1 /*
2 * Copyright (C) 2016 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include "code_generator_arm_vixl.h"
18
19 #include "arch/arm/asm_support_arm.h"
20 #include "arch/arm/instruction_set_features_arm.h"
21 #include "art_method.h"
22 #include "base/bit_utils.h"
23 #include "base/bit_utils_iterator.h"
24 #include "class_table.h"
25 #include "code_generator_utils.h"
26 #include "common_arm.h"
27 #include "compiled_method.h"
28 #include "entrypoints/quick/quick_entrypoints.h"
29 #include "gc/accounting/card_table.h"
30 #include "gc/space/image_space.h"
31 #include "heap_poisoning.h"
32 #include "intrinsics.h"
33 #include "intrinsics_arm_vixl.h"
34 #include "linker/linker_patch.h"
35 #include "mirror/array-inl.h"
36 #include "mirror/class-inl.h"
37 #include "thread.h"
38 #include "utils/arm/assembler_arm_vixl.h"
39 #include "utils/arm/managed_register_arm.h"
40 #include "utils/assembler.h"
41 #include "utils/stack_checks.h"
42
43 namespace art {
44 namespace arm {
45
46 namespace vixl32 = vixl::aarch32;
47 using namespace vixl32; // NOLINT(build/namespaces)
48
49 using helpers::DRegisterFrom;
50 using helpers::DWARFReg;
51 using helpers::HighRegisterFrom;
52 using helpers::InputDRegisterAt;
53 using helpers::InputOperandAt;
54 using helpers::InputRegister;
55 using helpers::InputRegisterAt;
56 using helpers::InputSRegisterAt;
57 using helpers::InputVRegister;
58 using helpers::InputVRegisterAt;
59 using helpers::Int32ConstantFrom;
60 using helpers::Int64ConstantFrom;
61 using helpers::LocationFrom;
62 using helpers::LowRegisterFrom;
63 using helpers::LowSRegisterFrom;
64 using helpers::OperandFrom;
65 using helpers::OutputRegister;
66 using helpers::OutputSRegister;
67 using helpers::OutputVRegister;
68 using helpers::RegisterFrom;
69 using helpers::SRegisterFrom;
70 using helpers::Uint64ConstantFrom;
71
72 using vixl::ExactAssemblyScope;
73 using vixl::CodeBufferCheckScope;
74
75 using RegisterList = vixl32::RegisterList;
76
ExpectedPairLayout(Location location)77 static bool ExpectedPairLayout(Location location) {
78 // We expected this for both core and fpu register pairs.
79 return ((location.low() & 1) == 0) && (location.low() + 1 == location.high());
80 }
81 // Use a local definition to prevent copying mistakes.
82 static constexpr size_t kArmWordSize = static_cast<size_t>(kArmPointerSize);
83 static constexpr size_t kArmBitsPerWord = kArmWordSize * kBitsPerByte;
84 static constexpr uint32_t kPackedSwitchCompareJumpThreshold = 7;
85
86 // Reference load (except object array loads) is using LDR Rt, [Rn, #offset] which can handle
87 // offset < 4KiB. For offsets >= 4KiB, the load shall be emitted as two or more instructions.
88 // For the Baker read barrier implementation using link-time generated thunks we need to split
89 // the offset explicitly.
90 constexpr uint32_t kReferenceLoadMinFarOffset = 4 * KB;
91
92 // Using a base helps identify when we hit Marking Register check breakpoints.
93 constexpr int kMarkingRegisterCheckBreakCodeBaseCode = 0x10;
94
95 #ifdef __
96 #error "ARM Codegen VIXL macro-assembler macro already defined."
97 #endif
98
99 // NOLINT on __ macro to suppress wrong warning/fix (misc-macro-parentheses) from clang-tidy.
100 #define __ down_cast<CodeGeneratorARMVIXL*>(codegen)->GetVIXLAssembler()-> // NOLINT
101 #define QUICK_ENTRY_POINT(x) QUICK_ENTRYPOINT_OFFSET(kArmPointerSize, x).Int32Value()
102
103 // Marker that code is yet to be, and must, be implemented.
104 #define TODO_VIXL32(level) LOG(level) << __PRETTY_FUNCTION__ << " unimplemented "
105
CanEmitNarrowLdr(vixl32::Register rt,vixl32::Register rn,uint32_t offset)106 static inline bool CanEmitNarrowLdr(vixl32::Register rt, vixl32::Register rn, uint32_t offset) {
107 return rt.IsLow() && rn.IsLow() && offset < 32u;
108 }
109
110 class EmitAdrCode {
111 public:
EmitAdrCode(ArmVIXLMacroAssembler * assembler,vixl32::Register rd,vixl32::Label * label)112 EmitAdrCode(ArmVIXLMacroAssembler* assembler, vixl32::Register rd, vixl32::Label* label)
113 : assembler_(assembler), rd_(rd), label_(label) {
114 DCHECK(!assembler->AllowMacroInstructions()); // In ExactAssemblyScope.
115 adr_location_ = assembler->GetCursorOffset();
116 assembler->adr(EncodingSize(Wide), rd, label);
117 }
118
~EmitAdrCode()119 ~EmitAdrCode() {
120 DCHECK(label_->IsBound());
121 // The ADR emitted by the assembler does not set the Thumb mode bit we need.
122 // TODO: Maybe extend VIXL to allow ADR for return address?
123 uint8_t* raw_adr = assembler_->GetBuffer()->GetOffsetAddress<uint8_t*>(adr_location_);
124 // Expecting ADR encoding T3 with `(offset & 1) == 0`.
125 DCHECK_EQ(raw_adr[1] & 0xfbu, 0xf2u); // Check bits 24-31, except 26.
126 DCHECK_EQ(raw_adr[0] & 0xffu, 0x0fu); // Check bits 16-23.
127 DCHECK_EQ(raw_adr[3] & 0x8fu, rd_.GetCode()); // Check bits 8-11 and 15.
128 DCHECK_EQ(raw_adr[2] & 0x01u, 0x00u); // Check bit 0, i.e. the `offset & 1`.
129 // Add the Thumb mode bit.
130 raw_adr[2] |= 0x01u;
131 }
132
133 private:
134 ArmVIXLMacroAssembler* const assembler_;
135 vixl32::Register rd_;
136 vixl32::Label* const label_;
137 int32_t adr_location_;
138 };
139
OneRegInReferenceOutSaveEverythingCallerSaves()140 static RegisterSet OneRegInReferenceOutSaveEverythingCallerSaves() {
141 InvokeRuntimeCallingConventionARMVIXL calling_convention;
142 RegisterSet caller_saves = RegisterSet::Empty();
143 caller_saves.Add(LocationFrom(calling_convention.GetRegisterAt(0)));
144 // TODO: Add GetReturnLocation() to the calling convention so that we can DCHECK()
145 // that the the kPrimNot result register is the same as the first argument register.
146 return caller_saves;
147 }
148
149 // SaveLiveRegisters and RestoreLiveRegisters from SlowPathCodeARM operate on sets of S registers,
150 // for each live D registers they treat two corresponding S registers as live ones.
151 //
152 // Two following functions (SaveContiguousSRegisterList, RestoreContiguousSRegisterList) build
153 // from a list of contiguous S registers a list of contiguous D registers (processing first/last
154 // S registers corner cases) and save/restore this new list treating them as D registers.
155 // - decreasing code size
156 // - avoiding hazards on Cortex-A57, when a pair of S registers for an actual live D register is
157 // restored and then used in regular non SlowPath code as D register.
158 //
159 // For the following example (v means the S register is live):
160 // D names: | D0 | D1 | D2 | D4 | ...
161 // S names: | S0 | S1 | S2 | S3 | S4 | S5 | S6 | S7 | ...
162 // Live? | | v | v | v | v | v | v | | ...
163 //
164 // S1 and S6 will be saved/restored independently; D registers list (D1, D2) will be processed
165 // as D registers.
166 //
167 // TODO(VIXL): All this code should be unnecessary once the VIXL AArch32 backend provides helpers
168 // for lists of floating-point registers.
SaveContiguousSRegisterList(size_t first,size_t last,CodeGenerator * codegen,size_t stack_offset)169 static size_t SaveContiguousSRegisterList(size_t first,
170 size_t last,
171 CodeGenerator* codegen,
172 size_t stack_offset) {
173 static_assert(kSRegSizeInBytes == kArmWordSize, "Broken assumption on reg/word sizes.");
174 static_assert(kDRegSizeInBytes == 2 * kArmWordSize, "Broken assumption on reg/word sizes.");
175 DCHECK_LE(first, last);
176 if ((first == last) && (first == 0)) {
177 __ Vstr(vixl32::SRegister(first), MemOperand(sp, stack_offset));
178 return stack_offset + kSRegSizeInBytes;
179 }
180 if (first % 2 == 1) {
181 __ Vstr(vixl32::SRegister(first++), MemOperand(sp, stack_offset));
182 stack_offset += kSRegSizeInBytes;
183 }
184
185 bool save_last = false;
186 if (last % 2 == 0) {
187 save_last = true;
188 --last;
189 }
190
191 if (first < last) {
192 vixl32::DRegister d_reg = vixl32::DRegister(first / 2);
193 DCHECK_EQ((last - first + 1) % 2, 0u);
194 size_t number_of_d_regs = (last - first + 1) / 2;
195
196 if (number_of_d_regs == 1) {
197 __ Vstr(d_reg, MemOperand(sp, stack_offset));
198 } else if (number_of_d_regs > 1) {
199 UseScratchRegisterScope temps(down_cast<CodeGeneratorARMVIXL*>(codegen)->GetVIXLAssembler());
200 vixl32::Register base = sp;
201 if (stack_offset != 0) {
202 base = temps.Acquire();
203 __ Add(base, sp, Operand::From(stack_offset));
204 }
205 __ Vstm(F64, base, NO_WRITE_BACK, DRegisterList(d_reg, number_of_d_regs));
206 }
207 stack_offset += number_of_d_regs * kDRegSizeInBytes;
208 }
209
210 if (save_last) {
211 __ Vstr(vixl32::SRegister(last + 1), MemOperand(sp, stack_offset));
212 stack_offset += kSRegSizeInBytes;
213 }
214
215 return stack_offset;
216 }
217
RestoreContiguousSRegisterList(size_t first,size_t last,CodeGenerator * codegen,size_t stack_offset)218 static size_t RestoreContiguousSRegisterList(size_t first,
219 size_t last,
220 CodeGenerator* codegen,
221 size_t stack_offset) {
222 static_assert(kSRegSizeInBytes == kArmWordSize, "Broken assumption on reg/word sizes.");
223 static_assert(kDRegSizeInBytes == 2 * kArmWordSize, "Broken assumption on reg/word sizes.");
224 DCHECK_LE(first, last);
225 if ((first == last) && (first == 0)) {
226 __ Vldr(vixl32::SRegister(first), MemOperand(sp, stack_offset));
227 return stack_offset + kSRegSizeInBytes;
228 }
229 if (first % 2 == 1) {
230 __ Vldr(vixl32::SRegister(first++), MemOperand(sp, stack_offset));
231 stack_offset += kSRegSizeInBytes;
232 }
233
234 bool restore_last = false;
235 if (last % 2 == 0) {
236 restore_last = true;
237 --last;
238 }
239
240 if (first < last) {
241 vixl32::DRegister d_reg = vixl32::DRegister(first / 2);
242 DCHECK_EQ((last - first + 1) % 2, 0u);
243 size_t number_of_d_regs = (last - first + 1) / 2;
244 if (number_of_d_regs == 1) {
245 __ Vldr(d_reg, MemOperand(sp, stack_offset));
246 } else if (number_of_d_regs > 1) {
247 UseScratchRegisterScope temps(down_cast<CodeGeneratorARMVIXL*>(codegen)->GetVIXLAssembler());
248 vixl32::Register base = sp;
249 if (stack_offset != 0) {
250 base = temps.Acquire();
251 __ Add(base, sp, Operand::From(stack_offset));
252 }
253 __ Vldm(F64, base, NO_WRITE_BACK, DRegisterList(d_reg, number_of_d_regs));
254 }
255 stack_offset += number_of_d_regs * kDRegSizeInBytes;
256 }
257
258 if (restore_last) {
259 __ Vldr(vixl32::SRegister(last + 1), MemOperand(sp, stack_offset));
260 stack_offset += kSRegSizeInBytes;
261 }
262
263 return stack_offset;
264 }
265
GetLoadOperandType(DataType::Type type)266 static LoadOperandType GetLoadOperandType(DataType::Type type) {
267 switch (type) {
268 case DataType::Type::kReference:
269 return kLoadWord;
270 case DataType::Type::kBool:
271 case DataType::Type::kUint8:
272 return kLoadUnsignedByte;
273 case DataType::Type::kInt8:
274 return kLoadSignedByte;
275 case DataType::Type::kUint16:
276 return kLoadUnsignedHalfword;
277 case DataType::Type::kInt16:
278 return kLoadSignedHalfword;
279 case DataType::Type::kInt32:
280 return kLoadWord;
281 case DataType::Type::kInt64:
282 return kLoadWordPair;
283 case DataType::Type::kFloat32:
284 return kLoadSWord;
285 case DataType::Type::kFloat64:
286 return kLoadDWord;
287 default:
288 LOG(FATAL) << "Unreachable type " << type;
289 UNREACHABLE();
290 }
291 }
292
GetStoreOperandType(DataType::Type type)293 static StoreOperandType GetStoreOperandType(DataType::Type type) {
294 switch (type) {
295 case DataType::Type::kReference:
296 return kStoreWord;
297 case DataType::Type::kBool:
298 case DataType::Type::kUint8:
299 case DataType::Type::kInt8:
300 return kStoreByte;
301 case DataType::Type::kUint16:
302 case DataType::Type::kInt16:
303 return kStoreHalfword;
304 case DataType::Type::kInt32:
305 return kStoreWord;
306 case DataType::Type::kInt64:
307 return kStoreWordPair;
308 case DataType::Type::kFloat32:
309 return kStoreSWord;
310 case DataType::Type::kFloat64:
311 return kStoreDWord;
312 default:
313 LOG(FATAL) << "Unreachable type " << type;
314 UNREACHABLE();
315 }
316 }
317
SaveLiveRegisters(CodeGenerator * codegen,LocationSummary * locations)318 void SlowPathCodeARMVIXL::SaveLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) {
319 size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath();
320 size_t orig_offset = stack_offset;
321
322 const uint32_t core_spills = codegen->GetSlowPathSpills(locations, /* core_registers= */ true);
323 for (uint32_t i : LowToHighBits(core_spills)) {
324 // If the register holds an object, update the stack mask.
325 if (locations->RegisterContainsObject(i)) {
326 locations->SetStackBit(stack_offset / kVRegSize);
327 }
328 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize());
329 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters);
330 saved_core_stack_offsets_[i] = stack_offset;
331 stack_offset += kArmWordSize;
332 }
333
334 CodeGeneratorARMVIXL* arm_codegen = down_cast<CodeGeneratorARMVIXL*>(codegen);
335 arm_codegen->GetAssembler()->StoreRegisterList(core_spills, orig_offset);
336
337 uint32_t fp_spills = codegen->GetSlowPathSpills(locations, /* core_registers= */ false);
338 orig_offset = stack_offset;
339 for (uint32_t i : LowToHighBits(fp_spills)) {
340 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters);
341 saved_fpu_stack_offsets_[i] = stack_offset;
342 stack_offset += kArmWordSize;
343 }
344
345 stack_offset = orig_offset;
346 while (fp_spills != 0u) {
347 uint32_t begin = CTZ(fp_spills);
348 uint32_t tmp = fp_spills + (1u << begin);
349 fp_spills &= tmp; // Clear the contiguous range of 1s.
350 uint32_t end = (tmp == 0u) ? 32u : CTZ(tmp); // CTZ(0) is undefined.
351 stack_offset = SaveContiguousSRegisterList(begin, end - 1, codegen, stack_offset);
352 }
353 DCHECK_LE(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize());
354 }
355
RestoreLiveRegisters(CodeGenerator * codegen,LocationSummary * locations)356 void SlowPathCodeARMVIXL::RestoreLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) {
357 size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath();
358 size_t orig_offset = stack_offset;
359
360 const uint32_t core_spills = codegen->GetSlowPathSpills(locations, /* core_registers= */ true);
361 for (uint32_t i : LowToHighBits(core_spills)) {
362 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize());
363 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters);
364 stack_offset += kArmWordSize;
365 }
366
367 // TODO(VIXL): Check the coherency of stack_offset after this with a test.
368 CodeGeneratorARMVIXL* arm_codegen = down_cast<CodeGeneratorARMVIXL*>(codegen);
369 arm_codegen->GetAssembler()->LoadRegisterList(core_spills, orig_offset);
370
371 uint32_t fp_spills = codegen->GetSlowPathSpills(locations, /* core_registers= */ false);
372 while (fp_spills != 0u) {
373 uint32_t begin = CTZ(fp_spills);
374 uint32_t tmp = fp_spills + (1u << begin);
375 fp_spills &= tmp; // Clear the contiguous range of 1s.
376 uint32_t end = (tmp == 0u) ? 32u : CTZ(tmp); // CTZ(0) is undefined.
377 stack_offset = RestoreContiguousSRegisterList(begin, end - 1, codegen, stack_offset);
378 }
379 DCHECK_LE(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize());
380 }
381
382 class NullCheckSlowPathARMVIXL : public SlowPathCodeARMVIXL {
383 public:
NullCheckSlowPathARMVIXL(HNullCheck * instruction)384 explicit NullCheckSlowPathARMVIXL(HNullCheck* instruction) : SlowPathCodeARMVIXL(instruction) {}
385
EmitNativeCode(CodeGenerator * codegen)386 void EmitNativeCode(CodeGenerator* codegen) override {
387 CodeGeneratorARMVIXL* arm_codegen = down_cast<CodeGeneratorARMVIXL*>(codegen);
388 __ Bind(GetEntryLabel());
389 if (instruction_->CanThrowIntoCatchBlock()) {
390 // Live registers will be restored in the catch block if caught.
391 SaveLiveRegisters(codegen, instruction_->GetLocations());
392 }
393 arm_codegen->InvokeRuntime(kQuickThrowNullPointer,
394 instruction_,
395 instruction_->GetDexPc(),
396 this);
397 CheckEntrypointTypes<kQuickThrowNullPointer, void, void>();
398 }
399
IsFatal() const400 bool IsFatal() const override { return true; }
401
GetDescription() const402 const char* GetDescription() const override { return "NullCheckSlowPathARMVIXL"; }
403
404 private:
405 DISALLOW_COPY_AND_ASSIGN(NullCheckSlowPathARMVIXL);
406 };
407
408 class DivZeroCheckSlowPathARMVIXL : public SlowPathCodeARMVIXL {
409 public:
DivZeroCheckSlowPathARMVIXL(HDivZeroCheck * instruction)410 explicit DivZeroCheckSlowPathARMVIXL(HDivZeroCheck* instruction)
411 : SlowPathCodeARMVIXL(instruction) {}
412
EmitNativeCode(CodeGenerator * codegen)413 void EmitNativeCode(CodeGenerator* codegen) override {
414 CodeGeneratorARMVIXL* arm_codegen = down_cast<CodeGeneratorARMVIXL*>(codegen);
415 __ Bind(GetEntryLabel());
416 arm_codegen->InvokeRuntime(kQuickThrowDivZero, instruction_, instruction_->GetDexPc(), this);
417 CheckEntrypointTypes<kQuickThrowDivZero, void, void>();
418 }
419
IsFatal() const420 bool IsFatal() const override { return true; }
421
GetDescription() const422 const char* GetDescription() const override { return "DivZeroCheckSlowPathARMVIXL"; }
423
424 private:
425 DISALLOW_COPY_AND_ASSIGN(DivZeroCheckSlowPathARMVIXL);
426 };
427
428 class SuspendCheckSlowPathARMVIXL : public SlowPathCodeARMVIXL {
429 public:
SuspendCheckSlowPathARMVIXL(HSuspendCheck * instruction,HBasicBlock * successor)430 SuspendCheckSlowPathARMVIXL(HSuspendCheck* instruction, HBasicBlock* successor)
431 : SlowPathCodeARMVIXL(instruction), successor_(successor) {}
432
EmitNativeCode(CodeGenerator * codegen)433 void EmitNativeCode(CodeGenerator* codegen) override {
434 CodeGeneratorARMVIXL* arm_codegen = down_cast<CodeGeneratorARMVIXL*>(codegen);
435 __ Bind(GetEntryLabel());
436 arm_codegen->InvokeRuntime(kQuickTestSuspend, instruction_, instruction_->GetDexPc(), this);
437 CheckEntrypointTypes<kQuickTestSuspend, void, void>();
438 if (successor_ == nullptr) {
439 __ B(GetReturnLabel());
440 } else {
441 __ B(arm_codegen->GetLabelOf(successor_));
442 }
443 }
444
GetReturnLabel()445 vixl32::Label* GetReturnLabel() {
446 DCHECK(successor_ == nullptr);
447 return &return_label_;
448 }
449
GetSuccessor() const450 HBasicBlock* GetSuccessor() const {
451 return successor_;
452 }
453
GetDescription() const454 const char* GetDescription() const override { return "SuspendCheckSlowPathARMVIXL"; }
455
456 private:
457 // If not null, the block to branch to after the suspend check.
458 HBasicBlock* const successor_;
459
460 // If `successor_` is null, the label to branch to after the suspend check.
461 vixl32::Label return_label_;
462
463 DISALLOW_COPY_AND_ASSIGN(SuspendCheckSlowPathARMVIXL);
464 };
465
466 class BoundsCheckSlowPathARMVIXL : public SlowPathCodeARMVIXL {
467 public:
BoundsCheckSlowPathARMVIXL(HBoundsCheck * instruction)468 explicit BoundsCheckSlowPathARMVIXL(HBoundsCheck* instruction)
469 : SlowPathCodeARMVIXL(instruction) {}
470
EmitNativeCode(CodeGenerator * codegen)471 void EmitNativeCode(CodeGenerator* codegen) override {
472 CodeGeneratorARMVIXL* arm_codegen = down_cast<CodeGeneratorARMVIXL*>(codegen);
473 LocationSummary* locations = instruction_->GetLocations();
474
475 __ Bind(GetEntryLabel());
476 if (instruction_->CanThrowIntoCatchBlock()) {
477 // Live registers will be restored in the catch block if caught.
