1 | // Copyright (c) 2014, the Dart project authors. Please see the AUTHORS file |
2 | // for details. All rights reserved. Use of this source code is governed by a |
3 | // BSD-style license that can be found in the LICENSE file. |
4 | |
5 | #include "vm/globals.h" // Needed here to get TARGET_ARCH_ARM64. |
6 | #if defined(TARGET_ARCH_ARM64) |
7 | |
8 | #include "vm/compiler/backend/flow_graph_compiler.h" |
9 | |
10 | #include "vm/compiler/api/type_check_mode.h" |
11 | #include "vm/compiler/backend/il_printer.h" |
12 | #include "vm/compiler/backend/locations.h" |
13 | #include "vm/compiler/jit/compiler.h" |
14 | #include "vm/cpu.h" |
15 | #include "vm/dart_entry.h" |
16 | #include "vm/deopt_instructions.h" |
17 | #include "vm/dispatch_table.h" |
18 | #include "vm/instructions.h" |
19 | #include "vm/object_store.h" |
20 | #include "vm/parser.h" |
21 | #include "vm/stack_frame.h" |
22 | #include "vm/stub_code.h" |
23 | #include "vm/symbols.h" |
24 | |
25 | namespace dart { |
26 | |
27 | DEFINE_FLAG(bool, trap_on_deoptimization, false, "Trap on deoptimization." ); |
28 | DECLARE_FLAG(bool, enable_simd_inline); |
29 | DEFINE_FLAG(bool, unbox_mints, true, "Optimize 64-bit integer arithmetic." ); |
30 | |
31 | void FlowGraphCompiler::ArchSpecificInitialization() { |
32 | if (FLAG_precompiled_mode && FLAG_use_bare_instructions) { |
33 | auto object_store = isolate()->object_store(); |
34 | |
35 | const auto& stub = |
36 | Code::ZoneHandle(object_store->write_barrier_wrappers_stub()); |
37 | if (CanPcRelativeCall(stub)) { |
38 | assembler_->generate_invoke_write_barrier_wrapper_ = [&](Register reg) { |
39 | const intptr_t offset_into_target = |
40 | Thread::WriteBarrierWrappersOffsetForRegister(reg); |
41 | assembler_->GenerateUnRelocatedPcRelativeCall(offset_into_target); |
42 | AddPcRelativeCallStubTarget(stub); |
43 | }; |
44 | } |
45 | |
46 | const auto& array_stub = |
47 | Code::ZoneHandle(object_store->array_write_barrier_stub()); |
48 | if (CanPcRelativeCall(stub)) { |
49 | assembler_->generate_invoke_array_write_barrier_ = [&]() { |
50 | assembler_->GenerateUnRelocatedPcRelativeCall(); |
51 | AddPcRelativeCallStubTarget(array_stub); |
52 | }; |
53 | } |
54 | } |
55 | } |
56 | |
57 | FlowGraphCompiler::~FlowGraphCompiler() { |
58 | // BlockInfos are zone-allocated, so their destructors are not called. |
59 | // Verify the labels explicitly here. |
60 | for (int i = 0; i < block_info_.length(); ++i) { |
61 | ASSERT(!block_info_[i]->jump_label()->IsLinked()); |
62 | } |
63 | } |
64 | |
65 | bool FlowGraphCompiler::SupportsUnboxedDoubles() { |
66 | return true; |
67 | } |
68 | |
69 | bool FlowGraphCompiler::SupportsUnboxedInt64() { |
70 | return FLAG_unbox_mints; |
71 | } |
72 | |
73 | bool FlowGraphCompiler::SupportsUnboxedSimd128() { |
74 | return FLAG_enable_simd_inline; |
75 | } |
76 | |
77 | bool FlowGraphCompiler::CanConvertInt64ToDouble() { |
78 | return true; |
79 | } |
80 | |
81 | bool FlowGraphCompiler::SupportsHardwareDivision() { |
82 | return true; |
83 | } |
84 | |
85 | void FlowGraphCompiler::EnterIntrinsicMode() { |
86 | ASSERT(!intrinsic_mode()); |
87 | intrinsic_mode_ = true; |
88 | ASSERT(!assembler()->constant_pool_allowed()); |
89 | } |
90 | |
91 | void FlowGraphCompiler::ExitIntrinsicMode() { |
92 | ASSERT(intrinsic_mode()); |
93 | intrinsic_mode_ = false; |
94 | } |
95 | |
96 | TypedDataPtr CompilerDeoptInfo::CreateDeoptInfo(FlowGraphCompiler* compiler, |
97 | DeoptInfoBuilder* builder, |
98 | const Array& deopt_table) { |
99 | if (deopt_env_ == NULL) { |
100 | ++builder->current_info_number_; |
101 | return TypedData::null(); |
102 | } |
103 | |
104 | intptr_t stack_height = compiler->StackSize(); |
105 | AllocateIncomingParametersRecursive(deopt_env_, &stack_height); |
106 | |
107 | intptr_t slot_ix = 0; |
108 | Environment* current = deopt_env_; |
109 | |
110 | // Emit all kMaterializeObject instructions describing objects to be |
111 | // materialized on the deoptimization as a prefix to the deoptimization info. |
112 | EmitMaterializations(deopt_env_, builder); |
113 | |
114 | // The real frame starts here. |
115 | builder->MarkFrameStart(); |
116 | |
117 | Zone* zone = compiler->zone(); |
118 | |
119 | builder->AddPp(current->function(), slot_ix++); |
120 | builder->AddPcMarker(Function::ZoneHandle(zone), slot_ix++); |
121 | builder->AddCallerFp(slot_ix++); |
122 | builder->AddReturnAddress(current->function(), deopt_id(), slot_ix++); |
123 | |
124 | // Emit all values that are needed for materialization as a part of the |
125 | // expression stack for the bottom-most frame. This guarantees that GC |
126 | // will be able to find them during materialization. |
127 | slot_ix = builder->EmitMaterializationArguments(slot_ix); |
128 | |
129 | // For the innermost environment, set outgoing arguments and the locals. |
130 | for (intptr_t i = current->Length() - 1; |
131 | i >= current->fixed_parameter_count(); i--) { |
132 | builder->AddCopy(current->ValueAt(i), current->LocationAt(i), slot_ix++); |
133 | } |
134 | |
135 | Environment* previous = current; |
136 | current = current->outer(); |
137 | while (current != NULL) { |
138 | builder->AddPp(current->function(), slot_ix++); |
139 | builder->AddPcMarker(previous->function(), slot_ix++); |
140 | builder->AddCallerFp(slot_ix++); |
141 | |
142 | // For any outer environment the deopt id is that of the call instruction |
143 | // which is recorded in the outer environment. |
144 | builder->AddReturnAddress(current->function(), |
145 | DeoptId::ToDeoptAfter(current->deopt_id()), |
146 | slot_ix++); |
147 | |
148 | // The values of outgoing arguments can be changed from the inlined call so |
149 | // we must read them from the previous environment. |
150 | for (intptr_t i = previous->fixed_parameter_count() - 1; i >= 0; i--) { |
151 | builder->AddCopy(previous->ValueAt(i), previous->LocationAt(i), |
152 | slot_ix++); |
153 | } |
154 | |
155 | // Set the locals, note that outgoing arguments are not in the environment. |
156 | for (intptr_t i = current->Length() - 1; |
157 | i >= current->fixed_parameter_count(); i--) { |
158 | builder->AddCopy(current->ValueAt(i), current->LocationAt(i), slot_ix++); |
159 | } |
160 | |
161 | // Iterate on the outer environment. |
162 | previous = current; |
163 | current = current->outer(); |
164 | } |
165 | // The previous pointer is now the outermost environment. |
166 | ASSERT(previous != NULL); |
167 | |
168 | // Add slots for the outermost environment. |
169 | builder->AddCallerPp(slot_ix++); |
170 | builder->AddPcMarker(previous->function(), slot_ix++); |
171 | builder->AddCallerFp(slot_ix++); |
172 | builder->AddCallerPc(slot_ix++); |
173 | |
174 | // For the outermost environment, set the incoming arguments. |
175 | for (intptr_t i = previous->fixed_parameter_count() - 1; i >= 0; i--) { |
176 | builder->AddCopy(previous->ValueAt(i), previous->LocationAt(i), slot_ix++); |
177 | } |
178 | |
179 | return builder->CreateDeoptInfo(deopt_table); |
180 | } |
181 | |
182 | void CompilerDeoptInfoWithStub::GenerateCode(FlowGraphCompiler* compiler, |
183 | intptr_t stub_ix) { |
184 | // Calls do not need stubs, they share a deoptimization trampoline. |
185 | ASSERT(reason() != ICData::kDeoptAtCall); |
186 | compiler::Assembler* assembler = compiler->assembler(); |
187 | #define __ assembler-> |
188 | __ Comment("%s" , Name()); |
189 | __ Bind(entry_label()); |
190 | if (FLAG_trap_on_deoptimization) { |
191 | __ brk(0); |
192 | } |
193 | |
194 | ASSERT(deopt_env() != NULL); |
195 | __ ldr(LR, compiler::Address(THR, Thread::deoptimize_entry_offset())); |
196 | __ blr(LR); |
197 | set_pc_offset(assembler->CodeSize()); |
198 | #undef __ |
199 | } |
200 | |
201 | #define __ assembler()-> |
202 | |
203 | // Fall through if bool_register contains null. |
204 | void FlowGraphCompiler::GenerateBoolToJump(Register bool_register, |
205 | compiler::Label* is_true, |
206 | compiler::Label* is_false) { |
207 | compiler::Label fall_through; |
208 | __ CompareObject(bool_register, Object::null_object()); |
209 | __ b(&fall_through, EQ); |
210 | BranchLabels labels = {is_true, is_false, &fall_through}; |
211 | Condition true_condition = |
212 | EmitBoolTest(bool_register, labels, /*invert=*/false); |
213 | ASSERT(true_condition == kInvalidCondition); |
214 | __ Bind(&fall_through); |
215 | } |
216 | |
217 | // R0: instance (must be preserved). |
218 | // R2: instantiator type arguments (if used). |
219 | // R1: function type arguments (if used). |
220 | SubtypeTestCachePtr FlowGraphCompiler::GenerateCallSubtypeTestStub( |
221 | TypeTestStubKind test_kind, |
222 | Register instance_reg, |
223 | Register instantiator_type_arguments_reg, |
224 | Register function_type_arguments_reg, |
225 | Register temp_reg, |
226 | compiler::Label* is_instance_lbl, |
227 | compiler::Label* is_not_instance_lbl) { |
228 | ASSERT(instance_reg == R0); |
229 | ASSERT(temp_reg == kNoRegister); // Unused on ARM64. |
230 | const SubtypeTestCache& type_test_cache = |
231 | SubtypeTestCache::ZoneHandle(zone(), SubtypeTestCache::New()); |
232 | __ LoadUniqueObject(R3, type_test_cache); |
233 | if (test_kind == kTestTypeOneArg) { |
234 | ASSERT(instantiator_type_arguments_reg == kNoRegister); |
235 | ASSERT(function_type_arguments_reg == kNoRegister); |
236 | __ BranchLink(StubCode::Subtype1TestCache()); |
237 | } else if (test_kind == kTestTypeTwoArgs) { |
238 | ASSERT(instantiator_type_arguments_reg == kNoRegister); |
239 | ASSERT(function_type_arguments_reg == kNoRegister); |
240 | __ BranchLink(StubCode::Subtype2TestCache()); |
241 | } else if (test_kind == kTestTypeFourArgs) { |
242 | ASSERT(instantiator_type_arguments_reg == |
