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