478 SaveLiveRegisters(codegen, instruction_->GetLocations());
479 }
480 // We're moving two locations to locations that could overlap, so we need a parallel
481 // move resolver.
482 InvokeRuntimeCallingConventionARMVIXL calling_convention;
483 codegen->EmitParallelMoves(
484 locations->InAt(0),
485 LocationFrom(calling_convention.GetRegisterAt(0)),
486 DataType::Type::kInt32,
487 locations->InAt(1),
488 LocationFrom(calling_convention.GetRegisterAt(1)),
489 DataType::Type::kInt32);
490 QuickEntrypointEnum entrypoint = instruction_->AsBoundsCheck()->IsStringCharAt()
491 ? kQuickThrowStringBounds
492 : kQuickThrowArrayBounds;
493 arm_codegen->InvokeRuntime(entrypoint, instruction_, instruction_->GetDexPc(), this);
494 CheckEntrypointTypes<kQuickThrowStringBounds, void, int32_t, int32_t>();
495 CheckEntrypointTypes<kQuickThrowArrayBounds, void, int32_t, int32_t>();
496 }
497
IsFatal() const498 bool IsFatal() const override { return true; }
499
GetDescription() const500 const char* GetDescription() const override { return "BoundsCheckSlowPathARMVIXL"; }
501
502 private:
503 DISALLOW_COPY_AND_ASSIGN(BoundsCheckSlowPathARMVIXL);
504 };
505
506 class LoadClassSlowPathARMVIXL : public SlowPathCodeARMVIXL {
507 public:
LoadClassSlowPathARMVIXL(HLoadClass * cls,HInstruction * at)508 LoadClassSlowPathARMVIXL(HLoadClass* cls, HInstruction* at)
509 : SlowPathCodeARMVIXL(at), cls_(cls) {
510 DCHECK(at->IsLoadClass() || at->IsClinitCheck());
511 DCHECK_EQ(instruction_->IsLoadClass(), cls_ == instruction_);
512 }
513
EmitNativeCode(CodeGenerator * codegen)514 void EmitNativeCode(CodeGenerator* codegen) override {
515 LocationSummary* locations = instruction_->GetLocations();
516 Location out = locations->Out();
517 const uint32_t dex_pc = instruction_->GetDexPc();
518 bool must_resolve_type = instruction_->IsLoadClass() && cls_->MustResolveTypeOnSlowPath();
519 bool must_do_clinit = instruction_->IsClinitCheck() || cls_->MustGenerateClinitCheck();
520
521 CodeGeneratorARMVIXL* arm_codegen = down_cast<CodeGeneratorARMVIXL*>(codegen);
522 __ Bind(GetEntryLabel());
523 SaveLiveRegisters(codegen, locations);
524
525 InvokeRuntimeCallingConventionARMVIXL calling_convention;
526 if (must_resolve_type) {
527 DCHECK(IsSameDexFile(cls_->GetDexFile(), arm_codegen->GetGraph()->GetDexFile()));
528 dex::TypeIndex type_index = cls_->GetTypeIndex();
529 __ Mov(calling_convention.GetRegisterAt(0), type_index.index_);
530 arm_codegen->InvokeRuntime(kQuickResolveType, instruction_, dex_pc, this);
531 CheckEntrypointTypes<kQuickResolveType, void*, uint32_t>();
532 // If we also must_do_clinit, the resolved type is now in the correct register.
533 } else {
534 DCHECK(must_do_clinit);
535 Location source = instruction_->IsLoadClass() ? out : locations->InAt(0);
536 arm_codegen->Move32(LocationFrom(calling_convention.GetRegisterAt(0)), source);
537 }
538 if (must_do_clinit) {
539 arm_codegen->InvokeRuntime(kQuickInitializeStaticStorage, instruction_, dex_pc, this);
540 CheckEntrypointTypes<kQuickInitializeStaticStorage, void*, mirror::Class*>();
541 }
542
543 // Move the class to the desired location.
544 if (out.IsValid()) {
545 DCHECK(out.IsRegister() && !locations->GetLiveRegisters()->ContainsCoreRegister(out.reg()));
546 arm_codegen->Move32(locations->Out(), LocationFrom(r0));
547 }
548 RestoreLiveRegisters(codegen, locations);
549 __ B(GetExitLabel());
550 }
551
GetDescription() const552 const char* GetDescription() const override { return "LoadClassSlowPathARMVIXL"; }
553
554 private:
555 // The class this slow path will load.
556 HLoadClass* const cls_;
557
558 DISALLOW_COPY_AND_ASSIGN(LoadClassSlowPathARMVIXL);
559 };
560
561 class LoadStringSlowPathARMVIXL : public SlowPathCodeARMVIXL {
562 public:
LoadStringSlowPathARMVIXL(HLoadString * instruction)563 explicit LoadStringSlowPathARMVIXL(HLoadString* instruction)
564 : SlowPathCodeARMVIXL(instruction) {}
565
EmitNativeCode(CodeGenerator * codegen)566 void EmitNativeCode(CodeGenerator* codegen) override {
567 DCHECK(instruction_->IsLoadString());
568 DCHECK_EQ(instruction_->AsLoadString()->GetLoadKind(), HLoadString::LoadKind::kBssEntry);
569 LocationSummary* locations = instruction_->GetLocations();
570 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg()));
571 const dex::StringIndex string_index = instruction_->AsLoadString()->GetStringIndex();
572
573 CodeGeneratorARMVIXL* arm_codegen = down_cast<CodeGeneratorARMVIXL*>(codegen);
574 __ Bind(GetEntryLabel());
575 SaveLiveRegisters(codegen, locations);
576
577 InvokeRuntimeCallingConventionARMVIXL calling_convention;
578 __ Mov(calling_convention.GetRegisterAt(0), string_index.index_);
579 arm_codegen->InvokeRuntime(kQuickResolveString, instruction_, instruction_->GetDexPc(), this);
580 CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>();
581
582 arm_codegen->Move32(locations->Out(), LocationFrom(r0));
583 RestoreLiveRegisters(codegen, locations);
584
585 __ B(GetExitLabel());
586 }
587
GetDescription() const588 const char* GetDescription() const override { return "LoadStringSlowPathARMVIXL"; }
589
590 private:
591 DISALLOW_COPY_AND_ASSIGN(LoadStringSlowPathARMVIXL);
592 };
593
594 class TypeCheckSlowPathARMVIXL : public SlowPathCodeARMVIXL {
595 public:
TypeCheckSlowPathARMVIXL(HInstruction * instruction,bool is_fatal)596 TypeCheckSlowPathARMVIXL(HInstruction* instruction, bool is_fatal)
597 : SlowPathCodeARMVIXL(instruction), is_fatal_(is_fatal) {}
598
EmitNativeCode(CodeGenerator * codegen)599 void EmitNativeCode(CodeGenerator* codegen) override {
600 LocationSummary* locations = instruction_->GetLocations();
601 DCHECK(instruction_->IsCheckCast()
602 || !locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg()));
603
604 CodeGeneratorARMVIXL* arm_codegen = down_cast<CodeGeneratorARMVIXL*>(codegen);
605 __ Bind(GetEntryLabel());
606
607 if (!is_fatal_ || instruction_->CanThrowIntoCatchBlock()) {
608 SaveLiveRegisters(codegen, locations);
609 }
610
611 // We're moving two locations to locations that could overlap, so we need a parallel
612 // move resolver.
613 InvokeRuntimeCallingConventionARMVIXL calling_convention;
614
615 codegen->EmitParallelMoves(locations->InAt(0),
616 LocationFrom(calling_convention.GetRegisterAt(0)),
617 DataType::Type::kReference,
618 locations->InAt(1),
619 LocationFrom(calling_convention.GetRegisterAt(1)),
620 DataType::Type::kReference);
621 if (instruction_->IsInstanceOf()) {
622 arm_codegen->InvokeRuntime(kQuickInstanceofNonTrivial,
623 instruction_,
624 instruction_->GetDexPc(),
625 this);
626 CheckEntrypointTypes<kQuickInstanceofNonTrivial, size_t, mirror::Object*, mirror::Class*>();
627 arm_codegen->Move32(locations->Out(), LocationFrom(r0));
628 } else {
629 DCHECK(instruction_->IsCheckCast());
630 arm_codegen->InvokeRuntime(kQuickCheckInstanceOf,
631 instruction_,
632 instruction_->GetDexPc(),
633 this);
634 CheckEntrypointTypes<kQuickCheckInstanceOf, void, mirror::Object*, mirror::Class*>();
635 }
636
637 if (!is_fatal_) {
638 RestoreLiveRegisters(codegen, locations);
639 __ B(GetExitLabel());
640 }
641 }
642
GetDescription() const643 const char* GetDescription() const override { return "TypeCheckSlowPathARMVIXL"; }
644
IsFatal() const645 bool IsFatal() const override { return is_fatal_; }
646
647 private:
648 const bool is_fatal_;
649
650 DISALLOW_COPY_AND_ASSIGN(TypeCheckSlowPathARMVIXL);
651 };
652
653 class DeoptimizationSlowPathARMVIXL : public SlowPathCodeARMVIXL {
654 public:
DeoptimizationSlowPathARMVIXL(HDeoptimize * instruction)655 explicit DeoptimizationSlowPathARMVIXL(HDeoptimize* instruction)
656 : SlowPathCodeARMVIXL(instruction) {}
657
EmitNativeCode(CodeGenerator * codegen)658 void EmitNativeCode(CodeGenerator* codegen) override {
659 CodeGeneratorARMVIXL* arm_codegen = down_cast<CodeGeneratorARMVIXL*>(codegen);
660 __ Bind(GetEntryLabel());
661 LocationSummary* locations = instruction_->GetLocations();
662 SaveLiveRegisters(codegen, locations);
663 InvokeRuntimeCallingConventionARMVIXL calling_convention;
664 __ Mov(calling_convention.GetRegisterAt(0),
665 static_cast<uint32_t>(instruction_->AsDeoptimize()->GetDeoptimizationKind()));
666
667 arm_codegen->InvokeRuntime(kQuickDeoptimize, instruction_, instruction_->GetDexPc(), this);
668 CheckEntrypointTypes<kQuickDeoptimize, void, DeoptimizationKind>();
669 }
670
GetDescription() const671 const char* GetDescription() const override { return "DeoptimizationSlowPathARMVIXL"; }
672
673 private:
674 DISALLOW_COPY_AND_ASSIGN(DeoptimizationSlowPathARMVIXL);
675 };
676
677 class ArraySetSlowPathARMVIXL : public SlowPathCodeARMVIXL {
678 public:
ArraySetSlowPathARMVIXL(HInstruction * instruction)679 explicit ArraySetSlowPathARMVIXL(HInstruction* instruction) : SlowPathCodeARMVIXL(instruction) {}
680
EmitNativeCode(CodeGenerator * codegen)681 void EmitNativeCode(CodeGenerator* codegen) override {
682 LocationSummary* locations = instruction_->GetLocations();
683 __ Bind(GetEntryLabel());
684 SaveLiveRegisters(codegen, locations);
685
686 InvokeRuntimeCallingConventionARMVIXL calling_convention;
687 HParallelMove parallel_move(codegen->GetGraph()->GetAllocator());
688 parallel_move.AddMove(
689 locations->InAt(0),
690 LocationFrom(calling_convention.GetRegisterAt(0)),
691 DataType::Type::kReference,
692 nullptr);
693 parallel_move.AddMove(
694 locations->InAt(1),
695 LocationFrom(calling_convention.GetRegisterAt(1)),
696 DataType::Type::kInt32,
697 nullptr);
698 parallel_move.AddMove(
699 locations->InAt(2),
700 LocationFrom(calling_convention.GetRegisterAt(2)),
701 DataType::Type::kReference,
702 nullptr);
703 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move);
704
705 CodeGeneratorARMVIXL* arm_codegen = down_cast<CodeGeneratorARMVIXL*>(codegen);
706 arm_codegen->InvokeRuntime(kQuickAputObject, instruction_, instruction_->GetDexPc(), this);
707 CheckEntrypointTypes<kQuickAputObject, void, mirror::Array*, int32_t, mirror::Object*>();
708 RestoreLiveRegisters(codegen, locations);
709 __ B(GetExitLabel());
710 }
711
GetDescription() const712 const char* GetDescription() const override { return "ArraySetSlowPathARMVIXL"; }
713
714 private:
715 DISALLOW_COPY_AND_ASSIGN(ArraySetSlowPathARMVIXL);
716 };
717
718 // Slow path generating a read barrier for a heap reference.
719 class ReadBarrierForHeapReferenceSlowPathARMVIXL : public SlowPathCodeARMVIXL {
720 public:
ReadBarrierForHeapReferenceSlowPathARMVIXL(HInstruction * instruction,Location out,Location ref,Location obj,uint32_t offset,Location index)721 ReadBarrierForHeapReferenceSlowPathARMVIXL(HInstruction* instruction,
722 Location out,
723 Location ref,
724 Location obj,
725 uint32_t offset,
726 Location index)
727 : SlowPathCodeARMVIXL(instruction),
728 out_(out),
729 ref_(ref),
730 obj_(obj),
731 offset_(offset),
732 index_(index) {
733 DCHECK(kEmitCompilerReadBarrier);
734 // If `obj` is equal to `out` or `ref`, it means the initial object
735 // has been overwritten by (or after) the heap object reference load
736 // to be instrumented, e.g.:
737 //
738 // __ LoadFromOffset(kLoadWord, out, out, offset);
739 // codegen_->GenerateReadBarrierSlow(instruction, out_loc, out_loc, out_loc, offset);
740 //
741 // In that case, we have lost the information about the original
742 // object, and the emitted read barrier cannot work properly.
743 DCHECK(!obj.Equals(out)) << "obj=" << obj << " out=" << out;
744 DCHECK(!obj.Equals(ref)) << "obj=" << obj << " ref=" << ref;
745 }
746
EmitNativeCode(CodeGenerator * codegen)747 void EmitNativeCode(CodeGenerator* codegen) override {
748 CodeGeneratorARMVIXL* arm_codegen = down_cast<CodeGeneratorARMVIXL*>(codegen);
749 LocationSummary* locations = instruction_->GetLocations();
750 vixl32::Register reg_out = RegisterFrom(out_);
751 DCHECK(locations->CanCall());
752 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out.GetCode()));
753 DCHECK(instruction_->IsInstanceFieldGet() ||
754 instruction_->IsStaticFieldGet() ||
755 instruction_->IsArrayGet() ||
756 instruction_->IsInstanceOf() ||
757 instruction_->IsCheckCast() ||
758 (instruction_->IsInvokeVirtual() && instruction_->GetLocations()->Intrinsified()))
759 << "Unexpected instruction in read barrier for heap reference slow path: "
760 << instruction_->DebugName();
761 // The read barrier instrumentation of object ArrayGet
762 // instructions does not support the HIntermediateAddress
763 // instruction.
764 DCHECK(!(instruction_->IsArrayGet() &&
765 instruction_->AsArrayGet()->GetArray()->IsIntermediateAddress()));
766
767 __ Bind(GetEntryLabel());
768 SaveLiveRegisters(codegen, locations);
769
770 // We may have to change the index's value, but as `index_` is a
771 // constant member (like other "inputs" of this slow path),
772 // introduce a copy of it, `index`.
773 Location index = index_;
774 if (index_.IsValid()) {
775 // Handle `index_` for HArrayGet and UnsafeGetObject/UnsafeGetObjectVolatile intrinsics.
776 if (instruction_->IsArrayGet()) {
777 // Compute the actual memory offset and store it in `index`.
778 vixl32::Register index_reg = RegisterFrom(index_);
779 DCHECK(locations->GetLiveRegisters()->ContainsCoreRegister(index_reg.GetCode()));
780 if (codegen->IsCoreCalleeSaveRegister(index_reg.GetCode())) {
781 // We are about to change the value of `index_reg` (see the
782 // calls to art::arm::ArmVIXLMacroAssembler::Lsl and
783 // art::arm::ArmVIXLMacroAssembler::Add below), but it has
784 // not been saved by the previous call to
785 // art::SlowPathCode::SaveLiveRegisters, as it is a
786 // callee-save register --
787 // art::SlowPathCode::SaveLiveRegisters does not consider
788 // callee-save registers, as it has been designed with the
789 // assumption that callee-save registers are supposed to be
790 // handled by the called function. So, as a callee-save
791 // register, `index_reg` _would_ eventually be saved onto
792 // the stack, but it would be too late: we would have
793 // changed its value earlier. Therefore, we manually save
794 // it here into another freely available register,
795 // `free_reg`, chosen of course among the caller-save
796 // registers (as a callee-save `free_reg` register would
797 // exhibit the same problem).
798 //
799 // Note we could have requested a temporary register from
800 // the register allocator instead; but we prefer not to, as
801 // this is a slow path, and we know we can find a
802 // caller-save register that is available.
803 vixl32::Register free_reg = FindAvailableCallerSaveRegister(codegen);
804 __ Mov(free_reg, index_reg);
805 index_reg = free_reg;
806 index = LocationFrom(index_reg);
807 } else {
808 // The initial register stored in `index_` has already been
809 // saved in the call to art::SlowPathCode::SaveLiveRegisters
810 // (as it is not a callee-save register), so we can freely
811 // use it.
812 }
813 // Shifting the index value contained in `index_reg` by the scale
814 // factor (2) cannot overflow in practice, as the runtime is
815 // unable to allocate object arrays with a size larger than
816 // 2^26 - 1 (that is, 2^28 - 4 bytes).
817 __ Lsl(index_reg, index_reg, TIMES_4);
818 static_assert(
819 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t),
820 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes.");
821 __ Add(index_reg, index_reg, offset_);
822 } else {
823 // In the case of the UnsafeGetObject/UnsafeGetObjectVolatile
824 // intrinsics, `index_` is not shifted by a scale factor of 2
825 // (as in the case of ArrayGet), as it is actually an offset
826 // to an object field within an object.
827 DCHECK(instruction_->IsInvoke()) << instruction_->DebugName();
828 DCHECK(instruction_->GetLocations()->Intrinsified());
829 DCHECK((instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObject) ||
830 (instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObjectVolatile))
831 << instruction_->AsInvoke()->GetIntrinsic();
832 DCHECK_EQ(offset_, 0U);
833 DCHECK(index_.IsRegisterPair());
834 // UnsafeGet's offset location is a register pair, the low
835 // part contains the correct offset.
836 index = index_.ToLow();
837 }
838 }
839
840 // We're moving two or three locations to locations that could
841 // overlap, so we need a parallel move resolver.
842 InvokeRuntimeCallingConventionARMVIXL calling_convention;
843 HParallelMove parallel_move(codegen->GetGraph()->GetAllocator());
844 parallel_move.AddMove(ref_,
845 LocationFrom(calling_convention.GetRegisterAt(0)),
846 DataType::Type::kReference,
847 nullptr);
848 parallel_move.AddMove(obj_,
849 LocationFrom(calling_convention.GetRegisterAt(1)),
850 DataType::Type::kReference,
851 nullptr);
852 if (index.IsValid()) {
853 parallel_move.AddMove(index,
854 LocationFrom(calling_convention.GetRegisterAt(2)),
855 DataType::Type::kInt32,
856 nullptr);
857 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move);
858 } else {
859 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move);
860 __ Mov(calling_convention.GetRegisterAt(2), offset_);
861 }
862 arm_codegen->InvokeRuntime(kQuickReadBarrierSlow, instruction_, instruction_->GetDexPc(), this);
863 CheckEntrypointTypes<
864 kQuickReadBarrierSlow, mirror::Object*, mirror::Object*, mirror::Object*, uint32_t>();
865 arm_codegen->Move32(out_, LocationFrom(r0));
866
867 RestoreLiveRegisters(codegen, locations);
868 __ B(GetExitLabel());
869 }
870
GetDescription() const871 const char* GetDescription() const override {
872 return "ReadBarrierForHeapReferenceSlowPathARMVIXL";
873 }
874
875 private:
FindAvailableCallerSaveRegister(CodeGenerator * codegen)876 vixl32::Register FindAvailableCallerSaveRegister(CodeGenerator* codegen) {
877 uint32_t ref = RegisterFrom(ref_).GetCode();
878 uint32_t obj = RegisterFrom(obj_).GetCode();
879 for (uint32_t i = 0, e = codegen->GetNumberOfCoreRegisters(); i < e; ++i) {
880 if (i != ref && i != obj && !codegen->IsCoreCalleeSaveRegister(i)) {
881 return vixl32::Register(i);
882 }
883 }
884 // We shall never fail to find a free caller-save register, as
885 // there are more than two core caller-save registers on ARM
886 // (meaning it is possible to find one which is different from
887 // `ref` and `obj`).
888 DCHECK_GT(codegen->GetNumberOfCoreCallerSaveRegisters(), 2u);
889 LOG(FATAL) << "Could not find a free caller-save register";
890 UNREACHABLE();
891 }
892
893 const Location out_;
894 const Location ref_;
895 const Location obj_;
896 const uint32_t offset_;
897 // An additional location containing an index to an array.
898 // Only used for HArrayGet and the UnsafeGetObject &
899 // UnsafeGetObjectVolatile intrinsics.
900 const Location index_;
901
902 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForHeapReferenceSlowPathARMVIXL);
903 };
904
905 // Slow path generating a read barrier for a GC root.