243 | TypeTestABI::kInstantiatorTypeArgumentsReg); |
244 | ASSERT(function_type_arguments_reg == |
245 | TypeTestABI::kFunctionTypeArgumentsReg); |
246 | __ BranchLink(StubCode::Subtype4TestCache()); |
247 | } else if (test_kind == kTestTypeSixArgs) { |
248 | ASSERT(instantiator_type_arguments_reg == |
249 | TypeTestABI::kInstantiatorTypeArgumentsReg); |
250 | ASSERT(function_type_arguments_reg == |
251 | TypeTestABI::kFunctionTypeArgumentsReg); |
252 | __ BranchLink(StubCode::Subtype6TestCache()); |
253 | } else { |
254 | UNREACHABLE(); |
255 | } |
256 | // Result is in R1: null -> not found, otherwise Bool::True or Bool::False. |
257 | GenerateBoolToJump(R1, is_instance_lbl, is_not_instance_lbl); |
258 | return type_test_cache.raw(); |
259 | } |
260 | |
261 | // Jumps to labels 'is_instance' or 'is_not_instance' respectively, if |
262 | // type test is conclusive, otherwise fallthrough if a type test could not |
263 | // be completed. |
264 | // R0: instance being type checked (preserved). |
265 | // Clobbers R1, R2. |
266 | SubtypeTestCachePtr |
267 | FlowGraphCompiler::GenerateInstantiatedTypeWithArgumentsTest( |
268 | TokenPosition token_pos, |
269 | const AbstractType& type, |
270 | compiler::Label* is_instance_lbl, |
271 | compiler::Label* is_not_instance_lbl) { |
272 | __ Comment("InstantiatedTypeWithArgumentsTest" ); |
273 | ASSERT(type.IsInstantiated()); |
274 | ASSERT(!type.IsFunctionType()); |
275 | const Class& type_class = Class::ZoneHandle(zone(), type.type_class()); |
276 | ASSERT(type_class.NumTypeArguments() > 0); |
277 | const Type& smi_type = Type::Handle(zone(), Type::SmiType()); |
278 | const bool smi_is_ok = smi_type.IsSubtypeOf(type, Heap::kOld); |
279 | // Fast case for type = FutureOr<int/num/top-type>. |
280 | __ BranchIfSmi(TypeTestABI::kInstanceReg, |
281 | smi_is_ok ? is_instance_lbl : is_not_instance_lbl); |
282 | |
283 | const intptr_t num_type_args = type_class.NumTypeArguments(); |
284 | const intptr_t num_type_params = type_class.NumTypeParameters(); |
285 | const intptr_t from_index = num_type_args - num_type_params; |
286 | const TypeArguments& type_arguments = |
287 | TypeArguments::ZoneHandle(zone(), type.arguments()); |
288 | const bool is_raw_type = type_arguments.IsNull() || |
289 | type_arguments.IsRaw(from_index, num_type_params); |
290 | if (is_raw_type) { |
291 | const Register kClassIdReg = R2; |
292 | // dynamic type argument, check only classes. |
293 | __ LoadClassId(kClassIdReg, TypeTestABI::kInstanceReg); |
294 | __ CompareImmediate(kClassIdReg, type_class.id()); |
295 | __ b(is_instance_lbl, EQ); |
296 | // List is a very common case. |
297 | if (IsListClass(type_class)) { |
298 | GenerateListTypeCheck(kClassIdReg, is_instance_lbl); |
299 | } |
300 | return GenerateSubtype1TestCacheLookup( |
301 | token_pos, type_class, is_instance_lbl, is_not_instance_lbl); |
302 | } |
303 | // If one type argument only, check if type argument is a top type. |
304 | if (type_arguments.Length() == 1) { |
305 | const AbstractType& tp_argument = |
306 | AbstractType::ZoneHandle(zone(), type_arguments.TypeAt(0)); |
307 | if (tp_argument.IsTopTypeForSubtyping()) { |
308 | // Instance class test only necessary. |
309 | return GenerateSubtype1TestCacheLookup( |
310 | token_pos, type_class, is_instance_lbl, is_not_instance_lbl); |
311 | } |
312 | } |
313 | // Regular subtype test cache involving instance's type arguments. |
314 | const Register kInstantiatorTypeArgumentsReg = kNoRegister; |
315 | const Register kFunctionTypeArgumentsReg = kNoRegister; |
316 | const Register kTempReg = kNoRegister; |
317 | // R0: instance (must be preserved). |
318 | return GenerateCallSubtypeTestStub( |
319 | kTestTypeTwoArgs, TypeTestABI::kInstanceReg, |
320 | kInstantiatorTypeArgumentsReg, kFunctionTypeArgumentsReg, kTempReg, |
321 | is_instance_lbl, is_not_instance_lbl); |
322 | } |
323 | |
324 | void FlowGraphCompiler::CheckClassIds(Register class_id_reg, |
325 | const GrowableArray<intptr_t>& class_ids, |
326 | compiler::Label* is_equal_lbl, |
327 | compiler::Label* is_not_equal_lbl) { |
328 | for (intptr_t i = 0; i < class_ids.length(); i++) { |
329 | __ CompareImmediate(class_id_reg, class_ids[i]); |
330 | __ b(is_equal_lbl, EQ); |
331 | } |
332 | __ b(is_not_equal_lbl); |
333 | } |
334 | |
335 | // Testing against an instantiated type with no arguments, without |
336 | // SubtypeTestCache. |
337 | // R0: instance being type checked (preserved). |
338 | // Clobbers R2, R3. |
339 | // Returns true if there is a fallthrough. |
340 | bool FlowGraphCompiler::GenerateInstantiatedTypeNoArgumentsTest( |
341 | TokenPosition token_pos, |
342 | const AbstractType& type, |
343 | compiler::Label* is_instance_lbl, |
344 | compiler::Label* is_not_instance_lbl) { |
345 | __ Comment("InstantiatedTypeNoArgumentsTest" ); |
346 | ASSERT(type.IsInstantiated()); |
347 | ASSERT(!type.IsFunctionType()); |
348 | const Class& type_class = Class::Handle(zone(), type.type_class()); |
349 | ASSERT(type_class.NumTypeArguments() == 0); |
350 | |
351 | // If instance is Smi, check directly. |
352 | const Class& smi_class = Class::Handle(zone(), Smi::Class()); |
353 | if (Class::IsSubtypeOf(smi_class, Object::null_type_arguments(), |
354 | Nullability::kNonNullable, type, Heap::kOld)) { |
355 | // Fast case for type = int/num/top-type. |
356 | __ BranchIfSmi(TypeTestABI::kInstanceReg, is_instance_lbl); |
357 | } else { |
358 | __ BranchIfSmi(TypeTestABI::kInstanceReg, is_not_instance_lbl); |
359 | } |
360 | const Register kClassIdReg = R2; |
361 | __ LoadClassId(kClassIdReg, TypeTestABI::kInstanceReg); |
362 | // Bool interface can be implemented only by core class Bool. |
363 | if (type.IsBoolType()) { |
364 | __ CompareImmediate(kClassIdReg, kBoolCid); |
365 | __ b(is_instance_lbl, EQ); |
366 | __ b(is_not_instance_lbl); |
367 | return false; |
368 | } |
369 | // Custom checking for numbers (Smi, Mint and Double). |
370 | // Note that instance is not Smi (checked above). |
371 | if (type.IsNumberType() || type.IsIntType() || type.IsDoubleType()) { |
372 | GenerateNumberTypeCheck(kClassIdReg, type, is_instance_lbl, |
373 | is_not_instance_lbl); |
374 | return false; |
375 | } |
376 | if (type.IsStringType()) { |
377 | GenerateStringTypeCheck(kClassIdReg, is_instance_lbl, is_not_instance_lbl); |
378 | return false; |
379 | } |
380 | if (type.IsDartFunctionType()) { |
381 | // Check if instance is a closure. |
382 | __ CompareImmediate(kClassIdReg, kClosureCid); |
383 | __ b(is_instance_lbl, EQ); |
384 | return true; // Fall through |
385 | } |
386 | |
387 | // Fast case for cid-range based checks. |
388 | // Warning: This code destroys the contents of [kClassIdReg]. |
389 | if (GenerateSubtypeRangeCheck(kClassIdReg, type_class, is_instance_lbl)) { |
390 | return false; |
391 | } |
392 | |
393 | // Otherwise fallthrough, result non-conclusive. |
394 | return true; |
395 | } |
396 | |
397 | // Uses SubtypeTestCache to store instance class and result. |
398 | // R0: instance to test. |
399 | // Clobbers R1-R5. |
400 | // Immediate class test already done. |
401 | // TODO(srdjan): Implement a quicker subtype check, as type test |
402 | // arrays can grow too high, but they may be useful when optimizing |
403 | // code (type-feedback). |
404 | SubtypeTestCachePtr FlowGraphCompiler::GenerateSubtype1TestCacheLookup( |
405 | TokenPosition token_pos, |
406 | const Class& type_class, |
407 | compiler::Label* is_instance_lbl, |
408 | compiler::Label* is_not_instance_lbl) { |
409 | __ Comment("Subtype1TestCacheLookup" ); |
410 | #if defined(DEBUG) |
411 | compiler::Label ok; |
412 | __ BranchIfNotSmi(TypeTestABI::kInstanceReg, &ok); |
413 | __ Breakpoint(); |
414 | __ Bind(&ok); |
415 | #endif |
416 | __ LoadClassId(TMP, TypeTestABI::kInstanceReg); |
417 | __ LoadClassById(R1, TMP); |
418 | // R1: instance class. |
419 | // Check immediate superclass equality. If type_class is Object, then testing |
420 | // supertype may yield a wrong result for Null in NNBD strong mode (because |
421 | // Null also extends Object). |
422 | if (!type_class.IsObjectClass() || !Isolate::Current()->null_safety()) { |
423 | __ LoadFieldFromOffset(R2, R1, Class::super_type_offset()); |
424 | __ LoadFieldFromOffset(R2, R2, Type::type_class_id_offset()); |
425 | __ CompareImmediate(R2, Smi::RawValue(type_class.id())); |
426 | __ b(is_instance_lbl, EQ); |
427 | } |
428 | |
429 | const Register kInstantiatorTypeArgumentsReg = kNoRegister; |
430 | const Register kFunctionTypeArgumentsReg = kNoRegister; |
431 | const Register kTempReg = kNoRegister; |
432 | return GenerateCallSubtypeTestStub(kTestTypeOneArg, TypeTestABI::kInstanceReg, |
433 | kInstantiatorTypeArgumentsReg, |
434 | kFunctionTypeArgumentsReg, kTempReg, |
435 | is_instance_lbl, is_not_instance_lbl); |
436 | } |
437 | |
438 | // Generates inlined check if 'type' is a type parameter or type itself |
439 | // R0: instance (preserved). |
440 | SubtypeTestCachePtr FlowGraphCompiler::GenerateUninstantiatedTypeTest( |
441 | TokenPosition token_pos, |
442 | const AbstractType& type, |
443 | compiler::Label* is_instance_lbl, |
444 | compiler::Label* is_not_instance_lbl) { |
445 | __ Comment("UninstantiatedTypeTest" ); |
446 | const Register kTempReg = kNoRegister; |
447 | ASSERT(!type.IsInstantiated()); |
448 | ASSERT(!type.IsFunctionType()); |
449 | // Skip check if destination is a dynamic type. |
450 | if (type.IsTypeParameter()) { |
451 | const TypeParameter& type_param = TypeParameter::Cast(type); |
452 | |
453 | // Get instantiator type args (high) and function type args (low). |
454 | __ ldp(TypeTestABI::kFunctionTypeArgumentsReg, |
455 | TypeTestABI::kInstantiatorTypeArgumentsReg, |
