ConvD2I:
2519 // case Op_ConvL2D: // handled by leaf call
2520 // case Op_ConvD2L: // handled by leaf call
2521 case Op_ConD:
2522 case Op_CmpD:
2523 case Op_CmpD3:
2524 frc.inc_double_count();
2525 break;
2526 case Op_Opaque1: // Remove Opaque Nodes before matching
2527 case Op_Opaque2: // Remove Opaque Nodes before matching
2528 n->subsume_by(n->in(1), this);
2529 break;
2530 case Op_CallStaticJava:
2531 case Op_CallJava:
2532 case Op_CallDynamicJava:
2533 frc.inc_java_call_count(); // Count java call site;
2534 case Op_CallRuntime:
2535 case Op_CallLeaf:
2536 case Op_CallLeafNoFP: {
2537 assert( n->is_Call(), "" );
2538 CallNode *call = n->as_Call();
2539 // Count call sites where the FP mode bit would have to be flipped.
2540 // Do not count uncommon runtime calls:
2541 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2542 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2543 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2544 frc.inc_call_count(); // Count the call site
2545 } else { // See if uncommon argument is shared
2546 Node *n = call->in(TypeFunc::Parms);
2547 int nop = n->Opcode();
2548 // Clone shared simple arguments to uncommon calls, item (1).
2549 if( n->outcnt() > 1 &&
2550 !n->is_Proj() &&
2551 nop != Op_CreateEx &&
2552 nop != Op_CheckCastPP &&
2553 nop != Op_DecodeN &&
2554 nop != Op_DecodeNKlass &&
2555 !n->is_Mem() ) {
2556 Node *x = n->clone();
2557 call->set_req( TypeFunc::Parms, x );
2558 }
2559 }
2560 break;
2561 }
2563 case Op_StoreD:
2564 case Op_LoadD:
2565 case Op_LoadD_unaligned:
2566 frc.inc_double_count();
2567 goto handle_mem;
2568 case Op_StoreF:
2569 case Op_LoadF:
2570 frc.inc_float_count();
2571 goto handle_mem;
2573 case Op_StoreCM:
2574 {
2575 // Convert OopStore dependence into precedence edge
2576 Node* prec = n->in(MemNode::OopStore);
2577 n->del_req(MemNode::OopStore);
2578 n->add_prec(prec);
2579 eliminate_redundant_card_marks(n);
2580 }
2582 // fall through
2584 case Op_StoreB:
2585 case Op_StoreC:
2586 case Op_StorePConditional:
2587 case Op_StoreI:
2588 case Op_StoreL:
2589 case Op_StoreIConditional:
2590 case Op_StoreLConditional:
2591 case Op_CompareAndSwapI:
2592 case Op_CompareAndSwapL:
2593 case Op_CompareAndSwapP:
2594 case Op_CompareAndSwapN:
2595 case Op_GetAndAddI:
2596 case Op_GetAndAddL:
2597 case Op_GetAndSetI:
2598 case Op_GetAndSetL:
2599 case Op_GetAndSetP:
2600 case Op_GetAndSetN:
2601 case Op_StoreP:
2602 case Op_StoreN:
2603 case Op_StoreNKlass:
2604 case Op_LoadB:
2605 case Op_LoadUB:
2606 case Op_LoadUS:
2607 case Op_LoadI:
2608 case Op_LoadKlass:
2609 case Op_LoadNKlass:
2610 case Op_LoadL:
2611 case Op_LoadL_unaligned:
2612 case Op_LoadPLocked:
2613 case Op_LoadP:
2614 case Op_LoadN:
2615 case Op_LoadRange:
2616 case Op_LoadS: {
2617 handle_mem:
2618 #ifdef ASSERT
2619 if( VerifyOptoOopOffsets ) {
2620 assert( n->is_Mem(), "" );
2621 MemNode *mem = (MemNode*)n;
2622 // Check to see if address types have grounded out somehow.
2623 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2624 assert( !tp || oop_offset_is_sane(tp), "" );
2625 }
2626 #endif
2627 break;
2628 }
2630 case Op_AddP: { // Assert sane base pointers
2631 Node *addp = n->in(AddPNode::Address);
2632 assert( !addp->is_AddP() ||
2633 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2634 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2635 "Base pointers must match" );
2636 #ifdef _LP64
2637 if ((UseCompressedOops || UseCompressedKlassPointers) &&
2638 addp->Opcode() == Op_ConP &&
2639 addp == n->in(AddPNode::Base) &&
2640 n->in(AddPNode::Offset)->is_Con()) {
2641 // Use addressing with narrow klass to load with offset on x86.
2642 // On sparc loading 32-bits constant and decoding it have less
2643 // instructions (4) then load 64-bits constant (7).
2644 // Do this transformation here since IGVN will convert ConN back to ConP.
2645 const Type* t = addp->bottom_type();
2646 if (t->isa_oopptr() || t->isa_klassptr()) {
2647 Node* nn = NULL;
2649 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2651 // Look for existing ConN node of the same exact type.
