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