456 | compiler::Address(SP, 0 * kWordSize, compiler::Address::PairOffset)); |
457 | const Register kTypeArgumentsReg = |
458 | type_param.IsClassTypeParameter() |
459 | ? TypeTestABI::kInstantiatorTypeArgumentsReg |
460 | : TypeTestABI::kFunctionTypeArgumentsReg; |
461 | // Check if type arguments are null, i.e. equivalent to vector of dynamic. |
462 | __ CompareObject(kTypeArgumentsReg, Object::null_object()); |
463 | __ b(is_instance_lbl, EQ); |
464 | __ LoadFieldFromOffset(R3, kTypeArgumentsReg, |
465 | TypeArguments::type_at_offset(type_param.index())); |
466 | // R3: concrete type of type. |
467 | // Check if type argument is dynamic, Object?, or void. |
468 | __ CompareObject(R3, Object::dynamic_type()); |
469 | __ b(is_instance_lbl, EQ); |
470 | __ CompareObject( |
471 | R3, Type::ZoneHandle( |
472 | zone(), isolate()->object_store()->nullable_object_type())); |
473 | __ b(is_instance_lbl, EQ); |
474 | __ CompareObject(R3, Object::void_type()); |
475 | __ b(is_instance_lbl, EQ); |
476 | |
477 | // For Smi check quickly against int and num interfaces. |
478 | compiler::Label not_smi; |
479 | __ BranchIfNotSmi(R0, ¬_smi); |
480 | __ CompareObject(R3, Type::ZoneHandle(zone(), Type::IntType())); |
481 | __ b(is_instance_lbl, EQ); |
482 | __ CompareObject(R3, Type::ZoneHandle(zone(), Type::Number())); |
483 | __ b(is_instance_lbl, EQ); |
484 | // Smi can be handled by type test cache. |
485 | __ Bind(¬_smi); |
486 | |
487 | const auto test_kind = GetTypeTestStubKindForTypeParameter(type_param); |
488 | const SubtypeTestCache& type_test_cache = SubtypeTestCache::ZoneHandle( |
489 | zone(), GenerateCallSubtypeTestStub( |
490 | test_kind, TypeTestABI::kInstanceReg, |
491 | TypeTestABI::kInstantiatorTypeArgumentsReg, |
492 | TypeTestABI::kFunctionTypeArgumentsReg, kTempReg, |
493 | is_instance_lbl, is_not_instance_lbl)); |
494 | return type_test_cache.raw(); |
495 | } |
496 | if (type.IsType()) { |
497 | // Smi is FutureOr<T>, when T is a top type or int or num. |
498 | if (!type.IsFutureOrType()) { |
499 | __ BranchIfSmi(TypeTestABI::kInstanceReg, is_not_instance_lbl); |
500 | } |
501 | __ ldp(TypeTestABI::kFunctionTypeArgumentsReg, |
502 | TypeTestABI::kInstantiatorTypeArgumentsReg, |
503 | compiler::Address(SP, 0 * kWordSize, compiler::Address::PairOffset)); |
504 | // Uninstantiated type class is known at compile time, but the type |
505 | // arguments are determined at runtime by the instantiator. |
506 | return GenerateCallSubtypeTestStub( |
507 | kTestTypeFourArgs, TypeTestABI::kInstanceReg, |
508 | TypeTestABI::kInstantiatorTypeArgumentsReg, |
509 | TypeTestABI::kFunctionTypeArgumentsReg, kTempReg, is_instance_lbl, |
510 | is_not_instance_lbl); |
511 | } |
512 | return SubtypeTestCache::null(); |
513 | } |
514 | |
515 | // Generates function type check. |
516 | // |
517 | // See [GenerateUninstantiatedTypeTest] for calling convention. |
518 | SubtypeTestCachePtr FlowGraphCompiler::GenerateFunctionTypeTest( |
519 | TokenPosition token_pos, |
520 | const AbstractType& type, |
521 | compiler::Label* is_instance_lbl, |
522 | compiler::Label* is_not_instance_lbl) { |
523 | __ BranchIfSmi(TypeTestABI::kInstanceReg, is_not_instance_lbl); |
524 | __ ldp(TypeTestABI::kFunctionTypeArgumentsReg, |
525 | TypeTestABI::kInstantiatorTypeArgumentsReg, |
526 | compiler::Address(SP, 0 * kWordSize, compiler::Address::PairOffset)); |
527 | // Uninstantiated type class is known at compile time, but the type |
528 | // arguments are determined at runtime by the instantiator(s). |
529 | const Register kTempReg = kNoRegister; |
530 | return GenerateCallSubtypeTestStub( |
531 | kTestTypeSixArgs, TypeTestABI::kInstanceReg, |
532 | TypeTestABI::kInstantiatorTypeArgumentsReg, |
533 | TypeTestABI::kFunctionTypeArgumentsReg, kTempReg, is_instance_lbl, |
534 | is_not_instance_lbl); |
535 | } |
536 | |
537 | // Inputs: |
538 | // - R0: instance being type checked (preserved). |
539 | // - R2: optional instantiator type arguments (preserved). |
540 | // - R1: optional function type arguments (preserved). |
541 | // Clobbers R3, R4, R8, R9. |
542 | // Returns: |
543 | // - preserved instance in R0, optional instantiator type arguments in R2, and |
544 | // optional function type arguments in R1. |
545 | // Note that this inlined code must be followed by the runtime_call code, as it |
546 | // may fall through to it. Otherwise, this inline code will jump to the label |
547 | // is_instance or to the label is_not_instance. |
548 | SubtypeTestCachePtr FlowGraphCompiler::GenerateInlineInstanceof( |
549 | TokenPosition token_pos, |
550 | const AbstractType& type, |
551 | compiler::Label* is_instance_lbl, |
552 | compiler::Label* is_not_instance_lbl) { |
553 | __ Comment("InlineInstanceof" ); |
554 | |
555 | if (type.IsFunctionType()) { |
556 | return GenerateFunctionTypeTest(token_pos, type, is_instance_lbl, |
557 | is_not_instance_lbl); |
558 | } |
559 | |
560 | if (type.IsInstantiated()) { |
561 | const Class& type_class = Class::ZoneHandle(zone(), type.type_class()); |
562 | // A class equality check is only applicable with a dst type (not a |
563 | // function type) of a non-parameterized class or with a raw dst type of |
564 | // a parameterized class. |
565 | if (type_class.NumTypeArguments() > 0) { |
566 | return GenerateInstantiatedTypeWithArgumentsTest( |
567 | token_pos, type, is_instance_lbl, is_not_instance_lbl); |
568 | // Fall through to runtime call. |
569 | } |
570 | const bool has_fall_through = GenerateInstantiatedTypeNoArgumentsTest( |
571 | token_pos, type, is_instance_lbl, is_not_instance_lbl); |
572 | if (has_fall_through) { |
573 | // If test non-conclusive so far, try the inlined type-test cache. |
574 | // 'type' is known at compile time. |
575 | return GenerateSubtype1TestCacheLookup( |
576 | token_pos, type_class, is_instance_lbl, is_not_instance_lbl); |
577 | } else { |
578 | return SubtypeTestCache::null(); |
579 | } |
580 | } |
581 | return GenerateUninstantiatedTypeTest(token_pos, type, is_instance_lbl, |
582 | is_not_instance_lbl); |
583 | } |
584 | |
585 | // If instanceof type test cannot be performed successfully at compile time and |
586 | // therefore eliminated, optimize it by adding inlined tests for: |
587 | // - Null -> see comment below. |
588 | // - Smi -> compile time subtype check (only if dst class is not parameterized). |
589 | // - Class equality (only if class is not parameterized). |
590 | // Inputs: |
591 | // - R0: object. |
592 | // - R2: instantiator type arguments or raw_null. |
593 | // - R1: function type arguments or raw_null. |
594 | // Returns: |
595 | // - true or false in R0. |
596 | void FlowGraphCompiler::GenerateInstanceOf(TokenPosition token_pos, |
597 | intptr_t deopt_id, |
598 | const AbstractType& type, |
599 | LocationSummary* locs) { |
600 | ASSERT(type.IsFinalized()); |
601 | ASSERT(!type.IsTopTypeForInstanceOf()); // Already checked. |
602 | __ PushPair(TypeTestABI::kFunctionTypeArgumentsReg, |
603 | TypeTestABI::kInstantiatorTypeArgumentsReg); |
604 | |
605 | compiler::Label is_instance, is_not_instance; |
606 | // 'null' is an instance of Null, Object*, Never*, void, and dynamic. |
607 | // In addition, 'null' is an instance of any nullable type. |
608 | // It is also an instance of FutureOr<T> if it is an instance of T. |
609 | const AbstractType& unwrapped_type = |
610 | AbstractType::Handle(type.UnwrapFutureOr()); |
611 | if (!unwrapped_type.IsTypeParameter() || unwrapped_type.IsNullable()) { |
612 | // Only nullable type parameter remains nullable after instantiation. |
613 | // See NullIsInstanceOf(). |
614 | __ CompareObject(TypeTestABI::kInstanceReg, Object::null_object()); |
615 | __ b((unwrapped_type.IsNullable() || |
616 | (unwrapped_type.IsLegacy() && unwrapped_type.IsNeverType())) |
617 | ? &is_instance |
618 | : &is_not_instance, |
619 | EQ); |
620 | } |
621 | |
622 | // Generate inline instanceof test. |
623 | SubtypeTestCache& test_cache = SubtypeTestCache::ZoneHandle(zone()); |
624 | test_cache = |
625 | GenerateInlineInstanceof(token_pos, type, &is_instance, &is_not_instance); |
626 | |
627 | // test_cache is null if there is no fall-through. |
628 | compiler::Label done; |
629 | if (!test_cache.IsNull()) { |
630 | // Generate runtime call. |
631 | __ ldp(TypeTestABI::kFunctionTypeArgumentsReg, |
632 | TypeTestABI::kInstantiatorTypeArgumentsReg, |
633 | compiler::Address(SP, 0 * kWordSize, compiler::Address::PairOffset)); |
634 | __ LoadUniqueObject(TypeTestABI::kDstTypeReg, type); |
635 | __ LoadUniqueObject(TypeTestABI::kSubtypeTestCacheReg, test_cache); |
636 | GenerateStubCall(token_pos, StubCode::InstanceOf(), |
637 | /*kind=*/PcDescriptorsLayout::kOther, locs, deopt_id); |
638 | __ b(&done); |
639 | } |
640 | __ Bind(&is_not_instance); |
641 | __ LoadObject(R0, Bool::Get(false)); |
642 | __ b(&done); |
643 | |
644 | __ Bind(&is_instance); |
645 | __ LoadObject(R0, Bool::Get(true)); |
646 | __ Bind(&done); |
647 | // Remove instantiator type arguments and function type arguments. |
648 | __ Drop(2); |
649 | } |
650 | |
651 | // Optimize assignable type check by adding inlined tests for: |
652 | // - NULL -> return NULL. |
653 | // - Smi -> compile time subtype check (only if dst class is not parameterized). |
654 | // - Class equality (only if class is not parameterized). |
655 | // Inputs: |
656 | // - R0: instance being type checked. |
657 | // - R8: destination type (if non-constant). |
658 | // - R2: instantiator type arguments or raw_null. |
659 | // - R1: function type arguments or raw_null. |
660 | // Returns: |
661 | // - object in R0 for successful assignable check (or throws TypeError). |