906 class ReadBarrierForRootSlowPathARMVIXL : public SlowPathCodeARMVIXL {
907 public:
ReadBarrierForRootSlowPathARMVIXL(HInstruction * instruction,Location out,Location root)908 ReadBarrierForRootSlowPathARMVIXL(HInstruction* instruction, Location out, Location root)
909 : SlowPathCodeARMVIXL(instruction), out_(out), root_(root) {
910 DCHECK(kEmitCompilerReadBarrier);
911 }
912
EmitNativeCode(CodeGenerator * codegen)913 void EmitNativeCode(CodeGenerator* codegen) override {
914 LocationSummary* locations = instruction_->GetLocations();
915 vixl32::Register reg_out = RegisterFrom(out_);
916 DCHECK(locations->CanCall());
917 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out.GetCode()));
918 DCHECK(instruction_->IsLoadClass() || instruction_->IsLoadString())
919 << "Unexpected instruction in read barrier for GC root slow path: "
920 << instruction_->DebugName();
921
922 __ Bind(GetEntryLabel());
923 SaveLiveRegisters(codegen, locations);
924
925 InvokeRuntimeCallingConventionARMVIXL calling_convention;
926 CodeGeneratorARMVIXL* arm_codegen = down_cast<CodeGeneratorARMVIXL*>(codegen);
927 arm_codegen->Move32(LocationFrom(calling_convention.GetRegisterAt(0)), root_);
928 arm_codegen->InvokeRuntime(kQuickReadBarrierForRootSlow,
929 instruction_,
930 instruction_->GetDexPc(),
931 this);
932 CheckEntrypointTypes<kQuickReadBarrierForRootSlow, mirror::Object*, GcRoot<mirror::Object>*>();
933 arm_codegen->Move32(out_, LocationFrom(r0));
934
935 RestoreLiveRegisters(codegen, locations);
936 __ B(GetExitLabel());
937 }
938
GetDescription() const939 const char* GetDescription() const override { return "ReadBarrierForRootSlowPathARMVIXL"; }
940
941 private:
942 const Location out_;
943 const Location root_;
944
945 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForRootSlowPathARMVIXL);
946 };
947
ARMCondition(IfCondition cond)948 inline vixl32::Condition ARMCondition(IfCondition cond) {
949 switch (cond) {
950 case kCondEQ: return eq;
951 case kCondNE: return ne;
952 case kCondLT: return lt;
953 case kCondLE: return le;
954 case kCondGT: return gt;
955 case kCondGE: return ge;
956 case kCondB: return lo;
957 case kCondBE: return ls;
958 case kCondA: return hi;
959 case kCondAE: return hs;
960 }
961 LOG(FATAL) << "Unreachable";
962 UNREACHABLE();
963 }
964
965 // Maps signed condition to unsigned condition.
ARMUnsignedCondition(IfCondition cond)966 inline vixl32::Condition ARMUnsignedCondition(IfCondition cond) {
967 switch (cond) {
968 case kCondEQ: return eq;
969 case kCondNE: return ne;
970 // Signed to unsigned.
971 case kCondLT: return lo;
972 case kCondLE: return ls;
973 case kCondGT: return hi;
974 case kCondGE: return hs;
975 // Unsigned remain unchanged.
976 case kCondB: return lo;
977 case kCondBE: return ls;
978 case kCondA: return hi;
979 case kCondAE: return hs;
980 }
981 LOG(FATAL) << "Unreachable";
982 UNREACHABLE();
983 }
984
ARMFPCondition(IfCondition cond,bool gt_bias)985 inline vixl32::Condition ARMFPCondition(IfCondition cond, bool gt_bias) {
986 // The ARM condition codes can express all the necessary branches, see the
987 // "Meaning (floating-point)" column in the table A8-1 of the ARMv7 reference manual.
988 // There is no dex instruction or HIR that would need the missing conditions
989 // "equal or unordered" or "not equal".
990 switch (cond) {
991 case kCondEQ: return eq;
992 case kCondNE: return ne /* unordered */;
993 case kCondLT: return gt_bias ? cc : lt /* unordered */;
994 case kCondLE: return gt_bias ? ls : le /* unordered */;
995 case kCondGT: return gt_bias ? hi /* unordered */ : gt;
996 case kCondGE: return gt_bias ? cs /* unordered */ : ge;
997 default:
998 LOG(FATAL) << "UNREACHABLE";
999 UNREACHABLE();
1000 }
1001 }
1002
ShiftFromOpKind(HDataProcWithShifterOp::OpKind op_kind)1003 inline ShiftType ShiftFromOpKind(HDataProcWithShifterOp::OpKind op_kind) {
1004 switch (op_kind) {
1005 case HDataProcWithShifterOp::kASR: return ShiftType::ASR;
1006 case HDataProcWithShifterOp::kLSL: return ShiftType::LSL;
1007 case HDataProcWithShifterOp::kLSR: return ShiftType::LSR;
1008 default:
1009 LOG(FATAL) << "Unexpected op kind " << op_kind;
1010 UNREACHABLE();
1011 }
1012 }
1013
DumpCoreRegister(std::ostream & stream,int reg) const1014 void CodeGeneratorARMVIXL::DumpCoreRegister(std::ostream& stream, int reg) const {
1015 stream << vixl32::Register(reg);
1016 }
1017
DumpFloatingPointRegister(std::ostream & stream,int reg) const1018 void CodeGeneratorARMVIXL::DumpFloatingPointRegister(std::ostream& stream, int reg) const {
1019 stream << vixl32::SRegister(reg);
1020 }
1021
GetInstructionSetFeatures() const1022 const ArmInstructionSetFeatures& CodeGeneratorARMVIXL::GetInstructionSetFeatures() const {
1023 return *GetCompilerOptions().GetInstructionSetFeatures()->AsArmInstructionSetFeatures();
1024 }
1025
ComputeSRegisterListMask(const SRegisterList & regs)1026 static uint32_t ComputeSRegisterListMask(const SRegisterList& regs) {
1027 uint32_t mask = 0;
1028 for (uint32_t i = regs.GetFirstSRegister().GetCode();
1029 i <= regs.GetLastSRegister().GetCode();
1030 ++i) {
1031 mask |= (1 << i);
1032 }
1033 return mask;
1034 }
1035
1036 // Saves the register in the stack. Returns the size taken on stack.
SaveCoreRegister(size_t stack_index ATTRIBUTE_UNUSED,uint32_t reg_id ATTRIBUTE_UNUSED)1037 size_t CodeGeneratorARMVIXL::SaveCoreRegister(size_t stack_index ATTRIBUTE_UNUSED,
1038 uint32_t reg_id ATTRIBUTE_UNUSED) {
1039 TODO_VIXL32(FATAL);
1040 UNREACHABLE();
1041 }
1042
1043 // Restores the register from the stack. Returns the size taken on stack.
RestoreCoreRegister(size_t stack_index ATTRIBUTE_UNUSED,uint32_t reg_id ATTRIBUTE_UNUSED)1044 size_t CodeGeneratorARMVIXL::RestoreCoreRegister(size_t stack_index ATTRIBUTE_UNUSED,
1045 uint32_t reg_id ATTRIBUTE_UNUSED) {
1046 TODO_VIXL32(FATAL);
1047 UNREACHABLE();
1048 }
1049
SaveFloatingPointRegister(size_t stack_index ATTRIBUTE_UNUSED,uint32_t reg_id ATTRIBUTE_UNUSED)1050 size_t CodeGeneratorARMVIXL::SaveFloatingPointRegister(size_t stack_index ATTRIBUTE_UNUSED,
1051 uint32_t reg_id ATTRIBUTE_UNUSED) {
1052 TODO_VIXL32(FATAL);
1053 UNREACHABLE();
1054 }
1055
RestoreFloatingPointRegister(size_t stack_index ATTRIBUTE_UNUSED,uint32_t reg_id ATTRIBUTE_UNUSED)1056 size_t CodeGeneratorARMVIXL::RestoreFloatingPointRegister(size_t stack_index ATTRIBUTE_UNUSED,
1057 uint32_t reg_id ATTRIBUTE_UNUSED) {
1058 TODO_VIXL32(FATAL);
1059 UNREACHABLE();
1060 }
1061
GenerateDataProcInstruction(HInstruction::InstructionKind kind,vixl32::Register out,vixl32::Register first,const Operand & second,CodeGeneratorARMVIXL * codegen)1062 static void GenerateDataProcInstruction(HInstruction::InstructionKind kind,
1063 vixl32::Register out,
1064 vixl32::Register first,
1065 const Operand& second,
1066 CodeGeneratorARMVIXL* codegen) {
1067 if (second.IsImmediate() && second.GetImmediate() == 0) {
1068 const Operand in = kind == HInstruction::kAnd
1069 ? Operand(0)
1070 : Operand(first);
1071
1072 __ Mov(out, in);
1073 } else {
1074 switch (kind) {
1075 case HInstruction::kAdd:
1076 __ Add(out, first, second);
1077 break;
1078 case HInstruction::kAnd:
1079 __ And(out, first, second);
1080 break;
1081 case HInstruction::kOr:
1082 __ Orr(out, first, second);
1083 break;
1084 case HInstruction::kSub:
1085 __ Sub(out, first, second);
1086 break;
1087 case HInstruction::kXor:
1088 __ Eor(out, first, second);
1089 break;
1090 default:
1091 LOG(FATAL) << "Unexpected instruction kind: " << kind;
1092 UNREACHABLE();
1093 }
1094 }
1095 }
1096
GenerateDataProc(HInstruction::InstructionKind kind,const Location & out,const Location & first,const Operand & second_lo,const Operand & second_hi,CodeGeneratorARMVIXL * codegen)1097 static void GenerateDataProc(HInstruction::InstructionKind kind,
1098 const Location& out,
1099 const Location& first,
1100 const Operand& second_lo,
1101 const Operand& second_hi,
1102 CodeGeneratorARMVIXL* codegen) {
1103 const vixl32::Register first_hi = HighRegisterFrom(first);
1104 const vixl32::Register first_lo = LowRegisterFrom(first);
1105 const vixl32::Register out_hi = HighRegisterFrom(out);
1106 const vixl32::Register out_lo = LowRegisterFrom(out);
1107
1108 if (kind == HInstruction::kAdd) {
1109 __ Adds(out_lo, first_lo, second_lo);
1110 __ Adc(out_hi, first_hi, second_hi);
1111 } else if (kind == HInstruction::kSub) {
1112 __ Subs(out_lo, first_lo, second_lo);
1113 __ Sbc(out_hi, first_hi, second_hi);
1114 } else {
1115 GenerateDataProcInstruction(kind, out_lo, first_lo, second_lo, codegen);
1116 GenerateDataProcInstruction(kind, out_hi, first_hi, second_hi, codegen);
1117 }
1118 }
1119
GetShifterOperand(vixl32::Register rm,ShiftType shift,uint32_t shift_imm)1120 static Operand GetShifterOperand(vixl32::Register rm, ShiftType shift, uint32_t shift_imm) {
1121 return shift_imm == 0 ? Operand(rm) : Operand(rm, shift, shift_imm);
1122 }
1123
GenerateLongDataProc(HDataProcWithShifterOp * instruction,CodeGeneratorARMVIXL * codegen)1124 static void GenerateLongDataProc(HDataProcWithShifterOp* instruction,
1125 CodeGeneratorARMVIXL* codegen) {
1126 DCHECK_EQ(instruction->GetType(), DataType::Type::kInt64);
1127 DCHECK(HDataProcWithShifterOp::IsShiftOp(instruction->GetOpKind()));
1128
1129 const LocationSummary* const locations = instruction->GetLocations();
1130 const uint32_t shift_value = instruction->GetShiftAmount();
1131 const HInstruction::InstructionKind kind = instruction->GetInstrKind();
1132 const Location first = locations->InAt(0);
1133 const Location second = locations->InAt(1);
1134 const Location out = locations->Out();
1135 const vixl32::Register first_hi = HighRegisterFrom(first);
1136 const vixl32::Register first_lo = LowRegisterFrom(first);
1137 const vixl32::Register out_hi = HighRegisterFrom(out);
1138 const vixl32::Register out_lo = LowRegisterFrom(out);
1139 const vixl32::Register second_hi = HighRegisterFrom(second);
1140 const vixl32::Register second_lo = LowRegisterFrom(second);
1141 const ShiftType shift = ShiftFromOpKind(instruction->GetOpKind());
1142
1143 if (shift_value >= 32) {
1144 if (shift == ShiftType::LSL) {
1145 GenerateDataProcInstruction(kind,
1146 out_hi,
1147 first_hi,
1148 Operand(second_lo, ShiftType::LSL, shift_value - 32),
1149 codegen);
1150 GenerateDataProcInstruction(kind, out_lo, first_lo, 0, codegen);
1151 } else if (shift == ShiftType::ASR) {
1152 GenerateDataProc(kind,
1153 out,
1154 first,
1155 GetShifterOperand(second_hi, ShiftType::ASR, shift_value - 32),
1156 Operand(second_hi, ShiftType::ASR, 31),
1157 codegen);
1158 } else {
1159 DCHECK_EQ(shift, ShiftType::LSR);
1160 GenerateDataProc(kind,
1161 out,
1162 first,
1163 GetShifterOperand(second_hi, ShiftType::LSR, shift_value - 32),
1164 0,
1165 codegen);
1166 }
1167 } else {
1168 DCHECK_GT(shift_value, 1U);
1169 DCHECK_LT(shift_value, 32U);
1170
1171 UseScratchRegisterScope temps(codegen->GetVIXLAssembler());
1172
1173 if (shift == ShiftType::LSL) {
1174 // We are not doing this for HInstruction::kAdd because the output will require
1175 // Location::kOutputOverlap; not applicable to other cases.
1176 if (kind == HInstruction::kOr || kind == HInstruction::kXor) {
1177 GenerateDataProcInstruction(kind,
1178 out_hi,
1179 first_hi,
1180 Operand(second_hi, ShiftType::LSL, shift_value),
1181 codegen);
1182 GenerateDataProcInstruction(kind,
1183 out_hi,
1184 out_hi,
1185 Operand(second_lo, ShiftType::LSR, 32 - shift_value),
1186 codegen);
1187 GenerateDataProcInstruction(kind,
1188 out_lo,
1189 first_lo,
1190 Operand(second_lo, ShiftType::LSL, shift_value),
1191 codegen);
1192 } else {
1193 const vixl32::Register temp = temps.Acquire();
1194
1195 __ Lsl(temp, second_hi, shift_value);
1196 __ Orr(temp, temp, Operand(second_lo, ShiftType::LSR, 32 - shift_value));
1197 GenerateDataProc(kind,
1198 out,
1199 first,
1200 Operand(second_lo, ShiftType::LSL, shift_value),
1201 temp,
1202 codegen);
1203 }
1204 } else {
1205 DCHECK(shift == ShiftType::ASR || shift == ShiftType::LSR);
1206
1207 // We are not doing this for HInstruction::kAdd because the output will require
1208 // Location::kOutputOverlap; not applicable to other cases.
1209 if (kind == HInstruction::kOr || kind == HInstruction::kXor) {
1210 GenerateDataProcInstruction(kind,
1211 out_lo,
1212 first_lo,
1213 Operand(second_lo, ShiftType::LSR, shift_value),
1214 codegen);
1215 GenerateDataProcInstruction(kind,
1216 out_lo,
1217 out_lo,
1218 Operand(second_hi, ShiftType::LSL, 32 - shift_value),
1219 codegen);
1220 GenerateDataProcInstruction(kind,
1221 out_hi,
1222 first_hi,
1223 Operand(second_hi, shift, shift_value),
1224 codegen);
1225 } else {
1226 const vixl32::Register temp = temps.Acquire();
1227
1228 __ Lsr(temp, second_lo, shift_value);
1229 __ Orr(temp, temp, Operand(second_hi, ShiftType::LSL, 32 - shift_value));
1230 GenerateDataProc(kind,
1231 out,
1232 first,
1233 temp,
1234 Operand(second_hi, shift, shift_value),
1235 codegen);
1236 }
1237 }
1238 }
1239 }
1240
GenerateVcmp(HInstruction * instruction,CodeGeneratorARMVIXL * codegen)1241 static void GenerateVcmp(HInstruction* instruction, CodeGeneratorARMVIXL* codegen) {
1242 const Location rhs_loc = instruction->GetLocations()->InAt(1);
1243 if (rhs_loc.IsConstant()) {
1244 // 0.0 is the only immediate that can be encoded directly in
1245 // a VCMP instruction.
1246 //
1247 // Both the JLS (section 15.20.1) and the JVMS (section 6.5)
1248 // specify that in a floating-point comparison, positive zero
1249 // and negative zero are considered equal, so we can use the
1250 // literal 0.0 for both cases here.
1251 //
1252 // Note however that some methods (Float.equal, Float.compare,
1253 // Float.compareTo, Double.equal, Double.compare,
1254 // Double.compareTo, Math.max, Math.min, StrictMath.max,
1255 // StrictMath.min) consider 0.0 to be (strictly) greater than
1256 // -0.0. So if we ever translate calls to these methods into a
1257 // HCompare instruction, we must handle the -0.0 case with
1258 // care here.
1259 DCHECK(rhs_loc.GetConstant()->IsArithmeticZero());
1260
1261 const DataType::Type type = instruction->InputAt(0)->GetType();
1262
1263 if (type == DataType::Type::kFloat32) {
1264 __ Vcmp(F32, InputSRegisterAt(instruction, 0), 0.0);
1265 } else {
1266 DCHECK_EQ(type, DataType::Type::kFloat64);
1267 __ Vcmp(F64, InputDRegisterAt(instruction, 0), 0.0);
1268 }
1269 } else {
1270 __ Vcmp(InputVRegisterAt(instruction, 0), InputVRegisterAt(instruction, 1));
1271 }
1272 }
1273
AdjustConstantForCondition(int64_t value,IfCondition * condition,IfCondition * opposite)1274 static int64_t AdjustConstantForCondition(int64_t value,
1275 IfCondition* condition,
1276 IfCondition* opposite) {
1277 if (value == 1) {
1278 if (*condition == kCondB) {
1279 value = 0;
1280 *condition = kCondEQ;
1281 *opposite = kCondNE;
1282 } else if (*condition == kCondAE) {
1283 value = 0;
1284 *condition = kCondNE;
1285 *opposite = kCondEQ;
1286 }
1287 } else if (value == -1) {
1288 if (*condition == kCondGT) {
1289 value = 0;
1290 *condition = kCondGE;
1291 *opposite = kCondLT;
1292 } else if (*condition == kCondLE) {
1293 value = 0;
1294 *condition = kCondLT;
1295 *opposite = kCondGE;
1296 }
1297 }
1298
1299 return value;
1300 }
1301
GenerateLongTestConstant(HCondition * condition,bool invert,CodeGeneratorARMVIXL * codegen)1302 static std::pair<vixl32::Condition, vixl32::Condition> GenerateLongTestConstant(
1303 HCondition* condition,
1304 bool invert,
1305 CodeGeneratorARMVIXL* codegen) {
1306 DCHECK_EQ(condition->GetLeft()->GetType(), DataType::Type::kInt64);
1307
1308 const LocationSummary* const locations = condition->GetLocations();
1309 IfCondition cond = condition->GetCondition();
1310 IfCondition opposite = condition->GetOppositeCondition();
1311
1312 if (invert) {
1313 std::swap(cond, opposite);
1314 }
1315
1316 std::pair<vixl32::Condition, vixl32::Condition> ret(eq, ne);
1317 const Location left = locations->InAt(0);
1318 const Location right = locations->InAt(1);
1319
1320 DCHECK(right.IsConstant());
1321
1322 const vixl32::Register left_high = HighRegisterFrom(left);
1323 const vixl32::Register left_low = LowRegisterFrom(left);
1324 int64_t value = AdjustConstantForCondition(Int64ConstantFrom(right), &cond, &opposite);
1325 UseScratchRegisterScope temps(codegen->GetVIXLAssembler());
1326
1327 // Comparisons against 0 are common enough to deserve special attention.
1328 if (value == 0) {
1329 switch (cond) {
1330 case kCondNE:
1331 // x > 0 iff x != 0 when the comparison is unsigned.
1332 case kCondA:
1333 ret = std::make_pair(ne, eq);
1334 FALLTHROUGH_INTENDED;
1335 case kCondEQ:
1336 // x <= 0 iff x == 0 when the comparison is unsigned.
1337 case kCondBE:
1338 __ Orrs(temps.Acquire(), left_low, left_high);
1339 return ret;
1340 case kCondLT:
1341 case kCondGE:
1342 __ Cmp(left_high, 0);
1343 return std::make_pair(ARMCondition(cond), ARMCondition(opposite));
1344 // Trivially true or false.
1345 case kCondB:
1346 ret = std::make_pair(ne, eq);
1347 FALLTHROUGH_INTENDED;
1348 case kCondAE:
1349 __ Cmp(left_low, left_low);
1350 return ret;
1351 default:
1352 break;
1353 }
1354 }
1355
1356 switch (cond) {
1357 case kCondEQ:
1358 case kCondNE:
1359 case kCondB:
1360 case kCondBE:
1361 case kCondA:
1362 case kCondAE: {
1363 const uint32_t value_low = Low32Bits(value);
1364 Operand operand_low(value_low);
1365
1366 __ Cmp(left_high, High32Bits(value));
1367
1368 // Since IT blocks longer than a 16-bit instruction are deprecated by ARMv8,
1369 // we must ensure that the operands corresponding to the least significant
1370 // halves of the inputs fit into a 16-bit CMP encoding.
1371 if (!left_low.IsLow() || !IsUint<8>(value_low)) {
1372 operand_low = Operand(temps.Acquire());
1373 __ Mov(LeaveFlags, operand_low.GetBaseRegister(), value_low);
1374 }
1375
1376 // We use the scope because of the IT block that follows.
1377 ExactAssemblyScope guard(codegen->GetVIXLAssembler(),
1378 2 * vixl32::k16BitT32InstructionSizeInBytes,
1379 CodeBufferCheckScope::kExactSize);
1380
1381 __ it(eq);
1382 __ cmp(eq, left_low, operand_low);
1383 ret = std::make_pair(ARMUnsignedCondition(cond), ARMUnsignedCondition(opposite));
1384 break;
1385 }
1386 case kCondLE:
1387 case kCondGT:
1388 // Trivially true or false.
