1 | // Copyright (c) 2013, 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_X64. |
6 | #if defined(TARGET_ARCH_X64) |
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/ffi/native_location.h" |
14 | #include "vm/compiler/jit/compiler.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 | DEFINE_FLAG(bool, unbox_mints, true, "Optimize 64-bit integer arithmetic." ); |
29 | DECLARE_FLAG(bool, enable_simd_inline); |
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 | ASSERT(!block_info_[i]->jump_label()->HasNear()); |
63 | } |
64 | } |
65 | |
66 | bool FlowGraphCompiler::SupportsUnboxedDoubles() { |
67 | return true; |
68 | } |
69 | |
70 | bool FlowGraphCompiler::SupportsUnboxedInt64() { |
71 | return FLAG_unbox_mints; |
72 | } |
73 | |
74 | bool FlowGraphCompiler::SupportsUnboxedSimd128() { |
75 | return FLAG_enable_simd_inline; |
76 | } |
77 | |
78 | bool FlowGraphCompiler::SupportsHardwareDivision() { |
79 | return true; |
80 | } |
81 | |
82 | bool FlowGraphCompiler::CanConvertInt64ToDouble() { |
83 | return true; |
84 | } |
85 | |
86 | void FlowGraphCompiler::EnterIntrinsicMode() { |
87 | ASSERT(!intrinsic_mode()); |
88 | intrinsic_mode_ = true; |
89 | ASSERT(!assembler()->constant_pool_allowed()); |
90 | } |
91 | |
92 | void FlowGraphCompiler::ExitIntrinsicMode() { |
93 | ASSERT(intrinsic_mode()); |
94 | intrinsic_mode_ = false; |
95 | } |
96 | |
97 | TypedDataPtr CompilerDeoptInfo::CreateDeoptInfo(FlowGraphCompiler* compiler, |
98 | DeoptInfoBuilder* builder, |
99 | const Array& deopt_table) { |
100 | if (deopt_env_ == NULL) { |
101 | ++builder->current_info_number_; |
102 | return TypedData::null(); |
103 | } |
104 | |
105 | intptr_t stack_height = compiler->StackSize(); |
106 | AllocateIncomingParametersRecursive(deopt_env_, &stack_height); |
107 | |
108 | intptr_t slot_ix = 0; |
109 | Environment* current = deopt_env_; |
110 | |
111 | // Emit all kMaterializeObject instructions describing objects to be |
112 | // materialized on the deoptimization as a prefix to the deoptimization info. |
113 | EmitMaterializations(deopt_env_, builder); |
114 | |
115 | // The real frame starts here. |
116 | builder->MarkFrameStart(); |
117 | |
118 | Zone* zone = compiler->zone(); |
119 | |
120 | builder->AddPp(current->function(), slot_ix++); |
121 | builder->AddPcMarker(Function::ZoneHandle(zone), slot_ix++); |
122 | builder->AddCallerFp(slot_ix++); |
123 | builder->AddReturnAddress(current->function(), deopt_id(), slot_ix++); |
124 | |
125 | // Emit all values that are needed for materialization as a part of the |
126 | // expression stack for the bottom-most frame. This guarantees that GC |
127 | // will be able to find them during materialization. |
128 | slot_ix = builder->EmitMaterializationArguments(slot_ix); |
129 | |
130 | // For the innermost environment, set outgoing arguments and the locals. |
131 | for (intptr_t i = current->Length() - 1; |
132 | i >= current->fixed_parameter_count(); i--) { |
133 | builder->AddCopy(current->ValueAt(i), current->LocationAt(i), slot_ix++); |
134 | } |
135 | |
136 | Environment* previous = current; |
137 | current = current->outer(); |
138 | while (current != NULL) { |
139 | builder->AddPp(current->function(), slot_ix++); |
140 | builder->AddPcMarker(previous->function(), slot_ix++); |
141 | builder->AddCallerFp(slot_ix++); |
142 | |
143 | // For any outer environment the deopt id is that of the call instruction |
144 | // which is recorded in the outer environment. |
145 | builder->AddReturnAddress(current->function(), |
146 | DeoptId::ToDeoptAfter(current->deopt_id()), |
147 | slot_ix++); |
148 | |
149 | // The values of outgoing arguments can be changed from the inlined call so |
150 | // we must read them from the previous environment. |
151 | for (intptr_t i = previous->fixed_parameter_count() - 1; i >= 0; i--) { |
152 | builder->AddCopy(previous->ValueAt(i), previous->LocationAt(i), |
153 | slot_ix++); |
154 | } |
155 | |
156 | // Set the locals, note that outgoing arguments are not in the environment. |
157 | for (intptr_t i = current->Length() - 1; |
158 | i >= current->fixed_parameter_count(); i--) { |
159 | builder->AddCopy(current->ValueAt(i), current->LocationAt(i), slot_ix++); |
160 | } |
161 | |
162 | // Iterate on the outer environment. |
163 | previous = current; |
164 | current = current->outer(); |
165 | } |
166 | // The previous pointer is now the outermost environment. |
167 | ASSERT(previous != NULL); |
168 | |
169 | // Set slots for the outermost environment. |
170 | builder->AddCallerPp(slot_ix++); |
171 | builder->AddPcMarker(previous->function(), slot_ix++); |
172 | builder->AddCallerFp(slot_ix++); |
173 | builder->AddCallerPc(slot_ix++); |
174 | |
175 | // For the outermost environment, set the incoming arguments. |
176 | for (intptr_t i = previous->fixed_parameter_count() - 1; i >= 0; i--) { |
177 | builder->AddCopy(previous->ValueAt(i), previous->LocationAt(i), slot_ix++); |
178 | } |
179 | |
180 | return builder->CreateDeoptInfo(deopt_table); |
181 | } |
182 | |
183 | void CompilerDeoptInfoWithStub::GenerateCode(FlowGraphCompiler* compiler, |
184 | intptr_t stub_ix) { |
185 | // Calls do not need stubs, they share a deoptimization trampoline. |
186 | ASSERT(reason() != ICData::kDeoptAtCall); |
187 | compiler::Assembler* assembler = compiler->assembler(); |
188 | #define __ assembler-> |
189 | __ Comment("%s" , Name()); |
190 | __ Bind(entry_label()); |
191 | if (FLAG_trap_on_deoptimization) { |
192 | __ int3(); |
193 | } |
194 | |
195 | ASSERT(deopt_env() != NULL); |
196 | __ call(compiler::Address(THR, Thread::deoptimize_entry_offset())); |
197 | set_pc_offset(assembler->CodeSize()); |
198 | __ int3(); |
199 | #undef __ |
200 | } |
201 | |
202 | #define __ assembler()-> |
203 | |
204 | // Fall through if bool_register contains null. |
205 | void FlowGraphCompiler::GenerateBoolToJump(Register bool_register, |
206 | compiler::Label* is_true, |
207 | compiler::Label* is_false) { |
208 | compiler::Label fall_through; |
209 | __ CompareObject(bool_register, Object::null_object()); |
210 | __ j(EQUAL, &fall_through, compiler::Assembler::kNearJump); |
211 | BranchLabels labels = {is_true, is_false, &fall_through}; |
212 | Condition true_condition = |
213 | EmitBoolTest(bool_register, labels, /*invert=*/false); |
214 | ASSERT(true_condition != kInvalidCondition); |
215 | __ j(true_condition, is_true); |
216 | __ jmp(is_false); |
217 | __ Bind(&fall_through); |
218 | } |
219 | |
220 | // Call stub to perform subtype test using a cache (see |
221 | // stub_code_x64.cc:GenerateSubtypeNTestCacheStub) |
222 | // |
223 | // Inputs: |
224 | // - RAX : instance to test against. |
225 | // - RDX : instantiator type arguments (if necessary). |
226 | // - RCX : function type arguments (if necessary). |
227 | // |
228 | // Preserves RAX/RCX/RDX. |
229 | SubtypeTestCachePtr FlowGraphCompiler::GenerateCallSubtypeTestStub( |
230 | TypeTestStubKind test_kind, |
231 | Register instance_reg, |
232 | Register instantiator_type_arguments_reg, |
233 | Register function_type_arguments_reg, |
234 | Register temp_reg, |
235 | compiler::Label* is_instance_lbl, |
236 | compiler::Label* is_not_instance_lbl) { |
237 | ASSERT(temp_reg == kNoRegister); |
238 | const SubtypeTestCache& type_test_cache = |
239 | SubtypeTestCache::ZoneHandle(zone(), SubtypeTestCache::New()); |
240 | __ LoadUniqueObject(R9, type_test_cache); |
241 | if (test_kind == kTestTypeOneArg) { |
242 | ASSERT(instantiator_type_arguments_reg == kNoRegister); |
243 | ASSERT(function_type_arguments_reg == kNoRegister); |
244 | __ Call(StubCode::Subtype1TestCache()); |
245 | } else if (test_kind == kTestTypeTwoArgs) { |
246 | ASSERT(instantiator_type_arguments_reg == kNoRegister); |
247 | ASSERT(function_type_arguments_reg == kNoRegister); |
248 | __ Call(StubCode::Subtype2TestCache()); |
249 | } else if (test_kind == kTestTypeFourArgs) { |
250 | ASSERT(instantiator_type_arguments_reg == |
251 | TypeTestABI::kInstantiatorTypeArgumentsReg); |
252 | ASSERT(function_type_arguments_reg == |
253 | TypeTestABI::kFunctionTypeArgumentsReg); |
254 | __ Call(StubCode::Subtype4TestCache()); |
255 | } else if (test_kind == kTestTypeSixArgs) { |
256 | ASSERT(instantiator_type_arguments_reg == |
257 | TypeTestABI::kInstantiatorTypeArgumentsReg); |
258 | ASSERT(function_type_arguments_reg == |
259 | TypeTestABI::kFunctionTypeArgumentsReg); |
260 | __ Call(StubCode::Subtype6TestCache()); |
261 | } else { |
262 | UNREACHABLE(); |
263 | } |
264 | // Result is in R8: null -> not found, otherwise Bool::True or Bool::False. |
265 | GenerateBoolToJump(R8, is_instance_lbl, is_not_instance_lbl); |
266 | return type_test_cache.raw(); |
267 | } |
268 | |
269 | // Jumps to labels 'is_instance' or 'is_not_instance' respectively, if |
270 | // type test is conclusive, otherwise fallthrough if a type test could not |
271 | // be completed. |
272 | // RAX: instance (must survive). |
273 | // Clobbers R10. |
274 | SubtypeTestCachePtr |
275 | FlowGraphCompiler::GenerateInstantiatedTypeWithArgumentsTest( |