2652 Node* r = root();
2653 uint cnt = r->outcnt();
2654 for (uint i = 0; i < cnt; i++) {
2655 Node* m = r->raw_out(i);
2656 if (m!= NULL && m->Opcode() == op &&
2657 m->bottom_type()->make_ptr() == t) {
2658 nn = m;
2659 break;
2660 }
2661 }
2662 if (nn != NULL) {
2663 // Decode a narrow oop to match address
2664 // [R12 + narrow_oop_reg<<3 + offset]
2665 if (t->isa_oopptr()) {
2666 nn = new (this) DecodeNNode(nn, t);
2667 } else {
2668 nn = new (this) DecodeNKlassNode(nn, t);
2669 }
2670 n->set_req(AddPNode::Base, nn);
2671 n->set_req(AddPNode::Address, nn);
2672 if (addp->outcnt() == 0) {
2673 addp->disconnect_inputs(NULL, this);
2674 }
2675 }
2676 }
2677 }
2678 #endif
2679 break;
2680 }
2682 #ifdef _LP64
2683 case Op_CastPP:
2684 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2685 Node* in1 = n->in(1);
2686 const Type* t = n->bottom_type();
2687 Node* new_in1 = in1->clone();
2688 new_in1->as_DecodeN()->set_type(t);
2690 if (!Matcher::narrow_oop_use_complex_address()) {
2691 //
2692 // x86, ARM and friends can handle 2 adds in addressing mode
2693 // and Matcher can fold a DecodeN node into address by using
2694 // a narrow oop directly and do implicit NULL check in address:
2695 //
2696 // [R12 + narrow_oop_reg<<3 + offset]
2697 // NullCheck narrow_oop_reg
2698 //
2699 // On other platforms (Sparc) we have to keep new DecodeN node and
2700 // use it to do implicit NULL check in address:
2701 //
2702 // decode_not_null narrow_oop_reg, base_reg
2703 // [base_reg + offset]
2704 // NullCheck base_reg
2705 //
2706 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2707 // to keep the information to which NULL check the new DecodeN node
2708 // corresponds to use it as value in implicit_null_check().
2709 //
2710 new_in1->set_req(0, n->in(0));
2711 }
2713 n->subsume_by(new_in1, this);
2714 if (in1->outcnt() == 0) {
2715 in1->disconnect_inputs(NULL, this);
2716 }
2717 }
2718 break;
2720 case Op_CmpP:
2721 // Do this transformation here to preserve CmpPNode::sub() and
2722 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2723 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2724 Node* in1 = n->in(1);
2725 Node* in2 = n->in(2);
2726 if (!in1->is_DecodeNarrowPtr()) {
2727 in2 = in1;
2728 in1 = n->in(2);
2729 }
2730 assert(in1->is_DecodeNarrowPtr(), "sanity");
2732 Node* new_in2 = NULL;
2733 if (in2->is_DecodeNarrowPtr()) {
2734 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2735 new_in2 = in2->in(1);
2736 } else if (in2->Opcode() == Op_ConP) {
2737 const Type* t = in2->bottom_type();
2738 if (t == TypePtr::NULL_PTR) {
2739 assert(in1->is_DecodeN(), "compare klass to null?");
2740 // Don't convert CmpP null check into CmpN if compressed
2741 // oops implicit null check is not generated.
2742 // This will allow to generate normal oop implicit null check.
2743 if (Matcher::gen_narrow_oop_implicit_null_checks())
2744 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2745 //
2746 // This transformation together with CastPP transformation above
2747 // will generated code for implicit NULL checks for compressed oops.
2748 //
2749 // The original code after Optimize()
2750 //
2751 // LoadN memory, narrow_oop_reg
2752 // decode narrow_oop_reg, base_reg
2753 // CmpP base_reg, NULL
2754 // CastPP base_reg // NotNull
2755 // Load [base_reg + offset], val_reg
2756 //
2757 // after these transformations will be
2758 //
2759 // LoadN memory, narrow_oop_reg
2760 // CmpN narrow_oop_reg, NULL
2761 // decode_not_null narrow_oop_reg, base_reg
2762 // Load [base_reg + offset], val_reg
2763 //
2764 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2765 // since narrow oops can be used in debug info now (see the code in
2766 // final_graph_reshaping_walk()).
2767 //
2768 // At the end the code will be matched to
2769 // on x86:
2770 //
2771 // Load_narrow_oop memory, narrow_oop_reg
2772 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2773 // NullCheck narrow_oop_reg
2774 //
2775 // and on sparc:
2776 //
2777 // Load_narrow_oop memory, narrow_oop_reg
2778 // decode_not_null narrow_oop_reg, base_reg
2779 // Load [base_reg + offset], val_reg
2780 // NullCheck base_reg
2781 //
2782 } else if (t->isa_oopptr()) {
2783 new_in2 = ConNode::make(this, t->make_narrowoop());
2784 } else if (t->isa_klassptr()) {
2785 new_in2 = ConNode::make(this, t->make_narrowklass());
2786 }
2787 }
2788 if (new_in2 != NULL) {
2789 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2790 n->subsume_by(cmpN, this);
2791 if (in1->outcnt() == 0) {
2792 in1->disconnect_inputs(NULL, this);
2793 }
2794 if (in2->outcnt() == 0) {
2795 in2->disconnect_inputs(NULL, this);
2796 }
2797 }
2798 }
2799 break;
2801 case Op_DecodeN:
2802 case Op_DecodeNKlass:
2803 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2804 // DecodeN could be pinned when it can't be fold into
2805 // an address expression, see the code for Op_CastPP above.