1389 if (value == std::numeric_limits<int64_t>::max()) {
1390 __ Cmp(left_low, left_low);
1391 ret = cond == kCondLE ? std::make_pair(eq, ne) : std::make_pair(ne, eq);
1392 break;
1393 }
1394
1395 if (cond == kCondLE) {
1396 DCHECK_EQ(opposite, kCondGT);
1397 cond = kCondLT;
1398 opposite = kCondGE;
1399 } else {
1400 DCHECK_EQ(cond, kCondGT);
1401 DCHECK_EQ(opposite, kCondLE);
1402 cond = kCondGE;
1403 opposite = kCondLT;
1404 }
1405
1406 value++;
1407 FALLTHROUGH_INTENDED;
1408 case kCondGE:
1409 case kCondLT: {
1410 __ Cmp(left_low, Low32Bits(value));
1411 __ Sbcs(temps.Acquire(), left_high, High32Bits(value));
1412 ret = std::make_pair(ARMCondition(cond), ARMCondition(opposite));
1413 break;
1414 }
1415 default:
1416 LOG(FATAL) << "Unreachable";
1417 UNREACHABLE();
1418 }
1419
1420 return ret;
1421 }
1422
GenerateLongTest(HCondition * condition,bool invert,CodeGeneratorARMVIXL * codegen)1423 static std::pair<vixl32::Condition, vixl32::Condition> GenerateLongTest(
1424 HCondition* condition,
1425 bool invert,
1426 CodeGeneratorARMVIXL* codegen) {
1427 DCHECK_EQ(condition->GetLeft()->GetType(), DataType::Type::kInt64);
1428
1429 const LocationSummary* const locations = condition->GetLocations();
1430 IfCondition cond = condition->GetCondition();
1431 IfCondition opposite = condition->GetOppositeCondition();
1432
1433 if (invert) {
1434 std::swap(cond, opposite);
1435 }
1436
1437 std::pair<vixl32::Condition, vixl32::Condition> ret(eq, ne);
1438 Location left = locations->InAt(0);
1439 Location right = locations->InAt(1);
1440
1441 DCHECK(right.IsRegisterPair());
1442
1443 switch (cond) {
1444 case kCondEQ:
1445 case kCondNE:
1446 case kCondB:
1447 case kCondBE:
1448 case kCondA:
1449 case kCondAE: {
1450 __ Cmp(HighRegisterFrom(left), HighRegisterFrom(right));
1451
1452 // We use the scope because of the IT block that follows.
1453 ExactAssemblyScope guard(codegen->GetVIXLAssembler(),
1454 2 * vixl32::k16BitT32InstructionSizeInBytes,
1455 CodeBufferCheckScope::kExactSize);
1456
1457 __ it(eq);
1458 __ cmp(eq, LowRegisterFrom(left), LowRegisterFrom(right));
1459 ret = std::make_pair(ARMUnsignedCondition(cond), ARMUnsignedCondition(opposite));
1460 break;
1461 }
1462 case kCondLE:
1463 case kCondGT:
1464 if (cond == kCondLE) {
1465 DCHECK_EQ(opposite, kCondGT);
1466 cond = kCondGE;
1467 opposite = kCondLT;
1468 } else {
1469 DCHECK_EQ(cond, kCondGT);
1470 DCHECK_EQ(opposite, kCondLE);
1471 cond = kCondLT;
1472 opposite = kCondGE;
1473 }
1474
1475 std::swap(left, right);
1476 FALLTHROUGH_INTENDED;
1477 case kCondGE:
1478 case kCondLT: {
1479 UseScratchRegisterScope temps(codegen->GetVIXLAssembler());
1480
1481 __ Cmp(LowRegisterFrom(left), LowRegisterFrom(right));
1482 __ Sbcs(temps.Acquire(), HighRegisterFrom(left), HighRegisterFrom(right));
1483 ret = std::make_pair(ARMCondition(cond), ARMCondition(opposite));
1484 break;
1485 }
1486 default:
1487 LOG(FATAL) << "Unreachable";
1488 UNREACHABLE();
1489 }
1490
1491 return ret;
1492 }
1493
GenerateTest(HCondition * condition,bool invert,CodeGeneratorARMVIXL * codegen)1494 static std::pair<vixl32::Condition, vixl32::Condition> GenerateTest(HCondition* condition,
1495 bool invert,
1496 CodeGeneratorARMVIXL* codegen) {
1497 const DataType::Type type = condition->GetLeft()->GetType();
1498 IfCondition cond = condition->GetCondition();
1499 IfCondition opposite = condition->GetOppositeCondition();
1500 std::pair<vixl32::Condition, vixl32::Condition> ret(eq, ne);
1501
1502 if (invert) {
1503 std::swap(cond, opposite);
1504 }
1505
1506 if (type == DataType::Type::kInt64) {
1507 ret = condition->GetLocations()->InAt(1).IsConstant()
1508 ? GenerateLongTestConstant(condition, invert, codegen)
1509 : GenerateLongTest(condition, invert, codegen);
1510 } else if (DataType::IsFloatingPointType(type)) {
1511 GenerateVcmp(condition, codegen);
1512 __ Vmrs(RegisterOrAPSR_nzcv(kPcCode), FPSCR);
1513 ret = std::make_pair(ARMFPCondition(cond, condition->IsGtBias()),
1514 ARMFPCondition(opposite, condition->IsGtBias()));
1515 } else {
1516 DCHECK(DataType::IsIntegralType(type) || type == DataType::Type::kReference) << type;
1517 __ Cmp(InputRegisterAt(condition, 0), InputOperandAt(condition, 1));
1518 ret = std::make_pair(ARMCondition(cond), ARMCondition(opposite));
1519 }
1520
1521 return ret;
1522 }
1523
GenerateConditionGeneric(HCondition * cond,CodeGeneratorARMVIXL * codegen)1524 static void GenerateConditionGeneric(HCondition* cond, CodeGeneratorARMVIXL* codegen) {
1525 const vixl32::Register out = OutputRegister(cond);
1526 const auto condition = GenerateTest(cond, false, codegen);
1527
1528 __ Mov(LeaveFlags, out, 0);
1529
1530 if (out.IsLow()) {
1531 // We use the scope because of the IT block that follows.
1532 ExactAssemblyScope guard(codegen->GetVIXLAssembler(),
1533 2 * vixl32::k16BitT32InstructionSizeInBytes,
1534 CodeBufferCheckScope::kExactSize);
1535
1536 __ it(condition.first);
1537 __ mov(condition.first, out, 1);
1538 } else {
1539 vixl32::Label done_label;
1540 vixl32::Label* const final_label = codegen->GetFinalLabel(cond, &done_label);
1541
1542 __ B(condition.second, final_label, /* is_far_target= */ false);
1543 __ Mov(out, 1);
1544
1545 if (done_label.IsReferenced()) {
1546 __ Bind(&done_label);
1547 }
1548 }
1549 }
1550
GenerateEqualLong(HCondition * cond,CodeGeneratorARMVIXL * codegen)1551 static void GenerateEqualLong(HCondition* cond, CodeGeneratorARMVIXL* codegen) {
1552 DCHECK_EQ(cond->GetLeft()->GetType(), DataType::Type::kInt64);
1553
1554 const LocationSummary* const locations = cond->GetLocations();
1555 IfCondition condition = cond->GetCondition();
1556 const vixl32::Register out = OutputRegister(cond);
1557 const Location left = locations->InAt(0);
1558 const Location right = locations->InAt(1);
1559 vixl32::Register left_high = HighRegisterFrom(left);
1560 vixl32::Register left_low = LowRegisterFrom(left);
1561 vixl32::Register temp;
1562 UseScratchRegisterScope temps(codegen->GetVIXLAssembler());
1563
1564 if (right.IsConstant()) {
1565 IfCondition opposite = cond->GetOppositeCondition();
1566 const int64_t value = AdjustConstantForCondition(Int64ConstantFrom(right),
1567 &condition,
1568 &opposite);
1569 Operand right_high = High32Bits(value);
1570 Operand right_low = Low32Bits(value);
1571
1572 // The output uses Location::kNoOutputOverlap.
1573 if (out.Is(left_high)) {
1574 std::swap(left_low, left_high);
1575 std::swap(right_low, right_high);
1576 }
1577
1578 __ Sub(out, left_low, right_low);
1579 temp = temps.Acquire();
1580 __ Sub(temp, left_high, right_high);
1581 } else {
1582 DCHECK(right.IsRegisterPair());
1583 temp = temps.Acquire();
1584 __ Sub(temp, left_high, HighRegisterFrom(right));
1585 __ Sub(out, left_low, LowRegisterFrom(right));
1586 }
1587
1588 // Need to check after calling AdjustConstantForCondition().
1589 DCHECK(condition == kCondEQ || condition == kCondNE) << condition;
1590
1591 if (condition == kCondNE && out.IsLow()) {
1592 __ Orrs(out, out, temp);
1593
1594 // We use the scope because of the IT block that follows.
1595 ExactAssemblyScope guard(codegen->GetVIXLAssembler(),
1596 2 * vixl32::k16BitT32InstructionSizeInBytes,
1597 CodeBufferCheckScope::kExactSize);
1598
1599 __ it(ne);
1600 __ mov(ne, out, 1);
1601 } else {
1602 __ Orr(out, out, temp);
1603 codegen->GenerateConditionWithZero(condition, out, out, temp);
1604 }
1605 }
1606
GenerateConditionLong(HCondition * cond,CodeGeneratorARMVIXL * codegen)1607 static void GenerateConditionLong(HCondition* cond, CodeGeneratorARMVIXL* codegen) {
1608 DCHECK_EQ(cond->GetLeft()->GetType(), DataType::Type::kInt64);
1609
1610 const LocationSummary* const locations = cond->GetLocations();
1611 IfCondition condition = cond->GetCondition();
1612 const vixl32::Register out = OutputRegister(cond);
1613 const Location left = locations->InAt(0);
1614 const Location right = locations->InAt(1);
1615
1616 if (right.IsConstant()) {
1617 IfCondition opposite = cond->GetOppositeCondition();
1618
1619 // Comparisons against 0 are common enough to deserve special attention.
1620 if (AdjustConstantForCondition(Int64ConstantFrom(right), &condition, &opposite) == 0) {
1621 switch (condition) {
1622 case kCondNE:
1623 case kCondA:
1624 if (out.IsLow()) {
1625 // We only care if both input registers are 0 or not.
1626 __ Orrs(out, LowRegisterFrom(left), HighRegisterFrom(left));
1627
1628 // We use the scope because of the IT block that follows.
1629 ExactAssemblyScope guard(codegen->GetVIXLAssembler(),
1630 2 * vixl32::k16BitT32InstructionSizeInBytes,
1631 CodeBufferCheckScope::kExactSize);
1632
1633 __ it(ne);
1634 __ mov(ne, out, 1);
1635 return;
1636 }
1637
1638 FALLTHROUGH_INTENDED;
1639 case kCondEQ:
1640 case kCondBE:
1641 // We only care if both input registers are 0 or not.
1642 __ Orr(out, LowRegisterFrom(left), HighRegisterFrom(left));
1643 codegen->GenerateConditionWithZero(condition, out, out);
1644 return;
1645 case kCondLT:
1646 case kCondGE:
1647 // We only care about the sign bit.
1648 FALLTHROUGH_INTENDED;
1649 case kCondAE:
1650 case kCondB:
1651 codegen->GenerateConditionWithZero(condition, out, HighRegisterFrom(left));
1652 return;
1653 case kCondLE:
1654 case kCondGT:
1655 default:
1656 break;
1657 }
1658 }
1659 }
1660
1661 // If `out` is a low register, then the GenerateConditionGeneric()
1662 // function generates a shorter code sequence that is still branchless.
1663 if ((condition == kCondEQ || condition == kCondNE) && !out.IsLow()) {
1664 GenerateEqualLong(cond, codegen);
1665 return;
1666 }
1667
1668 GenerateConditionGeneric(cond, codegen);
1669 }
1670
GenerateConditionIntegralOrNonPrimitive(HCondition * cond,CodeGeneratorARMVIXL * codegen)1671 static void GenerateConditionIntegralOrNonPrimitive(HCondition* cond,
1672 CodeGeneratorARMVIXL* codegen) {
1673 const DataType::Type type = cond->GetLeft()->GetType();
1674
1675 DCHECK(DataType::IsIntegralType(type) || type == DataType::Type::kReference) << type;
1676
1677 if (type == DataType::Type::kInt64) {
1678 GenerateConditionLong(cond, codegen);
1679 return;
1680 }
1681
1682 IfCondition condition = cond->GetCondition();
1683 vixl32::Register in = InputRegisterAt(cond, 0);
1684 const vixl32::Register out = OutputRegister(cond);
1685 const Location right = cond->GetLocations()->InAt(1);
1686 int64_t value;
1687
1688 if (right.IsConstant()) {
1689 IfCondition opposite = cond->GetOppositeCondition();
1690
1691 value = AdjustConstantForCondition(Int64ConstantFrom(right), &condition, &opposite);
1692
1693 // Comparisons against 0 are common enough to deserve special attention.
1694 if (value == 0) {
1695 switch (condition) {
1696 case kCondNE:
1697 case kCondA:
1698 if (out.IsLow() && out.Is(in)) {
1699 __ Cmp(out, 0);
1700
1701 // We use the scope because of the IT block that follows.
1702 ExactAssemblyScope guard(codegen->GetVIXLAssembler(),
1703 2 * vixl32::k16BitT32InstructionSizeInBytes,
1704 CodeBufferCheckScope::kExactSize);
1705
1706 __ it(ne);
1707 __ mov(ne, out, 1);
1708 return;
1709 }
1710
1711 FALLTHROUGH_INTENDED;
1712 case kCondEQ:
1713 case kCondBE:
1714 case kCondLT:
1715 case kCondGE:
1716 case kCondAE:
1717 case kCondB:
1718 codegen->GenerateConditionWithZero(condition, out, in);
1719 return;
1720 case kCondLE:
1721 case kCondGT:
1722 default:
1723 break;
1724 }
1725 }
1726 }
1727
1728 if (condition == kCondEQ || condition == kCondNE) {
1729 Operand operand(0);
1730
1731 if (right.IsConstant()) {
1732 operand = Operand::From(value);
1733 } else if (out.Is(RegisterFrom(right))) {
1734 // Avoid 32-bit instructions if possible.
1735 operand = InputOperandAt(cond, 0);
1736 in = RegisterFrom(right);
1737 } else {
1738 operand = InputOperandAt(cond, 1);
1739 }
1740
1741 if (condition == kCondNE && out.IsLow()) {
1742 __ Subs(out, in, operand);
1743
1744 // We use the scope because of the IT block that follows.
1745 ExactAssemblyScope guard(codegen->GetVIXLAssembler(),
1746 2 * vixl32::k16BitT32InstructionSizeInBytes,
1747 CodeBufferCheckScope::kExactSize);
1748
1749 __ it(ne);
1750 __ mov(ne, out, 1);
1751 } else {
1752 __ Sub(out, in, operand);
1753 codegen->GenerateConditionWithZero(condition, out, out);
1754 }
1755
1756 return;
1757 }
1758
1759 GenerateConditionGeneric(cond, codegen);
1760 }
1761
CanEncodeConstantAs8BitImmediate(HConstant * constant)1762 static bool CanEncodeConstantAs8BitImmediate(HConstant* constant) {
1763 const DataType::Type type = constant->GetType();
1764 bool ret = false;
1765
1766 DCHECK(DataType::IsIntegralType(type) || type == DataType::Type::kReference) << type;
1767
1768 if (type == DataType::Type::kInt64) {
1769 const uint64_t value = Uint64ConstantFrom(constant);
1770
1771 ret = IsUint<8>(Low32Bits(value)) && IsUint<8>(High32Bits(value));
1772 } else {
1773 ret = IsUint<8>(Int32ConstantFrom(constant));
1774 }
1775
1776 return ret;
1777 }
1778
Arm8BitEncodableConstantOrRegister(HInstruction * constant)1779 static Location Arm8BitEncodableConstantOrRegister(HInstruction* constant) {
1780 DCHECK(!DataType::IsFloatingPointType(constant->GetType()));
1781
1782 if (constant->IsConstant() && CanEncodeConstantAs8BitImmediate(constant->AsConstant())) {
1783 return Location::ConstantLocation(constant->AsConstant());
1784 }
1785
1786 return Location::RequiresRegister();
1787 }
1788
CanGenerateConditionalMove(const Location & out,const Location & src)1789 static bool CanGenerateConditionalMove(const Location& out, const Location& src) {
1790 // Since IT blocks longer than a 16-bit instruction are deprecated by ARMv8,
1791 // we check that we are not dealing with floating-point output (there is no
1792 // 16-bit VMOV encoding).
1793 if (!out.IsRegister() && !out.IsRegisterPair()) {
1794 return false;
1795 }
1796
1797 // For constants, we also check that the output is in one or two low registers,
1798 // and that the constants fit in an 8-bit unsigned integer, so that a 16-bit
1799 // MOV encoding can be used.
1800 if (src.IsConstant()) {
1801 if (!CanEncodeConstantAs8BitImmediate(src.GetConstant())) {
1802 return false;
1803 }
1804
1805 if (out.IsRegister()) {
1806 if (!RegisterFrom(out).IsLow()) {
1807 return false;
1808 }
1809 } else {
1810 DCHECK(out.IsRegisterPair());
1811
1812 if (!HighRegisterFrom(out).IsLow()) {
1813 return false;
1814 }
1815 }
1816 }
1817
1818 return true;
1819 }
1820
1821 #undef __
1822
GetFinalLabel(HInstruction * instruction,vixl32::Label * final_label)1823 vixl32::Label* CodeGeneratorARMVIXL::GetFinalLabel(HInstruction* instruction,
1824 vixl32::Label* final_label) {
1825 DCHECK(!instruction->IsControlFlow() && !instruction->IsSuspendCheck());
1826 DCHECK(!instruction->IsInvoke() || !instruction->GetLocations()->CanCall());
1827
1828 const HBasicBlock* const block = instruction->GetBlock();
1829 const HLoopInformation* const info = block->GetLoopInformation();
1830 HInstruction* const next = instruction->GetNext();
1831
1832 // Avoid a branch to a branch.
1833 if (next->IsGoto() && (info == nullptr ||
1834 !info->IsBackEdge(*block) ||
1835 !info->HasSuspendCheck())) {
1836 final_label = GetLabelOf(next->AsGoto()->GetSuccessor());
1837 }
1838
1839 return final_label;
1840 }
1841
CodeGeneratorARMVIXL(HGraph * graph,const CompilerOptions & compiler_options,OptimizingCompilerStats * stats)1842 CodeGeneratorARMVIXL::CodeGeneratorARMVIXL(HGraph* graph,
1843 const CompilerOptions& compiler_options,
1844 OptimizingCompilerStats* stats)
1845 : CodeGenerator(graph,
1846 kNumberOfCoreRegisters,
1847 kNumberOfSRegisters,
1848 kNumberOfRegisterPairs,
1849 kCoreCalleeSaves.GetList(),
1850 ComputeSRegisterListMask(kFpuCalleeSaves),
1851 compiler_options,
1852 stats),
1853 block_labels_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
1854 jump_tables_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
1855 location_builder_(graph, this),
1856 instruction_visitor_(graph, this),
1857 move_resolver_(graph->GetAllocator(), this),
1858 assembler_(graph->GetAllocator()),
1859 uint32_literals_(std::less<uint32_t>(),
1860 graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
1861 boot_image_method_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
1862 method_bss_entry_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
1863 boot_image_type_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
1864 type_bss_entry_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
1865 boot_image_string_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
1866 string_bss_entry_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
1867 boot_image_intrinsic_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
1868 baker_read_barrier_patches_(graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
1869 jit_string_patches_(StringReferenceValueComparator(),
1870 graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
1871 jit_class_patches_(TypeReferenceValueComparator(),
1872 graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)),
1873 jit_baker_read_barrier_slow_paths_(std::less<uint32_t>(),
1874 graph->GetAllocator()->Adapter(kArenaAllocCodeGenerator)) {
1875 // Always save the LR register to mimic Quick.
1876 AddAllocatedRegister(Location::RegisterLocation(LR));
1877 // Give D30 and D31 as scratch register to VIXL. The register allocator only works on
1878 // S0-S31, which alias to D0-D15.
1879 GetVIXLAssembler()->GetScratchVRegisterList()->Combine(d31);
1880 GetVIXLAssembler()->GetScratchVRegisterList()->Combine(d30);
1881 }
1882
EmitTable(CodeGeneratorARMVIXL * codegen)1883 void JumpTableARMVIXL::EmitTable(CodeGeneratorARMVIXL* codegen) {
1884 uint32_t num_entries = switch_instr_->GetNumEntries();
1885 DCHECK_GE(num_entries, kPackedSwitchCompareJumpThreshold);
1886
1887 // We are about to use the assembler to place literals directly. Make sure we have enough
1888 // underlying code buffer and we have generated a jump table of the right size, using
1889 // codegen->GetVIXLAssembler()->GetBuffer().Align();
1890 ExactAssemblyScope aas(codegen->GetVIXLAssembler(),
1891 num_entries * sizeof(int32_t),
1892 CodeBufferCheckScope::kMaximumSize);
1893 // TODO(VIXL): Check that using lower case bind is fine here.