276 | TokenPosition token_pos, |
277 | const AbstractType& type, |
278 | compiler::Label* is_instance_lbl, |
279 | compiler::Label* is_not_instance_lbl) { |
280 | __ Comment("InstantiatedTypeWithArgumentsTest" ); |
281 | ASSERT(type.IsInstantiated()); |
282 | ASSERT(!type.IsFunctionType()); |
283 | const Class& type_class = Class::ZoneHandle(zone(), type.type_class()); |
284 | ASSERT(type_class.NumTypeArguments() > 0); |
285 | const Type& smi_type = Type::Handle(zone(), Type::SmiType()); |
286 | const bool smi_is_ok = smi_type.IsSubtypeOf(type, Heap::kOld); |
287 | __ testq(TypeTestABI::kInstanceReg, compiler::Immediate(kSmiTagMask)); |
288 | if (smi_is_ok) { |
289 | // Fast case for type = FutureOr<int/num/top-type>. |
290 | __ j(ZERO, is_instance_lbl); |
291 | } else { |
292 | __ j(ZERO, is_not_instance_lbl); |
293 | } |
294 | |
295 | const intptr_t num_type_args = type_class.NumTypeArguments(); |
296 | const intptr_t num_type_params = type_class.NumTypeParameters(); |
297 | const intptr_t from_index = num_type_args - num_type_params; |
298 | const TypeArguments& type_arguments = |
299 | TypeArguments::ZoneHandle(zone(), type.arguments()); |
300 | const bool is_raw_type = type_arguments.IsNull() || |
301 | type_arguments.IsRaw(from_index, num_type_params); |
302 | if (is_raw_type) { |
303 | const Register kClassIdReg = R10; |
304 | // dynamic type argument, check only classes. |
305 | __ LoadClassId(kClassIdReg, TypeTestABI::kInstanceReg); |
306 | __ cmpl(kClassIdReg, compiler::Immediate(type_class.id())); |
307 | __ j(EQUAL, is_instance_lbl); |
308 | // List is a very common case. |
309 | if (IsListClass(type_class)) { |
310 | GenerateListTypeCheck(kClassIdReg, is_instance_lbl); |
311 | } |
312 | return GenerateSubtype1TestCacheLookup( |
313 | token_pos, type_class, is_instance_lbl, is_not_instance_lbl); |
314 | } |
315 | // If one type argument only, check if type argument is a top type. |
316 | if (type_arguments.Length() == 1) { |
317 | const AbstractType& tp_argument = |
318 | AbstractType::ZoneHandle(zone(), type_arguments.TypeAt(0)); |
319 | if (tp_argument.IsTopTypeForSubtyping()) { |
320 | // Instance class test only necessary. |
321 | return GenerateSubtype1TestCacheLookup( |
322 | token_pos, type_class, is_instance_lbl, is_not_instance_lbl); |
323 | } |
324 | } |
325 | |
326 | // Regular subtype test cache involving instance's type arguments. |
327 | const Register kInstantiatorTypeArgumentsReg = kNoRegister; |
328 | const Register kFunctionTypeArgumentsReg = kNoRegister; |
329 | const Register kTempReg = kNoRegister; |
330 | return GenerateCallSubtypeTestStub( |
331 | kTestTypeTwoArgs, TypeTestABI::kInstanceReg, |
332 | kInstantiatorTypeArgumentsReg, kFunctionTypeArgumentsReg, kTempReg, |
333 | is_instance_lbl, is_not_instance_lbl); |
334 | } |
335 | |
336 | void FlowGraphCompiler::CheckClassIds(Register class_id_reg, |
337 | const GrowableArray<intptr_t>& class_ids, |
338 | compiler::Label* is_equal_lbl, |
339 | compiler::Label* is_not_equal_lbl) { |
340 | for (intptr_t i = 0; i < class_ids.length(); i++) { |
341 | __ cmpl(class_id_reg, compiler::Immediate(class_ids[i])); |
342 | __ j(EQUAL, is_equal_lbl); |
343 | } |
344 | __ jmp(is_not_equal_lbl); |
345 | } |
346 | |
347 | // Testing against an instantiated type with no arguments, without |
348 | // SubtypeTestCache |
349 | // |
350 | // Inputs: |
351 | // - RAX : instance to test against |
352 | // |
353 | // Preserves RAX/RCX/RDX. |
354 | // |
355 | // Returns true if there is a fallthrough. |
356 | bool FlowGraphCompiler::GenerateInstantiatedTypeNoArgumentsTest( |
357 | TokenPosition token_pos, |
358 | const AbstractType& type, |
359 | compiler::Label* is_instance_lbl, |
360 | compiler::Label* is_not_instance_lbl) { |
361 | __ Comment("InstantiatedTypeNoArgumentsTest" ); |
362 | ASSERT(type.IsInstantiated()); |
363 | ASSERT(!type.IsFunctionType()); |
364 | const Class& type_class = Class::Handle(zone(), type.type_class()); |
365 | ASSERT(type_class.NumTypeArguments() == 0); |
366 | |
367 | __ testq(TypeTestABI::kInstanceReg, compiler::Immediate(kSmiTagMask)); |
368 | // If instance is Smi, check directly. |
369 | const Class& smi_class = Class::Handle(zone(), Smi::Class()); |
370 | if (Class::IsSubtypeOf(smi_class, Object::null_type_arguments(), |
371 | Nullability::kNonNullable, type, Heap::kOld)) { |
372 | // Fast case for type = int/num/top-type. |
373 | __ j(ZERO, is_instance_lbl); |
374 | } else { |
375 | __ j(ZERO, is_not_instance_lbl); |
376 | } |
377 | const Register kClassIdReg = R10; |
378 | __ LoadClassId(kClassIdReg, TypeTestABI::kInstanceReg); |
379 | // Bool interface can be implemented only by core class Bool. |
380 | if (type.IsBoolType()) { |
381 | __ cmpl(kClassIdReg, compiler::Immediate(kBoolCid)); |
382 | __ j(EQUAL, is_instance_lbl); |
383 | __ jmp(is_not_instance_lbl); |
384 | return false; |
385 | } |
386 | // Custom checking for numbers (Smi, Mint and Double). |
387 | // Note that instance is not Smi (checked above). |
388 | if (type.IsNumberType() || type.IsIntType() || type.IsDoubleType()) { |
389 | GenerateNumberTypeCheck(kClassIdReg, type, is_instance_lbl, |
390 | is_not_instance_lbl); |
391 | return false; |
392 | } |
393 | if (type.IsStringType()) { |
394 | GenerateStringTypeCheck(kClassIdReg, is_instance_lbl, is_not_instance_lbl); |
395 | return false; |
396 | } |
397 | if (type.IsDartFunctionType()) { |
398 | // Check if instance is a closure. |
399 | __ cmpq(kClassIdReg, compiler::Immediate(kClosureCid)); |
400 | __ j(EQUAL, is_instance_lbl); |
401 | return true; |
402 | } |
403 | |
404 | // Fast case for cid-range based checks. |
405 | // Warning: This code destroys the contents of [kClassIdReg]. |
406 | if (GenerateSubtypeRangeCheck(kClassIdReg, type_class, is_instance_lbl)) { |
407 | return false; |
408 | } |
409 | |
410 | // Otherwise fallthrough, result non-conclusive. |
411 | return true; |
412 | } |
413 | |
414 | // Uses SubtypeTestCache to store instance class and result. |
415 | // Immediate class test already done. |
416 | // |
417 | // Inputs: |
418 | // RAX : instance to test against. |
419 | // |
420 | // Preserves RAX/RCX/RDX. |
421 | // |
422 | // TODO(srdjan): Implement a quicker subtype check, as type test |
423 | // arrays can grow too high, but they may be useful when optimizing |
424 | // code (type-feedback). |
425 | SubtypeTestCachePtr FlowGraphCompiler::GenerateSubtype1TestCacheLookup( |
426 | TokenPosition token_pos, |
427 | const Class& type_class, |
428 | compiler::Label* is_instance_lbl, |
429 | compiler::Label* is_not_instance_lbl) { |
430 | __ Comment("Subtype1TestCacheLookup" ); |
431 | #if defined(DEBUG) |
432 | compiler::Label ok; |
433 | __ BranchIfNotSmi(TypeTestABI::kInstanceReg, &ok); |
434 | __ Breakpoint(); |
435 | __ Bind(&ok); |
436 | #endif |
437 | __ LoadClassId(TMP, TypeTestABI::kInstanceReg); |
438 | __ LoadClassById(R10, TMP); |
439 | // R10: instance class. |
440 | // Check immediate superclass equality. If type_class is Object, then testing |
441 | // supertype may yield a wrong result for Null in NNBD strong mode (because |
442 | // Null also extends Object). |
443 | if (!type_class.IsObjectClass() || !Isolate::Current()->null_safety()) { |
444 | __ movq(R13, compiler::FieldAddress(R10, Class::super_type_offset())); |
445 | __ movq(R13, compiler::FieldAddress(R13, Type::type_class_id_offset())); |
446 | __ CompareImmediate(R13, |
447 | compiler::Immediate(Smi::RawValue(type_class.id()))); |
448 | __ j(EQUAL, is_instance_lbl); |
449 | } |
450 | |
451 | const Register kInstantiatorTypeArgumentsReg = kNoRegister; |
452 | const Register kFunctionTypeArgumentsReg = kNoRegister; |
453 | const Register kTempReg = kNoRegister; |
454 | return GenerateCallSubtypeTestStub(kTestTypeOneArg, TypeTestABI::kInstanceReg, |
455 | kInstantiatorTypeArgumentsReg, |
456 | kFunctionTypeArgumentsReg, kTempReg, |
457 | is_instance_lbl, is_not_instance_lbl); |
458 | } |
459 | |
460 | // Generates inlined check if 'type' is a type parameter or type itself |
461 | // |
462 | // Inputs: |
463 | // - RAX : instance to test against. |
464 | // - RDX : instantiator type arguments (if necessary). |
465 | // - RCX : function type arguments (if necessary). |
466 | // |
467 | // Preserves RAX/RCX/RDX. |
468 | SubtypeTestCachePtr FlowGraphCompiler::GenerateUninstantiatedTypeTest( |
469 | TokenPosition token_pos, |
470 | const AbstractType& type, |
471 | compiler::Label* is_instance_lbl, |
472 | compiler::Label* is_not_instance_lbl) { |
473 | const Register kTempReg = kNoRegister; |
474 | __ Comment("UninstantiatedTypeTest" ); |
475 | ASSERT(!type.IsInstantiated()); |
476 | ASSERT(!type.IsFunctionType()); |
477 | // Skip check if destination is a dynamic type. |
478 | if (type.IsTypeParameter()) { |
479 | const TypeParameter& type_param = TypeParameter::Cast(type); |
480 | |
481 | const Register kTypeArgumentsReg = |
482 | type_param.IsClassTypeParameter() |
483 | ? TypeTestABI::kInstantiatorTypeArgumentsReg |