662 | // Performance notes: positive checks must be quick, negative checks can be slow |
663 | // as they throw an exception. |
664 | void FlowGraphCompiler::GenerateAssertAssignable(CompileType* receiver_type, |
665 | TokenPosition token_pos, |
666 | intptr_t deopt_id, |
667 | const String& dst_name, |
668 | LocationSummary* locs) { |
669 | ASSERT(!TokenPosition(token_pos).IsClassifying()); |
670 | ASSERT(CheckAssertAssignableTypeTestingABILocations(*locs)); |
671 | |
672 | compiler::Label is_assignable_fast, is_assignable, runtime_call; |
673 | // Generate inline type check, linking to runtime call if not assignable. |
674 | SubtypeTestCache& test_cache = SubtypeTestCache::ZoneHandle(zone()); |
675 | |
676 | if (locs->in(1).IsConstant()) { |
677 | const auto& dst_type = AbstractType::Cast(locs->in(1).constant()); |
678 | ASSERT(dst_type.IsFinalized()); |
679 | |
680 | if (dst_type.IsTopTypeForSubtyping()) return; // No code needed. |
681 | |
682 | if (ShouldUseTypeTestingStubFor(is_optimizing(), dst_type)) { |
683 | GenerateAssertAssignableViaTypeTestingStub(receiver_type, token_pos, |
684 | deopt_id, dst_name, locs); |
685 | return; |
686 | } |
687 | |
688 | if (Instance::NullIsAssignableTo(dst_type)) { |
689 | __ CompareObject(TypeTestABI::kInstanceReg, Object::null_object()); |
690 | __ b(&is_assignable_fast, EQ); |
691 | } |
692 | |
693 | __ PushPair(TypeTestABI::kFunctionTypeArgumentsReg, |
694 | TypeTestABI::kInstantiatorTypeArgumentsReg); |
695 | |
696 | test_cache = GenerateInlineInstanceof(token_pos, dst_type, &is_assignable, |
697 | &runtime_call); |
698 | } else { |
699 | // TODO(dartbug.com/40813): Handle setting up the non-constant case. |
700 | UNREACHABLE(); |
701 | } |
702 | |
703 | __ Bind(&runtime_call); |
704 | __ ldp(TypeTestABI::kFunctionTypeArgumentsReg, |
705 | TypeTestABI::kInstantiatorTypeArgumentsReg, |
706 | compiler::Address(SP, 0 * kWordSize, compiler::Address::PairOffset)); |
707 | // Make room for the result and push the source object. |
708 | __ PushPair(TypeTestABI::kInstanceReg, NULL_REG); |
709 | // Push the destination type and the instantiator type arguments. |
710 | if (locs->in(1).IsConstant()) { |
711 | __ LoadObject(TMP, locs->in(1).constant()); |
712 | __ PushPair(TypeTestABI::kInstantiatorTypeArgumentsReg, TMP); |
713 | } else { |
714 | // TODO(dartbug.com/40813): Handle setting up the non-constant case. |
715 | UNREACHABLE(); |
716 | } |
717 | // Push the function type arguments and the name of the destination. |
718 | __ LoadObject(TMP, dst_name); |
719 | __ PushPair(TMP, TypeTestABI::kFunctionTypeArgumentsReg); |
720 | |
721 | __ LoadUniqueObject(R0, test_cache); |
722 | __ LoadObject(TMP, Smi::ZoneHandle(zone(), Smi::New(kTypeCheckFromInline))); |
723 | __ PushPair(TMP, R0); |
724 | GenerateRuntimeCall(token_pos, deopt_id, kTypeCheckRuntimeEntry, 7, locs); |
725 | // Pop the parameters supplied to the runtime entry. The result of the |
726 | // type check runtime call is the checked value. |
727 | __ Drop(7); |
728 | __ Pop(TypeTestABI::kInstanceReg); |
729 | __ Bind(&is_assignable); |
730 | __ PopPair(TypeTestABI::kFunctionTypeArgumentsReg, |
731 | TypeTestABI::kInstantiatorTypeArgumentsReg); |
732 | __ Bind(&is_assignable_fast); |
733 | } |
734 | |
735 | void FlowGraphCompiler::GenerateAssertAssignableViaTypeTestingStub( |
736 | CompileType* receiver_type, |
737 | TokenPosition token_pos, |
738 | intptr_t deopt_id, |
739 | const String& dst_name, |
740 | LocationSummary* locs) { |
741 | ASSERT(CheckAssertAssignableTypeTestingABILocations(*locs)); |
742 | // We must have a constant dst_type for generating a call to the stub. |
743 | ASSERT(locs->in(1).IsConstant()); |
744 | const auto& dst_type = AbstractType::Cast(locs->in(1).constant()); |
745 | |
746 | // If the dst_type is instantiated we know the target TTS stub at |
747 | // compile-time and can therefore use a pc-relative call. |
748 | const bool use_pc_relative_call = |
749 | dst_type.IsInstantiated() && CanPcRelativeCall(dst_type); |
750 | |
751 | const Register kRegToCall = |
752 | use_pc_relative_call |
753 | ? kNoRegister |
754 | : (dst_type.IsTypeParameter() ? R9 : TypeTestABI::kDstTypeReg); |
755 | const Register kScratchReg = R4; |
756 | |
757 | compiler::Label done; |
758 | |
759 | GenerateAssertAssignableViaTypeTestingStub(receiver_type, dst_type, dst_name, |
760 | kRegToCall, kScratchReg, &done); |
761 | |
762 | // We use 2 consecutive entries in the pool for the subtype cache and the |
763 | // destination name. The second entry, namely [dst_name] seems to be unused, |
764 | // but it will be used by the code throwing a TypeError if the type test fails |
765 | // (see runtime/vm/runtime_entry.cc:TypeCheck). It will use pattern matching |
766 | // on the call site to find out at which pool index the destination name is |
767 | // located. |
768 | const intptr_t sub_type_cache_index = __ object_pool_builder().AddObject( |
769 | Object::null_object(), ObjectPool::Patchability::kPatchable); |
770 | const intptr_t sub_type_cache_offset = |
771 | ObjectPool::element_offset(sub_type_cache_index); |
772 | const intptr_t dst_name_index = __ object_pool_builder().AddObject( |
773 | dst_name, ObjectPool::Patchability::kPatchable); |
774 | ASSERT((sub_type_cache_index + 1) == dst_name_index); |
775 | ASSERT(__ constant_pool_allowed()); |
776 | |
777 | if (use_pc_relative_call) { |
778 | __ LoadWordFromPoolOffset(TypeTestABI::kSubtypeTestCacheReg, |
779 | sub_type_cache_offset); |
780 | __ GenerateUnRelocatedPcRelativeCall(); |
781 | AddPcRelativeTTSCallTypeTarget(dst_type); |
782 | } else { |
783 | __ LoadField( |
784 | R9, compiler::FieldAddress( |
785 | kRegToCall, AbstractType::type_test_stub_entry_point_offset())); |
786 | __ LoadWordFromPoolOffset(TypeTestABI::kSubtypeTestCacheReg, |
787 | sub_type_cache_offset); |
788 | __ blr(R9); |
789 | } |
790 | EmitCallsiteMetadata(token_pos, deopt_id, PcDescriptorsLayout::kOther, locs); |
791 | __ Bind(&done); |
792 | } |
793 | |
794 | void FlowGraphCompiler::EmitInstructionEpilogue(Instruction* instr) { |
795 | if (is_optimizing()) { |
796 | return; |
797 | } |
798 | Definition* defn = instr->AsDefinition(); |
799 | if ((defn != NULL) && defn->HasTemp()) { |
800 | __ Push(defn->locs()->out(0).reg()); |
801 | } |
802 | } |
803 | |
804 | void FlowGraphCompiler::GenerateMethodExtractorIntrinsic( |
805 | const Function& extracted_method, |
806 | intptr_t type_arguments_field_offset) { |
807 | // No frame has been setup here. |
808 | ASSERT(!__ constant_pool_allowed()); |
809 | ASSERT(extracted_method.IsZoneHandle()); |
810 | |
811 | const Code& build_method_extractor = Code::ZoneHandle( |
812 | isolate()->object_store()->build_method_extractor_code()); |
813 | |
814 | const intptr_t stub_index = __ object_pool_builder().AddObject( |
815 | build_method_extractor, ObjectPool::Patchability::kNotPatchable); |
816 | const intptr_t function_index = __ object_pool_builder().AddObject( |
817 | extracted_method, ObjectPool::Patchability::kNotPatchable); |
818 | |
819 | // We use a custom pool register to preserve caller PP. |
820 | Register kPoolReg = R0; |
821 | |
822 | // R1 = extracted function |
823 | // R4 = offset of type argument vector (or 0 if class is not generic) |
824 | intptr_t pp_offset = 0; |
825 | if (FLAG_precompiled_mode && FLAG_use_bare_instructions) { |
826 | // PP is not tagged on arm64. |
827 | kPoolReg = PP; |
828 | pp_offset = kHeapObjectTag; |
829 | } else { |
830 | __ LoadFieldFromOffset(kPoolReg, CODE_REG, Code::object_pool_offset()); |
831 | } |
832 | __ LoadImmediate(R4, type_arguments_field_offset); |
833 | __ LoadFieldFromOffset( |
834 | R1, kPoolReg, ObjectPool::element_offset(function_index) + pp_offset); |
835 | __ LoadFieldFromOffset(CODE_REG, kPoolReg, |
836 | ObjectPool::element_offset(stub_index) + pp_offset); |
837 | __ LoadFieldFromOffset(R0, CODE_REG, |
838 | Code::entry_point_offset(Code::EntryKind::kUnchecked)); |
839 | __ br(R0); |
840 | } |
841 | |
842 | void FlowGraphCompiler::EmitFrameEntry() { |
843 | const Function& function = parsed_function().function(); |
844 | if (CanOptimizeFunction() && function.IsOptimizable() && |
845 | (!is_optimizing() || may_reoptimize())) { |
846 | __ Comment("Invocation Count Check" ); |
847 | const Register function_reg = R6; |
848 | __ ldr(function_reg, |
849 | compiler::FieldAddress(CODE_REG, Code::owner_offset())); |
850 | |
851 | __ LoadFieldFromOffset(R7, function_reg, Function::usage_counter_offset(), |
852 | kWord); |
853 | // Reoptimization of an optimized function is triggered by counting in |
854 | // IC stubs, but not at the entry of the function. |
855 | if (!is_optimizing()) { |
856 | __ add(R7, R7, compiler::Operand(1)); |
857 | __ StoreFieldToOffset(R7, function_reg, Function::usage_counter_offset(), |
858 | kWord); |
859 | } |
860 | __ CompareImmediate(R7, GetOptimizationThreshold()); |
861 | ASSERT(function_reg == R6); |
862 | compiler::Label dont_optimize; |
863 | __ b(&dont_optimize, LT); |
864 | __ ldr(TMP, compiler::Address(THR, Thread::optimize_entry_offset())); |
865 | __ br(TMP); |
866 | __ Bind(&dont_optimize); |
867 | } |
868 | __ Comment("Enter frame" ); |
869 | if (flow_graph().IsCompiledForOsr()) { |
870 | const intptr_t extra_slots = ExtraStackSlotsOnOsrEntry(); |
871 | ASSERT(extra_slots >= 0); |
872 | __ EnterOsrFrame(extra_slots * kWordSize); |
873 | } else { |
874 | ASSERT(StackSize() >= 0); |
875 | __ EnterDartFrame(StackSize() * kWordSize); |
876 | } |
877 | } |
878 | |
879 | void FlowGraphCompiler::EmitPrologue() { |
880 | EmitFrameEntry(); |
881 | ASSERT(assembler()->constant_pool_allowed()); |
882 | |
883 | // In unoptimized code, initialize (non-argument) stack allocated slots. |