2806 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2807 break;
2809 case Op_EncodeP:
2810 case Op_EncodePKlass: {
2811 Node* in1 = n->in(1);
2812 if (in1->is_DecodeNarrowPtr()) {
2813 n->subsume_by(in1->in(1), this);
2814 } else if (in1->Opcode() == Op_ConP) {
2815 const Type* t = in1->bottom_type();
2816 if (t == TypePtr::NULL_PTR) {
2817 assert(t->isa_oopptr(), "null klass?");
2818 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2819 } else if (t->isa_oopptr()) {
2820 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2821 } else if (t->isa_klassptr()) {
2822 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2823 }
2824 }
2825 if (in1->outcnt() == 0) {
2826 in1->disconnect_inputs(NULL, this);
2827 }
2828 break;
2829 }
2831 case Op_Proj: {
2832 if (OptimizeStringConcat) {
2833 ProjNode* p = n->as_Proj();
2834 if (p->_is_io_use) {
2835 // Separate projections were used for the exception path which
2836 // are normally removed by a late inline. If it wasn't inlined
2837 // then they will hang around and should just be replaced with
2838 // the original one.
2839 Node* proj = NULL;
2840 // Replace with just one
2841 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2842 Node *use = i.get();
2843 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2844 proj = use;
2845 break;
2846 }
2847 }
2848 assert(proj != NULL, "must be found");
2849 p->subsume_by(proj, this);
2850 }
2851 }
2852 break;
2853 }
2855 case Op_Phi:
2856 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
2857 // The EncodeP optimization may create Phi with the same edges
2858 // for all paths. It is not handled well by Register Allocator.
2859 Node* unique_in = n->in(1);
2860 assert(unique_in != NULL, "");
2861 uint cnt = n->req();
2862 for (uint i = 2; i < cnt; i++) {
2863 Node* m = n->in(i);
2864 assert(m != NULL, "");
2865 if (unique_in != m)
2866 unique_in = NULL;
2867 }
2868 if (unique_in != NULL) {
2869 n->subsume_by(unique_in, this);
2870 }
2871 }
2872 break;
2874 #endif
2876 case Op_ModI:
2877 if (UseDivMod) {
2878 // Check if a%b and a/b both exist
2879 Node* d = n->find_similar(Op_DivI);
2880 if (d) {
2881 // Replace them with a fused divmod if supported
2882 if (Matcher::has_match_rule(Op_DivModI)) {
2883 DivModINode* divmod = DivModINode::make(this, n);
2884 d->subsume_by(divmod->div_proj(), this);
2885 n->subsume_by(divmod->mod_proj(), this);
2886 } else {
2887 // replace a%b with a-((a/b)*b)
2888 Node* mult = new (this) MulINode(d, d->in(2));
2889 Node* sub = new (this) SubINode(d->in(1), mult);
2890 n->subsume_by(sub, this);
2891 }
2892 }
2893 }
2894 break;
2896 case Op_ModL:
2897 if (UseDivMod) {
2898 // Check if a%b and a/b both exist
2899 Node* d = n->find_similar(Op_DivL);
2900 if (d) {
2901 // Replace them with a fused divmod if supported
2902 if (Matcher::has_match_rule(Op_DivModL)) {
2903 DivModLNode* divmod = DivModLNode::make(this, n);
2904 d->subsume_by(divmod->div_proj(), this);
2905 n->subsume_by(divmod->mod_proj(), this);
2906 } else {
2907 // replace a%b with a-((a/b)*b)
2908 Node* mult = new (this) MulLNode(d, d->in(2));
2909 Node* sub = new (this) SubLNode(d->in(1), mult);
2910 n->subsume_by(sub, this);
2911 }
2912 }
2913 }
2914 break;
2916 case Op_LoadVector:
2917 case Op_StoreVector:
2918 break;
2920 case Op_PackB:
2921 case Op_PackS:
2922 case Op_PackI:
2923 case Op_PackF:
2924 case Op_PackL:
2925 case Op_PackD:
2926 if (n->req()-1 > 2) {
2927 // Replace many operand PackNodes with a binary tree for matching
2928 PackNode* p = (PackNode*) n;
2929 Node* btp = p->binary_tree_pack(this, 1, n->req());
2930 n->subsume_by(btp, this);
2931 }
2932 break;
2933 case Op_Loop:
2934 case Op_CountedLoop:
2935 if (n->as_Loop()->is_inner_loop()) {
2936 frc.inc_inner_loop_count();
2937 }
2938 break;
2939 case Op_LShiftI:
2940 case Op_RShiftI:
2941 case Op_URShiftI:
2942 case Op_LShiftL:
2943 case Op_RShiftL:
2944 case Op_URShiftL:
2945 if (Matcher::need_masked_shift_count) {
2946 // The cpu's shift instructions don't restrict the count to the
2947 // lower 5/6 bits. We need to do the masking ourselves.
2948 Node* in2 = n->in(2);
2949 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2950 const TypeInt* t = in2->find_int_type();
2951 if (t != NULL && t->is_con()) {
2952 juint shift = t->get_con();
2953 if (shift > mask) { // Unsigned cmp
2954 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
2955 }
2956 } else {
2957 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2958 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
2959 n->set_req(2, shift);
2960 }
2961 }
2962 if (in2->outcnt() == 0) { // Remove dead node
2963 in2->disconnect_inputs(NULL, this);
2964 }
2965 }
2966 break;
2967 case Op_MemBarStoreStore:
2968 case Op_MemBarRelease:
2969 // Break the link with AllocateNode: it is no longer useful and
2970 // confuses register allocation.