1894 codegen->GetVIXLAssembler()->bind(&table_start_);
1895 for (uint32_t i = 0; i < num_entries; i++) {
1896 codegen->GetVIXLAssembler()->place(bb_addresses_[i].get());
1897 }
1898 }
1899
FixTable(CodeGeneratorARMVIXL * codegen)1900 void JumpTableARMVIXL::FixTable(CodeGeneratorARMVIXL* codegen) {
1901 uint32_t num_entries = switch_instr_->GetNumEntries();
1902 DCHECK_GE(num_entries, kPackedSwitchCompareJumpThreshold);
1903
1904 const ArenaVector<HBasicBlock*>& successors = switch_instr_->GetBlock()->GetSuccessors();
1905 for (uint32_t i = 0; i < num_entries; i++) {
1906 vixl32::Label* target_label = codegen->GetLabelOf(successors[i]);
1907 DCHECK(target_label->IsBound());
1908 int32_t jump_offset = target_label->GetLocation() - table_start_.GetLocation();
1909 // When doing BX to address we need to have lower bit set to 1 in T32.
1910 if (codegen->GetVIXLAssembler()->IsUsingT32()) {
1911 jump_offset++;
1912 }
1913 DCHECK_GT(jump_offset, std::numeric_limits<int32_t>::min());
1914 DCHECK_LE(jump_offset, std::numeric_limits<int32_t>::max());
1915
1916 bb_addresses_[i].get()->UpdateValue(jump_offset, codegen->GetVIXLAssembler()->GetBuffer());
1917 }
1918 }
1919
FixJumpTables()1920 void CodeGeneratorARMVIXL::FixJumpTables() {
1921 for (auto&& jump_table : jump_tables_) {
1922 jump_table->FixTable(this);
1923 }
1924 }
1925
1926 #define __ reinterpret_cast<ArmVIXLAssembler*>(GetAssembler())->GetVIXLAssembler()-> // NOLINT
1927
Finalize(CodeAllocator * allocator)1928 void CodeGeneratorARMVIXL::Finalize(CodeAllocator* allocator) {
1929 FixJumpTables();
1930
1931 // Emit JIT baker read barrier slow paths.
1932 DCHECK(Runtime::Current()->UseJitCompilation() || jit_baker_read_barrier_slow_paths_.empty());
1933 for (auto& entry : jit_baker_read_barrier_slow_paths_) {
1934 uint32_t encoded_data = entry.first;
1935 vixl::aarch32::Label* slow_path_entry = &entry.second.label;
1936 __ Bind(slow_path_entry);
1937 CompileBakerReadBarrierThunk(*GetAssembler(), encoded_data, /* debug_name= */ nullptr);
1938 }
1939
1940 GetAssembler()->FinalizeCode();
1941 CodeGenerator::Finalize(allocator);
1942
1943 // Verify Baker read barrier linker patches.
1944 if (kIsDebugBuild) {
1945 ArrayRef<const uint8_t> code = allocator->GetMemory();
1946 for (const BakerReadBarrierPatchInfo& info : baker_read_barrier_patches_) {
1947 DCHECK(info.label.IsBound());
1948 uint32_t literal_offset = info.label.GetLocation();
1949 DCHECK_ALIGNED(literal_offset, 2u);
1950
1951 auto GetInsn16 = [&code](uint32_t offset) {
1952 DCHECK_ALIGNED(offset, 2u);
1953 return (static_cast<uint32_t>(code[offset + 0]) << 0) +
1954 (static_cast<uint32_t>(code[offset + 1]) << 8);
1955 };
1956 auto GetInsn32 = [=](uint32_t offset) {
1957 return (GetInsn16(offset) << 16) + (GetInsn16(offset + 2u) << 0);
1958 };
1959
1960 uint32_t encoded_data = info.custom_data;
1961 BakerReadBarrierKind kind = BakerReadBarrierKindField::Decode(encoded_data);
1962 // Check that the next instruction matches the expected LDR.
1963 switch (kind) {
1964 case BakerReadBarrierKind::kField: {
1965 BakerReadBarrierWidth width = BakerReadBarrierWidthField::Decode(encoded_data);
1966 if (width == BakerReadBarrierWidth::kWide) {
1967 DCHECK_GE(code.size() - literal_offset, 8u);
1968 uint32_t next_insn = GetInsn32(literal_offset + 4u);
1969 // LDR (immediate), encoding T3, with correct base_reg.
1970 CheckValidReg((next_insn >> 12) & 0xfu); // Check destination register.
1971 const uint32_t base_reg = BakerReadBarrierFirstRegField::Decode(encoded_data);
1972 CHECK_EQ(next_insn & 0xffff0000u, 0xf8d00000u | (base_reg << 16));
1973 } else {
1974 DCHECK_GE(code.size() - literal_offset, 6u);
1975 uint32_t next_insn = GetInsn16(literal_offset + 4u);
1976 // LDR (immediate), encoding T1, with correct base_reg.
1977 CheckValidReg(next_insn & 0x7u); // Check destination register.
1978 const uint32_t base_reg = BakerReadBarrierFirstRegField::Decode(encoded_data);
1979 CHECK_EQ(next_insn & 0xf838u, 0x6800u | (base_reg << 3));
1980 }
1981 break;
1982 }
1983 case BakerReadBarrierKind::kArray: {
1984 DCHECK_GE(code.size() - literal_offset, 8u);
1985 uint32_t next_insn = GetInsn32(literal_offset + 4u);
1986 // LDR (register) with correct base_reg, S=1 and option=011 (LDR Wt, [Xn, Xm, LSL #2]).
1987 CheckValidReg((next_insn >> 12) & 0xfu); // Check destination register.
1988 const uint32_t base_reg = BakerReadBarrierFirstRegField::Decode(encoded_data);
1989 CHECK_EQ(next_insn & 0xffff0ff0u, 0xf8500020u | (base_reg << 16));
1990 CheckValidReg(next_insn & 0xf); // Check index register
1991 break;
1992 }
1993 case BakerReadBarrierKind::kGcRoot: {
1994 BakerReadBarrierWidth width = BakerReadBarrierWidthField::Decode(encoded_data);
1995 if (width == BakerReadBarrierWidth::kWide) {
1996 DCHECK_GE(literal_offset, 4u);
1997 uint32_t prev_insn = GetInsn32(literal_offset - 4u);
1998 // LDR (immediate), encoding T3, with correct root_reg.
1999 const uint32_t root_reg = BakerReadBarrierFirstRegField::Decode(encoded_data);
2000 CHECK_EQ(prev_insn & 0xfff0f000u, 0xf8d00000u | (root_reg << 12));
2001 } else {
2002 DCHECK_GE(literal_offset, 2u);
2003 uint32_t prev_insn = GetInsn16(literal_offset - 2u);
2004 // LDR (immediate), encoding T1, with correct root_reg.
2005 const uint32_t root_reg = BakerReadBarrierFirstRegField::Decode(encoded_data);
2006 CHECK_EQ(prev_insn & 0xf807u, 0x6800u | root_reg);
2007 }
2008 break;
2009 }
2010 case BakerReadBarrierKind::kUnsafeCas: {
2011 DCHECK_GE(literal_offset, 4u);
2012 uint32_t prev_insn = GetInsn32(literal_offset - 4u);
2013 // ADD (register), encoding T3, with correct root_reg.
2014 const uint32_t root_reg = BakerReadBarrierFirstRegField::Decode(encoded_data);
2015 CHECK_EQ(prev_insn & 0xfff0fff0u, 0xeb000000u | (root_reg << 8));
2016 break;
2017 }
2018 default:
2019 LOG(FATAL) << "Unexpected kind: " << static_cast<uint32_t>(kind);
2020 UNREACHABLE();
2021 }
2022 }
2023 }
2024 }
2025
SetupBlockedRegisters() const2026 void CodeGeneratorARMVIXL::SetupBlockedRegisters() const {
2027 // Stack register, LR and PC are always reserved.
2028 blocked_core_registers_[SP] = true;
2029 blocked_core_registers_[LR] = true;
2030 blocked_core_registers_[PC] = true;
2031
2032 if (kEmitCompilerReadBarrier && kUseBakerReadBarrier) {
2033 // Reserve marking register.
2034 blocked_core_registers_[MR] = true;
2035 }
2036
2037 // Reserve thread register.
2038 blocked_core_registers_[TR] = true;
2039
2040 // Reserve temp register.
2041 blocked_core_registers_[IP] = true;
2042
2043 if (GetGraph()->IsDebuggable()) {
2044 // Stubs do not save callee-save floating point registers. If the graph
2045 // is debuggable, we need to deal with these registers differently. For
2046 // now, just block them.
2047 for (uint32_t i = kFpuCalleeSaves.GetFirstSRegister().GetCode();
2048 i <= kFpuCalleeSaves.GetLastSRegister().GetCode();
2049 ++i) {
2050 blocked_fpu_registers_[i] = true;
2051 }
2052 }
2053 }
2054
InstructionCodeGeneratorARMVIXL(HGraph * graph,CodeGeneratorARMVIXL * codegen)2055 InstructionCodeGeneratorARMVIXL::InstructionCodeGeneratorARMVIXL(HGraph* graph,
2056 CodeGeneratorARMVIXL* codegen)
2057 : InstructionCodeGenerator(graph, codegen),
2058 assembler_(codegen->GetAssembler()),
2059 codegen_(codegen) {}
2060
ComputeSpillMask()2061 void CodeGeneratorARMVIXL::ComputeSpillMask() {
2062 core_spill_mask_ = allocated_registers_.GetCoreRegisters() & core_callee_save_mask_;
2063 DCHECK_NE(core_spill_mask_, 0u) << "At least the return address register must be saved";
2064 // There is no easy instruction to restore just the PC on thumb2. We spill and
2065 // restore another arbitrary register.
2066 core_spill_mask_ |= (1 << kCoreAlwaysSpillRegister.GetCode());
2067 fpu_spill_mask_ = allocated_registers_.GetFloatingPointRegisters() & fpu_callee_save_mask_;
2068 // We use vpush and vpop for saving and restoring floating point registers, which take
2069 // a SRegister and the number of registers to save/restore after that SRegister. We
2070 // therefore update the `fpu_spill_mask_` to also contain those registers not allocated,
2071 // but in the range.
2072 if (fpu_spill_mask_ != 0) {
2073 uint32_t least_significant_bit = LeastSignificantBit(fpu_spill_mask_);
2074 uint32_t most_significant_bit = MostSignificantBit(fpu_spill_mask_);
2075 for (uint32_t i = least_significant_bit + 1 ; i < most_significant_bit; ++i) {
2076 fpu_spill_mask_ |= (1 << i);
2077 }
2078 }
2079 }
2080
GenerateFrameEntry()2081 void CodeGeneratorARMVIXL::GenerateFrameEntry() {
2082 bool skip_overflow_check =
2083 IsLeafMethod() && !FrameNeedsStackCheck(GetFrameSize(), InstructionSet::kArm);
2084 DCHECK(GetCompilerOptions().GetImplicitStackOverflowChecks());
2085 __ Bind(&frame_entry_label_);
2086
2087 if (GetCompilerOptions().CountHotnessInCompiledCode()) {
2088 UseScratchRegisterScope temps(GetVIXLAssembler());
2089 vixl32::Register temp = temps.Acquire();
2090 __ Ldrh(temp, MemOperand(kMethodRegister, ArtMethod::HotnessCountOffset().Int32Value()));
2091 __ Add(temp, temp, 1);
2092 __ Strh(temp, MemOperand(kMethodRegister, ArtMethod::HotnessCountOffset().Int32Value()));
2093 }
2094
2095 if (HasEmptyFrame()) {
2096 // Ensure that the CFI opcode list is not empty.
2097 GetAssembler()->cfi().Nop();
2098 return;
2099 }
2100
2101 if (!skip_overflow_check) {
2102 // Using r4 instead of IP saves 2 bytes.
2103 UseScratchRegisterScope temps(GetVIXLAssembler());
2104 vixl32::Register temp;
2105 // TODO: Remove this check when R4 is made a callee-save register
2106 // in ART compiled code (b/72801708). Currently we need to make
2107 // sure r4 is not blocked, e.g. in special purpose
2108 // TestCodeGeneratorARMVIXL; also asserting that r4 is available
2109 // here.
2110 if (!blocked_core_registers_[R4]) {
2111 for (vixl32::Register reg : kParameterCoreRegistersVIXL) {
2112 DCHECK(!reg.Is(r4));
2113 }
2114 DCHECK(!kCoreCalleeSaves.Includes(r4));
2115 temp = r4;
2116 } else {
2117 temp = temps.Acquire();
2118 }
2119 __ Sub(temp, sp, Operand::From(GetStackOverflowReservedBytes(InstructionSet::kArm)));
2120 // The load must immediately precede RecordPcInfo.
2121 ExactAssemblyScope aas(GetVIXLAssembler(),
2122 vixl32::kMaxInstructionSizeInBytes,
2123 CodeBufferCheckScope::kMaximumSize);
2124 __ ldr(temp, MemOperand(temp));
2125 RecordPcInfo(nullptr, 0);
2126 }
2127
2128 __ Push(RegisterList(core_spill_mask_));
2129 GetAssembler()->cfi().AdjustCFAOffset(kArmWordSize * POPCOUNT(core_spill_mask_));
2130 GetAssembler()->cfi().RelOffsetForMany(DWARFReg(kMethodRegister),
2131 0,
2132 core_spill_mask_,
2133 kArmWordSize);
2134 if (fpu_spill_mask_ != 0) {
2135 uint32_t first = LeastSignificantBit(fpu_spill_mask_);
2136
2137 // Check that list is contiguous.
2138 DCHECK_EQ(fpu_spill_mask_ >> CTZ(fpu_spill_mask_), ~0u >> (32 - POPCOUNT(fpu_spill_mask_)));
2139
2140 __ Vpush(SRegisterList(vixl32::SRegister(first), POPCOUNT(fpu_spill_mask_)));
2141 GetAssembler()->cfi().AdjustCFAOffset(kArmWordSize * POPCOUNT(fpu_spill_mask_));
2142 GetAssembler()->cfi().RelOffsetForMany(DWARFReg(s0), 0, fpu_spill_mask_, kArmWordSize);
2143 }
2144
2145 int adjust = GetFrameSize() - FrameEntrySpillSize();
2146 __ Sub(sp, sp, adjust);
2147 GetAssembler()->cfi().AdjustCFAOffset(adjust);
2148
2149 // Save the current method if we need it. Note that we do not
2150 // do this in HCurrentMethod, as the instruction might have been removed
2151 // in the SSA graph.
2152 if (RequiresCurrentMethod()) {
2153 GetAssembler()->StoreToOffset(kStoreWord, kMethodRegister, sp, 0);
2154 }
2155
2156 if (GetGraph()->HasShouldDeoptimizeFlag()) {
2157 UseScratchRegisterScope temps(GetVIXLAssembler());
2158 vixl32::Register temp = temps.Acquire();
2159 // Initialize should_deoptimize flag to 0.
2160 __ Mov(temp, 0);
2161 GetAssembler()->StoreToOffset(kStoreWord, temp, sp, GetStackOffsetOfShouldDeoptimizeFlag());
2162 }
2163
2164 MaybeGenerateMarkingRegisterCheck(/* code= */ 1);
2165 }
2166
GenerateFrameExit()2167 void CodeGeneratorARMVIXL::GenerateFrameExit() {
2168 if (HasEmptyFrame()) {
2169 __ Bx(lr);
2170 return;
2171 }
2172 GetAssembler()->cfi().RememberState();
2173 int adjust = GetFrameSize() - FrameEntrySpillSize();
2174 __ Add(sp, sp, adjust);
2175 GetAssembler()->cfi().AdjustCFAOffset(-adjust);
2176 if (fpu_spill_mask_ != 0) {
2177 uint32_t first = LeastSignificantBit(fpu_spill_mask_);
2178
2179 // Check that list is contiguous.
2180 DCHECK_EQ(fpu_spill_mask_ >> CTZ(fpu_spill_mask_), ~0u >> (32 - POPCOUNT(fpu_spill_mask_)));
2181
2182 __ Vpop(SRegisterList(vixl32::SRegister(first), POPCOUNT(fpu_spill_mask_)));
2183 GetAssembler()->cfi().AdjustCFAOffset(
2184 -static_cast<int>(kArmWordSize) * POPCOUNT(fpu_spill_mask_));
2185 GetAssembler()->cfi().RestoreMany(DWARFReg(vixl32::SRegister(0)), fpu_spill_mask_);
2186 }
2187 // Pop LR into PC to return.
2188 DCHECK_NE(core_spill_mask_ & (1 << kLrCode), 0U);
2189 uint32_t pop_mask = (core_spill_mask_ & (~(1 << kLrCode))) | 1 << kPcCode;
2190 __ Pop(RegisterList(pop_mask));
2191 GetAssembler()->cfi().RestoreState();
2192 GetAssembler()->cfi().DefCFAOffset(GetFrameSize());
2193 }
2194
Bind(HBasicBlock * block)2195 void CodeGeneratorARMVIXL::Bind(HBasicBlock* block) {
2196 __ Bind(GetLabelOf(block));
2197 }
2198
GetNextLocation(DataType::Type type)2199 Location InvokeDexCallingConventionVisitorARMVIXL::GetNextLocation(DataType::Type type) {
2200 switch (type) {
2201 case DataType::Type::kReference:
2202 case DataType::Type::kBool:
2203 case DataType::Type::kUint8:
2204 case DataType::Type::kInt8:
2205 case DataType::Type::kUint16:
2206 case DataType::Type::kInt16:
2207 case DataType::Type::kInt32: {
2208 uint32_t index = gp_index_++;
2209 uint32_t stack_index = stack_index_++;
2210 if (index < calling_convention.GetNumberOfRegisters()) {
2211 return LocationFrom(calling_convention.GetRegisterAt(index));
2212 } else {
2213 return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index));
2214 }
2215 }
2216
2217 case DataType::Type::kInt64: {
2218 uint32_t index = gp_index_;
2219 uint32_t stack_index = stack_index_;
2220 gp_index_ += 2;
2221 stack_index_ += 2;
2222 if (index + 1 < calling_convention.GetNumberOfRegisters()) {
2223 if (calling_convention.GetRegisterAt(index).Is(r1)) {
2224 // Skip R1, and use R2_R3 instead.
2225 gp_index_++;
2226 index++;
2227 }
2228 }
2229 if (index + 1 < calling_convention.GetNumberOfRegisters()) {
2230 DCHECK_EQ(calling_convention.GetRegisterAt(index).GetCode() + 1,
2231 calling_convention.GetRegisterAt(index + 1).GetCode());
2232
2233 return LocationFrom(calling_convention.GetRegisterAt(index),
2234 calling_convention.GetRegisterAt(index + 1));
2235 } else {
2236 return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index));
2237 }
2238 }
2239
2240 case DataType::Type::kFloat32: {
2241 uint32_t stack_index = stack_index_++;
2242 if (float_index_ % 2 == 0) {
2243 float_index_ = std::max(double_index_, float_index_);
2244 }
2245 if (float_index_ < calling_convention.GetNumberOfFpuRegisters()) {
2246 return LocationFrom(calling_convention.GetFpuRegisterAt(float_index_++));
2247 } else {
2248 return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index));
2249 }
2250 }
2251
2252 case DataType::Type::kFloat64: {
2253 double_index_ = std::max(double_index_, RoundUp(float_index_, 2));
2254 uint32_t stack_index = stack_index_;
2255 stack_index_ += 2;
2256 if (double_index_ + 1 < calling_convention.GetNumberOfFpuRegisters()) {
2257 uint32_t index = double_index_;
2258 double_index_ += 2;
2259 Location result = LocationFrom(
2260 calling_convention.GetFpuRegisterAt(index),
2261 calling_convention.GetFpuRegisterAt(index + 1));
2262 DCHECK(ExpectedPairLayout(result));
2263 return result;
2264 } else {
2265 return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index));
2266 }
2267 }
2268
2269 case DataType::Type::kUint32:
2270 case DataType::Type::kUint64:
2271 case DataType::Type::kVoid:
2272 LOG(FATAL) << "Unexpected parameter type " << type;
2273 UNREACHABLE();
2274 }
2275 return Location::NoLocation();
2276 }
2277
GetReturnLocation(DataType::Type type) const2278 Location InvokeDexCallingConventionVisitorARMVIXL::GetReturnLocation(DataType::Type type) const {
2279 switch (type) {
2280 case DataType::Type::kReference:
2281 case DataType::Type::kBool:
2282 case DataType::Type::kUint8:
2283 case DataType::Type::kInt8:
2284 case DataType::Type::kUint16:
2285 case DataType::Type::kInt16:
2286 case DataType::Type::kUint32:
2287 case DataType::Type::kInt32: {
2288 return LocationFrom(r0);
2289 }
2290
2291 case DataType::Type::kFloat32: {
2292 return LocationFrom(s0);
2293 }
2294
2295 case DataType::Type::kUint64:
2296 case DataType::Type::kInt64: {
2297 return LocationFrom(r0, r1);
2298 }
2299
2300 case DataType::Type::kFloat64: {
2301 return LocationFrom(s0, s1);
2302 }
2303
2304 case DataType::Type::kVoid:
2305 return Location::NoLocation();
2306 }
2307
2308 UNREACHABLE();
2309 }
2310
GetMethodLocation() const2311 Location InvokeDexCallingConventionVisitorARMVIXL::GetMethodLocation() const {
2312 return LocationFrom(kMethodRegister);
2313 }
2314
Move32(Location destination,Location source)2315 void CodeGeneratorARMVIXL::Move32(Location destination, Location source) {
2316 if (source.Equals(destination)) {
2317 return;
2318 }
2319 if (destination.IsRegister()) {
2320 if (source.IsRegister()) {
2321 __ Mov(RegisterFrom(destination), RegisterFrom(source));
2322 } else if (source.IsFpuRegister()) {
2323 __ Vmov(RegisterFrom(destination), SRegisterFrom(source));
2324 } else {
2325 GetAssembler()->LoadFromOffset(kLoadWord,
2326 RegisterFrom(destination),
2327 sp,
2328 source.GetStackIndex());
2329 }
2330 } else if (destination.IsFpuRegister()) {
2331 if (source.IsRegister()) {
2332 __ Vmov(SRegisterFrom(destination), RegisterFrom(source));
2333 } else if (source.IsFpuRegister()) {
2334 __ Vmov(SRegisterFrom(destination), SRegisterFrom(source));
2335 } else {
2336 GetAssembler()->LoadSFromOffset(SRegisterFrom(destination), sp, source.GetStackIndex());
2337 }
2338 } else {
2339 DCHECK(destination.IsStackSlot()) << destination;
2340 if (source.IsRegister()) {
2341 GetAssembler()->StoreToOffset(kStoreWord,
2342 RegisterFrom(source),
2343 sp,
2344 destination.GetStackIndex());
2345 } else if (source.IsFpuRegister()) {
2346 GetAssembler()->StoreSToOffset(SRegisterFrom(source), sp, destination.GetStackIndex());
2347 } else {
2348 DCHECK(source.IsStackSlot()) << source;
2349 UseScratchRegisterScope temps(GetVIXLAssembler());
2350 vixl32::Register temp = temps.Acquire();
2351 GetAssembler()->LoadFromOffset(kLoadWord, temp, sp, source.GetStackIndex());
2352 GetAssembler()->StoreToOffset(kStoreWord, temp, sp, destination.GetStackIndex());
2353 }
2354 }
2355 }
2356
MoveConstant(Location location,int32_t value)2357 void CodeGeneratorARMVIXL::MoveConstant(Location location, int32_t value) {
2358 DCHECK(location.IsRegister());
2359 __ Mov(RegisterFrom(location), value);
2360 }
2361
MoveLocation(Location dst,Location src,DataType::Type dst_type)2362 void CodeGeneratorARMVIXL::MoveLocation(Location dst, Location src, DataType::Type dst_type) {
2363 // TODO(VIXL): Maybe refactor to have the 'move' implementation here and use it in
2364 // `ParallelMoveResolverARMVIXL::EmitMove`, as is done in the `arm64` backend.