884 | if (!is_optimizing()) { |
885 | const int num_locals = parsed_function().num_stack_locals(); |
886 | |
887 | intptr_t args_desc_slot = -1; |
888 | if (parsed_function().has_arg_desc_var()) { |
889 | args_desc_slot = compiler::target::frame_layout.FrameSlotForVariable( |
890 | parsed_function().arg_desc_var()); |
891 | } |
892 | |
893 | __ Comment("Initialize spill slots" ); |
894 | if (num_locals > 1 || (num_locals == 1 && args_desc_slot == -1)) { |
895 | __ LoadObject(R0, Object::null_object()); |
896 | } |
897 | for (intptr_t i = 0; i < num_locals; ++i) { |
898 | const intptr_t slot_index = |
899 | compiler::target::frame_layout.FrameSlotForVariableIndex(-i); |
900 | Register value_reg = slot_index == args_desc_slot ? ARGS_DESC_REG : R0; |
901 | __ StoreToOffset(value_reg, FP, slot_index * kWordSize); |
902 | } |
903 | } |
904 | |
905 | EndCodeSourceRange(TokenPosition::kDartCodePrologue); |
906 | } |
907 | |
908 | // Input parameters: |
909 | // LR: return address. |
910 | // SP: address of last argument. |
911 | // FP: caller's frame pointer. |
912 | // PP: caller's pool pointer. |
913 | // R4: arguments descriptor array. |
914 | void FlowGraphCompiler::CompileGraph() { |
915 | InitCompiler(); |
916 | |
917 | // For JIT we have multiple entrypoints functionality which moved the frame |
918 | // setup into the [TargetEntryInstr] (which will set the constant pool |
919 | // allowed bit to true). Despite this we still have to set the |
920 | // constant pool allowed bit to true here as well, because we can generate |
921 | // code for [CatchEntryInstr]s, which need the pool. |
922 | __ set_constant_pool_allowed(true); |
923 | |
924 | VisitBlocks(); |
925 | |
926 | #if defined(DEBUG) |
927 | __ brk(0); |
928 | #endif |
929 | |
930 | if (!skip_body_compilation()) { |
931 | ASSERT(assembler()->constant_pool_allowed()); |
932 | GenerateDeferredCode(); |
933 | } |
934 | |
935 | for (intptr_t i = 0; i < indirect_gotos_.length(); ++i) { |
936 | indirect_gotos_[i]->ComputeOffsetTable(this); |
937 | } |
938 | } |
939 | |
940 | void FlowGraphCompiler::EmitCallToStub(const Code& stub) { |
941 | ASSERT(!stub.IsNull()); |
942 | if (CanPcRelativeCall(stub)) { |
943 | __ GenerateUnRelocatedPcRelativeCall(); |
944 | AddPcRelativeCallStubTarget(stub); |
945 | } else { |
946 | __ BranchLink(stub); |
947 | AddStubCallTarget(stub); |
948 | } |
949 | } |
950 | |
951 | void FlowGraphCompiler::EmitTailCallToStub(const Code& stub) { |
952 | ASSERT(!stub.IsNull()); |
953 | if (CanPcRelativeCall(stub)) { |
954 | __ LeaveDartFrame(); |
955 | __ GenerateUnRelocatedPcRelativeTailCall(); |
956 | AddPcRelativeTailCallStubTarget(stub); |
957 | #if defined(DEBUG) |
958 | __ Breakpoint(); |
959 | #endif |
960 | } else { |
961 | __ LoadObject(CODE_REG, stub); |
962 | __ LeaveDartFrame(); |
963 | __ ldr(TMP, compiler::FieldAddress( |
964 | CODE_REG, compiler::target::Code::entry_point_offset())); |
965 | __ br(TMP); |
966 | AddStubCallTarget(stub); |
967 | } |
968 | } |
969 | |
970 | void FlowGraphCompiler::GeneratePatchableCall(TokenPosition token_pos, |
971 | const Code& stub, |
972 | PcDescriptorsLayout::Kind kind, |
973 | LocationSummary* locs) { |
974 | __ BranchLinkPatchable(stub); |
975 | EmitCallsiteMetadata(token_pos, DeoptId::kNone, kind, locs); |
976 | } |
977 | |
978 | void FlowGraphCompiler::GenerateDartCall(intptr_t deopt_id, |
979 | TokenPosition token_pos, |
980 | const Code& stub, |
981 | PcDescriptorsLayout::Kind kind, |
982 | LocationSummary* locs, |
983 | Code::EntryKind entry_kind) { |
984 | ASSERT(CanCallDart()); |
985 | __ BranchLinkPatchable(stub, entry_kind); |
986 | EmitCallsiteMetadata(token_pos, deopt_id, kind, locs); |
987 | } |
988 | |
989 | void FlowGraphCompiler::GenerateStaticDartCall(intptr_t deopt_id, |
990 | TokenPosition token_pos, |
991 | PcDescriptorsLayout::Kind kind, |
992 | LocationSummary* locs, |
993 | const Function& target, |
994 | Code::EntryKind entry_kind) { |
995 | ASSERT(CanCallDart()); |
996 | if (CanPcRelativeCall(target)) { |
997 | __ GenerateUnRelocatedPcRelativeCall(); |
998 | AddPcRelativeCallTarget(target, entry_kind); |
999 | EmitCallsiteMetadata(token_pos, deopt_id, kind, locs); |
1000 | } else { |
1001 | // Call sites to the same target can share object pool entries. These |
1002 | // call sites are never patched for breakpoints: the function is deoptimized |
1003 | // and the unoptimized code with IC calls for static calls is patched |
1004 | // instead. |
1005 | ASSERT(is_optimizing()); |
1006 | const auto& stub = StubCode::CallStaticFunction(); |
1007 | __ BranchLinkWithEquivalence(stub, target, entry_kind); |
1008 | EmitCallsiteMetadata(token_pos, deopt_id, kind, locs); |
1009 | AddStaticCallTarget(target, entry_kind); |
1010 | } |
1011 | } |
1012 | |
1013 | void FlowGraphCompiler::GenerateRuntimeCall(TokenPosition token_pos, |
1014 | intptr_t deopt_id, |
1015 | const RuntimeEntry& entry, |
1016 | intptr_t argument_count, |
1017 | LocationSummary* locs) { |
1018 | __ CallRuntime(entry, argument_count); |
1019 | EmitCallsiteMetadata(token_pos, deopt_id, PcDescriptorsLayout::kOther, locs); |
1020 | } |
1021 | |
1022 | void FlowGraphCompiler::EmitEdgeCounter(intptr_t edge_id) { |
1023 | // We do not check for overflow when incrementing the edge counter. The |
1024 | // function should normally be optimized long before the counter can |
1025 | // overflow; and though we do not reset the counters when we optimize or |
1026 | // deoptimize, there is a bound on the number of |
1027 | // optimization/deoptimization cycles we will attempt. |
1028 | ASSERT(!edge_counters_array_.IsNull()); |
1029 | ASSERT(assembler_->constant_pool_allowed()); |
1030 | __ Comment("Edge counter" ); |
1031 | __ LoadObject(R0, edge_counters_array_); |
1032 | __ LoadFieldFromOffset(TMP, R0, Array::element_offset(edge_id)); |
1033 | __ add(TMP, TMP, compiler::Operand(Smi::RawValue(1))); |
1034 | __ StoreFieldToOffset(TMP, R0, Array::element_offset(edge_id)); |
1035 | } |
1036 | |
1037 | void FlowGraphCompiler::EmitOptimizedInstanceCall(const Code& stub, |
1038 | const ICData& ic_data, |
1039 | intptr_t deopt_id, |
1040 | TokenPosition token_pos, |
1041 | LocationSummary* locs, |
1042 | Code::EntryKind entry_kind) { |
1043 | ASSERT(CanCallDart()); |
1044 | ASSERT(Array::Handle(zone(), ic_data.arguments_descriptor()).Length() > 0); |
1045 | // Each ICData propagated from unoptimized to optimized code contains the |
1046 | // function that corresponds to the Dart function of that IC call. Due |
1047 | // to inlining in optimized code, that function may not correspond to the |
1048 | // top-level function (parsed_function().function()) which could be |
1049 | // reoptimized and which counter needs to be incremented. |
1050 | // Pass the function explicitly, it is used in IC stub. |
1051 | |
1052 | __ LoadObject(R6, parsed_function().function()); |
1053 | __ LoadFromOffset(R0, SP, (ic_data.SizeWithoutTypeArgs() - 1) * kWordSize); |
1054 | __ LoadUniqueObject(R5, ic_data); |
1055 | GenerateDartCall(deopt_id, token_pos, stub, PcDescriptorsLayout::kIcCall, |
1056 | locs, entry_kind); |
1057 | __ Drop(ic_data.SizeWithTypeArgs()); |
1058 | } |
1059 | |
1060 | void FlowGraphCompiler::EmitInstanceCallJIT(const Code& stub, |
1061 | const ICData& ic_data, |
1062 | intptr_t deopt_id, |
1063 | TokenPosition token_pos, |
1064 | LocationSummary* locs, |
1065 | Code::EntryKind entry_kind) { |
1066 | ASSERT(CanCallDart()); |
1067 | ASSERT(entry_kind == Code::EntryKind::kNormal || |
1068 | entry_kind == Code::EntryKind::kUnchecked); |
1069 | ASSERT(Array::Handle(zone(), ic_data.arguments_descriptor()).Length() > 0); |
1070 | __ LoadFromOffset(R0, SP, (ic_data.SizeWithoutTypeArgs() - 1) * kWordSize); |
1071 | |
1072 | compiler::ObjectPoolBuilder& op = __ object_pool_builder(); |
1073 | const intptr_t ic_data_index = |
1074 | op.AddObject(ic_data, ObjectPool::Patchability::kPatchable); |
1075 | const intptr_t stub_index = |
1076 | op.AddObject(stub, ObjectPool::Patchability::kPatchable); |
1077 | ASSERT((ic_data_index + 1) == stub_index); |
1078 | __ LoadDoubleWordFromPoolOffset(R5, CODE_REG, |
1079 | ObjectPool::element_offset(ic_data_index)); |
1080 | const intptr_t entry_point_offset = |
1081 | entry_kind == Code::EntryKind::kNormal |
1082 | ? Code::entry_point_offset(Code::EntryKind::kMonomorphic) |
1083 | : Code::entry_point_offset(Code::EntryKind::kMonomorphicUnchecked); |
1084 | __ ldr(LR, compiler::FieldAddress(CODE_REG, entry_point_offset)); |
1085 | __ blr(LR); |
1086 | EmitCallsiteMetadata(token_pos, deopt_id, PcDescriptorsLayout::kIcCall, locs); |
1087 | __ Drop(ic_data.SizeWithTypeArgs()); |
1088 | } |
1089 | |
1090 | void FlowGraphCompiler::EmitMegamorphicInstanceCall( |
1091 | const String& name, |
1092 | const Array& arguments_descriptor, |
1093 | intptr_t deopt_id, |
1094 | TokenPosition token_pos, |
1095 | LocationSummary* locs, |
1096 | intptr_t try_index, |
1097 | intptr_t slow_path_argument_count) { |
1098 | ASSERT(CanCallDart()); |
1099 | ASSERT(!arguments_descriptor.IsNull() && (arguments_descriptor.Length() > 0)); |
1100 | const ArgumentsDescriptor args_desc(arguments_descriptor); |
1101 | const MegamorphicCache& cache = MegamorphicCache::ZoneHandle( |
1102 | zone(), |
1103 | MegamorphicCacheTable::Lookup(thread(), name, arguments_descriptor)); |
1104 | |
1105 | __ Comment("MegamorphicCall" ); |
1106 | // Load receiver into R0. |
1107 | __ LoadFromOffset(R0, SP, (args_desc.Count() - 1) * kWordSize); |
1108 | |
1109 | // Use same code pattern as instance call so it can be parsed by code patcher. |
1110 | compiler::ObjectPoolBuilder& op = __ object_pool_builder(); |