2971 if (n->req() > MemBarNode::Precedent) {
2972 n->set_req(MemBarNode::Precedent, top());
2973 }
2974 break;
2975 default:
2976 assert( !n->is_Call(), "" );
2977 assert( !n->is_Mem(), "" );
2978 break;
2979 }
2981 // Collect CFG split points
2982 if (n->is_MultiBranch())
2983 frc._tests.push(n);
2984 }
2986 //------------------------------final_graph_reshaping_walk---------------------
2987 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2988 // requires that the walk visits a node's inputs before visiting the node.
2989 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
2990 ResourceArea *area = Thread::current()->resource_area();
2991 Unique_Node_List sfpt(area);
2993 frc._visited.set(root->_idx); // first, mark node as visited
2994 uint cnt = root->req();
2995 Node *n = root;
2996 uint i = 0;
2997 while (true) {
2998 if (i < cnt) {
2999 // Place all non-visited non-null inputs onto stack
3000 Node* m = n->in(i);
3001 ++i;
3002 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3003 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
3004 sfpt.push(m);
3005 cnt = m->req();
3006 nstack.push(n, i); // put on stack parent and next input's index
3007 n = m;
3008 i = 0;
3009 }
3010 } else {
3011 // Now do post-visit work
3012 final_graph_reshaping_impl( n, frc );
3013 if (nstack.is_empty())
3014 break; // finished
3015 n = nstack.node(); // Get node from stack
3016 cnt = n->req();
3017 i = nstack.index();
3018 nstack.pop(); // Shift to the next node on stack
3019 }
3020 }
3022 // Skip next transformation if compressed oops are not used.
3023 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3024 (!UseCompressedOops && !UseCompressedKlassPointers))
3025 return;
3027 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3028 // It could be done for an uncommon traps or any safepoints/calls
3029 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3030 while (sfpt.size() > 0) {
3031 n = sfpt.pop();
3032 JVMState *jvms = n->as_SafePoint()->jvms();
3033 assert(jvms != NULL, "sanity");
3034 int start = jvms->debug_start();
3035 int end = n->req();
3036 bool is_uncommon = (n->is_CallStaticJava() &&
3037 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3038 for (int j = start; j < end; j++) {
3039 Node* in = n->in(j);
3040 if (in->is_DecodeNarrowPtr()) {
3041 bool safe_to_skip = true;
3042 if (!is_uncommon ) {
3043 // Is it safe to skip?
3044 for (uint i = 0; i < in->outcnt(); i++) {
3045 Node* u = in->raw_out(i);
3046 if (!u->is_SafePoint() ||
3047 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3048 safe_to_skip = false;
3049 }
3050 }
3051 }
3052 if (safe_to_skip) {
3053 n->set_req(j, in->in(1));
3054 }
3055 if (in->outcnt() == 0) {
3056 in->disconnect_inputs(NULL, this);
3057 }
3058 }
3059 }
3060 }
3061 }
3063 //------------------------------final_graph_reshaping--------------------------
3064 // Final Graph Reshaping.
3065 //
3066 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3067 // and not commoned up and forced early. Must come after regular
3068 // optimizations to avoid GVN undoing the cloning. Clone constant
3069 // inputs to Loop Phis; these will be split by the allocator anyways.
3070 // Remove Opaque nodes.
3071 // (2) Move last-uses by commutative operations to the left input to encourage
3072 // Intel update-in-place two-address operations and better register usage
3073 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3074 // calls canonicalizing them back.
3075 // (3) Count the number of double-precision FP ops, single-precision FP ops
3076 // and call sites. On Intel, we can get correct rounding either by
3077 // forcing singles to memory (requires extra stores and loads after each
3078 // FP bytecode) or we can set a rounding mode bit (requires setting and
3079 // clearing the mode bit around call sites). The mode bit is only used
3080 // if the relative frequency of single FP ops to calls is low enough.
3081 // This is a key transform for SPEC mpeg_audio.
3082 // (4) Detect infinite loops; blobs of code reachable from above but not
3083 // below. Several of the Code_Gen algorithms fail on such code shapes,
3084 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3085 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3086 // Detection is by looking for IfNodes where only 1 projection is
3087 // reachable from below or CatchNodes missing some targets.
3088 // (5) Assert for insane oop offsets in debug mode.
3090 bool Compile::final_graph_reshaping() {
3091 // an infinite loop may have been eliminated by the optimizer,
3092 // in which case the graph will be empty.
3093 if (root()->req() == 1) {
3094 record_method_not_compilable("trivial infinite loop");
3095 return true;
3096 }
3098 // Expensive nodes have their control input set to prevent the GVN
3099 // from freely commoning them. There's no GVN beyond this point so
3100 // no need to keep the control input. We want the expensive nodes to
3101 // be freely moved to the least frequent code path by gcm.
3102 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3103 for (int i = 0; i < expensive_count(); i++) {
3104 _expensive_nodes->at(i)->set_req(0, NULL);
3105 }
3107 Final_Reshape_Counts frc;
3109 // Visit everybody reachable!
3110 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3111 Node_Stack nstack(unique() >> 1);
3112 final_graph_reshaping_walk(nstack, root(), frc);
3114 // Check for unreachable (from below) code (i.e., infinite loops).
3115 for( uint i = 0; i < frc._tests.size(); i++ ) {
3116 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3117 // Get number of CFG targets.
3118 // Note that PCTables include exception targets after calls.
3119 uint required_outcnt = n->required_outcnt();
3120 if (n->outcnt() != required_outcnt) {
3121 // Check for a few special cases. Rethrow Nodes never take the
3122 // 'fall-thru' path, so expected kids is 1 less.