2365 HParallelMove move(GetGraph()->GetAllocator());
2366 move.AddMove(src, dst, dst_type, nullptr);
2367 GetMoveResolver()->EmitNativeCode(&move);
2368 }
2369
AddLocationAsTemp(Location location,LocationSummary * locations)2370 void CodeGeneratorARMVIXL::AddLocationAsTemp(Location location, LocationSummary* locations) {
2371 if (location.IsRegister()) {
2372 locations->AddTemp(location);
2373 } else if (location.IsRegisterPair()) {
2374 locations->AddTemp(LocationFrom(LowRegisterFrom(location)));
2375 locations->AddTemp(LocationFrom(HighRegisterFrom(location)));
2376 } else {
2377 UNIMPLEMENTED(FATAL) << "AddLocationAsTemp not implemented for location " << location;
2378 }
2379 }
2380
InvokeRuntime(QuickEntrypointEnum entrypoint,HInstruction * instruction,uint32_t dex_pc,SlowPathCode * slow_path)2381 void CodeGeneratorARMVIXL::InvokeRuntime(QuickEntrypointEnum entrypoint,
2382 HInstruction* instruction,
2383 uint32_t dex_pc,
2384 SlowPathCode* slow_path) {
2385 ValidateInvokeRuntime(entrypoint, instruction, slow_path);
2386 __ Ldr(lr, MemOperand(tr, GetThreadOffset<kArmPointerSize>(entrypoint).Int32Value()));
2387 // Ensure the pc position is recorded immediately after the `blx` instruction.
2388 // blx in T32 has only 16bit encoding that's why a stricter check for the scope is used.
2389 ExactAssemblyScope aas(GetVIXLAssembler(),
2390 vixl32::k16BitT32InstructionSizeInBytes,
2391 CodeBufferCheckScope::kExactSize);
2392 __ blx(lr);
2393 if (EntrypointRequiresStackMap(entrypoint)) {
2394 RecordPcInfo(instruction, dex_pc, slow_path);
2395 }
2396 }
2397
InvokeRuntimeWithoutRecordingPcInfo(int32_t entry_point_offset,HInstruction * instruction,SlowPathCode * slow_path)2398 void CodeGeneratorARMVIXL::InvokeRuntimeWithoutRecordingPcInfo(int32_t entry_point_offset,
2399 HInstruction* instruction,
2400 SlowPathCode* slow_path) {
2401 ValidateInvokeRuntimeWithoutRecordingPcInfo(instruction, slow_path);
2402 __ Ldr(lr, MemOperand(tr, entry_point_offset));
2403 __ Blx(lr);
2404 }
2405
HandleGoto(HInstruction * got,HBasicBlock * successor)2406 void InstructionCodeGeneratorARMVIXL::HandleGoto(HInstruction* got, HBasicBlock* successor) {
2407 if (successor->IsExitBlock()) {
2408 DCHECK(got->GetPrevious()->AlwaysThrows());
2409 return; // no code needed
2410 }
2411
2412 HBasicBlock* block = got->GetBlock();
2413 HInstruction* previous = got->GetPrevious();
2414 HLoopInformation* info = block->GetLoopInformation();
2415
2416 if (info != nullptr && info->IsBackEdge(*block) && info->HasSuspendCheck()) {
2417 if (codegen_->GetCompilerOptions().CountHotnessInCompiledCode()) {
2418 UseScratchRegisterScope temps(GetVIXLAssembler());
2419 vixl32::Register temp = temps.Acquire();
2420 __ Push(vixl32::Register(kMethodRegister));
2421 GetAssembler()->LoadFromOffset(kLoadWord, kMethodRegister, sp, kArmWordSize);
2422 __ Ldrh(temp, MemOperand(kMethodRegister, ArtMethod::HotnessCountOffset().Int32Value()));
2423 __ Add(temp, temp, 1);
2424 __ Strh(temp, MemOperand(kMethodRegister, ArtMethod::HotnessCountOffset().Int32Value()));
2425 __ Pop(vixl32::Register(kMethodRegister));
2426 }
2427 GenerateSuspendCheck(info->GetSuspendCheck(), successor);
2428 return;
2429 }
2430 if (block->IsEntryBlock() && (previous != nullptr) && previous->IsSuspendCheck()) {
2431 GenerateSuspendCheck(previous->AsSuspendCheck(), nullptr);
2432 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 2);
2433 }
2434 if (!codegen_->GoesToNextBlock(block, successor)) {
2435 __ B(codegen_->GetLabelOf(successor));
2436 }
2437 }
2438
VisitGoto(HGoto * got)2439 void LocationsBuilderARMVIXL::VisitGoto(HGoto* got) {
2440 got->SetLocations(nullptr);
2441 }
2442
VisitGoto(HGoto * got)2443 void InstructionCodeGeneratorARMVIXL::VisitGoto(HGoto* got) {
2444 HandleGoto(got, got->GetSuccessor());
2445 }
2446
VisitTryBoundary(HTryBoundary * try_boundary)2447 void LocationsBuilderARMVIXL::VisitTryBoundary(HTryBoundary* try_boundary) {
2448 try_boundary->SetLocations(nullptr);
2449 }
2450
VisitTryBoundary(HTryBoundary * try_boundary)2451 void InstructionCodeGeneratorARMVIXL::VisitTryBoundary(HTryBoundary* try_boundary) {
2452 HBasicBlock* successor = try_boundary->GetNormalFlowSuccessor();
2453 if (!successor->IsExitBlock()) {
2454 HandleGoto(try_boundary, successor);
2455 }
2456 }
2457
VisitExit(HExit * exit)2458 void LocationsBuilderARMVIXL::VisitExit(HExit* exit) {
2459 exit->SetLocations(nullptr);
2460 }
2461
VisitExit(HExit * exit ATTRIBUTE_UNUSED)2462 void InstructionCodeGeneratorARMVIXL::VisitExit(HExit* exit ATTRIBUTE_UNUSED) {
2463 }
2464
GenerateCompareTestAndBranch(HCondition * condition,vixl32::Label * true_target,vixl32::Label * false_target,bool is_far_target)2465 void InstructionCodeGeneratorARMVIXL::GenerateCompareTestAndBranch(HCondition* condition,
2466 vixl32::Label* true_target,
2467 vixl32::Label* false_target,
2468 bool is_far_target) {
2469 if (true_target == false_target) {
2470 DCHECK(true_target != nullptr);
2471 __ B(true_target);
2472 return;
2473 }
2474
2475 vixl32::Label* non_fallthrough_target;
2476 bool invert;
2477 bool emit_both_branches;
2478
2479 if (true_target == nullptr) {
2480 // The true target is fallthrough.
2481 DCHECK(false_target != nullptr);
2482 non_fallthrough_target = false_target;
2483 invert = true;
2484 emit_both_branches = false;
2485 } else {
2486 non_fallthrough_target = true_target;
2487 invert = false;
2488 // Either the false target is fallthrough, or there is no fallthrough
2489 // and both branches must be emitted.
2490 emit_both_branches = (false_target != nullptr);
2491 }
2492
2493 const auto cond = GenerateTest(condition, invert, codegen_);
2494
2495 __ B(cond.first, non_fallthrough_target, is_far_target);
2496
2497 if (emit_both_branches) {
2498 // No target falls through, we need to branch.
2499 __ B(false_target);
2500 }
2501 }
2502
GenerateTestAndBranch(HInstruction * instruction,size_t condition_input_index,vixl32::Label * true_target,vixl32::Label * false_target,bool far_target)2503 void InstructionCodeGeneratorARMVIXL::GenerateTestAndBranch(HInstruction* instruction,
2504 size_t condition_input_index,
2505 vixl32::Label* true_target,
2506 vixl32::Label* false_target,
2507 bool far_target) {
2508 HInstruction* cond = instruction->InputAt(condition_input_index);
2509
2510 if (true_target == nullptr && false_target == nullptr) {
2511 // Nothing to do. The code always falls through.
2512 return;
2513 } else if (cond->IsIntConstant()) {
2514 // Constant condition, statically compared against "true" (integer value 1).
2515 if (cond->AsIntConstant()->IsTrue()) {
2516 if (true_target != nullptr) {
2517 __ B(true_target);
2518 }
2519 } else {
2520 DCHECK(cond->AsIntConstant()->IsFalse()) << Int32ConstantFrom(cond);
2521 if (false_target != nullptr) {
2522 __ B(false_target);
2523 }
2524 }
2525 return;
2526 }
2527
2528 // The following code generates these patterns:
2529 // (1) true_target == nullptr && false_target != nullptr
2530 // - opposite condition true => branch to false_target
2531 // (2) true_target != nullptr && false_target == nullptr
2532 // - condition true => branch to true_target
2533 // (3) true_target != nullptr && false_target != nullptr
2534 // - condition true => branch to true_target
2535 // - branch to false_target
2536 if (IsBooleanValueOrMaterializedCondition(cond)) {
2537 // Condition has been materialized, compare the output to 0.
2538 if (kIsDebugBuild) {
2539 Location cond_val = instruction->GetLocations()->InAt(condition_input_index);
2540 DCHECK(cond_val.IsRegister());
2541 }
2542 if (true_target == nullptr) {
2543 __ CompareAndBranchIfZero(InputRegisterAt(instruction, condition_input_index),
2544 false_target,
2545 far_target);
2546 } else {
2547 __ CompareAndBranchIfNonZero(InputRegisterAt(instruction, condition_input_index),
2548 true_target,
2549 far_target);
2550 }
2551 } else {
2552 // Condition has not been materialized. Use its inputs as the comparison and
2553 // its condition as the branch condition.
2554 HCondition* condition = cond->AsCondition();
2555
2556 // If this is a long or FP comparison that has been folded into
2557 // the HCondition, generate the comparison directly.
2558 DataType::Type type = condition->InputAt(0)->GetType();
2559 if (type == DataType::Type::kInt64 || DataType::IsFloatingPointType(type)) {
2560 GenerateCompareTestAndBranch(condition, true_target, false_target, far_target);
2561 return;
2562 }
2563
2564 vixl32::Label* non_fallthrough_target;
2565 vixl32::Condition arm_cond = vixl32::Condition::None();
2566 const vixl32::Register left = InputRegisterAt(cond, 0);
2567 const Operand right = InputOperandAt(cond, 1);
2568
2569 if (true_target == nullptr) {
2570 arm_cond = ARMCondition(condition->GetOppositeCondition());
2571 non_fallthrough_target = false_target;
2572 } else {
2573 arm_cond = ARMCondition(condition->GetCondition());
2574 non_fallthrough_target = true_target;
2575 }
2576
2577 if (right.IsImmediate() && right.GetImmediate() == 0 && (arm_cond.Is(ne) || arm_cond.Is(eq))) {
2578 if (arm_cond.Is(eq)) {
2579 __ CompareAndBranchIfZero(left, non_fallthrough_target, far_target);
2580 } else {
2581 DCHECK(arm_cond.Is(ne));
2582 __ CompareAndBranchIfNonZero(left, non_fallthrough_target, far_target);
2583 }
2584 } else {
2585 __ Cmp(left, right);
2586 __ B(arm_cond, non_fallthrough_target, far_target);
2587 }
2588 }
2589
2590 // If neither branch falls through (case 3), the conditional branch to `true_target`
2591 // was already emitted (case 2) and we need to emit a jump to `false_target`.
2592 if (true_target != nullptr && false_target != nullptr) {
2593 __ B(false_target);
2594 }
2595 }
2596
VisitIf(HIf * if_instr)2597 void LocationsBuilderARMVIXL::VisitIf(HIf* if_instr) {
2598 LocationSummary* locations = new (GetGraph()->GetAllocator()) LocationSummary(if_instr);
2599 if (IsBooleanValueOrMaterializedCondition(if_instr->InputAt(0))) {
2600 locations->SetInAt(0, Location::RequiresRegister());
2601 }
2602 }
2603
VisitIf(HIf * if_instr)2604 void InstructionCodeGeneratorARMVIXL::VisitIf(HIf* if_instr) {
2605 HBasicBlock* true_successor = if_instr->IfTrueSuccessor();
2606 HBasicBlock* false_successor = if_instr->IfFalseSuccessor();
2607 vixl32::Label* true_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), true_successor) ?
2608 nullptr : codegen_->GetLabelOf(true_successor);
2609 vixl32::Label* false_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), false_successor) ?
2610 nullptr : codegen_->GetLabelOf(false_successor);
2611 GenerateTestAndBranch(if_instr, /* condition_input_index= */ 0, true_target, false_target);
2612 }
2613
VisitDeoptimize(HDeoptimize * deoptimize)2614 void LocationsBuilderARMVIXL::VisitDeoptimize(HDeoptimize* deoptimize) {
2615 LocationSummary* locations = new (GetGraph()->GetAllocator())
2616 LocationSummary(deoptimize, LocationSummary::kCallOnSlowPath);
2617 InvokeRuntimeCallingConventionARMVIXL calling_convention;
2618 RegisterSet caller_saves = RegisterSet::Empty();
2619 caller_saves.Add(LocationFrom(calling_convention.GetRegisterAt(0)));
2620 locations->SetCustomSlowPathCallerSaves(caller_saves);
2621 if (IsBooleanValueOrMaterializedCondition(deoptimize->InputAt(0))) {
2622 locations->SetInAt(0, Location::RequiresRegister());
2623 }
2624 }
2625
VisitDeoptimize(HDeoptimize * deoptimize)2626 void InstructionCodeGeneratorARMVIXL::VisitDeoptimize(HDeoptimize* deoptimize) {
2627 SlowPathCodeARMVIXL* slow_path =
2628 deopt_slow_paths_.NewSlowPath<DeoptimizationSlowPathARMVIXL>(deoptimize);
2629 GenerateTestAndBranch(deoptimize,
2630 /* condition_input_index= */ 0,
2631 slow_path->GetEntryLabel(),
2632 /* false_target= */ nullptr);
2633 }
2634
VisitShouldDeoptimizeFlag(HShouldDeoptimizeFlag * flag)2635 void LocationsBuilderARMVIXL::VisitShouldDeoptimizeFlag(HShouldDeoptimizeFlag* flag) {
2636 LocationSummary* locations = new (GetGraph()->GetAllocator())
2637 LocationSummary(flag, LocationSummary::kNoCall);
2638 locations->SetOut(Location::RequiresRegister());
2639 }
2640
VisitShouldDeoptimizeFlag(HShouldDeoptimizeFlag * flag)2641 void InstructionCodeGeneratorARMVIXL::VisitShouldDeoptimizeFlag(HShouldDeoptimizeFlag* flag) {
2642 GetAssembler()->LoadFromOffset(kLoadWord,
2643 OutputRegister(flag),
2644 sp,
2645 codegen_->GetStackOffsetOfShouldDeoptimizeFlag());
2646 }
2647
VisitSelect(HSelect * select)2648 void LocationsBuilderARMVIXL::VisitSelect(HSelect* select) {
2649 LocationSummary* locations = new (GetGraph()->GetAllocator()) LocationSummary(select);
2650 const bool is_floating_point = DataType::IsFloatingPointType(select->GetType());
2651
2652 if (is_floating_point) {
2653 locations->SetInAt(0, Location::RequiresFpuRegister());
2654 locations->SetInAt(1, Location::FpuRegisterOrConstant(select->GetTrueValue()));
2655 } else {
2656 locations->SetInAt(0, Location::RequiresRegister());
2657 locations->SetInAt(1, Arm8BitEncodableConstantOrRegister(select->GetTrueValue()));
2658 }
2659
2660 if (IsBooleanValueOrMaterializedCondition(select->GetCondition())) {
2661 locations->SetInAt(2, Location::RegisterOrConstant(select->GetCondition()));
2662 // The code generator handles overlap with the values, but not with the condition.
2663 locations->SetOut(Location::SameAsFirstInput());
2664 } else if (is_floating_point) {
2665 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
2666 } else {
2667 if (!locations->InAt(1).IsConstant()) {
2668 locations->SetInAt(0, Arm8BitEncodableConstantOrRegister(select->GetFalseValue()));
2669 }
2670
2671 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
2672 }
2673 }
2674
VisitSelect(HSelect * select)2675 void InstructionCodeGeneratorARMVIXL::VisitSelect(HSelect* select) {
2676 HInstruction* const condition = select->GetCondition();
2677 const LocationSummary* const locations = select->GetLocations();
2678 const DataType::Type type = select->GetType();
2679 const Location first = locations->InAt(0);
2680 const Location out = locations->Out();
2681 const Location second = locations->InAt(1);
2682
2683 // In the unlucky case the output of this instruction overlaps
2684 // with an input of an "emitted-at-use-site" condition, and
2685 // the output of this instruction is not one of its inputs, we'll
2686 // need to fallback to branches instead of conditional ARM instructions.
2687 bool output_overlaps_with_condition_inputs =
2688 !IsBooleanValueOrMaterializedCondition(condition) &&
2689 !out.Equals(first) &&
2690 !out.Equals(second) &&
2691 (condition->GetLocations()->InAt(0).Equals(out) ||
2692 condition->GetLocations()->InAt(1).Equals(out));
2693 DCHECK(!output_overlaps_with_condition_inputs || condition->IsCondition());
2694 Location src;
2695
2696 if (condition->IsIntConstant()) {
2697 if (condition->AsIntConstant()->IsFalse()) {
2698 src = first;
2699 } else {
2700 src = second;
2701 }
2702
2703 codegen_->MoveLocation(out, src, type);
2704 return;
2705 }
2706
2707 if (!DataType::IsFloatingPointType(type) && !output_overlaps_with_condition_inputs) {
2708 bool invert = false;
2709
2710 if (out.Equals(second)) {
2711 src = first;
2712 invert = true;
2713 } else if (out.Equals(first)) {
2714 src = second;
2715 } else if (second.IsConstant()) {
2716 DCHECK(CanEncodeConstantAs8BitImmediate(second.GetConstant()));
2717 src = second;
2718 } else if (first.IsConstant()) {
2719 DCHECK(CanEncodeConstantAs8BitImmediate(first.GetConstant()));
2720 src = first;
2721 invert = true;
2722 } else {
2723 src = second;
2724 }
2725
2726 if (CanGenerateConditionalMove(out, src)) {
2727 if (!out.Equals(first) && !out.Equals(second)) {
2728 codegen_->MoveLocation(out, src.Equals(first) ? second : first, type);
2729 }
2730
2731 std::pair<vixl32::Condition, vixl32::Condition> cond(eq, ne);
2732
2733 if (IsBooleanValueOrMaterializedCondition(condition)) {
2734 __ Cmp(InputRegisterAt(select, 2), 0);
2735 cond = invert ? std::make_pair(eq, ne) : std::make_pair(ne, eq);
2736 } else {
2737 cond = GenerateTest(condition->AsCondition(), invert, codegen_);
2738 }
2739
2740 const size_t instr_count = out.IsRegisterPair() ? 4 : 2;
2741 // We use the scope because of the IT block that follows.
2742 ExactAssemblyScope guard(GetVIXLAssembler(),
2743 instr_count * vixl32::k16BitT32InstructionSizeInBytes,
2744 CodeBufferCheckScope::kExactSize);
2745
2746 if (out.IsRegister()) {
2747 __ it(cond.first);
2748 __ mov(cond.first, RegisterFrom(out), OperandFrom(src, type));
2749 } else {
2750 DCHECK(out.IsRegisterPair());
2751
2752 Operand operand_high(0);
2753 Operand operand_low(0);
2754
2755 if (src.IsConstant()) {
2756 const int64_t value = Int64ConstantFrom(src);
2757
2758 operand_high = High32Bits(value);
2759 operand_low = Low32Bits(value);
2760 } else {
2761 DCHECK(src.IsRegisterPair());
2762 operand_high = HighRegisterFrom(src);
2763 operand_low = LowRegisterFrom(src);
2764 }
2765
2766 __ it(cond.first);
2767 __ mov(cond.first, LowRegisterFrom(out), operand_low);
2768 __ it(cond.first);
2769 __ mov(cond.first, HighRegisterFrom(out), operand_high);
2770 }
2771
2772 return;
2773 }
2774 }
2775
2776 vixl32::Label* false_target = nullptr;
2777 vixl32::Label* true_target = nullptr;
2778 vixl32::Label select_end;
2779 vixl32::Label other_case;
2780 vixl32::Label* const target = codegen_->GetFinalLabel(select, &select_end);
2781
2782 if (out.Equals(second)) {
2783 true_target = target;
2784 src = first;
2785 } else {
2786 false_target = target;
2787 src = second;
2788
2789 if (!out.Equals(first)) {
2790 if (output_overlaps_with_condition_inputs) {
2791 false_target = &other_case;
2792 } else {
2793 codegen_->MoveLocation(out, first, type);
2794 }
2795 }
2796 }
2797
2798 GenerateTestAndBranch(select, 2, true_target, false_target, /* far_target= */ false);
2799 codegen_->MoveLocation(out, src, type);
2800 if (output_overlaps_with_condition_inputs) {
2801 __ B(target);
2802 __ Bind(&other_case);
2803 codegen_->MoveLocation(out, first, type);
2804 }
2805
2806 if (select_end.IsReferenced()) {
2807 __ Bind(&select_end);
2808 }
2809 }
2810
VisitNativeDebugInfo(HNativeDebugInfo * info)2811 void LocationsBuilderARMVIXL::VisitNativeDebugInfo(HNativeDebugInfo* info) {
2812 new (GetGraph()->GetAllocator()) LocationSummary(info);
2813 }
2814
VisitNativeDebugInfo(HNativeDebugInfo *)2815 void InstructionCodeGeneratorARMVIXL::VisitNativeDebugInfo(HNativeDebugInfo*) {
2816 // MaybeRecordNativeDebugInfo is already called implicitly in CodeGenerator::Compile.