484 | : TypeTestABI::kFunctionTypeArgumentsReg; |
485 | // Check if type arguments are null, i.e. equivalent to vector of dynamic. |
486 | __ CompareObject(kTypeArgumentsReg, Object::null_object()); |
487 | __ j(EQUAL, is_instance_lbl); |
488 | __ movq(RDI, compiler::FieldAddress( |
489 | kTypeArgumentsReg, |
490 | TypeArguments::type_at_offset(type_param.index()))); |
491 | // RDI: Concrete type of type. |
492 | // Check if type argument is dynamic, Object?, or void. |
493 | __ CompareObject(RDI, Object::dynamic_type()); |
494 | __ j(EQUAL, is_instance_lbl); |
495 | __ CompareObject( |
496 | RDI, Type::ZoneHandle( |
497 | zone(), isolate()->object_store()->nullable_object_type())); |
498 | __ j(EQUAL, is_instance_lbl); |
499 | __ CompareObject(RDI, Object::void_type()); |
500 | __ j(EQUAL, is_instance_lbl); |
501 | |
502 | // For Smi check quickly against int and num interfaces. |
503 | compiler::Label not_smi; |
504 | __ testq(RAX, compiler::Immediate(kSmiTagMask)); // Value is Smi? |
505 | __ j(NOT_ZERO, ¬_smi, compiler::Assembler::kNearJump); |
506 | __ CompareObject(RDI, Type::ZoneHandle(zone(), Type::IntType())); |
507 | __ j(EQUAL, is_instance_lbl); |
508 | __ CompareObject(RDI, Type::ZoneHandle(zone(), Type::Number())); |
509 | __ j(EQUAL, is_instance_lbl); |
510 | // Smi can be handled by type test cache. |
511 | __ Bind(¬_smi); |
512 | |
513 | const auto test_kind = GetTypeTestStubKindForTypeParameter(type_param); |
514 | const SubtypeTestCache& type_test_cache = SubtypeTestCache::ZoneHandle( |
515 | zone(), GenerateCallSubtypeTestStub( |
516 | test_kind, TypeTestABI::kInstanceReg, |
517 | TypeTestABI::kInstantiatorTypeArgumentsReg, |
518 | TypeTestABI::kFunctionTypeArgumentsReg, kTempReg, |
519 | is_instance_lbl, is_not_instance_lbl)); |
520 | return type_test_cache.raw(); |
521 | } |
522 | if (type.IsType()) { |
523 | // Smi is FutureOr<T>, when T is a top type or int or num. |
524 | if (!type.IsFutureOrType()) { |
525 | __ testq(TypeTestABI::kInstanceReg, |
526 | compiler::Immediate(kSmiTagMask)); // Is instance Smi? |
527 | __ j(ZERO, is_not_instance_lbl); |
528 | } |
529 | // Uninstantiated type class is known at compile time, but the type |
530 | // arguments are determined at runtime by the instantiator(s). |
531 | return GenerateCallSubtypeTestStub( |
532 | kTestTypeFourArgs, TypeTestABI::kInstanceReg, |
533 | TypeTestABI::kInstantiatorTypeArgumentsReg, |
534 | TypeTestABI::kFunctionTypeArgumentsReg, kTempReg, is_instance_lbl, |
535 | is_not_instance_lbl); |
536 | } |
537 | return SubtypeTestCache::null(); |
538 | } |
539 | |
540 | // Generates function type check. |
541 | // |
542 | // See [GenerateUninstantiatedTypeTest] for calling convention. |
543 | SubtypeTestCachePtr FlowGraphCompiler::GenerateFunctionTypeTest( |
544 | TokenPosition token_pos, |
545 | const AbstractType& type, |
546 | compiler::Label* is_instance_lbl, |
547 | compiler::Label* is_not_instance_lbl) { |
548 | const Register kTempReg = kNoRegister; |
549 | __ Comment("FunctionTypeTest" ); |
550 | |
551 | __ testq(TypeTestABI::kInstanceReg, compiler::Immediate(kSmiTagMask)); |
552 | __ j(ZERO, is_not_instance_lbl); |
553 | return GenerateCallSubtypeTestStub( |
554 | kTestTypeSixArgs, TypeTestABI::kInstanceReg, |
555 | TypeTestABI::kInstantiatorTypeArgumentsReg, |
556 | TypeTestABI::kFunctionTypeArgumentsReg, kTempReg, is_instance_lbl, |
557 | is_not_instance_lbl); |
558 | } |
559 | |
560 | // Inputs: |
561 | // - RAX : instance to test against. |
562 | // - RDX : instantiator type arguments. |
563 | // - RCX : function type arguments. |
564 | // |
565 | // Preserves RAX/RCX/RDX. |
566 | // |
567 | // Note that this inlined code must be followed by the runtime_call code, as it |
568 | // may fall through to it. Otherwise, this inline code will jump to the label |
569 | // is_instance or to the label is_not_instance. |
570 | SubtypeTestCachePtr FlowGraphCompiler::GenerateInlineInstanceof( |
571 | TokenPosition token_pos, |
572 | const AbstractType& type, |
573 | compiler::Label* is_instance_lbl, |
574 | compiler::Label* is_not_instance_lbl) { |
575 | __ Comment("InlineInstanceof" ); |
576 | |
577 | if (type.IsFunctionType()) { |
578 | return GenerateFunctionTypeTest(token_pos, type, is_instance_lbl, |
579 | is_not_instance_lbl); |
580 | } |
581 | |
582 | if (type.IsInstantiated()) { |
583 | const Class& type_class = Class::ZoneHandle(zone(), type.type_class()); |
584 | // A class equality check is only applicable with a dst type (not a |
585 | // function type) of a non-parameterized class or with a raw dst type of |
586 | // a parameterized class. |
587 | if (type_class.NumTypeArguments() > 0) { |
588 | return GenerateInstantiatedTypeWithArgumentsTest( |
589 | token_pos, type, is_instance_lbl, is_not_instance_lbl); |
590 | // Fall through to runtime call. |
591 | } |
592 | const bool has_fall_through = GenerateInstantiatedTypeNoArgumentsTest( |
593 | token_pos, type, is_instance_lbl, is_not_instance_lbl); |
594 | if (has_fall_through) { |
595 | // If test non-conclusive so far, try the inlined type-test cache. |
596 | // 'type' is known at compile time. |
597 | return GenerateSubtype1TestCacheLookup( |
598 | token_pos, type_class, is_instance_lbl, is_not_instance_lbl); |
599 | } else { |
600 | return SubtypeTestCache::null(); |
601 | } |
602 | } |
603 | return GenerateUninstantiatedTypeTest(token_pos, type, is_instance_lbl, |
604 | is_not_instance_lbl); |
605 | } |
606 | |
607 | // If instanceof type test cannot be performed successfully at compile time and |
608 | // therefore eliminated, optimize it by adding inlined tests for: |
609 | // - Null -> see comment below. |
610 | // - Smi -> compile time subtype check (only if dst class is not parameterized). |
611 | // - Class equality (only if class is not parameterized). |
612 | // Inputs: |
613 | // - RAX: object. |
614 | // - RDX: instantiator type arguments or raw_null. |
615 | // - RCX: function type arguments or raw_null. |
616 | // Returns: |
617 | // - true or false in RAX. |
618 | void FlowGraphCompiler::GenerateInstanceOf(TokenPosition token_pos, |
619 | intptr_t deopt_id, |
620 | const AbstractType& type, |
621 | LocationSummary* locs) { |
622 | ASSERT(type.IsFinalized()); |
623 | ASSERT(!type.IsTopTypeForInstanceOf()); // Already checked. |
624 | |
625 | compiler::Label is_instance, is_not_instance; |
626 | // 'null' is an instance of Null, Object*, Never*, void, and dynamic. |
627 | // In addition, 'null' is an instance of any nullable type. |
628 | // It is also an instance of FutureOr<T> if it is an instance of T. |
629 | const AbstractType& unwrapped_type = |
630 | AbstractType::Handle(type.UnwrapFutureOr()); |
631 | if (!unwrapped_type.IsTypeParameter() || unwrapped_type.IsNullable()) { |
632 | // Only nullable type parameter remains nullable after instantiation. |
633 | // See NullIsInstanceOf(). |
634 | __ CompareObject(TypeTestABI::kInstanceReg, Object::null_object()); |
635 | __ j(EQUAL, (unwrapped_type.IsNullable() || |
636 | (unwrapped_type.IsLegacy() && unwrapped_type.IsNeverType())) |
637 | ? &is_instance |
638 | : &is_not_instance); |
639 | } |
640 | |
641 | // Generate inline instanceof test. |
642 | SubtypeTestCache& test_cache = SubtypeTestCache::ZoneHandle(zone()); |
643 | // The registers RAX, RCX, RDX are preserved across the call. |
644 | test_cache = |
645 | GenerateInlineInstanceof(token_pos, type, &is_instance, &is_not_instance); |
646 | |
647 | // test_cache is null if there is no fall-through. |
648 | compiler::Label done; |
649 | if (!test_cache.IsNull()) { |
650 | // Generate Runtime call. |
651 | __ LoadUniqueObject(TypeTestABI::kDstTypeReg, type); |
652 | __ LoadUniqueObject(TypeTestABI::kSubtypeTestCacheReg, test_cache); |
653 | GenerateStubCall(token_pos, StubCode::InstanceOf(), |
654 | /*kind=*/PcDescriptorsLayout::kOther, locs, deopt_id); |
655 | __ jmp(&done, compiler::Assembler::kNearJump); |
656 | } |
657 | __ Bind(&is_not_instance); |
658 | __ LoadObject(RAX, Bool::Get(false)); |
659 | __ jmp(&done, compiler::Assembler::kNearJump); |
660 | |
661 | __ Bind(&is_instance); |
662 | __ LoadObject(RAX, Bool::Get(true)); |
663 | __ Bind(&done); |
664 | } |
665 | |
666 | // Optimize assignable type check by adding inlined tests for: |
667 | // - NULL -> return NULL. |
668 | // - Smi -> compile time subtype check (only if dst class is not parameterized). |
669 | // - Class equality (only if class is not parameterized). |
670 | // Inputs: |
671 | // - RAX: object. |
672 | // - RBX: destination type (if non-constant). |
673 | // - RDX: instantiator type arguments or raw_null. |
674 | // - RCX: function type arguments or raw_null. |
675 | // Returns: |
676 | // - object in RAX for successful assignable check (or throws TypeError). |
677 | // Performance notes: positive checks must be quick, negative checks can be slow |
678 | // as they throw an exception. |
679 | void FlowGraphCompiler::GenerateAssertAssignable(CompileType* receiver_type, |
680 | TokenPosition token_pos, |
681 | intptr_t deopt_id, |