3123 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3124 if (n->in(0)->in(0)->is_Call()) {
3125 CallNode *call = n->in(0)->in(0)->as_Call();
3126 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3127 required_outcnt--; // Rethrow always has 1 less kid
3128 } else if (call->req() > TypeFunc::Parms &&
3129 call->is_CallDynamicJava()) {
3130 // Check for null receiver. In such case, the optimizer has
3131 // detected that the virtual call will always result in a null
3132 // pointer exception. The fall-through projection of this CatchNode
3133 // will not be populated.
3134 Node *arg0 = call->in(TypeFunc::Parms);
3135 if (arg0->is_Type() &&
3136 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3137 required_outcnt--;
3138 }
3139 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3140 call->req() > TypeFunc::Parms+1 &&
3141 call->is_CallStaticJava()) {
3142 // Check for negative array length. In such case, the optimizer has
3143 // detected that the allocation attempt will always result in an
3144 // exception. There is no fall-through projection of this CatchNode .
3145 Node *arg1 = call->in(TypeFunc::Parms+1);
3146 if (arg1->is_Type() &&
3147 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3148 required_outcnt--;
3149 }
3150 }
3151 }
3152 }
3153 // Recheck with a better notion of 'required_outcnt'
3154 if (n->outcnt() != required_outcnt) {
3155 record_method_not_compilable("malformed control flow");
3156 return true; // Not all targets reachable!
3157 }
3158 }
3159 // Check that I actually visited all kids. Unreached kids
3160 // must be infinite loops.
3161 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3162 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3163 record_method_not_compilable("infinite loop");
3164 return true; // Found unvisited kid; must be unreach
3165 }
3166 }
3168 // If original bytecodes contained a mixture of floats and doubles
3169 // check if the optimizer has made it homogenous, item (3).
3170 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3171 frc.get_float_count() > 32 &&
3172 frc.get_double_count() == 0 &&
3173 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3174 set_24_bit_selection_and_mode( false, true );
3175 }
3177 set_java_calls(frc.get_java_call_count());
3178 set_inner_loops(frc.get_inner_loop_count());
3180 // No infinite loops, no reason to bail out.
3181 return false;
3182 }
3184 //-----------------------------too_many_traps----------------------------------
3185 // Report if there are too many traps at the current method and bci.
3186 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3187 bool Compile::too_many_traps(ciMethod* method,
3188 int bci,
3189 Deoptimization::DeoptReason reason) {
3190 ciMethodData* md = method->method_data();
3191 if (md->is_empty()) {
3192 // Assume the trap has not occurred, or that it occurred only
3193 // because of a transient condition during start-up in the interpreter.
3194 return false;
3195 }
3196 if (md->has_trap_at(bci, reason) != 0) {
3197 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3198 // Also, if there are multiple reasons, or if there is no per-BCI record,
3199 // assume the worst.
3200 if (log())
3201 log()->elem("observe trap='%s' count='%d'",
3202 Deoptimization::trap_reason_name(reason),
3203 md->trap_count(reason));
3204 return true;
3205 } else {
3206 // Ignore method/bci and see if there have been too many globally.
3207 return too_many_traps(reason, md);
3208 }
3209 }
3211 // Less-accurate variant which does not require a method and bci.
3212 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3213 ciMethodData* logmd) {
3214 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
3215 // Too many traps globally.
3216 // Note that we use cumulative trap_count, not just md->trap_count.
3217 if (log()) {
3218 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3219 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3220 Deoptimization::trap_reason_name(reason),
3221 mcount, trap_count(reason));
3222 }
3223 return true;
3224 } else {
3225 // The coast is clear.
3226 return false;
3227 }
3228 }
3230 //--------------------------too_many_recompiles--------------------------------
3231 // Report if there are too many recompiles at the current method and bci.
3232 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3233 // Is not eager to return true, since this will cause the compiler to use
3234 // Action_none for a trap point, to avoid too many recompilations.
3235 bool Compile::too_many_recompiles(ciMethod* method,
3236 int bci,
3237 Deoptimization::DeoptReason reason) {
3238 ciMethodData* md = method->method_data();
3239 if (md->is_empty()) {
3240 // Assume the trap has not occurred, or that it occurred only
3241 // because of a transient condition during start-up in the interpreter.
3242 return false;
3243 }
3244 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3245 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3246 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3247 Deoptimization::DeoptReason per_bc_reason
3248 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3249 if ((per_bc_reason == Deoptimization::Reason_none
3250 || md->has_trap_at(bci, reason) != 0)
3251 // The trap frequency measure we care about is the recompile count:
3252 && md->trap_recompiled_at(bci)
3253 && md->overflow_recompile_count() >= bc_cutoff) {
3254 // Do not emit a trap here if it has already caused recompilations.
3255 // Also, if there are multiple reasons, or if there is no per-BCI record,
3256 // assume the worst.
3257 if (log())
3258 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3259 Deoptimization::trap_reason_name(reason),
3260 md->trap_count(reason),
3261 md->overflow_recompile_count());
3262 return true;
3263 } else if (trap_count(reason) != 0
3264 && decompile_count() >= m_cutoff) {
3265 // Too many recompiles globally, and we have seen this sort of trap.
3266 // Use cumulative decompile_count, not just md->decompile_count.
