1 /*
2 * Copyright (c) 2009, 2025, 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 */
24
25 #include "ci/ciSymbols.hpp"
26 #include "classfile/javaClasses.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "opto/callnode.hpp"
29 #include "opto/graphKit.hpp"
30 #include "opto/idealKit.hpp"
31 #include "opto/rootnode.hpp"
32 #include "opto/runtime.hpp"
33 #include "opto/stringopts.hpp"
34 #include "runtime/atomicAccess.hpp"
35 #include "runtime/stubRoutines.hpp"
36
37 #define __ kit.
38
39 class StringConcat : public ResourceObj {
40 private:
41 PhaseStringOpts* _stringopts;
42 AllocateNode* _begin; // The allocation the begins the pattern
43 CallStaticJavaNode* _end; // The final call of the pattern. Will either be
44 // SB.toString or String.<init>(SB.toString)
45 bool _multiple; // indicates this is a fusion of two or more
46 // separate StringBuilders
47
48 Node* _arguments; // The list of arguments to be concatenated
49 GrowableArray<int> _mode; // into a String along with a mode flag
50 // indicating how to treat the value.
51 Node_List _constructors; // List of constructors (many in case of stacked concat)
52 Node_List _control; // List of control nodes that will be deleted
53 Node_List _uncommon_traps; // Uncommon traps that needs to be rewritten
54 // to restart at the initial JVMState.
55
56 static constexpr uint STACKED_CONCAT_UPPER_BOUND = 256; // argument limit for a merged concat.
57 // The value 256 was derived by measuring
58 // compilation time on variable length sequences
59 // of stackable concatenations and chosen to keep
60 // a safe margin to any critical point.
61 public:
62 // Mode for converting arguments to Strings
63 enum {
64 StringMode,
65 IntMode,
66 CharMode,
67 StringNullCheckMode,
68 NegativeIntCheckMode
69 };
70
71 StringConcat(PhaseStringOpts* stringopts, CallStaticJavaNode* end):
72 _stringopts(stringopts),
73 _begin(nullptr),
74 _end(end),
75 _multiple(false) {
76 _arguments = new Node(1);
77 _arguments->del_req(0);
78 }
79
80 bool validate_mem_flow();
81 bool validate_control_flow();
82
83 StringConcat* merge(StringConcat* other, Node* arg);
84
85 void set_allocation(AllocateNode* alloc) {
86 _begin = alloc;
87 }
88
89 void append(Node* value, int mode) {
90 _arguments->add_req(value);
91 _mode.append(mode);
92 }
93 void push(Node* value, int mode) {
94 _arguments->ins_req(0, value);
95 _mode.insert_before(0, mode);
96 }
97
98 void push_string(Node* value) {
99 push(value, StringMode);
100 }
101
102 void push_string_null_check(Node* value) {
103 push(value, StringNullCheckMode);
104 }
105
106 void push_negative_int_check(Node* value) {
107 push(value, NegativeIntCheckMode);
108 }
109
110 void push_int(Node* value) {
111 push(value, IntMode);
112 }
113
114 void push_char(Node* value) {
115 push(value, CharMode);
116 }
117
118 static bool is_SB_toString(Node* call) {
119 if (call->is_CallStaticJava()) {
120 CallStaticJavaNode* csj = call->as_CallStaticJava();
121 ciMethod* m = csj->method();
122 if (m != nullptr &&
123 (m->intrinsic_id() == vmIntrinsics::_StringBuilder_toString ||
124 m->intrinsic_id() == vmIntrinsics::_StringBuffer_toString)) {
125 return true;
126 }
127 }
128 return false;
129 }
130
131 static Node* skip_string_null_check(Node* value) {
132 // Look for a diamond shaped Null check of toString() result
133 // (could be code from String.valueOf()):
134 // (Proj == nullptr) ? "null":"CastPP(Proj)#Notnull
135 if (value->is_Phi()) {
136 int true_path = value->as_Phi()->is_diamond_phi();
137 if (true_path != 0) {
138 // phi->region->if_proj->ifnode->bool
139 BoolNode* b = value->in(0)->in(1)->in(0)->in(1)->as_Bool();
140 Node* cmp = b->in(1);
141 Node* v1 = cmp->in(1);
142 Node* v2 = cmp->in(2);
143 // Null check of the return of toString which can simply be skipped.
144 if (b->_test._test == BoolTest::ne &&
145 v2->bottom_type() == TypePtr::NULL_PTR &&
146 value->in(true_path)->Opcode() == Op_CastPP &&
147 value->in(true_path)->in(1) == v1 &&
148 v1->is_Proj() && is_SB_toString(v1->in(0))) {
149 return v1;
150 }
151 }
152 }
153 return value;
154 }
155
156 Node* argument(int i) {
157 return _arguments->in(i);
158 }
159 Node* argument_uncast(int i) {
160 Node* arg = argument(i);
161 int amode = mode(i);
162 if (amode == StringConcat::StringMode ||
163 amode == StringConcat::StringNullCheckMode) {
164 arg = skip_string_null_check(arg);
165 }
166 return arg;
167 }
168 void set_argument(int i, Node* value) {
169 _arguments->set_req(i, value);
170 }
171 int num_arguments() {
172 return _mode.length();
173 }
174 int mode(int i) {
175 return _mode.at(i);
176 }
177 void add_control(Node* ctrl) {
178 assert(!_control.contains(ctrl), "only push once");
179 _control.push(ctrl);
180 }
181 void add_constructor(Node* init) {
182 assert(!_constructors.contains(init), "only push once");
183 _constructors.push(init);
184 }
185 CallStaticJavaNode* end() { return _end; }
186 AllocateNode* begin() { return _begin; }
187
188 void eliminate_unneeded_control();
189 void eliminate_initialize(InitializeNode* init);
190 void eliminate_call(CallNode* call);
191
192 void maybe_log_transform() {
193 CompileLog* log = _stringopts->C->log();
194 if (log != nullptr) {
195 log->head("replace_string_concat arguments='%d' multiple='%d'", num_arguments(), _multiple);
196 JVMState* p = _begin->jvms();
197 while (p != nullptr) {
198 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
199 p = p->caller();
200 }
201 log->tail("replace_string_concat");
202 }
203 }
204
205 void convert_uncommon_traps(GraphKit& kit, const JVMState* jvms) {
206 for (uint u = 0; u < _uncommon_traps.size(); u++) {
207 Node* uct = _uncommon_traps.at(u);
208
209 // Build a new call using the jvms state of the allocate
210 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
211 const TypeFunc* call_type = OptoRuntime::uncommon_trap_Type();
212 const TypePtr* no_memory_effects = nullptr;
213 Compile* C = _stringopts->C;
214 CallStaticJavaNode* call = new CallStaticJavaNode(call_type, call_addr, "uncommon_trap",
215 no_memory_effects);
216 for (int e = 0; e < TypeFunc::Parms; e++) {
217 call->init_req(e, uct->in(e));
218 }
219 // Set the trap request to record intrinsic failure if this trap
220 // is taken too many times. Ideally we would handle then traps by
221 // doing the original bookkeeping in the MDO so that if it caused
222 // the code to be thrown out we could still recompile and use the
223 // optimization. Failing the uncommon traps doesn't really mean
224 // that the optimization is a bad idea but there's no other way to
225 // do the MDO updates currently.
226 int trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_intrinsic,
227 Deoptimization::Action_make_not_entrant);
228 call->init_req(TypeFunc::Parms, __ intcon(trap_request));
229 kit.add_safepoint_edges(call);
230
231 _stringopts->gvn()->transform(call);
232 C->gvn_replace_by(uct, call);
233 uct->disconnect_inputs(C);
234 }
235 }
236
237 void cleanup() {
238 // disconnect the hook node
239 _arguments->disconnect_inputs(_stringopts->C);
240 }
241 };
242
243
244 void StringConcat::eliminate_unneeded_control() {
245 for (uint i = 0; i < _control.size(); i++) {
246 Node* n = _control.at(i);
247 if (n->is_Allocate()) {
248 eliminate_initialize(n->as_Allocate()->initialization());
249 }
250 if (n->is_Call()) {
251 if (n != _end) {
252 eliminate_call(n->as_Call());
253 }
254 } else if (n->is_IfTrue()) {
255 Compile* C = _stringopts->C;
256 C->gvn_replace_by(n, n->in(0)->in(0));
257 // get rid of the other projection
258 C->gvn_replace_by(n->in(0)->as_If()->proj_out(false), C->top());
259 } else if (n->is_Region()) {
260 Node* iff = n->in(1)->in(0);
261 assert(n->req() == 3 && n->in(2)->in(0) == iff, "not a diamond");
262 assert(iff->is_If(), "no if for the diamond");
263 Node* bol = iff->in(1);
264 if (bol->is_Con()) {
265 // A BoolNode shared by two diamond Region/If sub-graphs
266 // was replaced by a constant zero in a previous call to this method.
267 // Do nothing as the transformation in the previous call ensures both are folded away.
268 assert(bol == _stringopts->gvn()->intcon(0), "shared condition should have been set to false");
269 continue;
270 }
271 assert(bol->is_Bool(), "unexpected if shape");
272 Node* cmp = bol->in(1);
273 assert(cmp->is_Cmp(), "unexpected if shape");
274 if (cmp->in(1)->is_top() || cmp->in(2)->is_top()) {
275 // This region should lose its Phis. They are removed either in PhaseRemoveUseless (for data phis) or in IGVN
276 // (for memory phis). During IGVN, there is a chance that the If folds to top before the Region is processed
277 // which then causes a reachable part of the graph to become dead. To prevent this, set the boolean input of
278 // the If to a constant to nicely let the diamond Region/If fold away.
279 Compile* C = _stringopts->C;
280 C->gvn_replace_by(iff->in(1), _stringopts->gvn()->intcon(0));
281 }
282 }
283 }
284 }
285
286
287 StringConcat* StringConcat::merge(StringConcat* other, Node* arg) {
288 StringConcat* result = new StringConcat(_stringopts, _end);
289 for (uint x = 0; x < _control.size(); x++) {
290 Node* n = _control.at(x);
291 if (n->is_Call()) {
292 result->_control.push(n);
293 }
294 }
295 for (uint x = 0; x < other->_control.size(); x++) {
296 Node* n = other->_control.at(x);
297 if (n->is_Call()) {
298 result->_control.push(n);
299 }
300 }
301 assert(result->_control.contains(other->_end), "what?");
302 assert(result->_control.contains(_begin), "what?");
303
304 uint arguments_appended = 0;
305 for (int x = 0; x < num_arguments(); x++) {
306 Node* argx = argument_uncast(x);
307 if (argx == arg) {
308 // replace the toString result with the all the arguments that
309 // made up the other StringConcat
310 for (int y = 0; y < other->num_arguments(); y++) {
311 result->append(other->argument(y), other->mode(y));
312 }
313 arguments_appended += other->num_arguments();
314 } else {
315 result->append(argx, mode(x));
316 arguments_appended++;
317 }
318 // Check if this concatenation would result in an excessive number of arguments
319 // -- leading to high memory use, compilation time, and later, a large number of IR nodes
320 // -- and bail out in that case.