682 | const String& dst_name, |
683 | LocationSummary* locs) { |
684 | ASSERT(!token_pos.IsClassifying()); |
685 | ASSERT(CheckAssertAssignableTypeTestingABILocations(*locs)); |
686 | |
687 | compiler::Label is_assignable, runtime_call; |
688 | |
689 | // Generate inline type check, linking to runtime call if not assignable. |
690 | SubtypeTestCache& test_cache = SubtypeTestCache::ZoneHandle(zone()); |
691 | |
692 | if (locs->in(1).IsConstant()) { |
693 | const auto& dst_type = AbstractType::Cast(locs->in(1).constant()); |
694 | ASSERT(dst_type.IsFinalized()); |
695 | |
696 | if (dst_type.IsTopTypeForSubtyping()) return; // No code needed. |
697 | |
698 | if (ShouldUseTypeTestingStubFor(is_optimizing(), dst_type)) { |
699 | GenerateAssertAssignableViaTypeTestingStub(receiver_type, token_pos, |
700 | deopt_id, dst_name, locs); |
701 | return; |
702 | } |
703 | |
704 | if (Instance::NullIsAssignableTo(dst_type)) { |
705 | __ CompareObject(TypeTestABI::kInstanceReg, Object::null_object()); |
706 | __ j(EQUAL, &is_assignable); |
707 | } |
708 | |
709 | // The registers RAX, RCX, RDX are preserved across the call. |
710 | test_cache = GenerateInlineInstanceof(token_pos, dst_type, &is_assignable, |
711 | &runtime_call); |
712 | |
713 | } else { |
714 | // TODO(dartbug.com/40813): Handle setting up the non-constant case. |
715 | UNREACHABLE(); |
716 | } |
717 | |
718 | __ Bind(&runtime_call); |
719 | __ PushObject(Object::null_object()); // Make room for the result. |
720 | __ pushq(TypeTestABI::kInstanceReg); // Push the source object. |
721 | // Push the destination type. |
722 | if (locs->in(1).IsConstant()) { |
723 | __ PushObject(locs->in(1).constant()); |
724 | } else { |
725 | // TODO(dartbug.com/40813): Handle setting up the non-constant case. |
726 | UNREACHABLE(); |
727 | } |
728 | __ pushq(TypeTestABI::kInstantiatorTypeArgumentsReg); |
729 | __ pushq(TypeTestABI::kFunctionTypeArgumentsReg); |
730 | __ PushObject(dst_name); // Push the name of the destination. |
731 | __ LoadUniqueObject(RAX, test_cache); |
732 | __ pushq(RAX); |
733 | __ PushImmediate(compiler::Immediate(Smi::RawValue(kTypeCheckFromInline))); |
734 | GenerateRuntimeCall(token_pos, deopt_id, kTypeCheckRuntimeEntry, 7, locs); |
735 | // Pop the parameters supplied to the runtime entry. The result of the |
736 | // type check runtime call is the checked value. |
737 | __ Drop(7); |
738 | __ popq(TypeTestABI::kInstanceReg); |
739 | __ Bind(&is_assignable); |
740 | } |
741 | |
742 | void FlowGraphCompiler::GenerateAssertAssignableViaTypeTestingStub( |
743 | CompileType* receiver_type, |
744 | TokenPosition token_pos, |
745 | intptr_t deopt_id, |
746 | const String& dst_name, |
747 | LocationSummary* locs) { |
748 | ASSERT(CheckAssertAssignableTypeTestingABILocations(*locs)); |
749 | // We must have a constant dst_type for generating a call to the stub. |
750 | ASSERT(locs->in(1).IsConstant()); |
751 | const auto& dst_type = AbstractType::Cast(locs->in(1).constant()); |
752 | |
753 | // If the dst_type is instantiated we know the target TTS stub at |
754 | // compile-time and can therefore use a pc-relative call. |
755 | const bool use_pc_relative_call = |
756 | dst_type.IsInstantiated() && CanPcRelativeCall(dst_type); |
757 | const Register kScratchReg = |
758 | dst_type.IsTypeParameter() ? RSI : TypeTestABI::kDstTypeReg; |
759 | |
760 | const Register kRegToCall = use_pc_relative_call ? kNoRegister : kScratchReg; |
761 | |
762 | compiler::Label done; |
763 | |
764 | GenerateAssertAssignableViaTypeTestingStub(receiver_type, dst_type, dst_name, |
765 | kRegToCall, kScratchReg, &done); |
766 | |
767 | // We use 2 consecutive entries in the pool for the subtype cache and the |
768 | // destination name. The second entry, namely [dst_name] seems to be unused, |
769 | // but it will be used by the code throwing a TypeError if the type test fails |
770 | // (see runtime/vm/runtime_entry.cc:TypeCheck). It will use pattern matching |
771 | // on the call site to find out at which pool index the destination name is |
772 | // located. |
773 | const intptr_t sub_type_cache_index = __ object_pool_builder().AddObject( |
774 | Object::null_object(), compiler::ObjectPoolBuilderEntry::kPatchable); |
775 | const intptr_t sub_type_cache_offset = |
776 | ObjectPool::element_offset(sub_type_cache_index) - kHeapObjectTag; |
777 | const intptr_t dst_name_index = __ object_pool_builder().AddObject( |
778 | dst_name, compiler::ObjectPoolBuilderEntry::kPatchable); |
779 | ASSERT((sub_type_cache_index + 1) == dst_name_index); |
780 | ASSERT(__ constant_pool_allowed()); |
781 | |
782 | __ movq(TypeTestABI::kSubtypeTestCacheReg, |
783 | compiler::Address(PP, sub_type_cache_offset)); |
784 | if (use_pc_relative_call) { |
785 | __ GenerateUnRelocatedPcRelativeCall(); |
786 | AddPcRelativeTTSCallTypeTarget(dst_type); |
787 | } else { |
788 | __ call(compiler::FieldAddress( |
789 | kRegToCall, AbstractType::type_test_stub_entry_point_offset())); |
790 | } |
791 | EmitCallsiteMetadata(token_pos, deopt_id, PcDescriptorsLayout::kOther, locs); |
792 | __ Bind(&done); |
793 | } |
794 | |
795 | void FlowGraphCompiler::EmitInstructionEpilogue(Instruction* instr) { |
796 | if (is_optimizing()) { |
797 | return; |
798 | } |
799 | Definition* defn = instr->AsDefinition(); |
800 | if ((defn != NULL) && defn->HasTemp()) { |
801 | Location value = defn->locs()->out(0); |
802 | if (value.IsRegister()) { |
803 | __ pushq(value.reg()); |
804 | } else if (value.IsConstant()) { |
805 | __ PushObject(value.constant()); |
806 | } else { |
807 | ASSERT(value.IsStackSlot()); |
808 | __ pushq(LocationToStackSlotAddress(value)); |
809 | } |
810 | } |
811 | } |
812 | |
813 | void FlowGraphCompiler::( |
814 | const Function& , |
815 | intptr_t type_arguments_field_offset) { |
816 | // No frame has been setup here. |
817 | ASSERT(!__ constant_pool_allowed()); |
818 | ASSERT(extracted_method.IsZoneHandle()); |
819 | |
820 | const Code& = Code::ZoneHandle( |
821 | isolate()->object_store()->build_method_extractor_code()); |
822 | ASSERT(!build_method_extractor.IsNull()); |
823 | |
824 | const intptr_t stub_index = __ object_pool_builder().AddObject( |
825 | build_method_extractor, compiler::ObjectPoolBuilderEntry::kNotPatchable); |
826 | const intptr_t function_index = __ object_pool_builder().AddObject( |
827 | extracted_method, compiler::ObjectPoolBuilderEntry::kNotPatchable); |
828 | |
829 | // We use a custom pool register to preserve caller PP. |
830 | Register kPoolReg = RAX; |
831 | |
832 | // RBX = extracted function |
833 | // RDX = offset of type argument vector (or 0 if class is not generic) |
834 | if (FLAG_precompiled_mode && FLAG_use_bare_instructions) { |
835 | kPoolReg = PP; |
836 | } else { |
837 | __ movq(kPoolReg, |
838 | compiler::FieldAddress(CODE_REG, Code::object_pool_offset())); |
839 | } |
840 | __ movq(RDX, compiler::Immediate(type_arguments_field_offset)); |
841 | __ movq(RBX, compiler::FieldAddress( |
842 | kPoolReg, ObjectPool::element_offset(function_index))); |
843 | __ movq(CODE_REG, compiler::FieldAddress( |
844 | kPoolReg, ObjectPool::element_offset(stub_index))); |
845 | __ jmp(compiler::FieldAddress( |
846 | CODE_REG, Code::entry_point_offset(Code::EntryKind::kUnchecked))); |
847 | } |
848 | |
849 | // NOTE: If the entry code shape changes, ReturnAddressLocator in profiler.cc |
850 | // needs to be updated to match. |
851 | void FlowGraphCompiler::EmitFrameEntry() { |
852 | if (flow_graph().IsCompiledForOsr()) { |
853 | const intptr_t = ExtraStackSlotsOnOsrEntry(); |
854 | ASSERT(extra_slots >= 0); |
855 | __ EnterOsrFrame(extra_slots * kWordSize); |
856 | } else { |
857 | const Function& function = parsed_function().function(); |
858 | if (CanOptimizeFunction() && function.IsOptimizable() && |
859 | (!is_optimizing() || may_reoptimize())) { |
860 | __ Comment("Invocation Count Check" ); |
861 | const Register function_reg = RDI; |
862 | __ movq(function_reg, |
863 | compiler::FieldAddress(CODE_REG, Code::owner_offset())); |
864 | |
865 | // Reoptimization of an optimized function is triggered by counting in |
866 | // IC stubs, but not at the entry of the function. |
867 | if (!is_optimizing()) { |
868 | __ incl(compiler::FieldAddress(function_reg, |
869 | Function::usage_counter_offset())); |
870 | } |
871 | __ cmpl(compiler::FieldAddress(function_reg, |
872 | Function::usage_counter_offset()), |
873 | compiler::Immediate(GetOptimizationThreshold())); |
874 | ASSERT(function_reg == RDI); |
875 | compiler::Label dont_optimize; |
876 | __ j(LESS, &dont_optimize, compiler::Assembler::kNearJump); |
877 | __ jmp(compiler::Address(THR, Thread::optimize_entry_offset())); |
878 | __ Bind(&dont_optimize); |
879 | } |
880 | ASSERT(StackSize() >= 0); |
881 | __ Comment("Enter frame" ); |
882 | __ EnterDartFrame(StackSize() * kWordSize); |
883 | } |
884 | } |
885 | |
886 | void FlowGraphCompiler::EmitPrologue() { |
887 | BeginCodeSourceRange(); |
888 | |
889 | EmitFrameEntry(); |
890 | ASSERT(assembler()->constant_pool_allowed()); |
891 | |