3267 if (log())
3268 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3269 Deoptimization::trap_reason_name(reason),
3270 md->trap_count(reason), trap_count(reason),
3271 md->decompile_count(), decompile_count());
3272 return true;
3273 } else {
3274 // The coast is clear.
3275 return false;
3276 }
3277 }
3280 #ifndef PRODUCT
3281 //------------------------------verify_graph_edges---------------------------
3282 // Walk the Graph and verify that there is a one-to-one correspondence
3283 // between Use-Def edges and Def-Use edges in the graph.
3284 void Compile::verify_graph_edges(bool no_dead_code) {
3285 if (VerifyGraphEdges) {
3286 ResourceArea *area = Thread::current()->resource_area();
3287 Unique_Node_List visited(area);
3288 // Call recursive graph walk to check edges
3289 _root->verify_edges(visited);
3290 if (no_dead_code) {
3291 // Now make sure that no visited node is used by an unvisited node.
3292 bool dead_nodes = 0;
3293 Unique_Node_List checked(area);
3294 while (visited.size() > 0) {
3295 Node* n = visited.pop();
3296 checked.push(n);
3297 for (uint i = 0; i < n->outcnt(); i++) {
3298 Node* use = n->raw_out(i);
3299 if (checked.member(use)) continue; // already checked
3300 if (visited.member(use)) continue; // already in the graph
3301 if (use->is_Con()) continue; // a dead ConNode is OK
3302 // At this point, we have found a dead node which is DU-reachable.
3303 if (dead_nodes++ == 0)
3304 tty->print_cr("*** Dead nodes reachable via DU edges:");
3305 use->dump(2);
3306 tty->print_cr("---");
3307 checked.push(use); // No repeats; pretend it is now checked.
3308 }
3309 }
3310 assert(dead_nodes == 0, "using nodes must be reachable from root");
3311 }
3312 }
3313 }
3314 #endif
3316 // The Compile object keeps track of failure reasons separately from the ciEnv.
3317 // This is required because there is not quite a 1-1 relation between the
3318 // ciEnv and its compilation task and the Compile object. Note that one
3319 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3320 // to backtrack and retry without subsuming loads. Other than this backtracking
3321 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3322 // by the logic in C2Compiler.
3323 void Compile::record_failure(const char* reason) {
3324 if (log() != NULL) {
3325 log()->elem("failure reason='%s' phase='compile'", reason);
3326 }
3327 if (_failure_reason == NULL) {
3328 // Record the first failure reason.
3329 _failure_reason = reason;
3330 }
3332 EventCompilerFailure event;
3333 if (event.should_commit()) {
3334 event.set_compileID(Compile::compile_id());
3335 event.set_failure(reason);
3336 event.commit();
3337 }
3339 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3340 C->print_method(PHASE_FAILURE);
3341 }
3342 _root = NULL; // flush the graph, too
3343 }
3345 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3346 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3347 _phase_name(name), _dolog(dolog)
3348 {
3349 if (dolog) {
3350 C = Compile::current();
3351 _log = C->log();
3352 } else {
3353 C = NULL;
3354 _log = NULL;
3355 }
3356 if (_log != NULL) {
3357 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3358 _log->stamp();
3359 _log->end_head();
3360 }
3361 }
3363 Compile::TracePhase::~TracePhase() {
3365 C = Compile::current();
3366 if (_dolog) {
3367 _log = C->log();
3368 } else {
3369 _log = NULL;
3370 }
3372 #ifdef ASSERT
3373 if (PrintIdealNodeCount) {
3374 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3375 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3376 }
3378 if (VerifyIdealNodeCount) {
3379 Compile::current()->print_missing_nodes();
3380 }
3381 #endif
3383 if (_log != NULL) {
3384 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3385 }
3386 }
3388 //=============================================================================
3389 // Two Constant's are equal when the type and the value are equal.
3390 bool Compile::Constant::operator==(const Constant& other) {
3391 if (type() != other.type() ) return false;
3392 if (can_be_reused() != other.can_be_reused()) return false;
3393 // For floating point values we compare the bit pattern.
3394 switch (type()) {
3395 case T_FLOAT: return (_v._value.i == other._v._value.i);
3396 case T_LONG:
3397 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3398 case T_OBJECT:
3399 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3400 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3401 case T_METADATA: return (_v._metadata == other._v._metadata);
3402 default: ShouldNotReachHere();
3403 }
3404 return false;
3405 }
3407 static int type_to_size_in_bytes(BasicType t) {
3408 switch (t) {
3409 case T_LONG: return sizeof(jlong );
3410 case T_FLOAT: return sizeof(jfloat );
3411 case T_DOUBLE: return sizeof(jdouble);
3412 case T_METADATA: return sizeof(Metadata*);
3413 // We use T_VOID as marker for jump-table entries (labels) which
3414 // need an internal word relocation.
3415 case T_VOID:
3416 case T_ADDRESS:
3417 case T_OBJECT: return sizeof(jobject);
3418 }
3420 ShouldNotReachHere();
3421 return -1;
3422 }
3424 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3425 // sort descending
3426 if (a->freq() > b->freq()) return -1;
3427 if (a->freq() < b->freq()) return 1;
3428 return 0;
3429 }
3431 void Compile::ConstantTable::calculate_offsets_and_size() {
3432 // First, sort the array by frequencies.
3433 _constants.sort(qsort_comparator);
3435 #ifdef ASSERT
3436 // Make sure all jump-table entries were sorted to the end of the
3437 // array (they have a negative frequency).