321 if (arguments_appended > STACKED_CONCAT_UPPER_BOUND) {
322 #ifndef PRODUCT
323 if (PrintOptimizeStringConcat) {
324 tty->print_cr("Merge candidate of length %d exceeds argument limit", arguments_appended);
325 }
326 #endif
327 return nullptr;
328 }
329 }
330 result->set_allocation(other->_begin);
331 for (uint i = 0; i < _constructors.size(); i++) {
332 result->add_constructor(_constructors.at(i));
333 }
334 for (uint i = 0; i < other->_constructors.size(); i++) {
335 result->add_constructor(other->_constructors.at(i));
336 }
337 result->_multiple = true;
338 return result;
339 }
340
341
342 void StringConcat::eliminate_call(CallNode* call) {
343 Compile* C = _stringopts->C;
344 CallProjections* projs = call->extract_projections(false);
345 if (projs->fallthrough_catchproj != nullptr) {
346 C->gvn_replace_by(projs->fallthrough_catchproj, call->in(TypeFunc::Control));
347 }
348 if (projs->fallthrough_memproj != nullptr) {
349 C->gvn_replace_by(projs->fallthrough_memproj, call->in(TypeFunc::Memory));
350 }
351 if (projs->catchall_memproj != nullptr) {
352 C->gvn_replace_by(projs->catchall_memproj, C->top());
353 }
354 if (projs->fallthrough_ioproj != nullptr) {
355 C->gvn_replace_by(projs->fallthrough_ioproj, call->in(TypeFunc::I_O));
356 }
357 if (projs->catchall_ioproj != nullptr) {
358 C->gvn_replace_by(projs->catchall_ioproj, C->top());
359 }
360 if (projs->catchall_catchproj != nullptr) {
361 // EA can't cope with the partially collapsed graph this
362 // creates so put it on the worklist to be collapsed later.
363 for (SimpleDUIterator i(projs->catchall_catchproj); i.has_next(); i.next()) {
364 Node *use = i.get();
365 int opc = use->Opcode();
366 if (opc == Op_CreateEx || opc == Op_Region) {
367 _stringopts->record_dead_node(use);
368 }
369 }
370 C->gvn_replace_by(projs->catchall_catchproj, C->top());
371 }
372 if (projs->resproj[0] != nullptr) {
373 assert(projs->nb_resproj == 1, "unexpected number of results");
374 C->gvn_replace_by(projs->resproj[0], C->top());
375 }
376 C->gvn_replace_by(call, C->top());
377 }
378
379 void StringConcat::eliminate_initialize(InitializeNode* init) {
380 Compile* C = _stringopts->C;
381
382 // Eliminate Initialize node.
383 assert(init->outcnt() <= 2, "only a control and memory projection expected");
384 assert(init->req() <= InitializeNode::RawStores, "no pending inits");
385 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
386 if (ctrl_proj != nullptr) {
387 C->gvn_replace_by(ctrl_proj, init->in(TypeFunc::Control));
388 }
389 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
390 if (mem_proj != nullptr) {
391 Node *mem = init->in(TypeFunc::Memory);
392 C->gvn_replace_by(mem_proj, mem);
393 }
394 C->gvn_replace_by(init, C->top());
395 init->disconnect_inputs(C);
396 }
397
398 Node_List PhaseStringOpts::collect_toString_calls() {
399 Node_List string_calls;
400 Node_List worklist;
401
402 _visited.clear();
403
404 // Prime the worklist
405 for (uint i = 1; i < C->root()->len(); i++) {
406 Node* n = C->root()->in(i);
407 if (n != nullptr && !_visited.test_set(n->_idx)) {
408 worklist.push(n);
409 }
410 }
411
412 uint encountered = 0;
413 while (worklist.size() > 0) {
414 Node* ctrl = worklist.pop();
415 if (StringConcat::is_SB_toString(ctrl)) {
416 CallStaticJavaNode* csj = ctrl->as_CallStaticJava();
417 string_calls.push(csj);
418 encountered++;
419 }
420 if (ctrl->in(0) != nullptr && !_visited.test_set(ctrl->in(0)->_idx)) {
421 worklist.push(ctrl->in(0));
422 }
423 if (ctrl->is_Region()) {
424 for (uint i = 1; i < ctrl->len(); i++) {
425 if (ctrl->in(i) != nullptr && !_visited.test_set(ctrl->in(i)->_idx)) {
426 worklist.push(ctrl->in(i));
427 }
428 }
429 }
430 }
431 #ifndef PRODUCT
432 AtomicAccess::add(&_stropts_total, encountered);
433 #endif
434 return string_calls;
435 }
436
437 // Recognize a fluent-chain of StringBuilder/Buffer. They are either explicit usages
438 // of them or the legacy bytecodes of string concatenation prior to JEP-280. eg.
439 //
440 // String result = new StringBuilder()
441 // .append("foo")
442 // .append("bar")
443 // .append(123)
444 // .toString(); // "foobar123"
445 //
446 // PS: Only a certain subset of constructor and append methods are acceptable.
447 // The criterion is that the length of argument is easy to work out in this phrase.
448 // It will drop complex cases such as Object.
449 //
450 // Since it walks along the receivers of fluent-chain, it will give up if the codeshape is
451 // not "fluent" enough. eg.
452 // StringBuilder sb = new StringBuilder();
453 // sb.append("foo");
454 // sb.toString();
455 //
456 // The receiver of toString method is the result of Allocation Node(CheckCastPP).
457 // The append method is overlooked. It will fail at validate_control_flow() test.
458 //
459 StringConcat* PhaseStringOpts::build_candidate(CallStaticJavaNode* call) {
460 ciMethod* m = call->method();
461 ciSymbol* string_sig;
462 ciSymbol* int_sig;
463 ciSymbol* char_sig;
464 if (m->holder() == C->env()->StringBuilder_klass()) {
465 string_sig = ciSymbols::String_StringBuilder_signature();
466 int_sig = ciSymbols::int_StringBuilder_signature();
467 char_sig = ciSymbols::char_StringBuilder_signature();
468 } else if (m->holder() == C->env()->StringBuffer_klass()) {
469 string_sig = ciSymbols::String_StringBuffer_signature();
470 int_sig = ciSymbols::int_StringBuffer_signature();
471 char_sig = ciSymbols::char_StringBuffer_signature();
472 } else {
473 return nullptr;
474 }
475 #ifndef PRODUCT
476 if (PrintOptimizeStringConcat) {
477 tty->print("considering toString call in ");
478 call->jvms()->dump_spec(tty); tty->cr();
479 }
480 #endif
481
482 StringConcat* sc = new StringConcat(this, call);
483 AllocateNode* alloc = nullptr;
484
485 // possible opportunity for StringBuilder fusion
486 CallStaticJavaNode* cnode = call;
487 while (cnode) {
488 Node* recv = cnode->in(TypeFunc::Parms)->uncast();
489 if (recv->is_Proj()) {
490 recv = recv->in(0);
491 }
492 cnode = recv->isa_CallStaticJava();
493 if (cnode == nullptr) {
494 alloc = recv->isa_Allocate();
495 if (alloc == nullptr) {
496 break;
497 }
498 // Find the constructor call
499 Node* result = alloc->result_cast();
500 if (result == nullptr || !result->is_CheckCastPP() || alloc->in(TypeFunc::Memory)->is_top()) {
501 // strange looking allocation
502 #ifndef PRODUCT
503 if (PrintOptimizeStringConcat) {
504 tty->print("giving up because allocation looks strange ");
505 alloc->jvms()->dump_spec(tty); tty->cr();
506 }
507 #endif
508 break;
509 }
510 Node* constructor = nullptr;
511 for (SimpleDUIterator i(result); i.has_next(); i.next()) {
512 CallStaticJavaNode *use = i.get()->isa_CallStaticJava();
513 if (use != nullptr &&
514 use->method() != nullptr &&
515 !use->method()->is_static() &&
516 use->method()->name() == ciSymbols::object_initializer_name() &&
517 use->method()->holder() == m->holder()) {
518 // Matched the constructor.
519 ciSymbol* sig = use->method()->signature()->as_symbol();
520 if (sig == ciSymbols::void_method_signature() ||
521 sig == ciSymbols::int_void_signature() ||
522 sig == ciSymbols::string_void_signature()) {
523 if (sig == ciSymbols::string_void_signature()) {
524 // StringBuilder(String) so pick this up as the first argument
525 assert(use->in(TypeFunc::Parms + 1) != nullptr, "what?");
526 const Type* type = _gvn->type(use->in(TypeFunc::Parms + 1));
527 if (type == TypePtr::NULL_PTR) {
528 // StringBuilder(null) throws exception.
529 #ifndef PRODUCT
530 if (PrintOptimizeStringConcat) {
531 tty->print("giving up because StringBuilder(null) throws exception");
532 alloc->jvms()->dump_spec(tty);
533 tty->cr();
534 }
535 #endif
536 return nullptr;
537 }
538 // StringBuilder(str) argument needs null check.
539 sc->push_string_null_check(use->in(TypeFunc::Parms + 1));
540 } else if (sig == ciSymbols::int_void_signature()) {
541 // StringBuilder(int) case.
542 Node* parm = use->in(TypeFunc::Parms + 1);
543 assert(parm != nullptr, "must exist");
544 const TypeInt* type = _gvn->type(parm)->is_int();
545 if (type->_hi < 0) {
546 // Initial capacity argument is always negative in which case StringBuilder(int) throws
547 // a NegativeArraySizeException. Bail out from string opts.
548 #ifndef PRODUCT
549 if (PrintOptimizeStringConcat) {
550 tty->print("giving up because a negative argument is passed to StringBuilder(int) which "
551 "throws a NegativeArraySizeException");
552 alloc->jvms()->dump_spec(tty);
553 tty->cr();
554 }
555 #endif
556 return nullptr;
557 } else if (type->_lo < 0) {
558 // Argument could be negative: We need a runtime check to throw NegativeArraySizeException in that case.
559 sc->push_negative_int_check(parm);
560 }
561 }
562 // The int variant takes an initial size for the backing
563 // array so just treat it like the void version.
564 constructor = use;
565 } else {
566 #ifndef PRODUCT
567 if (PrintOptimizeStringConcat) {
568 tty->print("unexpected constructor signature: %s", sig->as_utf8());
569 }
570 #endif
571 }
572 break;
573 }
574 }
575 if (constructor == nullptr) {
576 // couldn't find constructor
577 #ifndef PRODUCT
578 if (PrintOptimizeStringConcat) {
579 tty->print("giving up because couldn't find constructor ");
580 alloc->jvms()->dump_spec(tty); tty->cr();
581 }
582 #endif
583 break;
584 }
585
586 // Walked all the way back and found the constructor call so see
587 // if this call converted into a direct string concatenation.