1111 | const intptr_t data_index = |
1112 | op.AddObject(cache, ObjectPool::Patchability::kPatchable); |
1113 | const intptr_t stub_index = op.AddObject( |
1114 | StubCode::MegamorphicCall(), ObjectPool::Patchability::kPatchable); |
1115 | ASSERT((data_index + 1) == stub_index); |
1116 | if (FLAG_precompiled_mode && FLAG_use_bare_instructions) { |
1117 | // The AOT runtime will replace the slot in the object pool with the |
1118 | // entrypoint address - see clustered_snapshot.cc. |
1119 | __ LoadDoubleWordFromPoolOffset(R5, LR, |
1120 | ObjectPool::element_offset(data_index)); |
1121 | } else { |
1122 | __ LoadDoubleWordFromPoolOffset(R5, CODE_REG, |
1123 | ObjectPool::element_offset(data_index)); |
1124 | __ ldr(LR, compiler::FieldAddress( |
1125 | CODE_REG, |
1126 | Code::entry_point_offset(Code::EntryKind::kMonomorphic))); |
1127 | } |
1128 | __ blr(LR); |
1129 | |
1130 | RecordSafepoint(locs, slow_path_argument_count); |
1131 | const intptr_t deopt_id_after = DeoptId::ToDeoptAfter(deopt_id); |
1132 | if (FLAG_precompiled_mode) { |
1133 | // Megamorphic calls may occur in slow path stubs. |
1134 | // If valid use try_index argument. |
1135 | if (try_index == kInvalidTryIndex) { |
1136 | try_index = CurrentTryIndex(); |
1137 | } |
1138 | AddDescriptor(PcDescriptorsLayout::kOther, assembler()->CodeSize(), |
1139 | DeoptId::kNone, token_pos, try_index); |
1140 | } else if (is_optimizing()) { |
1141 | AddCurrentDescriptor(PcDescriptorsLayout::kOther, DeoptId::kNone, |
1142 | token_pos); |
1143 | AddDeoptIndexAtCall(deopt_id_after); |
1144 | } else { |
1145 | AddCurrentDescriptor(PcDescriptorsLayout::kOther, DeoptId::kNone, |
1146 | token_pos); |
1147 | // Add deoptimization continuation point after the call and before the |
1148 | // arguments are removed. |
1149 | AddCurrentDescriptor(PcDescriptorsLayout::kDeopt, deopt_id_after, |
1150 | token_pos); |
1151 | } |
1152 | RecordCatchEntryMoves(pending_deoptimization_env_, try_index); |
1153 | __ Drop(args_desc.SizeWithTypeArgs()); |
1154 | } |
1155 | |
1156 | void FlowGraphCompiler::EmitInstanceCallAOT(const ICData& ic_data, |
1157 | intptr_t deopt_id, |
1158 | TokenPosition token_pos, |
1159 | LocationSummary* locs, |
1160 | Code::EntryKind entry_kind, |
1161 | bool receiver_can_be_smi) { |
1162 | ASSERT(CanCallDart()); |
1163 | ASSERT(ic_data.NumArgsTested() == 1); |
1164 | const Code& initial_stub = StubCode::SwitchableCallMiss(); |
1165 | const char* switchable_call_mode = "smiable" ; |
1166 | if (!receiver_can_be_smi) { |
1167 | switchable_call_mode = "non-smi" ; |
1168 | ic_data.set_receiver_cannot_be_smi(true); |
1169 | } |
1170 | const UnlinkedCall& data = |
1171 | UnlinkedCall::ZoneHandle(zone(), ic_data.AsUnlinkedCall()); |
1172 | |
1173 | compiler::ObjectPoolBuilder& op = __ object_pool_builder(); |
1174 | |
1175 | __ Comment("InstanceCallAOT (%s)" , switchable_call_mode); |
1176 | // Clear argument descriptor to keep gc happy when it gets pushed on to |
1177 | // the stack. |
1178 | __ LoadImmediate(R4, 0); |
1179 | __ LoadFromOffset(R0, SP, (ic_data.SizeWithoutTypeArgs() - 1) * kWordSize); |
1180 | |
1181 | const intptr_t data_index = |
1182 | op.AddObject(data, ObjectPool::Patchability::kPatchable); |
1183 | const intptr_t initial_stub_index = |
1184 | op.AddObject(initial_stub, ObjectPool::Patchability::kPatchable); |
1185 | ASSERT((data_index + 1) == initial_stub_index); |
1186 | |
1187 | if (FLAG_precompiled_mode && FLAG_use_bare_instructions) { |
1188 | // The AOT runtime will replace the slot in the object pool with the |
1189 | // entrypoint address - see clustered_snapshot.cc. |
1190 | __ LoadDoubleWordFromPoolOffset(R5, LR, |
1191 | ObjectPool::element_offset(data_index)); |
1192 | } else { |
1193 | __ LoadDoubleWordFromPoolOffset(R5, CODE_REG, |
1194 | ObjectPool::element_offset(data_index)); |
1195 | const intptr_t entry_point_offset = |
1196 | entry_kind == Code::EntryKind::kNormal |
1197 | ? compiler::target::Code::entry_point_offset( |
1198 | Code::EntryKind::kMonomorphic) |
1199 | : compiler::target::Code::entry_point_offset( |
1200 | Code::EntryKind::kMonomorphicUnchecked); |
1201 | __ ldr(LR, compiler::FieldAddress(CODE_REG, entry_point_offset)); |
1202 | } |
1203 | __ blr(LR); |
1204 | |
1205 | EmitCallsiteMetadata(token_pos, DeoptId::kNone, PcDescriptorsLayout::kOther, |
1206 | locs); |
1207 | __ Drop(ic_data.SizeWithTypeArgs()); |
1208 | } |
1209 | |
1210 | void FlowGraphCompiler::EmitUnoptimizedStaticCall(intptr_t size_with_type_args, |
1211 | intptr_t deopt_id, |
1212 | TokenPosition token_pos, |
1213 | LocationSummary* locs, |
1214 | const ICData& ic_data, |
1215 | Code::EntryKind entry_kind) { |
1216 | ASSERT(CanCallDart()); |
1217 | const Code& stub = |
1218 | StubCode::UnoptimizedStaticCallEntry(ic_data.NumArgsTested()); |
1219 | __ LoadObject(R5, ic_data); |
1220 | GenerateDartCall(deopt_id, token_pos, stub, |
1221 | PcDescriptorsLayout::kUnoptStaticCall, locs, entry_kind); |
1222 | __ Drop(size_with_type_args); |
1223 | } |
1224 | |
1225 | void FlowGraphCompiler::EmitOptimizedStaticCall( |
1226 | const Function& function, |
1227 | const Array& arguments_descriptor, |
1228 | intptr_t size_with_type_args, |
1229 | intptr_t deopt_id, |
1230 | TokenPosition token_pos, |
1231 | LocationSummary* locs, |
1232 | Code::EntryKind entry_kind) { |
1233 | ASSERT(CanCallDart()); |
1234 | ASSERT(!function.IsClosureFunction()); |
1235 | if (function.HasOptionalParameters() || function.IsGeneric()) { |
1236 | __ LoadObject(R4, arguments_descriptor); |
1237 | } else { |
1238 | if (!(FLAG_precompiled_mode && FLAG_use_bare_instructions)) { |
1239 | __ LoadImmediate(R4, 0); // GC safe smi zero because of stub. |
1240 | } |
1241 | } |
1242 | // Do not use the code from the function, but let the code be patched so that |
1243 | // we can record the outgoing edges to other code. |
1244 | GenerateStaticDartCall(deopt_id, token_pos, PcDescriptorsLayout::kOther, locs, |
1245 | function, entry_kind); |
1246 | __ Drop(size_with_type_args); |
1247 | } |
1248 | |
1249 | void FlowGraphCompiler::EmitDispatchTableCall( |
1250 | Register cid_reg, |
1251 | int32_t selector_offset, |
1252 | const Array& arguments_descriptor) { |
1253 | ASSERT(CanCallDart()); |
1254 | ASSERT(cid_reg != ARGS_DESC_REG); |
1255 | if (!arguments_descriptor.IsNull()) { |
1256 | __ LoadObject(ARGS_DESC_REG, arguments_descriptor); |
1257 | } |
1258 | const intptr_t offset = selector_offset - DispatchTable::OriginElement(); |
1259 | __ AddImmediate(cid_reg, cid_reg, offset); |
1260 | __ ldr(LR, compiler::Address(DISPATCH_TABLE_REG, cid_reg, UXTX, |
1261 | compiler::Address::Scaled)); |
1262 | __ blr(LR); |
1263 | } |
1264 | |
1265 | Condition FlowGraphCompiler::EmitEqualityRegConstCompare( |
1266 | Register reg, |
1267 | const Object& obj, |
1268 | bool needs_number_check, |
1269 | TokenPosition token_pos, |
1270 | intptr_t deopt_id) { |
1271 | if (needs_number_check) { |
1272 | ASSERT(!obj.IsMint() && !obj.IsDouble()); |
1273 | __ LoadObject(TMP, obj); |
1274 | __ PushPair(TMP, reg); |
1275 | if (is_optimizing()) { |
1276 | __ BranchLinkPatchable(StubCode::OptimizedIdenticalWithNumberCheck()); |
1277 | } else { |
1278 | __ BranchLinkPatchable(StubCode::UnoptimizedIdenticalWithNumberCheck()); |
1279 | } |
1280 | AddCurrentDescriptor(PcDescriptorsLayout::kRuntimeCall, deopt_id, |
1281 | token_pos); |
1282 | // Stub returns result in flags (result of a cmp, we need Z computed). |
1283 | // Discard constant. |
1284 | // Restore 'reg'. |
1285 | __ PopPair(ZR, reg); |
1286 | } else { |
1287 | __ CompareObject(reg, obj); |
1288 | } |
1289 | return EQ; |
1290 | } |
1291 | |
1292 | Condition FlowGraphCompiler::EmitEqualityRegRegCompare(Register left, |
1293 | Register right, |
1294 | bool needs_number_check, |
1295 | TokenPosition token_pos, |
1296 | intptr_t deopt_id) { |
1297 | if (needs_number_check) { |
1298 | __ PushPair(right, left); |
1299 | if (is_optimizing()) { |
1300 | __ BranchLinkPatchable(StubCode::OptimizedIdenticalWithNumberCheck()); |
1301 | } else { |
1302 | __ BranchLinkPatchable(StubCode::UnoptimizedIdenticalWithNumberCheck()); |
1303 | } |
1304 | AddCurrentDescriptor(PcDescriptorsLayout::kRuntimeCall, deopt_id, |
1305 | token_pos); |
1306 | // Stub returns result in flags (result of a cmp, we need Z computed). |
1307 | __ PopPair(right, left); |
1308 | } else { |
1309 | __ CompareRegisters(left, right); |
1310 | } |
1311 | return EQ; |
1312 | } |
1313 | |
1314 | Condition FlowGraphCompiler::EmitBoolTest(Register value, |
1315 | BranchLabels labels, |
1316 | bool invert) { |
1317 | __ Comment("BoolTest" ); |
1318 | if (labels.true_label == nullptr || labels.false_label == nullptr) { |
1319 | __ tsti(value, compiler::Immediate( |
1320 | compiler::target::ObjectAlignment::kBoolValueMask)); |
1321 | return invert ? NE : EQ; |
1322 | } |
1323 | const intptr_t bool_bit = |
1324 | compiler::target::ObjectAlignment::kBoolValueBitPosition; |
1325 | if (labels.fall_through == labels.false_label) { |
1326 | if (invert) { |
1327 | __ tbnz(labels.true_label, value, bool_bit); |
1328 | } else { |
1329 | __ tbz(labels.true_label, value, bool_bit); |
1330 | } |
1331 | } else { |
1332 | if (invert) { |
1333 | __ tbz(labels.false_label, value, bool_bit); |
1334 | } else { |
1335 | __ tbnz(labels.false_label, value, bool_bit); |
1336 | } |
1337 | if (labels.fall_through != labels.true_label) { |
1338 | __ b(labels.true_label); |
1339 | } |
1340 | } |
1341 | return kInvalidCondition; |
1342 | } |
1343 | |
1344 | // This function must be in sync with FlowGraphCompiler::RecordSafepoint and |
1345 | // FlowGraphCompiler::SlowPathEnvironmentFor. |