892 | // In unoptimized code, initialize (non-argument) stack allocated slots. |
893 | if (!is_optimizing()) { |
894 | const int num_locals = parsed_function().num_stack_locals(); |
895 | |
896 | intptr_t args_desc_slot = -1; |
897 | if (parsed_function().has_arg_desc_var()) { |
898 | args_desc_slot = compiler::target::frame_layout.FrameSlotForVariable( |
899 | parsed_function().arg_desc_var()); |
900 | } |
901 | |
902 | __ Comment("Initialize spill slots" ); |
903 | if (num_locals > 1 || (num_locals == 1 && args_desc_slot == -1)) { |
904 | __ LoadObject(RAX, Object::null_object()); |
905 | } |
906 | for (intptr_t i = 0; i < num_locals; ++i) { |
907 | const intptr_t slot_index = |
908 | compiler::target::frame_layout.FrameSlotForVariableIndex(-i); |
909 | Register value_reg = slot_index == args_desc_slot ? ARGS_DESC_REG : RAX; |
910 | __ movq(compiler::Address(RBP, slot_index * kWordSize), value_reg); |
911 | } |
912 | } |
913 | |
914 | EndCodeSourceRange(TokenPosition::kDartCodePrologue); |
915 | } |
916 | |
917 | void FlowGraphCompiler::CompileGraph() { |
918 | InitCompiler(); |
919 | |
920 | // We have multiple entrypoints functionality which moved the frame |
921 | // setup into the [FunctionEntryInstr] (which will set the constant pool |
922 | // allowed bit to true). Despite this we still have to set the |
923 | // constant pool allowed bit to true here as well, because we can generate |
924 | // code for [CatchEntryInstr]s, which need the pool. |
925 | __ set_constant_pool_allowed(true); |
926 | |
927 | ASSERT(!block_order().is_empty()); |
928 | VisitBlocks(); |
929 | |
930 | #if defined(DEBUG) |
931 | __ int3(); |
932 | #endif |
933 | |
934 | if (!skip_body_compilation()) { |
935 | ASSERT(assembler()->constant_pool_allowed()); |
936 | GenerateDeferredCode(); |
937 | } |
938 | |
939 | for (intptr_t i = 0; i < indirect_gotos_.length(); ++i) { |
940 | indirect_gotos_[i]->ComputeOffsetTable(this); |
941 | } |
942 | } |
943 | |
944 | void FlowGraphCompiler::EmitCallToStub(const Code& stub) { |
945 | ASSERT(!stub.IsNull()); |
946 | if (CanPcRelativeCall(stub)) { |
947 | __ GenerateUnRelocatedPcRelativeCall(); |
948 | AddPcRelativeCallStubTarget(stub); |
949 | } else { |
950 | __ Call(stub); |
951 | AddStubCallTarget(stub); |
952 | } |
953 | } |
954 | |
955 | void FlowGraphCompiler::EmitTailCallToStub(const Code& stub) { |
956 | ASSERT(!stub.IsNull()); |
957 | if (CanPcRelativeCall(stub)) { |
958 | __ LeaveDartFrame(); |
959 | __ GenerateUnRelocatedPcRelativeTailCall(); |
960 | AddPcRelativeTailCallStubTarget(stub); |
961 | #if defined(DEBUG) |
962 | __ Breakpoint(); |
963 | #endif |
964 | } else { |
965 | __ LoadObject(CODE_REG, stub); |
966 | __ LeaveDartFrame(); |
967 | __ jmp(compiler::FieldAddress( |
968 | CODE_REG, compiler::target::Code::entry_point_offset())); |
969 | AddStubCallTarget(stub); |
970 | } |
971 | } |
972 | |
973 | void FlowGraphCompiler::GeneratePatchableCall(TokenPosition token_pos, |
974 | const Code& stub, |
975 | PcDescriptorsLayout::Kind kind, |
976 | LocationSummary* locs) { |
977 | __ CallPatchable(stub); |
978 | EmitCallsiteMetadata(token_pos, DeoptId::kNone, kind, locs); |
979 | } |
980 | |
981 | void FlowGraphCompiler::GenerateDartCall(intptr_t deopt_id, |
982 | TokenPosition token_pos, |
983 | const Code& stub, |
984 | PcDescriptorsLayout::Kind kind, |
985 | LocationSummary* locs, |
986 | Code::EntryKind entry_kind) { |
987 | ASSERT(CanCallDart()); |
988 | __ CallPatchable(stub, entry_kind); |
989 | EmitCallsiteMetadata(token_pos, deopt_id, kind, locs); |
990 | } |
991 | |
992 | void FlowGraphCompiler::GenerateStaticDartCall(intptr_t deopt_id, |
993 | TokenPosition token_pos, |
994 | PcDescriptorsLayout::Kind kind, |
995 | LocationSummary* locs, |
996 | const Function& target, |
997 | Code::EntryKind entry_kind) { |
998 | ASSERT(CanCallDart()); |
999 | ASSERT(is_optimizing()); |
1000 | if (CanPcRelativeCall(target)) { |
1001 | __ GenerateUnRelocatedPcRelativeCall(); |
1002 | AddPcRelativeCallTarget(target, entry_kind); |
1003 | EmitCallsiteMetadata(token_pos, deopt_id, kind, locs); |
1004 | } else { |
1005 | // Call sites to the same target can share object pool entries. These |
1006 | // call sites are never patched for breakpoints: the function is deoptimized |
1007 | // and the unoptimized code with IC calls for static calls is patched |
1008 | // instead. |
1009 | const auto& stub_entry = StubCode::CallStaticFunction(); |
1010 | __ CallWithEquivalence(stub_entry, target, entry_kind); |
1011 | EmitCallsiteMetadata(token_pos, deopt_id, kind, locs); |
1012 | AddStaticCallTarget(target, entry_kind); |
1013 | } |
1014 | } |
1015 | |
1016 | void FlowGraphCompiler::GenerateRuntimeCall(TokenPosition token_pos, |
1017 | intptr_t deopt_id, |
1018 | const RuntimeEntry& entry, |
1019 | intptr_t argument_count, |
1020 | LocationSummary* locs) { |
1021 | __ CallRuntime(entry, argument_count); |
1022 | EmitCallsiteMetadata(token_pos, deopt_id, PcDescriptorsLayout::kOther, locs); |
1023 | } |
1024 | |
1025 | void FlowGraphCompiler::EmitUnoptimizedStaticCall(intptr_t size_with_type_args, |
1026 | intptr_t deopt_id, |
1027 | TokenPosition token_pos, |
1028 | LocationSummary* locs, |
1029 | const ICData& ic_data, |
1030 | Code::EntryKind entry_kind) { |
1031 | ASSERT(CanCallDart()); |
1032 | const Code& stub = |
1033 | StubCode::UnoptimizedStaticCallEntry(ic_data.NumArgsTested()); |
1034 | __ LoadObject(RBX, ic_data); |
1035 | GenerateDartCall(deopt_id, token_pos, stub, |
1036 | PcDescriptorsLayout::kUnoptStaticCall, locs, entry_kind); |
1037 | __ Drop(size_with_type_args, RCX); |
1038 | } |
1039 | |
1040 | void FlowGraphCompiler::EmitEdgeCounter(intptr_t edge_id) { |
1041 | // We do not check for overflow when incrementing the edge counter. The |
1042 | // function should normally be optimized long before the counter can |
1043 | // overflow; and though we do not reset the counters when we optimize or |
1044 | // deoptimize, there is a bound on the number of |
1045 | // optimization/deoptimization cycles we will attempt. |
1046 | ASSERT(!edge_counters_array_.IsNull()); |
1047 | ASSERT(assembler_->constant_pool_allowed()); |
1048 | __ Comment("Edge counter" ); |
1049 | __ LoadObject(RAX, edge_counters_array_); |
1050 | __ IncrementSmiField( |
1051 | compiler::FieldAddress(RAX, Array::element_offset(edge_id)), 1); |
1052 | } |
1053 | |
1054 | void FlowGraphCompiler::EmitOptimizedInstanceCall(const Code& stub, |
1055 | const ICData& ic_data, |
1056 | intptr_t deopt_id, |
1057 | TokenPosition token_pos, |
1058 | LocationSummary* locs, |
1059 | Code::EntryKind entry_kind) { |
1060 | ASSERT(CanCallDart()); |
1061 | ASSERT(Array::Handle(zone(), ic_data.arguments_descriptor()).Length() > 0); |
1062 | // Each ICData propagated from unoptimized to optimized code contains the |
1063 | // function that corresponds to the Dart function of that IC call. Due |
1064 | // to inlining in optimized code, that function may not correspond to the |
1065 | // top-level function (parsed_function().function()) which could be |
1066 | // reoptimized and which counter needs to be incremented. |
1067 | // Pass the function explicitly, it is used in IC stub. |
1068 | __ LoadObject(RDI, parsed_function().function()); |
1069 | // Load receiver into RDX. |
1070 | __ movq(RDX, compiler::Address( |
1071 | RSP, (ic_data.SizeWithoutTypeArgs() - 1) * kWordSize)); |
1072 | __ LoadUniqueObject(RBX, ic_data); |
1073 | GenerateDartCall(deopt_id, token_pos, stub, PcDescriptorsLayout::kIcCall, |
1074 | locs, entry_kind); |
1075 | __ Drop(ic_data.SizeWithTypeArgs(), RCX); |
1076 | } |
1077 | |
1078 | void FlowGraphCompiler::EmitInstanceCallJIT(const Code& stub, |
1079 | const ICData& ic_data, |
1080 | intptr_t deopt_id, |
1081 | TokenPosition token_pos, |
1082 | LocationSummary* locs, |
1083 | Code::EntryKind entry_kind) { |
1084 | ASSERT(CanCallDart()); |
1085 | ASSERT(entry_kind == Code::EntryKind::kNormal || |
1086 | entry_kind == Code::EntryKind::kUnchecked); |
1087 | ASSERT(Array::Handle(zone(), ic_data.arguments_descriptor()).Length() > 0); |
1088 | // Load receiver into RDX. |
1089 | __ movq(RDX, compiler::Address( |
1090 | RSP, (ic_data.SizeWithoutTypeArgs() - 1) * kWordSize)); |
1091 | __ LoadUniqueObject(RBX, ic_data); |
1092 | __ LoadUniqueObject(CODE_REG, stub); |
1093 | const intptr_t entry_point_offset = |
1094 | entry_kind == Code::EntryKind::kNormal |
1095 | ? Code::entry_point_offset(Code::EntryKind::kMonomorphic) |
1096 | : Code::entry_point_offset(Code::EntryKind::kMonomorphicUnchecked); |
1097 | __ call(compiler::FieldAddress(CODE_REG, entry_point_offset)); |
1098 | EmitCallsiteMetadata(token_pos, deopt_id, PcDescriptorsLayout::kIcCall, locs); |
1099 | __ Drop(ic_data.SizeWithTypeArgs(), RCX); |
1100 | } |
1101 | |
1102 | void FlowGraphCompiler::EmitMegamorphicInstanceCall( |
1103 | const String& name, |
1104 | const Array& arguments_descriptor, |
1105 | intptr_t deopt_id, |
1106 | TokenPosition token_pos, |
1107 | LocationSummary* locs, |
1108 | intptr_t try_index, |