2817 }
2818
GenerateNop()2819 void CodeGeneratorARMVIXL::GenerateNop() {
2820 __ Nop();
2821 }
2822
2823 // `temp` is an extra temporary register that is used for some conditions;
2824 // callers may not specify it, in which case the method will use a scratch
2825 // register instead.
GenerateConditionWithZero(IfCondition condition,vixl32::Register out,vixl32::Register in,vixl32::Register temp)2826 void CodeGeneratorARMVIXL::GenerateConditionWithZero(IfCondition condition,
2827 vixl32::Register out,
2828 vixl32::Register in,
2829 vixl32::Register temp) {
2830 switch (condition) {
2831 case kCondEQ:
2832 // x <= 0 iff x == 0 when the comparison is unsigned.
2833 case kCondBE:
2834 if (!temp.IsValid() || (out.IsLow() && !out.Is(in))) {
2835 temp = out;
2836 }
2837
2838 // Avoid 32-bit instructions if possible; note that `in` and `temp` must be
2839 // different as well.
2840 if (in.IsLow() && temp.IsLow() && !in.Is(temp)) {
2841 // temp = - in; only 0 sets the carry flag.
2842 __ Rsbs(temp, in, 0);
2843
2844 if (out.Is(in)) {
2845 std::swap(in, temp);
2846 }
2847
2848 // out = - in + in + carry = carry
2849 __ Adc(out, temp, in);
2850 } else {
2851 // If `in` is 0, then it has 32 leading zeros, and less than that otherwise.
2852 __ Clz(out, in);
2853 // Any number less than 32 logically shifted right by 5 bits results in 0;
2854 // the same operation on 32 yields 1.
2855 __ Lsr(out, out, 5);
2856 }
2857
2858 break;
2859 case kCondNE:
2860 // x > 0 iff x != 0 when the comparison is unsigned.
2861 case kCondA: {
2862 UseScratchRegisterScope temps(GetVIXLAssembler());
2863
2864 if (out.Is(in)) {
2865 if (!temp.IsValid() || in.Is(temp)) {
2866 temp = temps.Acquire();
2867 }
2868 } else if (!temp.IsValid() || !temp.IsLow()) {
2869 temp = out;
2870 }
2871
2872 // temp = in - 1; only 0 does not set the carry flag.
2873 __ Subs(temp, in, 1);
2874 // out = in + ~temp + carry = in + (-(in - 1) - 1) + carry = in - in + 1 - 1 + carry = carry
2875 __ Sbc(out, in, temp);
2876 break;
2877 }
2878 case kCondGE:
2879 __ Mvn(out, in);
2880 in = out;
2881 FALLTHROUGH_INTENDED;
2882 case kCondLT:
2883 // We only care about the sign bit.
2884 __ Lsr(out, in, 31);
2885 break;
2886 case kCondAE:
2887 // Trivially true.
2888 __ Mov(out, 1);
2889 break;
2890 case kCondB:
2891 // Trivially false.
2892 __ Mov(out, 0);
2893 break;
2894 default:
2895 LOG(FATAL) << "Unexpected condition " << condition;
2896 UNREACHABLE();
2897 }
2898 }
2899
HandleCondition(HCondition * cond)2900 void LocationsBuilderARMVIXL::HandleCondition(HCondition* cond) {
2901 LocationSummary* locations =
2902 new (GetGraph()->GetAllocator()) LocationSummary(cond, LocationSummary::kNoCall);
2903 const DataType::Type type = cond->InputAt(0)->GetType();
2904 if (DataType::IsFloatingPointType(type)) {
2905 locations->SetInAt(0, Location::RequiresFpuRegister());
2906 locations->SetInAt(1, ArithmeticZeroOrFpuRegister(cond->InputAt(1)));
2907 } else {
2908 locations->SetInAt(0, Location::RequiresRegister());
2909 locations->SetInAt(1, Location::RegisterOrConstant(cond->InputAt(1)));
2910 }
2911 if (!cond->IsEmittedAtUseSite()) {
2912 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
2913 }
2914 }
2915
HandleCondition(HCondition * cond)2916 void InstructionCodeGeneratorARMVIXL::HandleCondition(HCondition* cond) {
2917 if (cond->IsEmittedAtUseSite()) {
2918 return;
2919 }
2920
2921 const DataType::Type type = cond->GetLeft()->GetType();
2922
2923 if (DataType::IsFloatingPointType(type)) {
2924 GenerateConditionGeneric(cond, codegen_);
2925 return;
2926 }
2927
2928 DCHECK(DataType::IsIntegralType(type) || type == DataType::Type::kReference) << type;
2929
2930 const IfCondition condition = cond->GetCondition();
2931
2932 // A condition with only one boolean input, or two boolean inputs without being equality or
2933 // inequality results from transformations done by the instruction simplifier, and is handled
2934 // as a regular condition with integral inputs.
2935 if (type == DataType::Type::kBool &&
2936 cond->GetRight()->GetType() == DataType::Type::kBool &&
2937 (condition == kCondEQ || condition == kCondNE)) {
2938 vixl32::Register left = InputRegisterAt(cond, 0);
2939 const vixl32::Register out = OutputRegister(cond);
2940 const Location right_loc = cond->GetLocations()->InAt(1);
2941
2942 // The constant case is handled by the instruction simplifier.
2943 DCHECK(!right_loc.IsConstant());
2944
2945 vixl32::Register right = RegisterFrom(right_loc);
2946
2947 // Avoid 32-bit instructions if possible.
2948 if (out.Is(right)) {
2949 std::swap(left, right);
2950 }
2951
2952 __ Eor(out, left, right);
2953
2954 if (condition == kCondEQ) {
2955 __ Eor(out, out, 1);
2956 }
2957
2958 return;
2959 }
2960
2961 GenerateConditionIntegralOrNonPrimitive(cond, codegen_);
2962 }
2963
VisitEqual(HEqual * comp)2964 void LocationsBuilderARMVIXL::VisitEqual(HEqual* comp) {
2965 HandleCondition(comp);
2966 }
2967
VisitEqual(HEqual * comp)2968 void InstructionCodeGeneratorARMVIXL::VisitEqual(HEqual* comp) {
2969 HandleCondition(comp);
2970 }
2971
VisitNotEqual(HNotEqual * comp)2972 void LocationsBuilderARMVIXL::VisitNotEqual(HNotEqual* comp) {
2973 HandleCondition(comp);
2974 }
2975
VisitNotEqual(HNotEqual * comp)2976 void InstructionCodeGeneratorARMVIXL::VisitNotEqual(HNotEqual* comp) {
2977 HandleCondition(comp);
2978 }
2979
VisitLessThan(HLessThan * comp)2980 void LocationsBuilderARMVIXL::VisitLessThan(HLessThan* comp) {
2981 HandleCondition(comp);
2982 }
2983
VisitLessThan(HLessThan * comp)2984 void InstructionCodeGeneratorARMVIXL::VisitLessThan(HLessThan* comp) {
2985 HandleCondition(comp);
2986 }
2987
VisitLessThanOrEqual(HLessThanOrEqual * comp)2988 void LocationsBuilderARMVIXL::VisitLessThanOrEqual(HLessThanOrEqual* comp) {
2989 HandleCondition(comp);
2990 }
2991
VisitLessThanOrEqual(HLessThanOrEqual * comp)2992 void InstructionCodeGeneratorARMVIXL::VisitLessThanOrEqual(HLessThanOrEqual* comp) {
2993 HandleCondition(comp);
2994 }
2995
VisitGreaterThan(HGreaterThan * comp)2996 void LocationsBuilderARMVIXL::VisitGreaterThan(HGreaterThan* comp) {
2997 HandleCondition(comp);
2998 }
2999
VisitGreaterThan(HGreaterThan * comp)3000 void InstructionCodeGeneratorARMVIXL::VisitGreaterThan(HGreaterThan* comp) {
3001 HandleCondition(comp);
3002 }
3003
VisitGreaterThanOrEqual(HGreaterThanOrEqual * comp)3004 void LocationsBuilderARMVIXL::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) {
3005 HandleCondition(comp);
3006 }
3007
VisitGreaterThanOrEqual(HGreaterThanOrEqual * comp)3008 void InstructionCodeGeneratorARMVIXL::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) {
3009 HandleCondition(comp);
3010 }
3011
VisitBelow(HBelow * comp)3012 void LocationsBuilderARMVIXL::VisitBelow(HBelow* comp) {
3013 HandleCondition(comp);
3014 }
3015
VisitBelow(HBelow * comp)3016 void InstructionCodeGeneratorARMVIXL::VisitBelow(HBelow* comp) {
3017 HandleCondition(comp);
3018 }
3019
VisitBelowOrEqual(HBelowOrEqual * comp)3020 void LocationsBuilderARMVIXL::VisitBelowOrEqual(HBelowOrEqual* comp) {
3021 HandleCondition(comp);
3022 }
3023
VisitBelowOrEqual(HBelowOrEqual * comp)3024 void InstructionCodeGeneratorARMVIXL::VisitBelowOrEqual(HBelowOrEqual* comp) {
3025 HandleCondition(comp);
3026 }
3027
VisitAbove(HAbove * comp)3028 void LocationsBuilderARMVIXL::VisitAbove(HAbove* comp) {
3029 HandleCondition(comp);
3030 }
3031
VisitAbove(HAbove * comp)3032 void InstructionCodeGeneratorARMVIXL::VisitAbove(HAbove* comp) {
3033 HandleCondition(comp);
3034 }
3035
VisitAboveOrEqual(HAboveOrEqual * comp)3036 void LocationsBuilderARMVIXL::VisitAboveOrEqual(HAboveOrEqual* comp) {
3037 HandleCondition(comp);
3038 }
3039
VisitAboveOrEqual(HAboveOrEqual * comp)3040 void InstructionCodeGeneratorARMVIXL::VisitAboveOrEqual(HAboveOrEqual* comp) {
3041 HandleCondition(comp);
3042 }
3043
VisitIntConstant(HIntConstant * constant)3044 void LocationsBuilderARMVIXL::VisitIntConstant(HIntConstant* constant) {
3045 LocationSummary* locations =
3046 new (GetGraph()->GetAllocator()) LocationSummary(constant, LocationSummary::kNoCall);
3047 locations->SetOut(Location::ConstantLocation(constant));
3048 }
3049
VisitIntConstant(HIntConstant * constant ATTRIBUTE_UNUSED)3050 void InstructionCodeGeneratorARMVIXL::VisitIntConstant(HIntConstant* constant ATTRIBUTE_UNUSED) {
3051 // Will be generated at use site.
3052 }
3053
VisitNullConstant(HNullConstant * constant)3054 void LocationsBuilderARMVIXL::VisitNullConstant(HNullConstant* constant) {
3055 LocationSummary* locations =
3056 new (GetGraph()->GetAllocator()) LocationSummary(constant, LocationSummary::kNoCall);
3057 locations->SetOut(Location::ConstantLocation(constant));
3058 }
3059
VisitNullConstant(HNullConstant * constant ATTRIBUTE_UNUSED)3060 void InstructionCodeGeneratorARMVIXL::VisitNullConstant(HNullConstant* constant ATTRIBUTE_UNUSED) {
3061 // Will be generated at use site.
3062 }
3063
VisitLongConstant(HLongConstant * constant)3064 void LocationsBuilderARMVIXL::VisitLongConstant(HLongConstant* constant) {
3065 LocationSummary* locations =
3066 new (GetGraph()->GetAllocator()) LocationSummary(constant, LocationSummary::kNoCall);
3067 locations->SetOut(Location::ConstantLocation(constant));
3068 }
3069
VisitLongConstant(HLongConstant * constant ATTRIBUTE_UNUSED)3070 void InstructionCodeGeneratorARMVIXL::VisitLongConstant(HLongConstant* constant ATTRIBUTE_UNUSED) {
3071 // Will be generated at use site.
3072 }
3073
VisitFloatConstant(HFloatConstant * constant)3074 void LocationsBuilderARMVIXL::VisitFloatConstant(HFloatConstant* constant) {
3075 LocationSummary* locations =
3076 new (GetGraph()->GetAllocator()) LocationSummary(constant, LocationSummary::kNoCall);
3077 locations->SetOut(Location::ConstantLocation(constant));
3078 }
3079
VisitFloatConstant(HFloatConstant * constant ATTRIBUTE_UNUSED)3080 void InstructionCodeGeneratorARMVIXL::VisitFloatConstant(
3081 HFloatConstant* constant ATTRIBUTE_UNUSED) {
3082 // Will be generated at use site.
3083 }
3084
VisitDoubleConstant(HDoubleConstant * constant)3085 void LocationsBuilderARMVIXL::VisitDoubleConstant(HDoubleConstant* constant) {
3086 LocationSummary* locations =
3087 new (GetGraph()->GetAllocator()) LocationSummary(constant, LocationSummary::kNoCall);
3088 locations->SetOut(Location::ConstantLocation(constant));
3089 }
3090
VisitDoubleConstant(HDoubleConstant * constant ATTRIBUTE_UNUSED)3091 void InstructionCodeGeneratorARMVIXL::VisitDoubleConstant(
3092 HDoubleConstant* constant ATTRIBUTE_UNUSED) {
3093 // Will be generated at use site.
3094 }
3095
VisitConstructorFence(HConstructorFence * constructor_fence)3096 void LocationsBuilderARMVIXL::VisitConstructorFence(HConstructorFence* constructor_fence) {
3097 constructor_fence->SetLocations(nullptr);
3098 }
3099
VisitConstructorFence(HConstructorFence * constructor_fence ATTRIBUTE_UNUSED)3100 void InstructionCodeGeneratorARMVIXL::VisitConstructorFence(
3101 HConstructorFence* constructor_fence ATTRIBUTE_UNUSED) {
3102 codegen_->GenerateMemoryBarrier(MemBarrierKind::kStoreStore);
3103 }
3104
VisitMemoryBarrier(HMemoryBarrier * memory_barrier)3105 void LocationsBuilderARMVIXL::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) {
3106 memory_barrier->SetLocations(nullptr);
3107 }
3108
VisitMemoryBarrier(HMemoryBarrier * memory_barrier)3109 void InstructionCodeGeneratorARMVIXL::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) {
3110 codegen_->GenerateMemoryBarrier(memory_barrier->GetBarrierKind());
3111 }
3112
VisitReturnVoid(HReturnVoid * ret)3113 void LocationsBuilderARMVIXL::VisitReturnVoid(HReturnVoid* ret) {
3114 ret->SetLocations(nullptr);
3115 }
3116
VisitReturnVoid(HReturnVoid * ret ATTRIBUTE_UNUSED)3117 void InstructionCodeGeneratorARMVIXL::VisitReturnVoid(HReturnVoid* ret ATTRIBUTE_UNUSED) {
3118 codegen_->GenerateFrameExit();
3119 }
3120
VisitReturn(HReturn * ret)3121 void LocationsBuilderARMVIXL::VisitReturn(HReturn* ret) {
3122 LocationSummary* locations =
3123 new (GetGraph()->GetAllocator()) LocationSummary(ret, LocationSummary::kNoCall);
3124 locations->SetInAt(0, parameter_visitor_.GetReturnLocation(ret->InputAt(0)->GetType()));
3125 }
3126
VisitReturn(HReturn * ret ATTRIBUTE_UNUSED)3127 void InstructionCodeGeneratorARMVIXL::VisitReturn(HReturn* ret ATTRIBUTE_UNUSED) {
3128 codegen_->GenerateFrameExit();
3129 }
3130
VisitInvokeUnresolved(HInvokeUnresolved * invoke)3131 void LocationsBuilderARMVIXL::VisitInvokeUnresolved(HInvokeUnresolved* invoke) {
3132 // The trampoline uses the same calling convention as dex calling conventions,
3133 // except instead of loading arg0/r0 with the target Method*, arg0/r0 will contain
3134 // the method_idx.
3135 HandleInvoke(invoke);
3136 }
3137
VisitInvokeUnresolved(HInvokeUnresolved * invoke)3138 void InstructionCodeGeneratorARMVIXL::VisitInvokeUnresolved(HInvokeUnresolved* invoke) {
3139 codegen_->GenerateInvokeUnresolvedRuntimeCall(invoke);
3140 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 3);
3141 }
3142
VisitInvokeStaticOrDirect(HInvokeStaticOrDirect * invoke)3143 void LocationsBuilderARMVIXL::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
3144 // Explicit clinit checks triggered by static invokes must have been pruned by
3145 // art::PrepareForRegisterAllocation.
3146 DCHECK(!invoke->IsStaticWithExplicitClinitCheck());
3147
3148 IntrinsicLocationsBuilderARMVIXL intrinsic(codegen_);
3149 if (intrinsic.TryDispatch(invoke)) {
3150 return;
3151 }
3152
3153 HandleInvoke(invoke);
3154 }
3155
TryGenerateIntrinsicCode(HInvoke * invoke,CodeGeneratorARMVIXL * codegen)3156 static bool TryGenerateIntrinsicCode(HInvoke* invoke, CodeGeneratorARMVIXL* codegen) {
3157 if (invoke->GetLocations()->Intrinsified()) {
3158 IntrinsicCodeGeneratorARMVIXL intrinsic(codegen);
3159 intrinsic.Dispatch(invoke);
3160 return true;
3161 }
3162 return false;
3163 }
3164
VisitInvokeStaticOrDirect(HInvokeStaticOrDirect * invoke)3165 void InstructionCodeGeneratorARMVIXL::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
3166 // Explicit clinit checks triggered by static invokes must have been pruned by
3167 // art::PrepareForRegisterAllocation.
3168 DCHECK(!invoke->IsStaticWithExplicitClinitCheck());
3169
3170 if (TryGenerateIntrinsicCode(invoke, codegen_)) {
3171 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 4);
3172 return;
3173 }
3174
3175 LocationSummary* locations = invoke->GetLocations();
3176 codegen_->GenerateStaticOrDirectCall(
3177 invoke, locations->HasTemps() ? locations->GetTemp(0) : Location::NoLocation());
3178
3179 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 5);
3180 }
3181
HandleInvoke(HInvoke * invoke)3182 void LocationsBuilderARMVIXL::HandleInvoke(HInvoke* invoke) {
3183 InvokeDexCallingConventionVisitorARMVIXL calling_convention_visitor;
3184 CodeGenerator::CreateCommonInvokeLocationSummary(invoke, &calling_convention_visitor);
3185 }
3186
VisitInvokeVirtual(HInvokeVirtual * invoke)3187 void LocationsBuilderARMVIXL::VisitInvokeVirtual(HInvokeVirtual* invoke) {
3188 IntrinsicLocationsBuilderARMVIXL intrinsic(codegen_);
3189 if (intrinsic.TryDispatch(invoke)) {
3190 return;
3191 }
3192
3193 HandleInvoke(invoke);
3194 }
3195
VisitInvokeVirtual(HInvokeVirtual * invoke)3196 void InstructionCodeGeneratorARMVIXL::VisitInvokeVirtual(HInvokeVirtual* invoke) {
3197 if (TryGenerateIntrinsicCode(invoke, codegen_)) {
3198 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 6);
3199 return;
3200 }
3201
3202 codegen_->GenerateVirtualCall(invoke, invoke->GetLocations()->GetTemp(0));
3203 DCHECK(!codegen_->IsLeafMethod());
3204
3205 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 7);
3206 }
3207
VisitInvokeInterface(HInvokeInterface * invoke)3208 void LocationsBuilderARMVIXL::VisitInvokeInterface(HInvokeInterface* invoke) {
3209 HandleInvoke(invoke);
3210 // Add the hidden argument.
3211 invoke->GetLocations()->AddTemp(LocationFrom(r12));
3212 }
3213
VisitInvokeInterface(HInvokeInterface * invoke)3214 void InstructionCodeGeneratorARMVIXL::VisitInvokeInterface(HInvokeInterface* invoke) {
3215 // TODO: b/18116999, our IMTs can miss an IncompatibleClassChangeError.
3216 LocationSummary* locations = invoke->GetLocations();
3217 vixl32::Register temp = RegisterFrom(locations->GetTemp(0));
3218 vixl32::Register hidden_reg = RegisterFrom(locations->GetTemp(1));
3219 Location receiver = locations->InAt(0);
3220 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
3221
3222 DCHECK(!receiver.IsStackSlot());
3223
3224 // Ensure the pc position is recorded immediately after the `ldr` instruction.