1346 | void FlowGraphCompiler::SaveLiveRegisters(LocationSummary* locs) { |
1347 | #if defined(DEBUG) |
1348 | locs->CheckWritableInputs(); |
1349 | ClobberDeadTempRegisters(locs); |
1350 | #endif |
1351 | // TODO(vegorov): consider saving only caller save (volatile) registers. |
1352 | __ PushRegisters(*locs->live_registers()); |
1353 | } |
1354 | |
1355 | void FlowGraphCompiler::RestoreLiveRegisters(LocationSummary* locs) { |
1356 | __ PopRegisters(*locs->live_registers()); |
1357 | } |
1358 | |
1359 | #if defined(DEBUG) |
1360 | void FlowGraphCompiler::ClobberDeadTempRegisters(LocationSummary* locs) { |
1361 | // Clobber temporaries that have not been manually preserved. |
1362 | for (intptr_t i = 0; i < locs->temp_count(); ++i) { |
1363 | Location tmp = locs->temp(i); |
1364 | // TODO(zerny): clobber non-live temporary FPU registers. |
1365 | if (tmp.IsRegister() && |
1366 | !locs->live_registers()->ContainsRegister(tmp.reg())) { |
1367 | __ movz(tmp.reg(), compiler::Immediate(0xf7), 0); |
1368 | } |
1369 | } |
1370 | } |
1371 | #endif |
1372 | |
1373 | Register FlowGraphCompiler::EmitTestCidRegister() { |
1374 | return R2; |
1375 | } |
1376 | |
1377 | void FlowGraphCompiler::EmitTestAndCallLoadReceiver( |
1378 | intptr_t count_without_type_args, |
1379 | const Array& arguments_descriptor) { |
1380 | __ Comment("EmitTestAndCall" ); |
1381 | // Load receiver into R0. |
1382 | __ LoadFromOffset(R0, SP, (count_without_type_args - 1) * kWordSize); |
1383 | __ LoadObject(R4, arguments_descriptor); |
1384 | } |
1385 | |
1386 | void FlowGraphCompiler::EmitTestAndCallSmiBranch(compiler::Label* label, |
1387 | bool if_smi) { |
1388 | if (if_smi) { |
1389 | __ BranchIfSmi(R0, label); |
1390 | } else { |
1391 | __ BranchIfNotSmi(R0, label); |
1392 | } |
1393 | } |
1394 | |
1395 | void FlowGraphCompiler::EmitTestAndCallLoadCid(Register class_id_reg) { |
1396 | ASSERT(class_id_reg != R0); |
1397 | __ LoadClassId(class_id_reg, R0); |
1398 | } |
1399 | |
1400 | #undef __ |
1401 | #define __ assembler-> |
1402 | |
1403 | int FlowGraphCompiler::EmitTestAndCallCheckCid(compiler::Assembler* assembler, |
1404 | compiler::Label* label, |
1405 | Register class_id_reg, |
1406 | const CidRangeValue& range, |
1407 | int bias, |
1408 | bool jump_on_miss) { |
1409 | const intptr_t cid_start = range.cid_start; |
1410 | if (range.IsSingleCid()) { |
1411 | __ AddImmediateSetFlags(class_id_reg, class_id_reg, bias - cid_start); |
1412 | __ BranchIf(jump_on_miss ? NOT_EQUAL : EQUAL, label); |
1413 | bias = cid_start; |
1414 | } else { |
1415 | __ AddImmediate(class_id_reg, bias - cid_start); |
1416 | bias = cid_start; |
1417 | __ CompareImmediate(class_id_reg, range.Extent()); |
1418 | __ BranchIf(jump_on_miss ? UNSIGNED_GREATER : UNSIGNED_LESS_EQUAL, label); |
1419 | } |
1420 | return bias; |
1421 | } |
1422 | |
1423 | #undef __ |
1424 | #define __ assembler()-> |
1425 | |
1426 | void FlowGraphCompiler::EmitMove(Location destination, |
1427 | Location source, |
1428 | TemporaryRegisterAllocator* allocator) { |
1429 | if (destination.Equals(source)) return; |
1430 | |
1431 | if (source.IsRegister()) { |
1432 | if (destination.IsRegister()) { |
1433 | __ mov(destination.reg(), source.reg()); |
1434 | } else { |
1435 | ASSERT(destination.IsStackSlot()); |
1436 | const intptr_t dest_offset = destination.ToStackSlotOffset(); |
1437 | __ StoreToOffset(source.reg(), destination.base_reg(), dest_offset); |
1438 | } |
1439 | } else if (source.IsStackSlot()) { |
1440 | if (destination.IsRegister()) { |
1441 | const intptr_t source_offset = source.ToStackSlotOffset(); |
1442 | __ LoadFromOffset(destination.reg(), source.base_reg(), source_offset); |
1443 | } else if (destination.IsFpuRegister()) { |
1444 | const intptr_t src_offset = source.ToStackSlotOffset(); |
1445 | VRegister dst = destination.fpu_reg(); |
1446 | __ LoadDFromOffset(dst, source.base_reg(), src_offset); |
1447 | } else { |
1448 | ASSERT(destination.IsStackSlot()); |
1449 | const intptr_t source_offset = source.ToStackSlotOffset(); |
1450 | const intptr_t dest_offset = destination.ToStackSlotOffset(); |
1451 | Register tmp = allocator->AllocateTemporary(); |
1452 | __ LoadFromOffset(tmp, source.base_reg(), source_offset); |
1453 | __ StoreToOffset(tmp, destination.base_reg(), dest_offset); |
1454 | allocator->ReleaseTemporary(); |
1455 | } |
1456 | } else if (source.IsFpuRegister()) { |
1457 | if (destination.IsFpuRegister()) { |
1458 | __ vmov(destination.fpu_reg(), source.fpu_reg()); |
1459 | } else { |
1460 | if (destination.IsStackSlot() /*32-bit float*/ || |
1461 | destination.IsDoubleStackSlot()) { |
1462 | const intptr_t dest_offset = destination.ToStackSlotOffset(); |
1463 | VRegister src = source.fpu_reg(); |
1464 | __ StoreDToOffset(src, destination.base_reg(), dest_offset); |
1465 | } else { |
1466 | ASSERT(destination.IsQuadStackSlot()); |
1467 | const intptr_t dest_offset = destination.ToStackSlotOffset(); |
1468 | __ StoreQToOffset(source.fpu_reg(), destination.base_reg(), |
1469 | dest_offset); |
1470 | } |
1471 | } |
1472 | } else if (source.IsDoubleStackSlot()) { |
1473 | if (destination.IsFpuRegister()) { |
1474 | const intptr_t source_offset = source.ToStackSlotOffset(); |
1475 | const VRegister dst = destination.fpu_reg(); |
1476 | __ LoadDFromOffset(dst, source.base_reg(), source_offset); |
1477 | } else { |
1478 | ASSERT(destination.IsDoubleStackSlot() || |
1479 | destination.IsStackSlot() /*32-bit float*/); |
1480 | const intptr_t source_offset = source.ToStackSlotOffset(); |
1481 | const intptr_t dest_offset = destination.ToStackSlotOffset(); |
1482 | __ LoadDFromOffset(VTMP, source.base_reg(), source_offset); |
1483 | __ StoreDToOffset(VTMP, destination.base_reg(), dest_offset); |
1484 | } |
1485 | } else if (source.IsQuadStackSlot()) { |
1486 | if (destination.IsFpuRegister()) { |
1487 | const intptr_t source_offset = source.ToStackSlotOffset(); |
1488 | __ LoadQFromOffset(destination.fpu_reg(), source.base_reg(), |
1489 | source_offset); |
1490 | } else { |
1491 | ASSERT(destination.IsQuadStackSlot()); |
1492 | const intptr_t source_offset = source.ToStackSlotOffset(); |
1493 | const intptr_t dest_offset = destination.ToStackSlotOffset(); |
1494 | __ LoadQFromOffset(VTMP, source.base_reg(), source_offset); |
1495 | __ StoreQToOffset(VTMP, destination.base_reg(), dest_offset); |
1496 | } |
1497 | } else { |
1498 | ASSERT(source.IsConstant()); |
1499 | if (destination.IsStackSlot()) { |
1500 | Register tmp = allocator->AllocateTemporary(); |
1501 | source.constant_instruction()->EmitMoveToLocation(this, destination, tmp); |
1502 | allocator->ReleaseTemporary(); |
1503 | } else { |
1504 | source.constant_instruction()->EmitMoveToLocation(this, destination); |
1505 | } |
1506 | } |
1507 | } |
1508 | |
1509 | static OperandSize BytesToOperandSize(intptr_t bytes) { |
1510 | switch (bytes) { |
1511 | case 8: |
1512 | return OperandSize::kDoubleWord; |
1513 | case 4: |
1514 | return OperandSize::kWord; |
1515 | case 2: |
1516 | return OperandSize::kHalfword; |
1517 | case 1: |
1518 | return OperandSize::kByte; |
1519 | default: |
1520 | UNIMPLEMENTED(); |
1521 | } |
1522 | } |
1523 | |
1524 | void FlowGraphCompiler::EmitNativeMoveArchitecture( |
1525 | const compiler::ffi::NativeLocation& destination, |
1526 | const compiler::ffi::NativeLocation& source) { |
1527 | const auto& src_type = source.payload_type(); |
1528 | const auto& dst_type = destination.payload_type(); |
1529 | ASSERT(src_type.IsFloat() == dst_type.IsFloat()); |
1530 | ASSERT(src_type.IsInt() == dst_type.IsInt()); |
1531 | ASSERT(src_type.IsSigned() == dst_type.IsSigned()); |
1532 | ASSERT(src_type.IsFundamental()); |
1533 | ASSERT(dst_type.IsFundamental()); |
1534 | const intptr_t src_size = src_type.SizeInBytes(); |
1535 | const intptr_t dst_size = dst_type.SizeInBytes(); |
1536 | const bool sign_or_zero_extend = dst_size > src_size; |
1537 | |
1538 | if (source.IsRegisters()) { |
1539 | const auto& src = source.AsRegisters(); |
1540 | ASSERT(src.num_regs() == 1); |
1541 | const auto src_reg = src.reg_at(0); |
1542 | |
1543 | if (destination.IsRegisters()) { |
1544 | const auto& dst = destination.AsRegisters(); |
1545 | ASSERT(dst.num_regs() == 1); |
1546 | const auto dst_reg = dst.reg_at(0); |
1547 | if (!sign_or_zero_extend) { |
1548 | switch (dst_size) { |
1549 | case 8: |
1550 | __ mov(dst_reg, src_reg); |
1551 | return; |
1552 | case 4: |
1553 | __ movw(dst_reg, src_reg); |
1554 | return; |
1555 | default: |
1556 | UNIMPLEMENTED(); |
1557 | } |
1558 | } else { |
1559 | switch (src_type.AsFundamental().representation()) { |
1560 | case compiler::ffi::kInt8: // Sign extend operand. |
1561 | __ sxtb(dst_reg, src_reg); |
1562 | return; |
1563 | case compiler::ffi::kInt16: |
1564 | __ sxth(dst_reg, src_reg); |
1565 | return; |
1566 | case compiler::ffi::kUint8: // Zero extend operand. |
1567 | __ uxtb(dst_reg, src_reg); |
1568 | return; |
1569 | case compiler::ffi::kUint16: |
1570 | __ uxth(dst_reg, src_reg); |
1571 | return; |
1572 | default: |
1573 | // 32 to 64 bit is covered in IL by Representation conversions. |
1574 | UNIMPLEMENTED(); |
1575 | } |
1576 | } |
1577 | |
1578 | } else if (destination.IsFpuRegisters()) { |
1579 | // Fpu Registers should only contain doubles and registers only ints. |
1580 | UNIMPLEMENTED(); |
1581 | |
1582 | } else { |
1583 | ASSERT(destination.IsStack()); |
1584 | const auto& dst = destination.AsStack(); |
1585 | ASSERT(!sign_or_zero_extend); |
1586 | const OperandSize op_size = BytesToOperandSize(dst_size); |
1587 | __ StoreToOffset(src.reg_at(0), dst.base_register(), |
1588 | dst.offset_in_bytes(), op_size); |
1589 | } |
1590 | |
1591 | } else if (source.IsFpuRegisters()) { |