3438 bool found_void = false;
3439 for (int i = 0; i < _constants.length(); i++) {
3440 Constant con = _constants.at(i);
3441 if (con.type() == T_VOID)
3442 found_void = true; // jump-tables
3443 else
3444 assert(!found_void, "wrong sorting");
3445 }
3446 #endif
3448 int offset = 0;
3449 for (int i = 0; i < _constants.length(); i++) {
3450 Constant* con = _constants.adr_at(i);
3452 // Align offset for type.
3453 int typesize = type_to_size_in_bytes(con->type());
3454 offset = align_size_up(offset, typesize);
3455 con->set_offset(offset); // set constant's offset
3457 if (con->type() == T_VOID) {
3458 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3459 offset = offset + typesize * n->outcnt(); // expand jump-table
3460 } else {
3461 offset = offset + typesize;
3462 }
3463 }
3465 // Align size up to the next section start (which is insts; see
3466 // CodeBuffer::align_at_start).
3467 assert(_size == -1, "already set?");
3468 _size = align_size_up(offset, CodeEntryAlignment);
3469 }
3471 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3472 MacroAssembler _masm(&cb);
3473 for (int i = 0; i < _constants.length(); i++) {
3474 Constant con = _constants.at(i);
3475 address constant_addr;
3476 switch (con.type()) {
3477 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3478 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3479 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3480 case T_OBJECT: {
3481 jobject obj = con.get_jobject();
3482 int oop_index = _masm.oop_recorder()->find_index(obj);
3483 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3484 break;
3485 }
3486 case T_ADDRESS: {
3487 address addr = (address) con.get_jobject();
3488 constant_addr = _masm.address_constant(addr);
3489 break;
3490 }
3491 // We use T_VOID as marker for jump-table entries (labels) which
3492 // need an internal word relocation.
3493 case T_VOID: {
3494 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3495 // Fill the jump-table with a dummy word. The real value is
3496 // filled in later in fill_jump_table.
3497 address dummy = (address) n;
3498 constant_addr = _masm.address_constant(dummy);
3499 // Expand jump-table
3500 for (uint i = 1; i < n->outcnt(); i++) {
3501 address temp_addr = _masm.address_constant(dummy + i);
3502 assert(temp_addr, "consts section too small");
3503 }
3504 break;
3505 }
3506 case T_METADATA: {
3507 Metadata* obj = con.get_metadata();
3508 int metadata_index = _masm.oop_recorder()->find_index(obj);
3509 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3510 break;
3511 }
3512 default: ShouldNotReachHere();
3513 }
3514 assert(constant_addr, "consts section too small");
3515 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3516 }
3517 }
3519 int Compile::ConstantTable::find_offset(Constant& con) const {
3520 int idx = _constants.find(con);
3521 assert(idx != -1, "constant must be in constant table");
3522 int offset = _constants.at(idx).offset();
3523 assert(offset != -1, "constant table not emitted yet?");
3524 return offset;
3525 }
3527 void Compile::ConstantTable::add(Constant& con) {
3528 if (con.can_be_reused()) {
3529 int idx = _constants.find(con);
3530 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3531 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3532 return;
3533 }
3534 }
3535 (void) _constants.append(con);
3536 }
3538 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3539 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3540 Constant con(type, value, b->_freq);
3541 add(con);
3542 return con;
3543 }
3545 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3546 Constant con(metadata);
3547 add(con);
3548 return con;
3549 }
3551 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3552 jvalue value;
3553 BasicType type = oper->type()->basic_type();
3554 switch (type) {
3555 case T_LONG: value.j = oper->constantL(); break;
3556 case T_FLOAT: value.f = oper->constantF(); break;
3557 case T_DOUBLE: value.d = oper->constantD(); break;
3558 case T_OBJECT:
3559 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3560 case T_METADATA: return add((Metadata*)oper->constant()); break;
3561 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3562 }
3563 return add(n, type, value);
3564 }
3566 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3567 jvalue value;
3568 // We can use the node pointer here to identify the right jump-table
3569 // as this method is called from Compile::Fill_buffer right before
3570 // the MachNodes are emitted and the jump-table is filled (means the
3571 // MachNode pointers do not change anymore).
3572 value.l = (jobject) n;
3573 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3574 add(con);
3575 return con;
3576 }
3578 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3579 // If called from Compile::scratch_emit_size do nothing.
3580 if (Compile::current()->in_scratch_emit_size()) return;
3582 assert(labels.is_nonempty(), "must be");
3583 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3585 // Since MachConstantNode::constant_offset() also contains
3586 // table_base_offset() we need to subtract the table_base_offset()
3587 // to get the plain offset into the constant table.