588 sc->add_control(call);
589 sc->add_control(constructor);
590 sc->add_control(alloc);
591 sc->set_allocation(alloc);
592 sc->add_constructor(constructor);
593 if (sc->validate_control_flow() && sc->validate_mem_flow()) {
594 return sc;
595 } else {
596 return nullptr;
597 }
598 } else if (cnode->method() == nullptr) {
599 break;
600 } else if (!cnode->method()->is_static() &&
601 cnode->method()->holder() == m->holder() &&
602 cnode->method()->name() == ciSymbols::append_name() &&
603 (cnode->method()->signature()->as_symbol() == string_sig ||
604 cnode->method()->signature()->as_symbol() == char_sig ||
605 cnode->method()->signature()->as_symbol() == int_sig)) {
606 sc->add_control(cnode);
607 Node* arg = cnode->in(TypeFunc::Parms + 1);
608 if (arg == nullptr || arg->is_top()) {
609 #ifndef PRODUCT
610 if (PrintOptimizeStringConcat) {
611 tty->print("giving up because the call is effectively dead");
612 cnode->jvms()->dump_spec(tty); tty->cr();
613 }
614 #endif
615 break;
616 }
617 if (cnode->method()->signature()->as_symbol() == int_sig) {
618 sc->push_int(arg);
619 } else if (cnode->method()->signature()->as_symbol() == char_sig) {
620 sc->push_char(arg);
621 } else {
622 if (arg->is_Proj() && arg->in(0)->is_CallStaticJava()) {
623 CallStaticJavaNode* csj = arg->in(0)->as_CallStaticJava();
624 if (csj->method() != nullptr &&
625 csj->method()->intrinsic_id() == vmIntrinsics::_Integer_toString &&
626 arg->outcnt() == 1) {
627 // _control is the list of StringBuilder calls nodes which
628 // will be replaced by new String code after this optimization.
629 // Integer::toString() call is not part of StringBuilder calls
630 // chain. It could be eliminated only if its result is used
631 // only by this SB calls chain.
632 // Another limitation: it should be used only once because
633 // it is unknown that it is used only by this SB calls chain
634 // until all related SB calls nodes are collected.
635 assert(arg->unique_out() == cnode, "sanity");
636 sc->add_control(csj);
637 sc->push_int(csj->in(TypeFunc::Parms));
638 continue;
639 }
640 }
641 sc->push_string(arg);
642 }
643 continue;
644 } else {
645 // some unhandled signature
646 #ifndef PRODUCT
647 if (PrintOptimizeStringConcat) {
648 tty->print("giving up because encountered unexpected signature ");
649 cnode->tf()->dump(); tty->cr();
650 cnode->in(TypeFunc::Parms + 1)->dump();
651 }
652 #endif
653 break;
654 }
655 }
656 return nullptr;
657 }
658
659
660 PhaseStringOpts::PhaseStringOpts(PhaseGVN* gvn):
661 Phase(StringOpts),
662 _gvn(gvn) {
663
664 assert(OptimizeStringConcat, "shouldn't be here");
665
666 // Collect the types needed to talk about the various slices of memory
667 byte_adr_idx = C->get_alias_index(TypeAryPtr::BYTES);
668
669 // For each locally allocated StringBuffer see if the usages can be
670 // collapsed into a single String construction.
671
672 // Run through the list of allocation looking for SB.toString to see
673 // if it's possible to fuse the usage of the SB into a single String
674 // construction.
675 GrowableArray<StringConcat*> concats;
676 Node_List toStrings = collect_toString_calls();
677 while (toStrings.size() > 0) {
678 StringConcat* sc = build_candidate(toStrings.pop()->as_CallStaticJava());
679 if (sc != nullptr) {
680 concats.push(sc);
681 }
682 }
683
684 // try to coalesce separate concats
685 restart:
686 for (int c = 0; c < concats.length(); c++) {
687 StringConcat* sc = concats.at(c);
688 for (int i = 0; i < sc->num_arguments(); i++) {
689 Node* arg = sc->argument_uncast(i);
690 if (arg->is_Proj() && StringConcat::is_SB_toString(arg->in(0))) {
691 CallStaticJavaNode* csj = arg->in(0)->as_CallStaticJava();
692 for (int o = 0; o < concats.length(); o++) {
693 if (c == o) continue;
694 StringConcat* other = concats.at(o);
695 if (other->end() == csj) {
696 #ifndef PRODUCT
697 if (PrintOptimizeStringConcat) {
698 tty->print_cr("considering stacked concats");
699 }
700 #endif
701
702 StringConcat* merged = sc->merge(other, arg);
703 if (merged != nullptr && merged->validate_control_flow() && merged->validate_mem_flow()) {
704 #ifndef PRODUCT
705 AtomicAccess::inc(&_stropts_merged);
706 if (PrintOptimizeStringConcat) {
707 tty->print_cr("stacking would succeed");
708 }
709 #endif
710 if (c < o) {
711 concats.remove_at(o);
712 concats.at_put(c, merged);
713 } else {
714 concats.remove_at(c);
715 concats.at_put(o, merged);
716 }
717 goto restart;
718 } else {
719 #ifndef PRODUCT
720 if (PrintOptimizeStringConcat) {
721 tty->print_cr("stacking would fail");
722 }
723 #endif
724 }
725 }
726 }
727 }
728 }
729 }
730
731
732 for (int c = 0; c < concats.length(); c++) {
733 StringConcat* sc = concats.at(c);
734 replace_string_concat(sc);
735 }
736
737 remove_dead_nodes();
738 }
739
740 void PhaseStringOpts::record_dead_node(Node* dead) {
741 dead_worklist.push(dead);
742 }
743
744 void PhaseStringOpts::remove_dead_nodes() {
745 // Delete any dead nodes to make things clean enough that escape
746 // analysis doesn't get unhappy.
747 while (dead_worklist.size() > 0) {
748 Node* use = dead_worklist.pop();
749 int opc = use->Opcode();
750 switch (opc) {
751 case Op_Region: {
752 uint i = 1;
753 for (i = 1; i < use->req(); i++) {
754 if (use->in(i) != C->top()) {
755 break;
756 }
757 }
758 if (i >= use->req()) {
759 for (SimpleDUIterator i(use); i.has_next(); i.next()) {
760 Node* m = i.get();
761 if (m->is_Phi()) {
762 dead_worklist.push(m);
763 }
764 }
765 C->gvn_replace_by(use, C->top());
766 }
767 break;
768 }
769 case Op_AddP:
770 case Op_CreateEx: {
771 // Recursively clean up references to CreateEx so EA doesn't
772 // get unhappy about the partially collapsed graph.
773 for (SimpleDUIterator i(use); i.has_next(); i.next()) {
774 Node* m = i.get();
775 if (m->is_AddP()) {
776 dead_worklist.push(m);
777 }
778 }
779 C->gvn_replace_by(use, C->top());
780 break;
781 }
782 case Op_Phi:
783 if (use->in(0) == C->top()) {
784 C->gvn_replace_by(use, C->top());
785 }
786 break;
787 }
788 }
789 }
790
791
792 bool StringConcat::validate_mem_flow() {
793 Compile* C = _stringopts->C;
794
795 for (uint i = 0; i < _control.size(); i++) {
796 #ifndef PRODUCT
797 Node_List path;
798 #endif
799 Node* curr = _control.at(i);
800 if (curr->is_Call() && curr != _begin) { // For all calls except the first allocation
801 // Now here's the main invariant in our case:
802 // For memory between the constructor, and appends, and toString we should only see bottom memory,
803 // produced by the previous call we know about.
804 if (!_constructors.contains(curr)) {
805 NOT_PRODUCT(path.push(curr);)
806 Node* mem = curr->in(TypeFunc::Memory);
807 assert(mem != nullptr, "calls should have memory edge");
808 assert(!mem->is_Phi(), "should be handled by control flow validation");
809 NOT_PRODUCT(path.push(mem);)
810 while (mem->is_MergeMem()) {
811 for (uint i = 1; i < mem->req(); i++) {
812 if (i != Compile::AliasIdxBot && mem->in(i) != C->top()) {
813 #ifndef PRODUCT
814 if (PrintOptimizeStringConcat) {
815 tty->print("fusion has incorrect memory flow (side effects) for ");
816 _begin->jvms()->dump_spec(tty); tty->cr();
817 path.dump();
818 }
819 #endif
820 return false;
821 }
822 }
823 // skip through a potential MergeMem chain, linked through Bot
824 mem = mem->in(Compile::AliasIdxBot);
825 NOT_PRODUCT(path.push(mem);)
826 }
827 // now let it fall through, and see if we have a projection
828 if (mem->is_Proj()) {
829 // Should point to a previous known call
830 Node *prev = mem->in(0);
831 NOT_PRODUCT(path.push(prev);)
832 if (!prev->is_Call() || !_control.contains(prev)) {
833 #ifndef PRODUCT
834 if (PrintOptimizeStringConcat) {
835 tty->print("fusion has incorrect memory flow (unknown call) for ");
836 _begin->jvms()->dump_spec(tty); tty->cr();
837 path.dump();
838 }
839 #endif
840 return false;
841 }
842 } else {
843 assert(mem->is_Store() || mem->is_LoadStore(), "unexpected node type: %s", mem->Name());
844 #ifndef PRODUCT
845 if (PrintOptimizeStringConcat) {
846 tty->print("fusion has incorrect memory flow (unexpected source) for ");
847 _begin->jvms()->dump_spec(tty); tty->cr();
848 path.dump();
849 }
850 #endif
851 return false;
852 }
853 } else {
854 // For memory that feeds into constructors it's more complicated.
855 // However the advantage is that any side effect that happens between the Allocate/Initialize and
856 // the constructor will have to be control-dependent on Initialize.
857 // So we actually don't have to do anything, since it's going to be caught by the control flow
858 // analysis.