1592 | const auto& src = source.AsFpuRegisters(); |
1593 | // We have not implemented conversions here, use IL convert instructions. |
1594 | ASSERT(src_type.Equals(dst_type)); |
1595 | |
1596 | if (destination.IsRegisters()) { |
1597 | // Fpu Registers should only contain doubles and registers only ints. |
1598 | UNIMPLEMENTED(); |
1599 | |
1600 | } else if (destination.IsFpuRegisters()) { |
1601 | const auto& dst = destination.AsFpuRegisters(); |
1602 | __ vmov(dst.fpu_reg(), src.fpu_reg()); |
1603 | |
1604 | } else { |
1605 | ASSERT(destination.IsStack()); |
1606 | ASSERT(src_type.IsFloat()); |
1607 | const auto& dst = destination.AsStack(); |
1608 | switch (dst_size) { |
1609 | case 8: |
1610 | __ StoreDToOffset(src.fpu_reg(), dst.base_register(), |
1611 | dst.offset_in_bytes()); |
1612 | return; |
1613 | case 4: |
1614 | __ StoreSToOffset(src.fpu_reg(), dst.base_register(), |
1615 | dst.offset_in_bytes()); |
1616 | return; |
1617 | default: |
1618 | UNREACHABLE(); |
1619 | } |
1620 | } |
1621 | |
1622 | } else { |
1623 | ASSERT(source.IsStack()); |
1624 | const auto& src = source.AsStack(); |
1625 | if (destination.IsRegisters()) { |
1626 | const auto& dst = destination.AsRegisters(); |
1627 | ASSERT(dst.num_regs() == 1); |
1628 | const auto dst_reg = dst.reg_at(0); |
1629 | ASSERT(!sign_or_zero_extend); |
1630 | const OperandSize op_size = BytesToOperandSize(dst_size); |
1631 | __ LoadFromOffset(dst_reg, src.base_register(), src.offset_in_bytes(), |
1632 | op_size); |
1633 | |
1634 | } else if (destination.IsFpuRegisters()) { |
1635 | ASSERT(src_type.Equals(dst_type)); |
1636 | ASSERT(src_type.IsFloat()); |
1637 | const auto& dst = destination.AsFpuRegisters(); |
1638 | switch (src_size) { |
1639 | case 8: |
1640 | __ LoadDFromOffset(dst.fpu_reg(), src.base_register(), |
1641 | src.offset_in_bytes()); |
1642 | return; |
1643 | case 4: |
1644 | __ LoadSFromOffset(dst.fpu_reg(), src.base_register(), |
1645 | src.offset_in_bytes()); |
1646 | return; |
1647 | default: |
1648 | UNIMPLEMENTED(); |
1649 | } |
1650 | |
1651 | } else { |
1652 | ASSERT(destination.IsStack()); |
1653 | UNREACHABLE(); |
1654 | } |
1655 | } |
1656 | } |
1657 | |
1658 | void FlowGraphCompiler::LoadBSSEntry(BSS::Relocation relocation, |
1659 | Register dst, |
1660 | Register tmp) { |
1661 | compiler::Label skip_reloc; |
1662 | __ b(&skip_reloc); |
1663 | InsertBSSRelocation(relocation); |
1664 | __ Bind(&skip_reloc); |
1665 | |
1666 | __ adr(tmp, compiler::Immediate(-compiler::target::kWordSize)); |
1667 | |
1668 | // tmp holds the address of the relocation. |
1669 | __ ldr(dst, compiler::Address(tmp)); |
1670 | |
1671 | // dst holds the relocation itself: tmp - bss_start. |
1672 | // tmp = tmp + (bss_start - tmp) = bss_start |
1673 | __ add(tmp, tmp, compiler::Operand(dst)); |
1674 | |
1675 | // tmp holds the start of the BSS section. |
1676 | // Load the "get-thread" routine: *bss_start. |
1677 | __ ldr(dst, compiler::Address(tmp)); |
1678 | } |
1679 | |
1680 | #undef __ |
1681 | #define __ compiler_->assembler()-> |
1682 | |
1683 | void ParallelMoveResolver::EmitSwap(int index) { |
1684 | MoveOperands* move = moves_[index]; |
1685 | const Location source = move->src(); |
1686 | const Location destination = move->dest(); |
1687 | |
1688 | if (source.IsRegister() && destination.IsRegister()) { |
1689 | ASSERT(source.reg() != TMP); |
1690 | ASSERT(destination.reg() != TMP); |
1691 | __ mov(TMP, source.reg()); |
1692 | __ mov(source.reg(), destination.reg()); |
1693 | __ mov(destination.reg(), TMP); |
1694 | } else if (source.IsRegister() && destination.IsStackSlot()) { |
1695 | Exchange(source.reg(), destination.base_reg(), |
1696 | destination.ToStackSlotOffset()); |
1697 | } else if (source.IsStackSlot() && destination.IsRegister()) { |
1698 | Exchange(destination.reg(), source.base_reg(), source.ToStackSlotOffset()); |
1699 | } else if (source.IsStackSlot() && destination.IsStackSlot()) { |
1700 | Exchange(source.base_reg(), source.ToStackSlotOffset(), |
1701 | destination.base_reg(), destination.ToStackSlotOffset()); |
1702 | } else if (source.IsFpuRegister() && destination.IsFpuRegister()) { |
1703 | const VRegister dst = destination.fpu_reg(); |
1704 | const VRegister src = source.fpu_reg(); |
1705 | __ vmov(VTMP, src); |
1706 | __ vmov(src, dst); |
1707 | __ vmov(dst, VTMP); |
1708 | } else if (source.IsFpuRegister() || destination.IsFpuRegister()) { |
1709 | ASSERT(destination.IsDoubleStackSlot() || destination.IsQuadStackSlot() || |
1710 | source.IsDoubleStackSlot() || source.IsQuadStackSlot()); |
1711 | bool double_width = |
1712 | destination.IsDoubleStackSlot() || source.IsDoubleStackSlot(); |
1713 | VRegister reg = |
1714 | source.IsFpuRegister() ? source.fpu_reg() : destination.fpu_reg(); |
1715 | Register base_reg = |
1716 | source.IsFpuRegister() ? destination.base_reg() : source.base_reg(); |
1717 | const intptr_t slot_offset = source.IsFpuRegister() |
1718 | ? destination.ToStackSlotOffset() |
1719 | : source.ToStackSlotOffset(); |
1720 | |
1721 | if (double_width) { |
1722 | __ LoadDFromOffset(VTMP, base_reg, slot_offset); |
1723 | __ StoreDToOffset(reg, base_reg, slot_offset); |
1724 | __ fmovdd(reg, VTMP); |
1725 | } else { |
1726 | __ LoadQFromOffset(VTMP, base_reg, slot_offset); |
1727 | __ StoreQToOffset(reg, base_reg, slot_offset); |
1728 | __ vmov(reg, VTMP); |
1729 | } |
1730 | } else if (source.IsDoubleStackSlot() && destination.IsDoubleStackSlot()) { |
1731 | const intptr_t source_offset = source.ToStackSlotOffset(); |
1732 | const intptr_t dest_offset = destination.ToStackSlotOffset(); |
1733 | |
1734 | ScratchFpuRegisterScope ensure_scratch(this, kNoFpuRegister); |
1735 | VRegister scratch = ensure_scratch.reg(); |
1736 | __ LoadDFromOffset(VTMP, source.base_reg(), source_offset); |
1737 | __ LoadDFromOffset(scratch, destination.base_reg(), dest_offset); |
1738 | __ StoreDToOffset(VTMP, destination.base_reg(), dest_offset); |
1739 | __ StoreDToOffset(scratch, source.base_reg(), source_offset); |
1740 | } else if (source.IsQuadStackSlot() && destination.IsQuadStackSlot()) { |
1741 | const intptr_t source_offset = source.ToStackSlotOffset(); |
1742 | const intptr_t dest_offset = destination.ToStackSlotOffset(); |
1743 | |
1744 | ScratchFpuRegisterScope ensure_scratch(this, kNoFpuRegister); |
1745 | VRegister scratch = ensure_scratch.reg(); |
1746 | __ LoadQFromOffset(VTMP, source.base_reg(), source_offset); |
1747 | __ LoadQFromOffset(scratch, destination.base_reg(), dest_offset); |
1748 | __ StoreQToOffset(VTMP, destination.base_reg(), dest_offset); |
1749 | __ StoreQToOffset(scratch, source.base_reg(), source_offset); |
1750 | } else { |
1751 | UNREACHABLE(); |
1752 | } |
1753 | |
1754 | // The swap of source and destination has executed a move from source to |
1755 | // destination. |
1756 | move->Eliminate(); |
1757 | |
1758 | // Any unperformed (including pending) move with a source of either |
1759 | // this move's source or destination needs to have their source |
1760 | // changed to reflect the state of affairs after the swap. |
1761 | for (int i = 0; i < moves_.length(); ++i) { |
1762 | const MoveOperands& other_move = *moves_[i]; |
1763 | if (other_move.Blocks(source)) { |
1764 | moves_[i]->set_src(destination); |
1765 | } else if (other_move.Blocks(destination)) { |
1766 | moves_[i]->set_src(source); |
1767 | } |
1768 | } |
1769 | } |
1770 | |
1771 | void ParallelMoveResolver::MoveMemoryToMemory(const compiler::Address& dst, |
1772 | const compiler::Address& src) { |
1773 | UNREACHABLE(); |
1774 | } |
1775 | |
1776 | // Do not call or implement this function. Instead, use the form below that |
1777 | // uses an offset from the frame pointer instead of an Address. |
1778 | void ParallelMoveResolver::Exchange(Register reg, |
1779 | const compiler::Address& mem) { |
1780 | UNREACHABLE(); |
1781 | } |
1782 | |
1783 | // Do not call or implement this function. Instead, use the form below that |
1784 | // uses offsets from the frame pointer instead of Addresses. |
1785 | void ParallelMoveResolver::Exchange(const compiler::Address& mem1, |
1786 | const compiler::Address& mem2) { |
1787 | UNREACHABLE(); |
1788 | } |
1789 | |
1790 | void ParallelMoveResolver::Exchange(Register reg, |
1791 | Register base_reg, |
1792 | intptr_t stack_offset) { |
1793 | ScratchRegisterScope tmp(this, reg); |
1794 | __ mov(tmp.reg(), reg); |
1795 | __ LoadFromOffset(reg, base_reg, stack_offset); |
1796 | __ StoreToOffset(tmp.reg(), base_reg, stack_offset); |
1797 | } |
1798 | |
1799 | void ParallelMoveResolver::Exchange(Register base_reg1, |
1800 | intptr_t stack_offset1, |
1801 | Register base_reg2, |
1802 | intptr_t stack_offset2) { |
1803 | ScratchRegisterScope tmp1(this, kNoRegister); |
1804 | ScratchRegisterScope tmp2(this, tmp1.reg()); |
1805 | __ LoadFromOffset(tmp1.reg(), base_reg1, stack_offset1); |
1806 | __ LoadFromOffset(tmp2.reg(), base_reg2, stack_offset2); |
1807 | __ StoreToOffset(tmp1.reg(), base_reg2, stack_offset2); |
1808 | __ StoreToOffset(tmp2.reg(), base_reg1, stack_offset1); |
1809 | } |
1810 | |
1811 | void ParallelMoveResolver::SpillScratch(Register reg) { |
1812 | __ Push(reg); |
1813 | } |
1814 | |
1815 | void ParallelMoveResolver::RestoreScratch(Register reg) { |
1816 | __ Pop(reg); |
1817 | } |
1818 | |
1819 | void ParallelMoveResolver::SpillFpuScratch(FpuRegister reg) { |
1820 | __ PushDouble(reg); |
1821 | } |
1822 | |
1823 | void ParallelMoveResolver::RestoreFpuScratch(FpuRegister reg) { |
1824 | __ PopDouble(reg); |
1825 | } |
1826 | |
1827 | #undef __ |
1828 | |
1829 | } // namespace dart |
1830 | |
1831 | #endif // defined(TARGET_ARCH_ARM64) |
1832 | |