1109 | intptr_t slow_path_argument_count) { |
1110 | ASSERT(CanCallDart()); |
1111 | ASSERT(!arguments_descriptor.IsNull() && (arguments_descriptor.Length() > 0)); |
1112 | const ArgumentsDescriptor args_desc(arguments_descriptor); |
1113 | const MegamorphicCache& cache = MegamorphicCache::ZoneHandle( |
1114 | zone(), |
1115 | MegamorphicCacheTable::Lookup(thread(), name, arguments_descriptor)); |
1116 | __ Comment("MegamorphicCall" ); |
1117 | // Load receiver into RDX. |
1118 | __ movq(RDX, compiler::Address(RSP, (args_desc.Count() - 1) * kWordSize)); |
1119 | |
1120 | // Use same code pattern as instance call so it can be parsed by code patcher. |
1121 | if (FLAG_precompiled_mode) { |
1122 | if (FLAG_use_bare_instructions) { |
1123 | // The AOT runtime will replace the slot in the object pool with the |
1124 | // entrypoint address - see clustered_snapshot.cc. |
1125 | __ LoadUniqueObject(RCX, StubCode::MegamorphicCall()); |
1126 | } else { |
1127 | __ LoadUniqueObject(CODE_REG, StubCode::MegamorphicCall()); |
1128 | __ movq(RCX, compiler::FieldAddress(CODE_REG, |
1129 | Code::entry_point_offset( |
1130 | Code::EntryKind::kMonomorphic))); |
1131 | } |
1132 | __ LoadUniqueObject(RBX, cache); |
1133 | __ call(RCX); |
1134 | } else { |
1135 | __ LoadUniqueObject(RBX, cache); |
1136 | __ LoadUniqueObject(CODE_REG, StubCode::MegamorphicCall()); |
1137 | __ call(compiler::FieldAddress( |
1138 | CODE_REG, Code::entry_point_offset(Code::EntryKind::kMonomorphic))); |
1139 | } |
1140 | |
1141 | RecordSafepoint(locs, slow_path_argument_count); |
1142 | const intptr_t deopt_id_after = DeoptId::ToDeoptAfter(deopt_id); |
1143 | if (FLAG_precompiled_mode) { |
1144 | // Megamorphic calls may occur in slow path stubs. |
1145 | // If valid use try_index argument. |
1146 | if (try_index == kInvalidTryIndex) { |
1147 | try_index = CurrentTryIndex(); |
1148 | } |
1149 | AddDescriptor(PcDescriptorsLayout::kOther, assembler()->CodeSize(), |
1150 | DeoptId::kNone, token_pos, try_index); |
1151 | } else if (is_optimizing()) { |
1152 | AddCurrentDescriptor(PcDescriptorsLayout::kOther, DeoptId::kNone, |
1153 | token_pos); |
1154 | AddDeoptIndexAtCall(deopt_id_after); |
1155 | } else { |
1156 | AddCurrentDescriptor(PcDescriptorsLayout::kOther, DeoptId::kNone, |
1157 | token_pos); |
1158 | // Add deoptimization continuation point after the call and before the |
1159 | // arguments are removed. |
1160 | AddCurrentDescriptor(PcDescriptorsLayout::kDeopt, deopt_id_after, |
1161 | token_pos); |
1162 | } |
1163 | RecordCatchEntryMoves(pending_deoptimization_env_, try_index); |
1164 | __ Drop(args_desc.SizeWithTypeArgs(), RCX); |
1165 | } |
1166 | |
1167 | void FlowGraphCompiler::EmitInstanceCallAOT(const ICData& ic_data, |
1168 | intptr_t deopt_id, |
1169 | TokenPosition token_pos, |
1170 | LocationSummary* locs, |
1171 | Code::EntryKind entry_kind, |
1172 | bool receiver_can_be_smi) { |
1173 | ASSERT(CanCallDart()); |
1174 | ASSERT(entry_kind == Code::EntryKind::kNormal || |
1175 | entry_kind == Code::EntryKind::kUnchecked); |
1176 | ASSERT(ic_data.NumArgsTested() == 1); |
1177 | const Code& initial_stub = StubCode::SwitchableCallMiss(); |
1178 | const char* switchable_call_mode = "smiable" ; |
1179 | if (!receiver_can_be_smi) { |
1180 | switchable_call_mode = "non-smi" ; |
1181 | ic_data.set_receiver_cannot_be_smi(true); |
1182 | } |
1183 | const UnlinkedCall& data = |
1184 | UnlinkedCall::ZoneHandle(zone(), ic_data.AsUnlinkedCall()); |
1185 | |
1186 | __ Comment("InstanceCallAOT (%s)" , switchable_call_mode); |
1187 | __ movq(RDX, compiler::Address( |
1188 | RSP, (ic_data.SizeWithoutTypeArgs() - 1) * kWordSize)); |
1189 | if (FLAG_precompiled_mode && FLAG_use_bare_instructions) { |
1190 | // The AOT runtime will replace the slot in the object pool with the |
1191 | // entrypoint address - see clustered_snapshot.cc. |
1192 | __ LoadUniqueObject(RCX, initial_stub); |
1193 | } else { |
1194 | const intptr_t entry_point_offset = |
1195 | entry_kind == Code::EntryKind::kNormal |
1196 | ? Code::entry_point_offset(Code::EntryKind::kMonomorphic) |
1197 | : Code::entry_point_offset(Code::EntryKind::kMonomorphicUnchecked); |
1198 | __ LoadUniqueObject(CODE_REG, initial_stub); |
1199 | __ movq(RCX, compiler::FieldAddress(CODE_REG, entry_point_offset)); |
1200 | } |
1201 | __ LoadUniqueObject(RBX, data); |
1202 | __ call(RCX); |
1203 | |
1204 | EmitCallsiteMetadata(token_pos, deopt_id, PcDescriptorsLayout::kOther, locs); |
1205 | __ Drop(ic_data.SizeWithTypeArgs(), RCX); |
1206 | } |
1207 | |
1208 | void FlowGraphCompiler::EmitOptimizedStaticCall( |
1209 | const Function& function, |
1210 | const Array& arguments_descriptor, |
1211 | intptr_t size_with_type_args, |
1212 | intptr_t deopt_id, |
1213 | TokenPosition token_pos, |
1214 | LocationSummary* locs, |
1215 | Code::EntryKind entry_kind) { |
1216 | ASSERT(CanCallDart()); |
1217 | ASSERT(!function.IsClosureFunction()); |
1218 | if (function.HasOptionalParameters() || function.IsGeneric()) { |
1219 | __ LoadObject(R10, arguments_descriptor); |
1220 | } else { |
1221 | if (!(FLAG_precompiled_mode && FLAG_use_bare_instructions)) { |
1222 | __ xorl(R10, R10); // GC safe smi zero because of stub. |
1223 | } |
1224 | } |
1225 | // Do not use the code from the function, but let the code be patched so that |
1226 | // we can record the outgoing edges to other code. |
1227 | GenerateStaticDartCall(deopt_id, token_pos, PcDescriptorsLayout::kOther, locs, |
1228 | function, entry_kind); |
1229 | __ Drop(size_with_type_args, RCX); |
1230 | } |
1231 | |
1232 | void FlowGraphCompiler::EmitDispatchTableCall( |
1233 | Register cid_reg, |
1234 | int32_t selector_offset, |
1235 | const Array& arguments_descriptor) { |
1236 | ASSERT(CanCallDart()); |
1237 | const Register table_reg = RAX; |
1238 | ASSERT(cid_reg != table_reg); |
1239 | ASSERT(cid_reg != ARGS_DESC_REG); |
1240 | if (!arguments_descriptor.IsNull()) { |
1241 | __ LoadObject(ARGS_DESC_REG, arguments_descriptor); |
1242 | } |
1243 | const intptr_t offset = (selector_offset - DispatchTable::OriginElement()) * |
1244 | compiler::target::kWordSize; |
1245 | __ LoadDispatchTable(table_reg); |
1246 | __ call(compiler::Address(table_reg, cid_reg, TIMES_8, offset)); |
1247 | } |
1248 | |
1249 | Condition FlowGraphCompiler::EmitEqualityRegConstCompare( |
1250 | Register reg, |
1251 | const Object& obj, |
1252 | bool needs_number_check, |
1253 | TokenPosition token_pos, |
1254 | intptr_t deopt_id) { |
1255 | ASSERT(!needs_number_check || (!obj.IsMint() && !obj.IsDouble())); |
1256 | |
1257 | if (obj.IsSmi() && (Smi::Cast(obj).Value() == 0)) { |
1258 | ASSERT(!needs_number_check); |
1259 | __ testq(reg, reg); |
1260 | return EQUAL; |
1261 | } |
1262 | |
1263 | if (needs_number_check) { |
1264 | __ pushq(reg); |
1265 | __ PushObject(obj); |
1266 | if (is_optimizing()) { |
1267 | __ CallPatchable(StubCode::OptimizedIdenticalWithNumberCheck()); |
1268 | } else { |
1269 | __ CallPatchable(StubCode::UnoptimizedIdenticalWithNumberCheck()); |
1270 | } |
1271 | AddCurrentDescriptor(PcDescriptorsLayout::kRuntimeCall, deopt_id, |
1272 | token_pos); |
1273 | // Stub returns result in flags (result of a cmpq, we need ZF computed). |
1274 | __ popq(reg); // Discard constant. |
1275 | __ popq(reg); // Restore 'reg'. |
1276 | } else { |
1277 | __ CompareObject(reg, obj); |
1278 | } |
1279 | return EQUAL; |
1280 | } |
1281 | |
1282 | Condition FlowGraphCompiler::EmitEqualityRegRegCompare(Register left, |
1283 | Register right, |
1284 | bool needs_number_check, |
1285 | TokenPosition token_pos, |
1286 | intptr_t deopt_id) { |
1287 | if (needs_number_check) { |
1288 | __ pushq(left); |
1289 | __ pushq(right); |
1290 | if (is_optimizing()) { |
1291 | __ CallPatchable(StubCode::OptimizedIdenticalWithNumberCheck()); |
1292 | } else { |
1293 | __ CallPatchable(StubCode::UnoptimizedIdenticalWithNumberCheck()); |
1294 | } |
1295 | AddCurrentDescriptor(PcDescriptorsLayout::kRuntimeCall, deopt_id, |
1296 | token_pos); |
1297 | // Stub returns result in flags (result of a cmpq, we need ZF computed). |
1298 | __ popq(right); |
1299 | __ popq(left); |
1300 | } else { |
1301 | __ CompareRegisters(left, right); |
1302 | } |
1303 | return EQUAL; |
1304 | } |
1305 | |
1306 | Condition FlowGraphCompiler::EmitBoolTest(Register value, |
1307 | BranchLabels labels, |
1308 | bool invert) { |
1309 | __ Comment("BoolTest" ); |
1310 | __ testq(value, compiler::Immediate( |
1311 | compiler::target::ObjectAlignment::kBoolValueMask)); |
1312 | return invert ? NOT_EQUAL : EQUAL; |
1313 | } |
1314 | |
1315 | // This function must be in sync with FlowGraphCompiler::RecordSafepoint and |
1316 | // FlowGraphCompiler::SlowPathEnvironmentFor. |
1317 | void FlowGraphCompiler::SaveLiveRegisters(LocationSummary* locs) { |
1318 | #if defined(DEBUG) |
1319 | locs->CheckWritableInputs(); |
1320 | ClobberDeadTempRegisters(locs); |
1321 | #endif |
1322 | |
1323 | // TODO(vegorov): avoid saving non-volatile registers. |
1324 | __ PushRegisters(locs->live_registers()->cpu_registers(), |