859 #ifdef ASSERT
860 // Do a quick verification of the control pattern between the constructor and the initialize node
861 assert(curr->is_Call(), "constructor should be a call");
862 // Go up the control starting from the constructor call
863 Node* ctrl = curr->in(0);
864 IfNode* iff = nullptr;
865 RegionNode* copy = nullptr;
866
867 while (true) {
868 // skip known check patterns
869 if (ctrl->is_Region()) {
870 if (ctrl->as_Region()->is_copy()) {
871 copy = ctrl->as_Region();
872 ctrl = copy->is_copy();
873 } else { // a cast
874 assert(ctrl->req() == 3 &&
875 ctrl->in(1) != nullptr && ctrl->in(1)->is_Proj() &&
876 ctrl->in(2) != nullptr && ctrl->in(2)->is_Proj() &&
877 ctrl->in(1)->in(0) == ctrl->in(2)->in(0) &&
878 ctrl->in(1)->in(0) != nullptr && ctrl->in(1)->in(0)->is_If(),
879 "must be a simple diamond");
880 Node* true_proj = ctrl->in(1)->is_IfTrue() ? ctrl->in(1) : ctrl->in(2);
881 for (SimpleDUIterator i(true_proj); i.has_next(); i.next()) {
882 Node* use = i.get();
883 assert(use == ctrl || use->is_ConstraintCast(),
884 "unexpected user: %s", use->Name());
885 }
886
887 iff = ctrl->in(1)->in(0)->as_If();
888 ctrl = iff->in(0);
889 }
890 } else if (ctrl->is_IfTrue()) { // null checks, class checks
891 iff = ctrl->in(0)->as_If();
892 // Verify that the other arm is an uncommon trap
893 Node* otherproj = iff->proj_out(1 - ctrl->as_Proj()->_con);
894 CallStaticJavaNode* call = otherproj->unique_out()->isa_CallStaticJava();
895 assert(strcmp(call->_name, "uncommon_trap") == 0, "must be uncommon trap");
896 ctrl = iff->in(0);
897 } else {
898 break;
899 }
900 }
901
902 assert(ctrl->is_Proj(), "must be a projection");
903 assert(ctrl->in(0)->is_Initialize(), "should be initialize");
904 for (SimpleDUIterator i(ctrl); i.has_next(); i.next()) {
905 Node* use = i.get();
906 assert(use == copy || use == iff || use == curr || use->is_CheckCastPP() || use->is_Load(),
907 "unexpected user: %s", use->Name());
908 }
909 #endif // ASSERT
910 }
911 }
912 }
913
914 #ifndef PRODUCT
915 if (PrintOptimizeStringConcat) {
916 tty->print("fusion has correct memory flow for ");
917 _begin->jvms()->dump_spec(tty); tty->cr();
918 tty->cr();
919 }
920 #endif
921 return true;
922 }
923
924 bool StringConcat::validate_control_flow() {
925 // We found all the calls and arguments now lets see if it's
926 // safe to transform the graph as we would expect.
927
928 // Check to see if this resulted in too many uncommon traps previously
929 if (Compile::current()->too_many_traps(_begin->jvms()->method(), _begin->jvms()->bci(),
930 Deoptimization::Reason_intrinsic)) {
931 return false;
932 }
933
934 // Walk backwards over the control flow from toString to the
935 // allocation and make sure all the control flow is ok. This
936 // means it's either going to be eliminated once the calls are
937 // removed or it can safely be transformed into an uncommon
938 // trap.
939
940 int null_check_count = 0;
941 Unique_Node_List ctrl_path;
942
943 assert(_control.contains(_begin), "missing");
944 assert(_control.contains(_end), "missing");
945
946 // Collect the nodes that we know about and will eliminate into ctrl_path
947 for (uint i = 0; i < _control.size(); i++) {
948 // Push the call and it's control projection
949 Node* n = _control.at(i);
950 if (n->is_Allocate()) {
951 AllocateNode* an = n->as_Allocate();
952 InitializeNode* init = an->initialization();
953 ctrl_path.push(init);
954 ctrl_path.push(init->as_Multi()->proj_out(0));
955 }
956 if (n->is_Call()) {
957 CallNode* cn = n->as_Call();
958 ctrl_path.push(cn);
959 ctrl_path.push(cn->proj_out(0));
960 ctrl_path.push(cn->proj_out(0)->unique_out());
961 Node* catchproj = cn->proj_out(0)->unique_out()->as_Catch()->proj_out_or_null(0);
962 if (catchproj != nullptr) {
963 ctrl_path.push(catchproj);
964 }
965 } else {
966 ShouldNotReachHere();
967 }
968 }
969
970 // Skip backwards through the control checking for unexpected control flow
971 Node* ptr = _end;
972 bool fail = false;
973 while (ptr != _begin) {
974 if (ptr->is_Call() && ctrl_path.member(ptr)) {
975 ptr = ptr->in(0);
976 } else if (ptr->is_CatchProj() && ctrl_path.member(ptr)) {
977 ptr = ptr->in(0)->in(0)->in(0);
978 assert(ctrl_path.member(ptr), "should be a known piece of control");
979 } else if (ptr->is_IfTrue()) {
980 IfNode* iff = ptr->in(0)->as_If();
981 BoolNode* b = iff->in(1)->isa_Bool();
982
983 if (b == nullptr) {
984 #ifndef PRODUCT
985 if (PrintOptimizeStringConcat) {
986 tty->print_cr("unexpected input to IfNode");
987 iff->in(1)->dump();
988 tty->cr();
989 }
990 #endif
991 fail = true;
992 break;
993 }
994
995 Node* cmp = b->in(1);
996 Node* v1 = cmp->in(1);
997 Node* v2 = cmp->in(2);
998 Node* otherproj = iff->proj_out(1 - ptr->as_Proj()->_con);
999
1000 // Null check of the return of append which can simply be eliminated
1001 if (b->_test._test == BoolTest::ne &&
1002 v2->bottom_type() == TypePtr::NULL_PTR &&
1003 v1->is_Proj() && ctrl_path.member(v1->in(0))) {
1004 // null check of the return value of the append
1005 null_check_count++;
1006 if (otherproj->outcnt() == 1) {
1007 CallStaticJavaNode* call = otherproj->unique_out()->isa_CallStaticJava();
1008 if (call != nullptr && call->_name != nullptr && strcmp(call->_name, "uncommon_trap") == 0) {
1009 ctrl_path.push(call);
1010 }
1011 }
1012 _control.push(ptr);
1013 ptr = ptr->in(0)->in(0);
1014 continue;
1015 }
1016
1017 // A test which leads to an uncommon trap. It is safe to convert the trap
1018 // into a trap that restarts at the beginning as long as its test does not
1019 // depend on intermediate results of the candidate chain.
1020 // at the beginning.
1021 if (otherproj->outcnt() == 1) {
1022 CallStaticJavaNode* call = otherproj->unique_out()->isa_CallStaticJava();
1023 if (call != nullptr && call->_name != nullptr && strcmp(call->_name, "uncommon_trap") == 0) {
1024 // First check for dependency on a toString that is going away during stacked concats.
1025 if (_multiple &&
1026 ((v1->is_Proj() && is_SB_toString(v1->in(0)) && ctrl_path.member(v1->in(0))) ||
1027 (v2->is_Proj() && is_SB_toString(v2->in(0)) && ctrl_path.member(v2->in(0))))) {
1028 // iftrue -> if -> bool -> cmpp -> resproj -> tostring
1029 fail = true;
1030 break;
1031 }
1032 // control flow leads to uct so should be ok
1033 _uncommon_traps.push(call);
1034 ctrl_path.push(call);
1035 ptr = ptr->in(0)->in(0);
1036 continue;
1037 }
1038 }
1039
1040 #ifndef PRODUCT
1041 // Some unexpected control flow we don't know how to handle.
1042 if (PrintOptimizeStringConcat) {
1043 tty->print_cr("failing with unknown test");
1044 b->dump();
1045 cmp->dump();
1046 v1->dump();
1047 v2->dump();
1048 tty->cr();
1049 }
1050 #endif
1051 fail = true;
1052 break;
1053 } else if (ptr->is_Proj() && ptr->in(0)->is_Initialize()) {
1054 // Check for side effect between Initialize and the constructor
1055 for (SimpleDUIterator iter(ptr); iter.has_next(); iter.next()) {
1056 Node* use = iter.get();
1057 if (!use->is_CFG() && !use->is_CheckCastPP() && !use->is_Load()) {
1058 #ifndef PRODUCT
1059 if (PrintOptimizeStringConcat) {
1060 tty->print_cr("unexpected control use of Initialize");
1061 ptr->in(0)->dump(); // Initialize node
1062 use->dump(1);
1063 }
1064 #endif
1065 fail = true;
1066 break;
1067 }
1068 }
1069 ptr = ptr->in(0)->in(0);
1070 } else if (ptr->is_Region()) {
1071 Node* copy = ptr->as_Region()->is_copy();
1072 if (copy != nullptr) {
1073 ptr = copy;
1074 continue;
1075 }
1076 if (ptr->req() == 3 &&
1077 ptr->in(1) != nullptr && ptr->in(1)->is_Proj() &&
1078 ptr->in(2) != nullptr && ptr->in(2)->is_Proj() &&
1079 ptr->in(1)->in(0) == ptr->in(2)->in(0) &&
1080 ptr->in(1)->in(0) != nullptr && ptr->in(1)->in(0)->is_If()) {
1081 // Simple diamond.
1082 // XXX should check for possibly merging stores. simple data merges are ok.
1083 // The IGVN will make this simple diamond go away when it
1084 // transforms the Region. Make sure it sees it.
1085 Compile::current()->record_for_igvn(ptr);
1086 _control.push(ptr);
1087 ptr = ptr->in(1)->in(0)->in(0);
1088 continue;
1089 }
1090 #ifndef PRODUCT
1091 if (PrintOptimizeStringConcat) {
1092 tty->print_cr("fusion would fail for region");
1093 _begin->dump();
1094 ptr->dump(2);
1095 }
1096 #endif
1097 fail = true;
1098 break;
1099 } else {
1100 // other unknown control
1101 if (!fail) {
1102 #ifndef PRODUCT
1103 if (PrintOptimizeStringConcat) {
1104 tty->print_cr("fusion would fail for");
1105 _begin->dump();
1106 }
1107 #endif
1108 fail = true;
1109 }
1110 #ifndef PRODUCT
1111 if (PrintOptimizeStringConcat) {
1112 ptr->dump();
1113 }
1114 #endif
1115 ptr = ptr->in(0);
1116 }
1117 }
1118 #ifndef PRODUCT
1119 if (PrintOptimizeStringConcat && fail) {
1120 tty->cr();
1121 }
1122 #endif
1123 if (fail) return !fail;
1124
1125 // Validate that all these results produced are contained within
1126 // this cluster of objects. First collect all the results produced
1127 // by calls in the region.