3588 int offset = n->constant_offset() - table_base_offset();
3590 MacroAssembler _masm(&cb);
3591 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3593 for (uint i = 0; i < n->outcnt(); i++) {
3594 address* constant_addr = &jump_table_base[i];
3595 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i)));
3596 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3597 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3598 }
3599 }
3601 void Compile::dump_inlining() {
3602 if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) {
3603 // Print inlining message for candidates that we couldn't inline
3604 // for lack of space or non constant receiver
3605 for (int i = 0; i < _late_inlines.length(); i++) {
3606 CallGenerator* cg = _late_inlines.at(i);
3607 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3608 }
3609 Unique_Node_List useful;
3610 useful.push(root());
3611 for (uint next = 0; next < useful.size(); ++next) {
3612 Node* n = useful.at(next);
3613 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3614 CallNode* call = n->as_Call();
3615 CallGenerator* cg = call->generator();
3616 cg->print_inlining_late("receiver not constant");
3617 }
3618 uint max = n->len();
3619 for ( uint i = 0; i < max; ++i ) {
3620 Node *m = n->in(i);
3621 if ( m == NULL ) continue;
3622 useful.push(m);
3623 }
3624 }
3625 for (int i = 0; i < _print_inlining_list->length(); i++) {
3626 tty->print(_print_inlining_list->at(i).ss()->as_string());
3627 }
3628 }
3629 }
3631 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3632 if (n1->Opcode() < n2->Opcode()) return -1;
3633 else if (n1->Opcode() > n2->Opcode()) return 1;
3635 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()));
3636 for (uint i = 1; i < n1->req(); i++) {
3637 if (n1->in(i) < n2->in(i)) return -1;
3638 else if (n1->in(i) > n2->in(i)) return 1;
3639 }
3641 return 0;
3642 }
3644 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3645 Node* n1 = *n1p;
3646 Node* n2 = *n2p;
3648 return cmp_expensive_nodes(n1, n2);
3649 }
3651 void Compile::sort_expensive_nodes() {
3652 if (!expensive_nodes_sorted()) {
3653 _expensive_nodes->sort(cmp_expensive_nodes);
3654 }
3655 }
3657 bool Compile::expensive_nodes_sorted() const {
3658 for (int i = 1; i < _expensive_nodes->length(); i++) {
3659 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3660 return false;
3661 }
3662 }
3663 return true;
3664 }
3666 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3667 if (_expensive_nodes->length() == 0) {
3668 return false;
3669 }
3671 assert(OptimizeExpensiveOps, "optimization off?");
3673 // Take this opportunity to remove dead nodes from the list
3674 int j = 0;
3675 for (int i = 0; i < _expensive_nodes->length(); i++) {
3676 Node* n = _expensive_nodes->at(i);
3677 if (!n->is_unreachable(igvn)) {
3678 assert(n->is_expensive(), "should be expensive");
3679 _expensive_nodes->at_put(j, n);
3680 j++;
3681 }
3682 }
3683 _expensive_nodes->trunc_to(j);
3685 // Then sort the list so that similar nodes are next to each other
3686 // and check for at least two nodes of identical kind with same data
3687 // inputs.
3688 sort_expensive_nodes();
3690 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3691 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3692 return true;
3693 }
3694 }
3696 return false;
3697 }
3699 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3700 if (_expensive_nodes->length() == 0) {
3701 return;
3702 }
3704 assert(OptimizeExpensiveOps, "optimization off?");
3706 // Sort to bring similar nodes next to each other and clear the
3707 // control input of nodes for which there's only a single copy.
3708 sort_expensive_nodes();
3710 int j = 0;
3711 int identical = 0;
3712 int i = 0;
3713 for (; i < _expensive_nodes->length()-1; i++) {
3714 assert(j <= i, "can't write beyond current index");
3715 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3716 identical++;
3717 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3718 continue;
3719 }
3720 if (identical > 0) {
3721 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3722 identical = 0;
3723 } else {
3724 Node* n = _expensive_nodes->at(i);
3725 igvn.hash_delete(n);
3726 n->set_req(0, NULL);
3727 igvn.hash_insert(n);
3728 }
3729 }
3730 if (identical > 0) {
3731 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3732 } else if (_expensive_nodes->length() >= 1) {
3733 Node* n = _expensive_nodes->at(i);
3734 igvn.hash_delete(n);
3735 n->set_req(0, NULL);
3736 igvn.hash_insert(n);
3737 }
3738 _expensive_nodes->trunc_to(j);
3739 }
3741 void Compile::add_expensive_node(Node * n) {
3742 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3743 assert(n->is_expensive(), "expensive nodes with non-null control here only");
3744 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
3745 if (OptimizeExpensiveOps) {
3746 _expensive_nodes->append(n);
3747 } else {
3748 // Clear control input and let IGVN optimize expensive nodes if
3749 // OptimizeExpensiveOps is off.
3750 n->set_req(0, NULL);
3751 }
3752 }
3754 // Auxiliary method to support randomized stressing/fuzzing.
3755 //
3756 // This method can be called the arbitrary number of times, with current count
3757 // as the argument. The logic allows selecting a single candidate from the
3758 // running list of candidates as follows:
3759 // int count = 0;
3760 // Cand* selected = null;
3761 // while(cand = cand->next()) {
3762 // if (randomized_select(++count)) {
3763 // selected = cand;
3764 // }
3765 // }
3766 //
3767 // Including count equalizes the chances any candidate is "selected".
3768 // This is useful when we don't have the complete list of candidates to choose
3769 // from uniformly. In this case, we need to adjust the randomicity of the
3770 // selection, or else we will end up biasing the selection towards the latter
3771 // candidates.
3772 //
3773 // Quick back-envelope calculation shows that for the list of n candidates
3774 // the equal probability for the candidate to persist as "best" can be
3775 // achieved by replacing it with "next" k-th candidate with the probability
3776 // of 1/k. It can be easily shown that by the end of the run, the
3777 // probability for any candidate is converged to 1/n, thus giving the
3778 // uniform distribution among all the candidates.
3779 //
3780 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
3781 #define RANDOMIZED_DOMAIN_POW 29
3782 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
3783 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
3784 bool Compile::randomized_select(int count) {
3785 assert(count > 0, "only positive");
3786 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
3787 }