1325 | locs->live_registers()->fpu_registers()); |
1326 | } |
1327 | |
1328 | void FlowGraphCompiler::RestoreLiveRegisters(LocationSummary* locs) { |
1329 | __ PopRegisters(locs->live_registers()->cpu_registers(), |
1330 | locs->live_registers()->fpu_registers()); |
1331 | } |
1332 | |
1333 | #if defined(DEBUG) |
1334 | void FlowGraphCompiler::ClobberDeadTempRegisters(LocationSummary* locs) { |
1335 | // Clobber temporaries that have not been manually preserved. |
1336 | for (intptr_t i = 0; i < locs->temp_count(); ++i) { |
1337 | Location tmp = locs->temp(i); |
1338 | // TODO(zerny): clobber non-live temporary FPU registers. |
1339 | if (tmp.IsRegister() && |
1340 | !locs->live_registers()->ContainsRegister(tmp.reg())) { |
1341 | __ movq(tmp.reg(), compiler::Immediate(0xf7)); |
1342 | } |
1343 | } |
1344 | } |
1345 | #endif |
1346 | |
1347 | Register FlowGraphCompiler::EmitTestCidRegister() { |
1348 | return RDI; |
1349 | } |
1350 | |
1351 | void FlowGraphCompiler::EmitTestAndCallLoadReceiver( |
1352 | intptr_t count_without_type_args, |
1353 | const Array& arguments_descriptor) { |
1354 | __ Comment("EmitTestAndCall" ); |
1355 | // Load receiver into RAX. |
1356 | __ movq(RAX, |
1357 | compiler::Address(RSP, (count_without_type_args - 1) * kWordSize)); |
1358 | __ LoadObject(R10, arguments_descriptor); |
1359 | } |
1360 | |
1361 | void FlowGraphCompiler::EmitTestAndCallSmiBranch(compiler::Label* label, |
1362 | bool if_smi) { |
1363 | __ testq(RAX, compiler::Immediate(kSmiTagMask)); |
1364 | // Jump if receiver is (not) Smi. |
1365 | __ j(if_smi ? ZERO : NOT_ZERO, label); |
1366 | } |
1367 | |
1368 | void FlowGraphCompiler::EmitTestAndCallLoadCid(Register class_id_reg) { |
1369 | ASSERT(class_id_reg != RAX); |
1370 | __ LoadClassId(class_id_reg, RAX); |
1371 | } |
1372 | |
1373 | #undef __ |
1374 | #define __ assembler-> |
1375 | |
1376 | int FlowGraphCompiler::EmitTestAndCallCheckCid(compiler::Assembler* assembler, |
1377 | compiler::Label* label, |
1378 | Register class_id_reg, |
1379 | const CidRangeValue& range, |
1380 | int bias, |
1381 | bool jump_on_miss) { |
1382 | // Note of WARNING: Due to smaller instruction encoding we use the 32-bit |
1383 | // instructions on x64, which means the compare instruction has to be |
1384 | // 32-bit (since the subtraction instruction is as well). |
1385 | intptr_t cid_start = range.cid_start; |
1386 | if (range.IsSingleCid()) { |
1387 | __ cmpl(class_id_reg, compiler::Immediate(cid_start - bias)); |
1388 | __ BranchIf(jump_on_miss ? NOT_EQUAL : EQUAL, label); |
1389 | } else { |
1390 | __ addl(class_id_reg, compiler::Immediate(bias - cid_start)); |
1391 | bias = cid_start; |
1392 | __ cmpl(class_id_reg, compiler::Immediate(range.Extent())); |
1393 | __ BranchIf(jump_on_miss ? UNSIGNED_GREATER : UNSIGNED_LESS_EQUAL, label); |
1394 | } |
1395 | return bias; |
1396 | } |
1397 | |
1398 | #undef __ |
1399 | #define __ assembler()-> |
1400 | |
1401 | void FlowGraphCompiler::EmitMove(Location destination, |
1402 | Location source, |
1403 | TemporaryRegisterAllocator* tmp) { |
1404 | if (destination.Equals(source)) return; |
1405 | |
1406 | if (source.IsRegister()) { |
1407 | if (destination.IsRegister()) { |
1408 | __ movq(destination.reg(), source.reg()); |
1409 | } else { |
1410 | ASSERT(destination.IsStackSlot()); |
1411 | __ movq(LocationToStackSlotAddress(destination), source.reg()); |
1412 | } |
1413 | } else if (source.IsStackSlot()) { |
1414 | if (destination.IsRegister()) { |
1415 | __ movq(destination.reg(), LocationToStackSlotAddress(source)); |
1416 | } else if (destination.IsFpuRegister()) { |
1417 | // 32-bit float |
1418 | __ movq(TMP, LocationToStackSlotAddress(source)); |
1419 | __ movq(destination.fpu_reg(), TMP); |
1420 | } else { |
1421 | ASSERT(destination.IsStackSlot()); |
1422 | __ MoveMemoryToMemory(LocationToStackSlotAddress(destination), |
1423 | LocationToStackSlotAddress(source)); |
1424 | } |
1425 | } else if (source.IsFpuRegister()) { |
1426 | if (destination.IsFpuRegister()) { |
1427 | // Optimization manual recommends using MOVAPS for register |
1428 | // to register moves. |
1429 | __ movaps(destination.fpu_reg(), source.fpu_reg()); |
1430 | } else { |
1431 | if (destination.IsDoubleStackSlot()) { |
1432 | __ movsd(LocationToStackSlotAddress(destination), source.fpu_reg()); |
1433 | } else { |
1434 | ASSERT(destination.IsQuadStackSlot()); |
1435 | __ movups(LocationToStackSlotAddress(destination), source.fpu_reg()); |
1436 | } |
1437 | } |
1438 | } else if (source.IsDoubleStackSlot()) { |
1439 | if (destination.IsFpuRegister()) { |
1440 | __ movsd(destination.fpu_reg(), LocationToStackSlotAddress(source)); |
1441 | } else { |
1442 | ASSERT(destination.IsDoubleStackSlot() || |
1443 | destination.IsStackSlot() /*32-bit float*/); |
1444 | __ movsd(FpuTMP, LocationToStackSlotAddress(source)); |
1445 | __ movsd(LocationToStackSlotAddress(destination), FpuTMP); |
1446 | } |
1447 | } else if (source.IsQuadStackSlot()) { |
1448 | if (destination.IsFpuRegister()) { |
1449 | __ movups(destination.fpu_reg(), LocationToStackSlotAddress(source)); |
1450 | } else { |
1451 | ASSERT(destination.IsQuadStackSlot()); |
1452 | __ movups(FpuTMP, LocationToStackSlotAddress(source)); |
1453 | __ movups(LocationToStackSlotAddress(destination), FpuTMP); |
1454 | } |
1455 | } else { |
1456 | ASSERT(!source.IsInvalid()); |
1457 | ASSERT(source.IsConstant()); |
1458 | if (destination.IsFpuRegister() || destination.IsDoubleStackSlot()) { |
1459 | Register scratch = tmp->AllocateTemporary(); |
1460 | source.constant_instruction()->EmitMoveToLocation(this, destination, |
1461 | scratch); |
1462 | tmp->ReleaseTemporary(); |
1463 | } else { |
1464 | source.constant_instruction()->EmitMoveToLocation(this, destination); |
1465 | } |
1466 | } |
1467 | } |
1468 | |
1469 | void FlowGraphCompiler::EmitNativeMoveArchitecture( |
1470 | const compiler::ffi::NativeLocation& destination, |
1471 | const compiler::ffi::NativeLocation& source) { |
1472 | const auto& src_type = source.payload_type(); |
1473 | const auto& dst_type = destination.payload_type(); |
1474 | ASSERT(src_type.IsFloat() == dst_type.IsFloat()); |
1475 | ASSERT(src_type.IsInt() == dst_type.IsInt()); |
1476 | ASSERT(src_type.IsSigned() == dst_type.IsSigned()); |
1477 | ASSERT(src_type.IsFundamental()); |
1478 | ASSERT(dst_type.IsFundamental()); |
1479 | const intptr_t src_size = src_type.SizeInBytes(); |
1480 | const intptr_t dst_size = dst_type.SizeInBytes(); |
1481 | const bool sign_or_zero_extend = dst_size > src_size; |
1482 | |
1483 | if (source.IsRegisters()) { |
1484 | const auto& src = source.AsRegisters(); |
1485 | ASSERT(src.num_regs() == 1); |
1486 | const auto src_reg = src.reg_at(0); |
1487 | |
1488 | if (destination.IsRegisters()) { |
1489 | const auto& dst = destination.AsRegisters(); |
1490 | ASSERT(dst.num_regs() == 1); |
1491 | const auto dst_reg = dst.reg_at(0); |
1492 | if (!sign_or_zero_extend) { |
1493 | switch (dst_size) { |
1494 | case 8: |
1495 | __ movq(dst_reg, src_reg); |
1496 | return; |
1497 | case 4: |
1498 | __ movl(dst_reg, src_reg); |
1499 | return; |
1500 | default: |
1501 | UNIMPLEMENTED(); |
1502 | } |
1503 | } else { |
1504 | switch (src_type.AsFundamental().representation()) { |
1505 | case compiler::ffi::kInt8: // Sign extend operand. |
1506 | __ movsxb(dst_reg, src_reg); |
1507 | return; |
1508 | case compiler::ffi::kInt16: |
1509 | __ movsxw(dst_reg, src_reg); |
1510 | return; |
1511 | case compiler::ffi::kUint8: // Zero extend operand. |
1512 | __ movzxb(dst_reg, src_reg); |
1513 | return; |
1514 | case compiler::ffi::kUint16: |
1515 | __ movzxw(dst_reg, src_reg); |
1516 | return; |
1517 | default: |
1518 | // 32 to 64 bit is covered in IL by Representation conversions. |
1519 | UNIMPLEMENTED(); |
1520 | } |
1521 | } |
1522 | |
1523 | } else if (destination.IsFpuRegisters()) { |
1524 | // Fpu Registers should only contain doubles and registers only ints. |
1525 | UNIMPLEMENTED(); |
1526 | |
1527 | } else { |
1528 | ASSERT(destination.IsStack()); |
1529 | const auto& dst = destination.AsStack(); |
1530 | const auto dst_addr = NativeLocationToStackSlotAddress(dst); |
1531 | ASSERT(!sign_or_zero_extend); |
1532 | switch (dst_size) { |
1533 | case 8: |
1534 | __ movq(dst_addr, src_reg); |
1535 | return; |
1536 | case 4: |
1537 | __ movl(dst_addr, src_reg); |
1538 | return; |
1539 | case 2: |
1540 | __ movw(dst_addr, src_reg); |
1541 | return; |
1542 | case 1: |
1543 | __ movb(dst_addr, src_reg); |
1544 | return; |
1545 | default: |
1546 | UNREACHABLE(); |
1547 | } |
1548 | } |
1549 | |
1550 | } else if (source.IsFpuRegisters()) { |
1551 | const auto& src = source.AsFpuRegisters(); |
1552 | // We have not implemented conversions here, use IL convert instructions. |
1553 | ASSERT(src_type.Equals(dst_type)); |
1554 | |
1555 | if (destination.IsRegisters()) { |
1556 | // Fpu Registers should only contain doubles and registers only ints. |
1557 | UNIMPLEMENTED(); |
1558 | |
1559 | } else if (destination.IsFpuRegisters()) { |