1128 _stringopts->_visited.clear();
1129 Node_List worklist;
1130 Node* final_result = _end->proj_out_or_null(TypeFunc::Parms);
1131 for (uint i = 0; i < _control.size(); i++) {
1132 CallNode* cnode = _control.at(i)->isa_Call();
1133 if (cnode != nullptr) {
1134 _stringopts->_visited.test_set(cnode->_idx);
1135 }
1136 Node* result = cnode != nullptr ? cnode->proj_out_or_null(TypeFunc::Parms) : nullptr;
1137 if (result != nullptr && result != final_result) {
1138 worklist.push(result);
1139 }
1140 }
1141
1142 Node* last_result = nullptr;
1143 while (worklist.size() > 0) {
1144 Node* result = worklist.pop();
1145 if (_stringopts->_visited.test_set(result->_idx))
1146 continue;
1147 for (SimpleDUIterator i(result); i.has_next(); i.next()) {
1148 Node *use = i.get();
1149 if (ctrl_path.member(use)) {
1150 // already checked this
1151 continue;
1152 }
1153 int opc = use->Opcode();
1154 if (opc == Op_CmpP || opc == Op_Node) {
1155 ctrl_path.push(use);
1156 continue;
1157 }
1158 if (opc == Op_CastPP || opc == Op_CheckCastPP) {
1159 if (opc == Op_CheckCastPP) {
1160 worklist.push(use);
1161 }
1162 for (SimpleDUIterator j(use); j.has_next(); j.next()) {
1163 worklist.push(j.get());
1164 }
1165 worklist.push(use->in(1));
1166 ctrl_path.push(use);
1167 continue;
1168 }
1169 #ifndef PRODUCT
1170 if (PrintOptimizeStringConcat) {
1171 if (result != last_result) {
1172 last_result = result;
1173 tty->print_cr("extra uses for result:");
1174 last_result->dump();
1175 }
1176 use->dump();
1177 }
1178 #endif
1179 fail = true;
1180 break;
1181 }
1182 }
1183
1184 #ifndef PRODUCT
1185 if (PrintOptimizeStringConcat && !fail) {
1186 ttyLocker ttyl;
1187 tty->cr();
1188 tty->print("fusion has correct control flow (%d %d) for ", null_check_count, _uncommon_traps.size());
1189 _begin->jvms()->dump_spec(tty); tty->cr();
1190 for (int i = 0; i < num_arguments(); i++) {
1191 argument(i)->dump();
1192 }
1193 _control.dump();
1194 tty->cr();
1195 }
1196 #endif
1197
1198 return !fail;
1199 }
1200
1201 // Mirror of Integer.stringSize() method, return the count of digits in integer,
1202 Node* PhaseStringOpts::int_stringSize(GraphKit& kit, Node* arg) {
1203 if (arg->is_Con()) {
1204 // Constant integer. Compute constant length
1205 jint arg_val = arg->get_int();
1206 jint d = 1;
1207 if (arg_val >= 0) {
1208 d = 0;
1209 arg_val = -arg_val;
1210 }
1211 jint p = -10;
1212 for (int i = 1; i < 10; i++) {
1213 if (arg_val > p) {
1214 return __ intcon(i + d);
1215 }
1216 p = java_multiply(10, p);
1217 }
1218 return __ intcon(10 + d);
1219 }
1220
1221 // int d = 1;
1222 // if (x >= 0) {
1223 // d = 0;
1224 // x = -x;
1225 // }
1226 RegionNode* sign_merge = new RegionNode(3);
1227 kit.gvn().set_type(sign_merge, Type::CONTROL);
1228 Node* digit_cnt = new PhiNode(sign_merge, TypeInt::INT);
1229 kit.gvn().set_type(digit_cnt, TypeInt::INT);
1230 Node* val = new PhiNode(sign_merge, TypeInt::INT);
1231 kit.gvn().set_type(val, TypeInt::INT);
1232
1233 IfNode* iff = kit.create_and_map_if(kit.control(),
1234 __ Bool(__ CmpI(arg, __ intcon(0)), BoolTest::ge),
1235 PROB_FAIR, COUNT_UNKNOWN);
1236 sign_merge->init_req(1, __ IfTrue(iff));
1237 sign_merge->init_req(2, __ IfFalse(iff));
1238 digit_cnt->init_req(1, __ intcon(0));
1239 digit_cnt->init_req(2, __ intcon(1));
1240 val->init_req(1, __ SubI(__ intcon(0), arg));
1241 val->init_req(2, arg);
1242 kit.set_control(sign_merge);
1243
1244 // int p = -10;
1245 // for (int i = 1; i < 10; i++) {
1246 // if (x > p)
1247 // return i + d;
1248 // p = 10 * p;
1249 // }
1250 RegionNode* final_merge = new RegionNode(3);
1251 kit.gvn().set_type(final_merge, Type::CONTROL);
1252 Node* final_size = new PhiNode(final_merge, TypeInt::INT);
1253 kit.gvn().set_type(final_size, TypeInt::INT);
1254
1255 kit.add_parse_predicates();
1256 C->set_has_loops(true);
1257
1258 RegionNode* loop = new RegionNode(3);
1259 kit.gvn().set_type(loop, Type::CONTROL);
1260 Node* index = new PhiNode(loop, TypeInt::INT);
1261 kit.gvn().set_type(index, TypeInt::INT);
1262 Node* temp = new PhiNode(loop, TypeInt::INT);
1263 kit.gvn().set_type(temp, TypeInt::INT);
1264
1265 loop->init_req(1, kit.control());
1266 index->init_req(1, __ intcon(1));
1267 temp->init_req(1, __ intcon(-10));
1268 kit.set_control(loop);
1269
1270 Node* limit = __ CmpI(index, __ intcon(10));
1271 Node* limitb = __ Bool(limit, BoolTest::lt);
1272 IfNode* iff2 = kit.create_and_map_if(kit.control(), limitb, PROB_MIN, COUNT_UNKNOWN);
1273 Node* limit_less = __ IfTrue(iff2);
1274 kit.set_control(limit_less);
1275
1276 Node* cmp = __ CmpI(val, temp);
1277 Node* cmpb = __ Bool(cmp, BoolTest::gt);
1278 IfNode* iff3 = kit.create_and_map_if(kit.control(), cmpb, PROB_MIN, COUNT_UNKNOWN);
1279 Node* cmp_le = __ IfFalse(iff3);
1280 kit.set_control(cmp_le);
1281
1282 loop->init_req(2, kit.control());
1283 index->init_req(2, __ AddI(index, __ intcon(1)));
1284 temp->init_req(2, __ MulI(temp, __ intcon(10)));
1285
1286 final_merge->init_req(1, __ IfFalse(iff2));
1287 final_merge->init_req(2, __ IfTrue(iff3));
1288 final_size->init_req(1, __ AddI(digit_cnt, __ intcon(10)));
1289 final_size->init_req(2, __ AddI(digit_cnt, index));
1290 kit.set_control(final_merge);
1291
1292 C->record_for_igvn(sign_merge);
1293 C->record_for_igvn(digit_cnt);
1294 C->record_for_igvn(val);
1295 C->record_for_igvn(final_merge);
1296 C->record_for_igvn(final_size);
1297 C->record_for_igvn(loop);
1298 C->record_for_igvn(index);
1299 C->record_for_igvn(temp);
1300 return final_size;
1301 }
1302
1303 // Simplified version of Integer.getChars
1304 void PhaseStringOpts::getChars(GraphKit& kit, Node* arg, Node* dst_array, BasicType bt, Node* end, Node* final_merge, Node* final_mem, int merge_index) {
1305 // if (i < 0) {
1306 // sign = '-';
1307 // i = -i;
1308 // }
1309 IfNode* iff = kit.create_and_map_if(kit.control(), __ Bool(__ CmpI(arg, __ intcon(0)), BoolTest::lt),
1310 PROB_FAIR, COUNT_UNKNOWN);
1311
1312 RegionNode* merge = new RegionNode(3);
1313 kit.gvn().set_type(merge, Type::CONTROL);
1314 Node* i = new PhiNode(merge, TypeInt::INT);
1315 kit.gvn().set_type(i, TypeInt::INT);
1316 Node* sign = new PhiNode(merge, TypeInt::INT);
1317 kit.gvn().set_type(sign, TypeInt::INT);
1318
1319 merge->init_req(1, __ IfTrue(iff));
1320 i->init_req(1, __ SubI(__ intcon(0), arg));
1321 sign->init_req(1, __ intcon('-'));
1322 merge->init_req(2, __ IfFalse(iff));
1323 i->init_req(2, arg);
1324 sign->init_req(2, __ intcon(0));
1325
1326 kit.set_control(merge);
1327
1328 C->record_for_igvn(merge);
1329 C->record_for_igvn(i);
1330 C->record_for_igvn(sign);
1331
1332 // for (;;) {
1333 // q = i / 10;
1334 // r = i - ((q << 3) + (q << 1)); // r = i-(q*10) ...
1335 // buf [--charPos] = digits [r];
1336 // i = q;
1337 // if (i == 0) break;
1338 // }
1339
1340 // Add Parse Predicates first.
1341 kit.add_parse_predicates();
1342
1343 C->set_has_loops(true);
1344 RegionNode* head = new RegionNode(3);
1345 head->init_req(1, kit.control());
1346
1347 kit.gvn().set_type(head, Type::CONTROL);
1348 Node* i_phi = new PhiNode(head, TypeInt::INT);
1349 i_phi->init_req(1, i);
1350 kit.gvn().set_type(i_phi, TypeInt::INT);
1351 Node* charPos = new PhiNode(head, TypeInt::INT);
1352 charPos->init_req(1, end);
1353 kit.gvn().set_type(charPos, TypeInt::INT);
1354 Node* mem = PhiNode::make(head, kit.memory(byte_adr_idx), Type::MEMORY, TypeAryPtr::BYTES);
1355 kit.gvn().set_type(mem, Type::MEMORY);
1356
1357 kit.set_control(head);
1358 kit.set_memory(mem, byte_adr_idx);
1359
1360 Node* q = __ DivI(kit.null(), i_phi, __ intcon(10));
1361 Node* r = __ SubI(i_phi, __ AddI(__ LShiftI(q, __ intcon(3)),
1362 __ LShiftI(q, __ intcon(1))));
1363 Node* index = __ SubI(charPos, __ intcon((bt == T_BYTE) ? 1 : 2));
1364 Node* ch = __ AddI(r, __ intcon('0'));
1365 Node* st = __ store_to_memory(kit.control(), kit.array_element_address(dst_array, index, T_BYTE),
1366 ch, bt, MemNode::unordered, false /* require_atomic_access */,
1367 false /* unaligned */, (bt != T_BYTE) /* mismatched */);
1368
1369 iff = kit.create_and_map_if(head, __ Bool(__ CmpI(q, __ intcon(0)), BoolTest::ne),
1370 PROB_FAIR, COUNT_UNKNOWN);
1371 Node* ne = __ IfTrue(iff);
1372 Node* eq = __ IfFalse(iff);
1373
1374 head->init_req(2, ne);
1375 mem->init_req(2, st);
1376
1377 i_phi->init_req(2, q);
1378 charPos->init_req(2, index);
1379 charPos = index;
1380
1381 kit.set_control(eq);
1382 kit.set_memory(st, byte_adr_idx);
1383
1384 C->record_for_igvn(head);
1385 C->record_for_igvn(mem);
1386 C->record_for_igvn(i_phi);
1387 C->record_for_igvn(charPos);
1388
1389 // if (sign != 0) {
1390 // buf [--charPos] = sign;
1391 // }
1392 iff = kit.create_and_map_if(kit.control(), __ Bool(__ CmpI(sign, __ intcon(0)), BoolTest::ne),
1393 PROB_FAIR, COUNT_UNKNOWN);
1394
1395 final_merge->init_req(merge_index + 2, __ IfFalse(iff));
1396 final_mem->init_req(merge_index + 2, kit.memory(byte_adr_idx));
1397
1398 kit.set_control(__ IfTrue(iff));
1399 if (kit.stopped()) {
1400 final_merge->init_req(merge_index + 1, C->top());
1401 final_mem->init_req(merge_index + 1, C->top());
1402 } else {
1403 Node* index = __ SubI(charPos, __ intcon((bt == T_BYTE) ? 1 : 2));
1404 st = __ store_to_memory(kit.control(), kit.array_element_address(dst_array, index, T_BYTE),
1405 sign, bt, MemNode::unordered, false /* require_atomic_access */, false /* unaligned */,
1406 (bt != T_BYTE) /* mismatched */);
1407
1408 final_merge->init_req(merge_index + 1, kit.control());
1409 final_mem->init_req(merge_index + 1, st);
1410 }
1411 }
1412
1413 // Copy the characters representing arg into dst_array starting at start
1414 Node* PhaseStringOpts::int_getChars(GraphKit& kit, Node* arg, Node* dst_array, Node* dst_coder, Node* start, Node* size) {
1415 bool dcon = dst_coder->is_Con();
1416 bool dbyte = dcon ? (dst_coder->get_int() == java_lang_String::CODER_LATIN1) : false;
1417 Node* end = __ AddI(start, __ LShiftI(size, dst_coder));
1418
1419 // The final_merge node has 4 entries in case the encoding is known:
1420 // (0) Control, (1) result w/ sign, (2) result w/o sign, (3) result for Integer.min_value
1421 // or 6 entries in case the encoding is not known:
1422 // (0) Control, (1) Latin1 w/ sign, (2) Latin1 w/o sign, (3) min_value, (4) UTF16 w/ sign, (5) UTF16 w/o sign
1423 RegionNode* final_merge = new RegionNode(dcon ? 4 : 6);
1424 kit.gvn().set_type(final_merge, Type::CONTROL);
1425
1426 Node* final_mem = PhiNode::make(final_merge, kit.memory(byte_adr_idx), Type::MEMORY, TypeAryPtr::BYTES);
1427 kit.gvn().set_type(final_mem, Type::MEMORY);
1428
1429 // need to handle arg == Integer.MIN_VALUE specially because negating doesn't make it positive
1430 IfNode* iff = kit.create_and_map_if(kit.control(), __ Bool(__ CmpI(arg, __ intcon(0x80000000)), BoolTest::ne),
1431 PROB_FAIR, COUNT_UNKNOWN);
1432
1433 Node* old_mem = kit.memory(byte_adr_idx);
1434
1435 kit.set_control(__ IfFalse(iff));
1436 if (kit.stopped()) {
1437 // Statically not equal to MIN_VALUE so this path is dead
1438 final_merge->init_req(3, kit.control());
1439 } else {
1440 copy_string(kit, __ makecon(TypeInstPtr::make(C->env()->the_min_jint_string())),
1441 dst_array, dst_coder, start);
1442 final_merge->init_req(3, kit.control());
1443 final_mem->init_req(3, kit.memory(byte_adr_idx));
1444 }
1445
1446 kit.set_control(__ IfTrue(iff));
1447 kit.set_memory(old_mem, byte_adr_idx);
1448
1449 if (!dcon) {
1450 // Check encoding of destination
1451 iff = kit.create_and_map_if(kit.control(), __ Bool(__ CmpI(dst_coder, __ intcon(0)), BoolTest::eq),
1452 PROB_FAIR, COUNT_UNKNOWN);
1453 old_mem = kit.memory(byte_adr_idx);
1454 }
1455 if (!dcon || dbyte) {
1456 // Destination is Latin1,
1457 if (!dcon) {
1458 kit.set_control(__ IfTrue(iff));
1459 }
1460 getChars(kit, arg, dst_array, T_BYTE, end, final_merge, final_mem);
1461 }
1462 if (!dcon || !dbyte) {
1463 // Destination is UTF16
1464 int merge_index = 0;
1465 if (!dcon) {
1466 kit.set_control(__ IfFalse(iff));
1467 kit.set_memory(old_mem, byte_adr_idx);
1468 merge_index = 3; // Account for Latin1 case
1469 }
1470 getChars(kit, arg, dst_array, T_CHAR, end, final_merge, final_mem, merge_index);
1471 }
1472
1473 // Final merge point for Latin1 and UTF16 case
1474 kit.set_control(final_merge);
1475 kit.set_memory(final_mem, byte_adr_idx);
1476
1477 C->record_for_igvn(final_merge);
1478 C->record_for_igvn(final_mem);
1479 return end;
1480 }
1481
1482 // Copy 'count' bytes/chars from src_array to dst_array starting at index start
1483 void PhaseStringOpts::arraycopy(GraphKit& kit, IdealKit& ideal, Node* src_array, Node* dst_array, BasicType elembt, Node* start, Node* count) {
1484 assert(elembt == T_BYTE || elembt == T_CHAR, "Invalid type for arraycopy");
1485
1486 if (elembt == T_CHAR) {
1487 // Get number of chars
1488 count = __ RShiftI(count, __ intcon(1));
1489 }
1490
1491 Node* extra = nullptr;
1492 #ifdef _LP64
1493 count = __ ConvI2L(count);
1494 extra = C->top();
1495 #endif
1496
1497 Node* src_ptr = __ array_element_address(src_array, __ intcon(0), T_BYTE);
1498 Node* dst_ptr = __ array_element_address(dst_array, start, T_BYTE);
1499 // Check if src array address is aligned to HeapWordSize
1500 bool aligned = (arrayOopDesc::base_offset_in_bytes(T_BYTE) % HeapWordSize == 0);
1501 // If true, then check if dst array address is aligned to HeapWordSize
1502 if (aligned) {
1503 const TypeInt* tdst = __ gvn().type(start)->is_int();
1504 aligned = tdst->is_con() && ((arrayOopDesc::base_offset_in_bytes(T_BYTE) +
1505 tdst->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0);
1506 }
1507 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1508 const char* copyfunc_name = "arraycopy";
1509 address copyfunc_addr = StubRoutines::select_arraycopy_function(elembt, aligned, true, copyfunc_name, true);
1510 ideal.make_leaf_call_no_fp(OptoRuntime::fast_arraycopy_Type(), copyfunc_addr, copyfunc_name,
1511 TypeAryPtr::BYTES, src_ptr, dst_ptr, count, extra);
1512 }
1513
1514 #undef __
1515 #define __ ideal.
1516
1517 // Copy contents of a Latin1 encoded string from src_array to dst_array
1518 void PhaseStringOpts::copy_latin1_string(GraphKit& kit, IdealKit& ideal, Node* src_array, IdealVariable& count,
1519 Node* dst_array, Node* dst_coder, Node* start) {
1520 bool dcon = dst_coder->is_Con();
1521 bool dbyte = dcon ? (dst_coder->get_int() == java_lang_String::CODER_LATIN1) : false;
1522
1523 if (!dcon) {
1524 __ if_then(dst_coder, BoolTest::eq, __ ConI(java_lang_String::CODER_LATIN1));
1525 }
1526 if (!dcon || dbyte) {
1527 // Destination is Latin1. Simply emit a byte arraycopy.
1528 arraycopy(kit, ideal, src_array, dst_array, T_BYTE, start, __ value(count));
1529 }
1530 if (!dcon) {
1531 __ else_();
1532 }
1533 if (!dcon || !dbyte) {
1534 // Destination is UTF16. Inflate src_array into dst_array.
1535 kit.sync_kit(ideal);
1536 if (Matcher::match_rule_supported(Op_StrInflatedCopy)) {
1537 // Use fast intrinsic
1538 Node* src = kit.array_element_address(src_array, kit.intcon(0), T_BYTE);
1539 Node* dst = kit.array_element_address(dst_array, start, T_BYTE);
1540 kit.inflate_string(src, dst, TypeAryPtr::BYTES, __ value(count));
1541 } else {
1542 // No intrinsic available, use slow method
1543 kit.inflate_string_slow(src_array, dst_array, start, __ value(count));
1544 }
1545 ideal.sync_kit(&kit);
1546 // Multiply count by two since we now need two bytes per char
1547 __ set(count, __ LShiftI(__ value(count), __ ConI(1)));
1548 }
1549 if (!dcon) {
1550 __ end_if();
1551 }
1552 }
1553
1554 // Read two bytes from index and index+1 and convert them to a char
1555 static jchar readChar(ciTypeArray* array, int index) {
1556 int shift_high, shift_low;
1557 #ifdef VM_LITTLE_ENDIAN
1558 shift_high = 0;
1559 shift_low = 8;
1560 #else
1561 shift_high = 8;
1562 shift_low = 0;
1563 #endif
1564
1565 jchar b1 = ((jchar) array->byte_at(index)) & 0xff;
1566 jchar b2 = ((jchar) array->byte_at(index+1)) & 0xff;
1567 return (b1 << shift_high) | (b2 << shift_low);
1568 }
1569
1570 // Copy contents of constant src_array to dst_array by emitting individual stores
1571 void PhaseStringOpts::copy_constant_string(GraphKit& kit, IdealKit& ideal, ciTypeArray* src_array, IdealVariable& count,
1572 bool src_is_byte, Node* dst_array, Node* dst_coder, Node* start) {
1573 bool dcon = dst_coder->is_Con();
1574 bool dbyte = dcon ? (dst_coder->get_int() == java_lang_String::CODER_LATIN1) : false;
1575 int length = src_array->length();
1576
1577 if (!dcon) {
1578 __ if_then(dst_coder, BoolTest::eq, __ ConI(java_lang_String::CODER_LATIN1));
1579 }
1580 if (!dcon || dbyte) {
1581 // Destination is Latin1. Copy each byte of src_array into dst_array.
1582 Node* index = start;
1583 for (int i = 0; i < length; i++) {
1584 Node* adr = kit.array_element_address(dst_array, index, T_BYTE);
1585 Node* val = __ ConI(src_array->byte_at(i));
1586 __ store(__ ctrl(), adr, val, T_BYTE, byte_adr_idx, MemNode::unordered);
1587 index = __ AddI(index, __ ConI(1));
1588 }
1589 }
1590 if (!dcon) {
1591 __ else_();
1592 }
1593 if (!dcon || !dbyte) {
1594 // Destination is UTF16. Copy each char of src_array into dst_array.
1595 Node* index = start;
1596 for (int i = 0; i < length; i++) {
1597 Node* adr = kit.array_element_address(dst_array, index, T_BYTE);
1598 jchar val;
1599 if (src_is_byte) {
1600 val = src_array->byte_at(i) & 0xff;
1601 } else {
1602 val = readChar(src_array, i++);
1603 }
1604 __ store(__ ctrl(), adr, __ ConI(val), T_CHAR, byte_adr_idx, MemNode::unordered, false /* require_atomic_access */,
1605 true /* mismatched */);
1606 index = __ AddI(index, __ ConI(2));
1607 }
1608 if (src_is_byte) {
1609 // Multiply count by two since we now need two bytes per char
1610 __ set(count, __ ConI(2 * length));
1611 }
1612 }
1613 if (!dcon) {
1614 __ end_if();
1615 }
1616 }
1617
1618 // Compress copy contents of the byte/char String str into dst_array starting at index start.