1560 | const auto& dst = destination.AsFpuRegisters(); |
1561 | // Optimization manual recommends using MOVAPS for register |
1562 | // to register moves. |
1563 | __ movaps(dst.fpu_reg(), src.fpu_reg()); |
1564 | |
1565 | } else { |
1566 | ASSERT(destination.IsStack()); |
1567 | ASSERT(src_type.IsFloat()); |
1568 | const auto& dst = destination.AsStack(); |
1569 | const auto dst_addr = NativeLocationToStackSlotAddress(dst); |
1570 | switch (dst_size) { |
1571 | case 8: |
1572 | __ movsd(dst_addr, src.fpu_reg()); |
1573 | return; |
1574 | case 4: |
1575 | __ movss(dst_addr, src.fpu_reg()); |
1576 | return; |
1577 | default: |
1578 | UNREACHABLE(); |
1579 | } |
1580 | } |
1581 | |
1582 | } else { |
1583 | ASSERT(source.IsStack()); |
1584 | const auto& src = source.AsStack(); |
1585 | const auto src_addr = NativeLocationToStackSlotAddress(src); |
1586 | if (destination.IsRegisters()) { |
1587 | const auto& dst = destination.AsRegisters(); |
1588 | ASSERT(dst.num_regs() == 1); |
1589 | const auto dst_reg = dst.reg_at(0); |
1590 | if (!sign_or_zero_extend) { |
1591 | switch (dst_size) { |
1592 | case 8: |
1593 | __ movq(dst_reg, src_addr); |
1594 | return; |
1595 | case 4: |
1596 | __ movl(dst_reg, src_addr); |
1597 | return; |
1598 | default: |
1599 | UNIMPLEMENTED(); |
1600 | } |
1601 | } else { |
1602 | switch (src_type.AsFundamental().representation()) { |
1603 | case compiler::ffi::kInt8: // Sign extend operand. |
1604 | __ movsxb(dst_reg, src_addr); |
1605 | return; |
1606 | case compiler::ffi::kInt16: |
1607 | __ movsxw(dst_reg, src_addr); |
1608 | return; |
1609 | case compiler::ffi::kUint8: // Zero extend operand. |
1610 | __ movzxb(dst_reg, src_addr); |
1611 | return; |
1612 | case compiler::ffi::kUint16: |
1613 | __ movzxw(dst_reg, src_addr); |
1614 | return; |
1615 | default: |
1616 | // 32 to 64 bit is covered in IL by Representation conversions. |
1617 | UNIMPLEMENTED(); |
1618 | } |
1619 | } |
1620 | |
1621 | } else if (destination.IsFpuRegisters()) { |
1622 | ASSERT(src_type.Equals(dst_type)); |
1623 | ASSERT(src_type.IsFloat()); |
1624 | const auto& dst = destination.AsFpuRegisters(); |
1625 | switch (dst_size) { |
1626 | case 8: |
1627 | __ movsd(dst.fpu_reg(), src_addr); |
1628 | return; |
1629 | case 4: |
1630 | __ movss(dst.fpu_reg(), src_addr); |
1631 | return; |
1632 | default: |
1633 | UNREACHABLE(); |
1634 | } |
1635 | |
1636 | } else { |
1637 | ASSERT(destination.IsStack()); |
1638 | UNREACHABLE(); |
1639 | } |
1640 | } |
1641 | } |
1642 | |
1643 | void FlowGraphCompiler::LoadBSSEntry(BSS::Relocation relocation, |
1644 | Register dst, |
1645 | Register tmp) { |
1646 | compiler::Label skip_reloc; |
1647 | __ jmp(&skip_reloc); |
1648 | InsertBSSRelocation(relocation); |
1649 | const intptr_t reloc_end = __ CodeSize(); |
1650 | __ Bind(&skip_reloc); |
1651 | |
1652 | const intptr_t kLeaqLength = 7; |
1653 | __ leaq(dst, compiler::Address::AddressRIPRelative( |
1654 | -kLeaqLength - compiler::target::kWordSize)); |
1655 | ASSERT((__ CodeSize() - reloc_end) == kLeaqLength); |
1656 | |
1657 | // dst holds the address of the relocation. |
1658 | __ movq(tmp, compiler::Address(dst, 0)); |
1659 | |
1660 | // tmp holds the relocation itself: dst - bss_start. |
1661 | // dst = dst + (bss_start - dst) = bss_start |
1662 | __ addq(dst, tmp); |
1663 | |
1664 | // dst holds the start of the BSS section. |
1665 | // Load the routine. |
1666 | __ movq(dst, compiler::Address(dst, 0)); |
1667 | } |
1668 | |
1669 | #undef __ |
1670 | #define __ compiler_->assembler()-> |
1671 | |
1672 | void ParallelMoveResolver::EmitSwap(int index) { |
1673 | MoveOperands* move = moves_[index]; |
1674 | const Location source = move->src(); |
1675 | const Location destination = move->dest(); |
1676 | |
1677 | if (source.IsRegister() && destination.IsRegister()) { |
1678 | __ xchgq(destination.reg(), source.reg()); |
1679 | } else if (source.IsRegister() && destination.IsStackSlot()) { |
1680 | Exchange(source.reg(), LocationToStackSlotAddress(destination)); |
1681 | } else if (source.IsStackSlot() && destination.IsRegister()) { |
1682 | Exchange(destination.reg(), LocationToStackSlotAddress(source)); |
1683 | } else if (source.IsStackSlot() && destination.IsStackSlot()) { |
1684 | Exchange(LocationToStackSlotAddress(destination), |
1685 | LocationToStackSlotAddress(source)); |
1686 | } else if (source.IsFpuRegister() && destination.IsFpuRegister()) { |
1687 | __ movaps(FpuTMP, source.fpu_reg()); |
1688 | __ movaps(source.fpu_reg(), destination.fpu_reg()); |
1689 | __ movaps(destination.fpu_reg(), FpuTMP); |
1690 | } else if (source.IsFpuRegister() || destination.IsFpuRegister()) { |
1691 | ASSERT(destination.IsDoubleStackSlot() || destination.IsQuadStackSlot() || |
1692 | source.IsDoubleStackSlot() || source.IsQuadStackSlot()); |
1693 | bool double_width = |
1694 | destination.IsDoubleStackSlot() || source.IsDoubleStackSlot(); |
1695 | XmmRegister reg = |
1696 | source.IsFpuRegister() ? source.fpu_reg() : destination.fpu_reg(); |
1697 | compiler::Address slot_address = |
1698 | source.IsFpuRegister() ? LocationToStackSlotAddress(destination) |
1699 | : LocationToStackSlotAddress(source); |
1700 | |
1701 | if (double_width) { |
1702 | __ movsd(FpuTMP, slot_address); |
1703 | __ movsd(slot_address, reg); |
1704 | } else { |
1705 | __ movups(FpuTMP, slot_address); |
1706 | __ movups(slot_address, reg); |
1707 | } |
1708 | __ movaps(reg, FpuTMP); |
1709 | } else if (source.IsDoubleStackSlot() && destination.IsDoubleStackSlot()) { |
1710 | const compiler::Address& source_slot_address = |
1711 | LocationToStackSlotAddress(source); |
1712 | const compiler::Address& destination_slot_address = |
1713 | LocationToStackSlotAddress(destination); |
1714 | |
1715 | ScratchFpuRegisterScope ensure_scratch(this, FpuTMP); |
1716 | __ movsd(FpuTMP, source_slot_address); |
1717 | __ movsd(ensure_scratch.reg(), destination_slot_address); |
1718 | __ movsd(destination_slot_address, FpuTMP); |
1719 | __ movsd(source_slot_address, ensure_scratch.reg()); |
1720 | } else if (source.IsQuadStackSlot() && destination.IsQuadStackSlot()) { |
1721 | const compiler::Address& source_slot_address = |
1722 | LocationToStackSlotAddress(source); |
1723 | const compiler::Address& destination_slot_address = |
1724 | LocationToStackSlotAddress(destination); |
1725 | |
1726 | ScratchFpuRegisterScope ensure_scratch(this, FpuTMP); |
1727 | __ movups(FpuTMP, source_slot_address); |
1728 | __ movups(ensure_scratch.reg(), destination_slot_address); |
1729 | __ movups(destination_slot_address, FpuTMP); |
1730 | __ movups(source_slot_address, ensure_scratch.reg()); |
1731 | } else { |
1732 | UNREACHABLE(); |
1733 | } |
1734 | |
1735 | // The swap of source and destination has executed a move from source to |
1736 | // destination. |
1737 | move->Eliminate(); |
1738 | |
1739 | // Any unperformed (including pending) move with a source of either |
1740 | // this move's source or destination needs to have their source |
1741 | // changed to reflect the state of affairs after the swap. |
1742 | for (int i = 0; i < moves_.length(); ++i) { |
1743 | const MoveOperands& other_move = *moves_[i]; |
1744 | if (other_move.Blocks(source)) { |
1745 | moves_[i]->set_src(destination); |
1746 | } else if (other_move.Blocks(destination)) { |
1747 | moves_[i]->set_src(source); |
1748 | } |
1749 | } |
1750 | } |
1751 | |
1752 | void ParallelMoveResolver::MoveMemoryToMemory(const compiler::Address& dst, |
1753 | const compiler::Address& src) { |
1754 | __ MoveMemoryToMemory(dst, src); |
1755 | } |
1756 | |
1757 | void ParallelMoveResolver::Exchange(Register reg, |
1758 | const compiler::Address& mem) { |
1759 | __ Exchange(reg, mem); |
1760 | } |
1761 | |
1762 | void ParallelMoveResolver::Exchange(const compiler::Address& mem1, |
1763 | const compiler::Address& mem2) { |
1764 | __ Exchange(mem1, mem2); |
1765 | } |
1766 | |
1767 | void ParallelMoveResolver::Exchange(Register reg, |
1768 | Register base_reg, |
1769 | intptr_t stack_offset) { |
1770 | UNREACHABLE(); |
1771 | } |
1772 | |
1773 | void ParallelMoveResolver::Exchange(Register base_reg1, |
1774 | intptr_t stack_offset1, |
1775 | Register base_reg2, |
1776 | intptr_t stack_offset2) { |
1777 | UNREACHABLE(); |
1778 | } |
1779 | |
1780 | void ParallelMoveResolver::SpillScratch(Register reg) { |
1781 | __ pushq(reg); |
1782 | } |
1783 | |
1784 | void ParallelMoveResolver::RestoreScratch(Register reg) { |
1785 | __ popq(reg); |
1786 | } |
1787 | |
1788 | void ParallelMoveResolver::SpillFpuScratch(FpuRegister reg) { |
1789 | __ AddImmediate(RSP, compiler::Immediate(-kFpuRegisterSize)); |
1790 | __ movups(compiler::Address(RSP, 0), reg); |
1791 | } |
1792 | |
1793 | void ParallelMoveResolver::RestoreFpuScratch(FpuRegister reg) { |
1794 | __ movups(reg, compiler::Address(RSP, 0)); |
1795 | __ AddImmediate(RSP, compiler::Immediate(kFpuRegisterSize)); |
1796 | } |
1797 | |
1798 | #undef __ |
1799 | |
1800 | } // namespace dart |
1801 | |
1802 | #endif // defined(TARGET_ARCH_X64) |
1803 | |