1619 Node* PhaseStringOpts::copy_string(GraphKit& kit, Node* str, Node* dst_array, Node* dst_coder, Node* start) {
1620 Node* src_array = kit.load_String_value(str, true);
1621
1622 IdealKit ideal(&kit, true, true);
1623 IdealVariable count(ideal); __ declarations_done();
1624
1625 if (str->is_Con()) {
1626 // Constant source string
1627 ciTypeArray* src_array_type = get_constant_value(kit, str);
1628
1629 // Check encoding of constant string
1630 bool src_is_byte = (get_constant_coder(kit, str) == java_lang_String::CODER_LATIN1);
1631
1632 // For small constant strings just emit individual stores.
1633 // A length of 6 seems like a good space/speed tradeof.
1634 __ set(count, __ ConI(src_array_type->length()));
1635 int src_len = src_array_type->length() / (src_is_byte ? 1 : 2);
1636 if (src_len < unroll_string_copy_length) {
1637 // Small constant string
1638 copy_constant_string(kit, ideal, src_array_type, count, src_is_byte, dst_array, dst_coder, start);
1639 } else if (src_is_byte) {
1640 // Source is Latin1
1641 copy_latin1_string(kit, ideal, src_array, count, dst_array, dst_coder, start);
1642 } else {
1643 // Source is UTF16 (destination too). Simply emit a char arraycopy.
1644 arraycopy(kit, ideal, src_array, dst_array, T_CHAR, start, __ value(count));
1645 }
1646 } else {
1647 Node* size = kit.load_array_length(src_array);
1648 __ set(count, size);
1649 // Non-constant source string
1650 if (CompactStrings) {
1651 // Emit runtime check for coder
1652 Node* coder = kit.load_String_coder(str, true);
1653 __ if_then(coder, BoolTest::eq, __ ConI(java_lang_String::CODER_LATIN1)); {
1654 // Source is Latin1
1655 copy_latin1_string(kit, ideal, src_array, count, dst_array, dst_coder, start);
1656 } __ else_();
1657 }
1658 // Source is UTF16 (destination too). Simply emit a char arraycopy.
1659 arraycopy(kit, ideal, src_array, dst_array, T_CHAR, start, __ value(count));
1660
1661 if (CompactStrings) {
1662 __ end_if();
1663 }
1664 }
1665
1666 // Finally sync IdealKit and GraphKit.
1667 kit.sync_kit(ideal);
1668 return __ AddI(start, __ value(count));
1669 }
1670
1671 // Compress copy the char into dst_array at index start.
1672 Node* PhaseStringOpts::copy_char(GraphKit& kit, Node* val, Node* dst_array, Node* dst_coder, Node* start) {
1673 bool dcon = (dst_coder != nullptr) && dst_coder->is_Con();
1674 bool dbyte = dcon ? (dst_coder->get_int() == java_lang_String::CODER_LATIN1) : false;
1675
1676 IdealKit ideal(&kit, true, true);
1677 IdealVariable end(ideal); __ declarations_done();
1678 Node* adr = kit.array_element_address(dst_array, start, T_BYTE);
1679 if (!dcon){
1680 __ if_then(dst_coder, BoolTest::eq, __ ConI(java_lang_String::CODER_LATIN1));
1681 }
1682 if (!dcon || dbyte) {
1683 // Destination is Latin1. Store a byte.
1684 __ store(__ ctrl(), adr, val, T_BYTE, byte_adr_idx, MemNode::unordered);
1685 __ set(end, __ AddI(start, __ ConI(1)));
1686 }
1687 if (!dcon) {
1688 __ else_();
1689 }
1690 if (!dcon || !dbyte) {
1691 // Destination is UTF16. Store a char.
1692 __ store(__ ctrl(), adr, val, T_CHAR, byte_adr_idx, MemNode::unordered, false /* require_atomic_access */,
1693 true /* mismatched */);
1694 __ set(end, __ AddI(start, __ ConI(2)));
1695 }
1696 if (!dcon) {
1697 __ end_if();
1698 }
1699 // Finally sync IdealKit and GraphKit.
1700 kit.sync_kit(ideal);
1701 return __ value(end);
1702 }
1703
1704 #undef __
1705 #define __ kit.
1706
1707 // Allocate a byte array of specified length.
1708 Node* PhaseStringOpts::allocate_byte_array(GraphKit& kit, IdealKit* ideal, Node* length) {
1709 if (ideal != nullptr) {
1710 // Sync IdealKit and graphKit.
1711 kit.sync_kit(*ideal);
1712 }
1713 Node* byte_array = nullptr;
1714 {
1715 PreserveReexecuteState preexecs(&kit);
1716 // The original jvms is for an allocation of either a String or
1717 // StringBuffer so no stack adjustment is necessary for proper
1718 // reexecution. If we deoptimize in the slow path the bytecode
1719 // will be reexecuted and the char[] allocation will be thrown away.
1720 kit.jvms()->set_should_reexecute(true);
1721 byte_array = kit.new_array(__ makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))),
1722 length, 1);
1723 }
1724
1725 // Mark the allocation so that zeroing is skipped since the code
1726 // below will overwrite the entire array
1727 AllocateArrayNode* byte_alloc = AllocateArrayNode::Ideal_array_allocation(byte_array);
1728 byte_alloc->maybe_set_complete(_gvn);
1729
1730 if (ideal != nullptr) {
1731 // Sync IdealKit and graphKit.
1732 ideal->sync_kit(&kit);
1733 }
1734 return byte_array;
1735 }
1736
1737 jbyte PhaseStringOpts::get_constant_coder(GraphKit& kit, Node* str) {
1738 assert(str->is_Con(), "String must be constant");
1739 const TypeOopPtr* str_type = kit.gvn().type(str)->isa_oopptr();
1740 ciInstance* str_instance = str_type->const_oop()->as_instance();
1741 jbyte coder = str_instance->field_value_by_offset(java_lang_String::coder_offset()).as_byte();
1742 assert(CompactStrings || (coder == java_lang_String::CODER_UTF16), "Strings must be UTF16 encoded");
1743 return coder;
1744 }
1745
1746 int PhaseStringOpts::get_constant_length(GraphKit& kit, Node* str) {
1747 assert(str->is_Con(), "String must be constant");
1748 return get_constant_value(kit, str)->length();
1749 }
1750
1751 ciTypeArray* PhaseStringOpts::get_constant_value(GraphKit& kit, Node* str) {
1752 assert(str->is_Con(), "String must be constant");
1753 const TypeOopPtr* str_type = kit.gvn().type(str)->isa_oopptr();
1754 ciInstance* str_instance = str_type->const_oop()->as_instance();
1755 ciObject* src_array = str_instance->field_value_by_offset(java_lang_String::value_offset()).as_object();
1756 return src_array->as_type_array();
1757 }
1758
1759 void PhaseStringOpts::replace_string_concat(StringConcat* sc) {
1760 // Log a little info about the transformation
1761 sc->maybe_log_transform();
1762
1763 // pull the JVMState of the allocation into a SafePointNode to serve as
1764 // as a shim for the insertion of the new code.
1765 JVMState* jvms = sc->begin()->jvms()->clone_shallow(C);
1766 uint size = sc->begin()->req();
1767 SafePointNode* map = new SafePointNode(size, jvms);
1768
1769 // copy the control and memory state from the final call into our
1770 // new starting state. This allows any preceding tests to feed
1771 // into the new section of code.
1772 for (uint i1 = 0; i1 < TypeFunc::Parms; i1++) {
1773 map->init_req(i1, sc->end()->in(i1));
1774 }
1775 // blow away old allocation arguments
1776 for (uint i1 = TypeFunc::Parms; i1 < jvms->debug_start(); i1++) {
1777 map->init_req(i1, C->top());
1778 }
1779 // Copy the rest of the inputs for the JVMState
1780 for (uint i1 = jvms->debug_start(); i1 < sc->begin()->req(); i1++) {
1781 map->init_req(i1, sc->begin()->in(i1));
1782 }
1783 // Make sure the memory state is a MergeMem for parsing.
1784 if (!map->in(TypeFunc::Memory)->is_MergeMem()) {
1785 map->set_req(TypeFunc::Memory, MergeMemNode::make(map->in(TypeFunc::Memory)));
1786 }
1787
1788 jvms->set_map(map);
1789 map->ensure_stack(jvms, jvms->method()->max_stack());
1790
1791 // disconnect all the old StringBuilder calls from the graph
1792 sc->eliminate_unneeded_control();
1793
1794 // At this point all the old work has been completely removed from
1795 // the graph and the saved JVMState exists at the point where the
1796 // final toString call used to be.
1797 GraphKit kit(jvms);
1798
1799 // There may be uncommon traps which are still using the
1800 // intermediate states and these need to be rewritten to point at
1801 // the JVMState at the beginning of the transformation.
1802 sc->convert_uncommon_traps(kit, jvms);
1803
1804 // Now insert the logic to compute the size of the string followed
1805 // by all the logic to construct array and resulting string.
1806
1807 Node* null_string = __ makecon(TypeInstPtr::make(C->env()->the_null_string()));
1808
1809 // Create a region for the overflow checks to merge into.
1810 int args = MAX2(sc->num_arguments(), 1);
1811 RegionNode* overflow = new RegionNode(args);
1812 kit.gvn().set_type(overflow, Type::CONTROL);
1813
1814 // Create a hook node to hold onto the individual sizes since they
1815 // are need for the copying phase.
1816 Node* string_sizes = new Node(args);
1817
1818 Node* coder = __ intcon(0);
1819 Node* length = __ intcon(0);
1820 // If at least one argument is UTF16 encoded, we can fix the encoding.
1821 bool coder_fixed = false;
1822
1823 if (!CompactStrings) {
1824 // Fix encoding of result string to UTF16
1825 coder_fixed = true;
1826 coder = __ intcon(java_lang_String::CODER_UTF16);
1827 }
1828
1829 for (int argi = 0; argi < sc->num_arguments(); argi++) {
1830 Node* arg = sc->argument(argi);
1831 switch (sc->mode(argi)) {
1832 case StringConcat::NegativeIntCheckMode: {
1833 // Initial capacity argument might be negative in which case StringBuilder(int) throws
1834 // a NegativeArraySizeException. Insert a runtime check with an uncommon trap.
1835 const TypeInt* type = kit.gvn().type(arg)->is_int();
1836 assert(type->_hi >= 0 && type->_lo < 0, "no runtime int check needed");
1837 Node* p = __ Bool(__ CmpI(arg, kit.intcon(0)), BoolTest::ge);
1838 IfNode* iff = kit.create_and_map_if(kit.control(), p, PROB_MIN, COUNT_UNKNOWN);
1839 {
1840 // Negative int -> uncommon trap.
1841 PreserveJVMState pjvms(&kit);
1842 kit.set_control(__ IfFalse(iff));
1843 kit.uncommon_trap(Deoptimization::Reason_intrinsic,
1844 Deoptimization::Action_maybe_recompile);
1845 }
1846 kit.set_control(__ IfTrue(iff));
1847 break;
1848 }
1849 case StringConcat::IntMode: {
1850 Node* string_size = int_stringSize(kit, arg);
1851
1852 // accumulate total
1853 length = __ AddI(length, string_size);
1854
1855 // Cache this value for the use by int_toString
1856 string_sizes->init_req(argi, string_size);
1857 break;
1858 }
1859 case StringConcat::StringNullCheckMode: {
1860 const Type* type = kit.gvn().type(arg);
1861 assert(type != TypePtr::NULL_PTR, "missing check");
1862 if (!type->higher_equal(TypeInstPtr::NOTNULL)) {
1863 // Null check with uncommon trap since
1864 // StringBuilder(null) throws exception.
1865 // Use special uncommon trap instead of
1866 // calling normal do_null_check().
1867 Node* p = __ Bool(__ CmpP(arg, kit.null()), BoolTest::ne);
1868 IfNode* iff = kit.create_and_map_if(kit.control(), p, PROB_MIN, COUNT_UNKNOWN);
1869 overflow->add_req(__ IfFalse(iff));
1870 Node* notnull = __ IfTrue(iff);
1871 kit.set_control(notnull); // set control for the cast_not_null
1872 arg = kit.cast_not_null(arg, false);
1873 sc->set_argument(argi, arg);
1874 }
1875 assert(kit.gvn().type(arg)->higher_equal(TypeInstPtr::NOTNULL), "sanity");
1876 // Fallthrough to add string length.
1877 }
1878 case StringConcat::StringMode: {
1879 const Type* type = kit.gvn().type(arg);
1880 Node* count = nullptr;
1881 Node* arg_coder = nullptr;
1882 if (type == TypePtr::NULL_PTR) {
1883 // replace the argument with the null checked version
1884 arg = null_string;
1885 sc->set_argument(argi, arg);
1886 count = kit.load_String_length(arg, true);
1887 arg_coder = kit.load_String_coder(arg, true);
1888 } else if (!type->higher_equal(TypeInstPtr::NOTNULL)) {
1889 // s = s != null ? s : "null";
1890 // length = length + (s.count - s.offset);
1891 RegionNode *r = new RegionNode(3);
1892 kit.gvn().set_type(r, Type::CONTROL);
1893 Node *phi = new PhiNode(r, type);
1894 kit.gvn().set_type(phi, phi->bottom_type());
1895 Node* p = __ Bool(__ CmpP(arg, kit.null()), BoolTest::ne);
1896 IfNode* iff = kit.create_and_map_if(kit.control(), p, PROB_MIN, COUNT_UNKNOWN);
1897 Node* notnull = __ IfTrue(iff);
1898 Node* isnull = __ IfFalse(iff);
1899 kit.set_control(notnull); // set control for the cast_not_null
1900 r->init_req(1, notnull);
1901 phi->init_req(1, kit.cast_not_null(arg, false));
1902 r->init_req(2, isnull);
1903 phi->init_req(2, null_string);
1904 kit.set_control(r);
1905 C->record_for_igvn(r);
1906 C->record_for_igvn(phi);
1907 // replace the argument with the null checked version
1908 arg = phi;
1909 sc->set_argument(argi, arg);
1910 count = kit.load_String_length(arg, true);
1911 arg_coder = kit.load_String_coder(arg, true);
1912 } else {
1913 // A corresponding nullcheck will be connected during IGVN MemNode::Ideal_common_DU_postCCP
1914 // kit.control might be a different test, that can be hoisted above the actual nullcheck
1915 // in case, that the control input is not null, Ideal_common_DU_postCCP will not look for a nullcheck.
1916 count = kit.load_String_length(arg, false);
1917 arg_coder = kit.load_String_coder(arg, false);
1918 }
1919 if (arg->is_Con()) {
1920 // Constant string. Get constant coder and length.
1921 jbyte const_coder = get_constant_coder(kit, arg);
1922 int const_length = get_constant_length(kit, arg);
1923 if (const_coder == java_lang_String::CODER_LATIN1) {
1924 // Can be latin1 encoded
1925 arg_coder = __ intcon(const_coder);
1926 count = __ intcon(const_length);
1927 } else {
1928 // Found UTF16 encoded string. Fix result array encoding to UTF16.
1929 coder_fixed = true;
1930 coder = __ intcon(const_coder);
1931 count = __ intcon(const_length / 2);
1932 }
1933 }
1934
1935 if (!coder_fixed) {
1936 coder = __ OrI(coder, arg_coder);
1937 }
1938 length = __ AddI(length, count);
1939 string_sizes->init_req(argi, nullptr);
1940 break;
1941 }
1942 case StringConcat::CharMode: {
1943 // one character only
1944 const TypeInt* t = kit.gvn().type(arg)->is_int();
1945 if (!coder_fixed && t->is_con()) {
1946 // Constant char
1947 if (t->get_con() <= 255) {
1948 // Can be latin1 encoded
1949 coder = __ OrI(coder, __ intcon(java_lang_String::CODER_LATIN1));
1950 } else {
1951 // Must be UTF16 encoded. Fix result array encoding to UTF16.
1952 coder_fixed = true;
1953 coder = __ intcon(java_lang_String::CODER_UTF16);
1954 }
1955 } else if (!coder_fixed) {
1956 // Not constant
1957 #undef __
1958 #define __ ideal.
1959 IdealKit ideal(&kit, true, true);
1960 IdealVariable char_coder(ideal); __ declarations_done();
1961 // Check if character can be latin1 encoded
1962 __ if_then(arg, BoolTest::le, __ ConI(0xFF));
1963 __ set(char_coder, __ ConI(java_lang_String::CODER_LATIN1));
1964 __ else_();
1965 __ set(char_coder, __ ConI(java_lang_String::CODER_UTF16));
1966 __ end_if();
1967 kit.sync_kit(ideal);
1968 coder = __ OrI(coder, __ value(char_coder));
1969 #undef __
1970 #define __ kit.
1971 }
1972 length = __ AddI(length, __ intcon(1));
1973 break;
1974 }
1975 default:
1976 ShouldNotReachHere();
1977 }
1978 if (argi > 0) {
1979 // Check that the sum hasn't overflowed
1980 IfNode* iff = kit.create_and_map_if(kit.control(),
1981 __ Bool(__ CmpI(length, __ intcon(0)), BoolTest::lt),
1982 PROB_MIN, COUNT_UNKNOWN);
1983 kit.set_control(__ IfFalse(iff));
1984 overflow->set_req(argi, __ IfTrue(iff));
1985 }
1986 }
1987
1988 {
1989 // Hook
1990 PreserveJVMState pjvms(&kit);
1991 kit.set_control(overflow);
1992 C->record_for_igvn(overflow);
1993 kit.uncommon_trap(Deoptimization::Reason_intrinsic,
1994 Deoptimization::Action_make_not_entrant);
1995 }
1996
1997 Node* result;
1998 if (!kit.stopped()) {
1999 assert(CompactStrings || (coder->is_Con() && coder->get_int() == java_lang_String::CODER_UTF16),
2000 "Result string must be UTF16 encoded if CompactStrings is disabled");
2001
2002 Node* dst_array = nullptr;
2003 if (sc->num_arguments() == 1 &&
2004 (sc->mode(0) == StringConcat::StringMode ||
2005 sc->mode(0) == StringConcat::StringNullCheckMode)) {
2006 // Handle the case when there is only a single String argument.
2007 // In this case, we can just pull the value from the String itself.
2008 dst_array = kit.load_String_value(sc->argument(0), true);
2009 } else {
2010 // Allocate destination byte array according to coder
2011 dst_array = allocate_byte_array(kit, nullptr, __ LShiftI(length, coder));
2012
2013 // Now copy the string representations into the final byte[]
2014 Node* start = __ intcon(0);
2015 for (int argi = 0; argi < sc->num_arguments(); argi++) {
2016 Node* arg = sc->argument(argi);
2017 switch (sc->mode(argi)) {
2018 case StringConcat::NegativeIntCheckMode:
2019 break; // Nothing to do, was only needed to add a runtime check earlier.
2020 case StringConcat::IntMode: {
2021 start = int_getChars(kit, arg, dst_array, coder, start, string_sizes->in(argi));
2022 break;
2023 }
2024 case StringConcat::StringNullCheckMode:
2025 case StringConcat::StringMode: {
2026 start = copy_string(kit, arg, dst_array, coder, start);
2027 break;
2028 }
2029 case StringConcat::CharMode: {
2030 start = copy_char(kit, arg, dst_array, coder, start);
2031 break;
2032 }
2033 default:
2034 ShouldNotReachHere();
2035 }
2036 }
2037 }
2038
2039 {
2040 PreserveReexecuteState preexecs(&kit);
2041 // The original jvms is for an allocation of either a String or
2042 // StringBuffer so no stack adjustment is necessary for proper
2043 // reexecution.
2044 kit.jvms()->set_should_reexecute(true);
2045 result = kit.new_instance(__ makecon(TypeKlassPtr::make(C->env()->String_klass())));
2046 }
2047
2048 // Initialize the string
2049 kit.store_String_value(result, dst_array);
2050 kit.store_String_coder(result, coder);
2051
2052 // The value field is final. Emit a barrier here to ensure that the effect
2053 // of the initialization is committed to memory before any code publishes
2054 // a reference to the newly constructed object (see Parse::do_exits()).
2055 assert(AllocateNode::Ideal_allocation(result) != nullptr, "should be newly allocated");
2056 kit.insert_mem_bar(UseStoreStoreForCtor ? Op_MemBarStoreStore : Op_MemBarRelease, result);
2057 } else {
2058 result = C->top();
2059 }
2060 // hook up the outgoing control and result
2061 kit.replace_call(sc->end(), result);
2062
2063 // Unhook any hook nodes
2064 string_sizes->disconnect_inputs(C);
2065 sc->cleanup();
2066 #ifndef PRODUCT
2067 AtomicAccess::inc(&_stropts_replaced);
2068 #endif
2069 }
2070
2071 #ifndef PRODUCT
2072 uint PhaseStringOpts::_stropts_replaced = 0;
2073 uint PhaseStringOpts::_stropts_merged = 0;
2074 uint PhaseStringOpts::_stropts_total = 0;
2075
2076 void PhaseStringOpts::print_statistics() {
2077 tty->print_cr("StringConcat: %4d/%4d/%4d(replaced/merged/total)", _stropts_replaced, _stropts_merged, _stropts_total);
2078 }
2079 #endif