3225 {
3226 ExactAssemblyScope aas(GetVIXLAssembler(),
3227 vixl32::kMaxInstructionSizeInBytes,
3228 CodeBufferCheckScope::kMaximumSize);
3229 // /* HeapReference<Class> */ temp = receiver->klass_
3230 __ ldr(temp, MemOperand(RegisterFrom(receiver), class_offset));
3231 codegen_->MaybeRecordImplicitNullCheck(invoke);
3232 }
3233 // Instead of simply (possibly) unpoisoning `temp` here, we should
3234 // emit a read barrier for the previous class reference load.
3235 // However this is not required in practice, as this is an
3236 // intermediate/temporary reference and because the current
3237 // concurrent copying collector keeps the from-space memory
3238 // intact/accessible until the end of the marking phase (the
3239 // concurrent copying collector may not in the future).
3240 GetAssembler()->MaybeUnpoisonHeapReference(temp);
3241 GetAssembler()->LoadFromOffset(kLoadWord,
3242 temp,
3243 temp,
3244 mirror::Class::ImtPtrOffset(kArmPointerSize).Uint32Value());
3245 uint32_t method_offset = static_cast<uint32_t>(ImTable::OffsetOfElement(
3246 invoke->GetImtIndex(), kArmPointerSize));
3247 // temp = temp->GetImtEntryAt(method_offset);
3248 GetAssembler()->LoadFromOffset(kLoadWord, temp, temp, method_offset);
3249 uint32_t entry_point =
3250 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmPointerSize).Int32Value();
3251 // LR = temp->GetEntryPoint();
3252 GetAssembler()->LoadFromOffset(kLoadWord, lr, temp, entry_point);
3253
3254 // Set the hidden (in r12) argument. It is done here, right before a BLX to prevent other
3255 // instruction from clobbering it as they might use r12 as a scratch register.
3256 DCHECK(hidden_reg.Is(r12));
3257
3258 {
3259 // The VIXL macro assembler may clobber any of the scratch registers that are available to it,
3260 // so it checks if the application is using them (by passing them to the macro assembler
3261 // methods). The following application of UseScratchRegisterScope corrects VIXL's notion of
3262 // what is available, and is the opposite of the standard usage: Instead of requesting a
3263 // temporary location, it imposes an external constraint (i.e. a specific register is reserved
3264 // for the hidden argument). Note that this works even if VIXL needs a scratch register itself
3265 // (to materialize the constant), since the destination register becomes available for such use
3266 // internally for the duration of the macro instruction.
3267 UseScratchRegisterScope temps(GetVIXLAssembler());
3268 temps.Exclude(hidden_reg);
3269 __ Mov(hidden_reg, invoke->GetDexMethodIndex());
3270 }
3271 {
3272 // Ensure the pc position is recorded immediately after the `blx` instruction.
3273 // blx in T32 has only 16bit encoding that's why a stricter check for the scope is used.
3274 ExactAssemblyScope aas(GetVIXLAssembler(),
3275 vixl32::k16BitT32InstructionSizeInBytes,
3276 CodeBufferCheckScope::kExactSize);
3277 // LR();
3278 __ blx(lr);
3279 codegen_->RecordPcInfo(invoke, invoke->GetDexPc());
3280 DCHECK(!codegen_->IsLeafMethod());
3281 }
3282
3283 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 8);
3284 }
3285
VisitInvokePolymorphic(HInvokePolymorphic * invoke)3286 void LocationsBuilderARMVIXL::VisitInvokePolymorphic(HInvokePolymorphic* invoke) {
3287 HandleInvoke(invoke);
3288 }
3289
VisitInvokePolymorphic(HInvokePolymorphic * invoke)3290 void InstructionCodeGeneratorARMVIXL::VisitInvokePolymorphic(HInvokePolymorphic* invoke) {
3291 codegen_->GenerateInvokePolymorphicCall(invoke);
3292 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 9);
3293 }
3294
VisitInvokeCustom(HInvokeCustom * invoke)3295 void LocationsBuilderARMVIXL::VisitInvokeCustom(HInvokeCustom* invoke) {
3296 HandleInvoke(invoke);
3297 }
3298
VisitInvokeCustom(HInvokeCustom * invoke)3299 void InstructionCodeGeneratorARMVIXL::VisitInvokeCustom(HInvokeCustom* invoke) {
3300 codegen_->GenerateInvokeCustomCall(invoke);
3301 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 10);
3302 }
3303
VisitNeg(HNeg * neg)3304 void LocationsBuilderARMVIXL::VisitNeg(HNeg* neg) {
3305 LocationSummary* locations =
3306 new (GetGraph()->GetAllocator()) LocationSummary(neg, LocationSummary::kNoCall);
3307 switch (neg->GetResultType()) {
3308 case DataType::Type::kInt32: {
3309 locations->SetInAt(0, Location::RequiresRegister());
3310 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
3311 break;
3312 }
3313 case DataType::Type::kInt64: {
3314 locations->SetInAt(0, Location::RequiresRegister());
3315 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
3316 break;
3317 }
3318
3319 case DataType::Type::kFloat32:
3320 case DataType::Type::kFloat64:
3321 locations->SetInAt(0, Location::RequiresFpuRegister());
3322 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
3323 break;
3324
3325 default:
3326 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType();
3327 }
3328 }
3329
VisitNeg(HNeg * neg)3330 void InstructionCodeGeneratorARMVIXL::VisitNeg(HNeg* neg) {
3331 LocationSummary* locations = neg->GetLocations();
3332 Location out = locations->Out();
3333 Location in = locations->InAt(0);
3334 switch (neg->GetResultType()) {
3335 case DataType::Type::kInt32:
3336 __ Rsb(OutputRegister(neg), InputRegisterAt(neg, 0), 0);
3337 break;
3338
3339 case DataType::Type::kInt64:
3340 // out.lo = 0 - in.lo (and update the carry/borrow (C) flag)
3341 __ Rsbs(LowRegisterFrom(out), LowRegisterFrom(in), 0);
3342 // We cannot emit an RSC (Reverse Subtract with Carry)
3343 // instruction here, as it does not exist in the Thumb-2
3344 // instruction set. We use the following approach
3345 // using SBC and SUB instead.
3346 //
3347 // out.hi = -C
3348 __ Sbc(HighRegisterFrom(out), HighRegisterFrom(out), HighRegisterFrom(out));
3349 // out.hi = out.hi - in.hi
3350 __ Sub(HighRegisterFrom(out), HighRegisterFrom(out), HighRegisterFrom(in));
3351 break;
3352
3353 case DataType::Type::kFloat32:
3354 case DataType::Type::kFloat64:
3355 __ Vneg(OutputVRegister(neg), InputVRegister(neg));
3356 break;
3357
3358 default:
3359 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType();
3360 }
3361 }
3362
VisitTypeConversion(HTypeConversion * conversion)3363 void LocationsBuilderARMVIXL::VisitTypeConversion(HTypeConversion* conversion) {
3364 DataType::Type result_type = conversion->GetResultType();
3365 DataType::Type input_type = conversion->GetInputType();
3366 DCHECK(!DataType::IsTypeConversionImplicit(input_type, result_type))
3367 << input_type << " -> " << result_type;
3368
3369 // The float-to-long, double-to-long and long-to-float type conversions
3370 // rely on a call to the runtime.
3371 LocationSummary::CallKind call_kind =
3372 (((input_type == DataType::Type::kFloat32 || input_type == DataType::Type::kFloat64)
3373 && result_type == DataType::Type::kInt64)
3374 || (input_type == DataType::Type::kInt64 && result_type == DataType::Type::kFloat32))
3375 ? LocationSummary::kCallOnMainOnly
3376 : LocationSummary::kNoCall;
3377 LocationSummary* locations =
3378 new (GetGraph()->GetAllocator()) LocationSummary(conversion, call_kind);
3379
3380 switch (result_type) {
3381 case DataType::Type::kUint8:
3382 case DataType::Type::kInt8:
3383 case DataType::Type::kUint16:
3384 case DataType::Type::kInt16:
3385 DCHECK(DataType::IsIntegralType(input_type)) << input_type;
3386 locations->SetInAt(0, Location::RequiresRegister());
3387 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
3388 break;
3389
3390 case DataType::Type::kInt32:
3391 switch (input_type) {
3392 case DataType::Type::kInt64:
3393 locations->SetInAt(0, Location::Any());
3394 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
3395 break;
3396
3397 case DataType::Type::kFloat32:
3398 locations->SetInAt(0, Location::RequiresFpuRegister());
3399 locations->SetOut(Location::RequiresRegister());
3400 locations->AddTemp(Location::RequiresFpuRegister());
3401 break;
3402
3403 case DataType::Type::kFloat64:
3404 locations->SetInAt(0, Location::RequiresFpuRegister());
3405 locations->SetOut(Location::RequiresRegister());
3406 locations->AddTemp(Location::RequiresFpuRegister());
3407 break;
3408
3409 default:
3410 LOG(FATAL) << "Unexpected type conversion from " << input_type
3411 << " to " << result_type;
3412 }
3413 break;
3414
3415 case DataType::Type::kInt64:
3416 switch (input_type) {
3417 case DataType::Type::kBool:
3418 case DataType::Type::kUint8:
3419 case DataType::Type::kInt8:
3420 case DataType::Type::kUint16:
3421 case DataType::Type::kInt16:
3422 case DataType::Type::kInt32:
3423 locations->SetInAt(0, Location::RequiresRegister());
3424 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
3425 break;
3426
3427 case DataType::Type::kFloat32: {
3428 InvokeRuntimeCallingConventionARMVIXL calling_convention;
3429 locations->SetInAt(0, LocationFrom(calling_convention.GetFpuRegisterAt(0)));
3430 locations->SetOut(LocationFrom(r0, r1));
3431 break;
3432 }
3433
3434 case DataType::Type::kFloat64: {
3435 InvokeRuntimeCallingConventionARMVIXL calling_convention;
3436 locations->SetInAt(0, LocationFrom(calling_convention.GetFpuRegisterAt(0),
3437 calling_convention.GetFpuRegisterAt(1)));
3438 locations->SetOut(LocationFrom(r0, r1));
3439 break;
3440 }
3441
3442 default:
3443 LOG(FATAL) << "Unexpected type conversion from " << input_type
3444 << " to " << result_type;
3445 }
3446 break;
3447
3448 case DataType::Type::kFloat32:
3449 switch (input_type) {
3450 case DataType::Type::kBool:
3451 case DataType::Type::kUint8:
3452 case DataType::Type::kInt8:
3453 case DataType::Type::kUint16:
3454 case DataType::Type::kInt16:
3455 case DataType::Type::kInt32:
3456 locations->SetInAt(0, Location::RequiresRegister());
3457 locations->SetOut(Location::RequiresFpuRegister());
3458 break;
3459
3460 case DataType::Type::kInt64: {
3461 InvokeRuntimeCallingConventionARMVIXL calling_convention;
3462 locations->SetInAt(0, LocationFrom(calling_convention.GetRegisterAt(0),
3463 calling_convention.GetRegisterAt(1)));
3464 locations->SetOut(LocationFrom(calling_convention.GetFpuRegisterAt(0)));
3465 break;
3466 }
3467
3468 case DataType::Type::kFloat64:
3469 locations->SetInAt(0, Location::RequiresFpuRegister());
3470 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
3471 break;
3472
3473 default:
3474 LOG(FATAL) << "Unexpected type conversion from " << input_type
3475 << " to " << result_type;
3476 }
3477 break;
3478
3479 case DataType::Type::kFloat64:
3480 switch (input_type) {
3481 case DataType::Type::kBool:
3482 case DataType::Type::kUint8:
3483 case DataType::Type::kInt8:
3484 case DataType::Type::kUint16:
3485 case DataType::Type::kInt16:
3486 case DataType::Type::kInt32:
3487 locations->SetInAt(0, Location::RequiresRegister());
3488 locations->SetOut(Location::RequiresFpuRegister());
3489 break;
3490
3491 case DataType::Type::kInt64:
3492 locations->SetInAt(0, Location::RequiresRegister());
3493 locations->SetOut(Location::RequiresFpuRegister());
3494 locations->AddTemp(Location::RequiresFpuRegister());
3495 locations->AddTemp(Location::RequiresFpuRegister());
3496 break;
3497
3498 case DataType::Type::kFloat32:
3499 locations->SetInAt(0, Location::RequiresFpuRegister());
3500 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
3501 break;
3502
3503 default:
3504 LOG(FATAL) << "Unexpected type conversion from " << input_type
3505 << " to " << result_type;
3506 }
3507 break;
3508
3509 default:
3510 LOG(FATAL) << "Unexpected type conversion from " << input_type
3511 << " to " << result_type;
3512 }
3513 }
3514
VisitTypeConversion(HTypeConversion * conversion)3515 void InstructionCodeGeneratorARMVIXL::VisitTypeConversion(HTypeConversion* conversion) {
3516 LocationSummary* locations = conversion->GetLocations();
3517 Location out = locations->Out();
3518 Location in = locations->InAt(0);
3519 DataType::Type result_type = conversion->GetResultType();
3520 DataType::Type input_type = conversion->GetInputType();
3521 DCHECK(!DataType::IsTypeConversionImplicit(input_type, result_type))
3522 << input_type << " -> " << result_type;
3523 switch (result_type) {
3524 case DataType::Type::kUint8:
3525 switch (input_type) {
3526 case DataType::Type::kInt8:
3527 case DataType::Type::kUint16:
3528 case DataType::Type::kInt16:
3529 case DataType::Type::kInt32:
3530 __ Ubfx(OutputRegister(conversion), InputRegisterAt(conversion, 0), 0, 8);
3531 break;
3532 case DataType::Type::kInt64:
3533 __ Ubfx(OutputRegister(conversion), LowRegisterFrom(in), 0, 8);
3534 break;
3535
3536 default:
3537 LOG(FATAL) << "Unexpected type conversion from " << input_type
3538 << " to " << result_type;
3539 }
3540 break;
3541
3542 case DataType::Type::kInt8:
3543 switch (input_type) {
3544 case DataType::Type::kUint8:
3545 case DataType::Type::kUint16:
3546 case DataType::Type::kInt16:
3547 case DataType::Type::kInt32:
3548 __ Sbfx(OutputRegister(conversion), InputRegisterAt(conversion, 0), 0, 8);
3549 break;
3550 case DataType::Type::kInt64:
3551 __ Sbfx(OutputRegister(conversion), LowRegisterFrom(in), 0, 8);
3552 break;
3553
3554 default:
3555 LOG(FATAL) << "Unexpected type conversion from " << input_type
3556 << " to " << result_type;
3557 }
3558 break;
3559
3560 case DataType::Type::kUint16:
3561 switch (input_type) {
3562 case DataType::Type::kInt8:
3563 case DataType::Type::kInt16:
3564 case DataType::Type::kInt32:
3565 __ Ubfx(OutputRegister(conversion), InputRegisterAt(conversion, 0), 0, 16);
3566 break;
3567 case DataType::Type::kInt64:
3568 __ Ubfx(OutputRegister(conversion), LowRegisterFrom(in), 0, 16);
3569 break;
3570
3571 default:
3572 LOG(FATAL) << "Unexpected type conversion from " << input_type
3573 << " to " << result_type;
3574 }
3575 break;
3576
3577 case DataType::Type::kInt16:
3578 switch (input_type) {
3579 case DataType::Type::kUint16:
3580 case DataType::Type::kInt32:
3581 __ Sbfx(OutputRegister(conversion), InputRegisterAt(conversion, 0), 0, 16);
3582 break;
3583 case DataType::Type::kInt64:
3584 __ Sbfx(OutputRegister(conversion), LowRegisterFrom(in), 0, 16);
3585 break;
3586
3587 default:
3588 LOG(FATAL) << "Unexpected type conversion from " << input_type
3589 << " to " << result_type;
3590 }
3591 break;
3592
3593 case DataType::Type::kInt32:
3594 switch (input_type) {
3595 case DataType::Type::kInt64:
3596 DCHECK(out.IsRegister());
3597 if (in.IsRegisterPair()) {
3598 __ Mov(OutputRegister(conversion), LowRegisterFrom(in));
3599 } else if (in.IsDoubleStackSlot()) {
3600 GetAssembler()->LoadFromOffset(kLoadWord,
3601 OutputRegister(conversion),
3602 sp,
3603 in.GetStackIndex());
3604 } else {
3605 DCHECK(in.IsConstant());
3606 DCHECK(in.GetConstant()->IsLongConstant());
3607 int64_t value = in.GetConstant()->AsLongConstant()->GetValue();
3608 __ Mov(OutputRegister(conversion), static_cast<int32_t>(value));
3609 }
3610 break;
3611
3612 case DataType::Type::kFloat32: {
3613 vixl32::SRegister temp = LowSRegisterFrom(locations->GetTemp(0));
3614 __ Vcvt(S32, F32, temp, InputSRegisterAt(conversion, 0));
3615 __ Vmov(OutputRegister(conversion), temp);
3616 break;
3617 }
3618
3619 case DataType::Type::kFloat64: {
3620 vixl32::SRegister temp_s = LowSRegisterFrom(locations->GetTemp(0));
3621 __ Vcvt(S32, F64, temp_s, DRegisterFrom(in));
3622 __ Vmov(OutputRegister(conversion), temp_s);
3623 break;
3624 }
3625
3626 default:
3627 LOG(FATAL) << "Unexpected type conversion from " << input_type
3628 << " to " << result_type;
3629 }
3630 break;
3631
3632 case DataType::Type::kInt64:
3633 switch (input_type) {
3634 case DataType::Type::kBool:
3635 case DataType::Type::kUint8:
3636 case DataType::Type::kInt8:
3637 case DataType::Type::kUint16:
3638 case DataType::Type::kInt16:
3639 case DataType::Type::kInt32:
3640 DCHECK(out.IsRegisterPair());
3641 DCHECK(in.IsRegister());
3642 __ Mov(LowRegisterFrom(out), InputRegisterAt(conversion, 0));
3643 // Sign extension.
3644 __ Asr(HighRegisterFrom(out), LowRegisterFrom(out), 31);
3645 break;
3646
3647 case DataType::Type::kFloat32:
3648 codegen_->InvokeRuntime(kQuickF2l, conversion, conversion->GetDexPc());
3649 CheckEntrypointTypes<kQuickF2l, int64_t, float>();
3650 break;
3651
3652 case DataType::Type::kFloat64:
3653 codegen_->InvokeRuntime(kQuickD2l, conversion, conversion->GetDexPc());
3654 CheckEntrypointTypes<kQuickD2l, int64_t, double>();
3655 break;
3656
3657 default:
3658 LOG(FATAL) << "Unexpected type conversion from " << input_type
3659 << " to " << result_type;
3660 }
3661 break;
3662
3663 case DataType::Type::kFloat32:
3664 switch (input_type) {
3665 case DataType::Type::kBool:
3666 case DataType::Type::kUint8:
3667 case DataType::Type::kInt8:
3668 case DataType::Type::kUint16:
3669 case DataType::Type::kInt16:
3670 case DataType::Type::kInt32:
3671 __ Vmov(OutputSRegister(conversion), InputRegisterAt(conversion, 0));
3672 __ Vcvt(F32, S32, OutputSRegister(conversion), OutputSRegister(conversion));
3673 break;
3674
3675 case DataType::Type::kInt64:
3676 codegen_->InvokeRuntime(kQuickL2f, conversion, conversion->GetDexPc());
3677 CheckEntrypointTypes<kQuickL2f, float, int64_t>();
3678 break;
3679
3680 case DataType::Type::kFloat64:
3681 __ Vcvt(F32, F64, OutputSRegister(conversion), DRegisterFrom(in));
3682 break;
3683
3684 default:
3685 LOG(FATAL) << "Unexpected type conversion from " << input_type
3686 << " to " << result_type;
3687 }
3688 break;
3689
3690 case DataType::Type::kFloat64:
3691 switch (input_type) {
3692 case DataType::Type::kBool:
3693 case DataType::Type::kUint8:
3694 case DataType::Type::kInt8:
3695 case DataType::Type::kUint16:
3696 case DataType::Type::kInt16:
3697 case DataType::Type::kInt32:
3698 __ Vmov(LowSRegisterFrom(out), InputRegisterAt(conversion, 0));
3699 __ Vcvt(F64, S32, DRegisterFrom(out), LowSRegisterFrom(out));
3700 break;
3701
3702 case DataType::Type::kInt64: {
3703 vixl32::Register low = LowRegisterFrom(in);
3704 vixl32::Register high = HighRegisterFrom(in);
3705 vixl32::SRegister out_s = LowSRegisterFrom(out);
3706 vixl32::DRegister out_d = DRegisterFrom(out);
3707 vixl32::SRegister temp_s = LowSRegisterFrom(locations->GetTemp(0));
3708 vixl32::DRegister temp_d = DRegisterFrom(locations->GetTemp(0));
3709 vixl32::DRegister constant_d = DRegisterFrom(locations->GetTemp(1));
3710
3711 // temp_d = int-to-double(high)
3712 __ Vmov(temp_s, high);
3713 __ Vcvt(F64, S32, temp_d, temp_s);
3714 // constant_d = k2Pow32EncodingForDouble
3715 __ Vmov(constant_d, bit_cast<double, int64_t>(k2Pow32EncodingForDouble));
3716 // out_d = unsigned-to-double(low)
3717 __ Vmov(out_s, low);
3718 __ Vcvt(F64, U32, out_d, out_s);
3719 // out_d += temp_d * constant_d
3720 __ Vmla(F64, out_d, temp_d, constant_d);
3721 break;
3722 }
3723
3724 case DataType::Type::kFloat32:
3725 __ Vcvt(F64, F32, DRegisterFrom(out), InputSRegisterAt(conversion, 0));
3726 break;
3727
3728 default:
3729 LOG(FATAL) << "Unexpected type conversion from " << input_type
3730 << " to " << result_type;
3731 }
3732 break;
3733
3734 default:
3735 LOG(FATAL) << "Unexpected type conversion from " << input_type
3736 << " to " << result_type;
3737 }
3738 }
3739
VisitAdd(HAdd * add)3740 void LocationsBuilderARMVIXL::VisitAdd(HAdd* add) {
3741 LocationSummary* locations =
3742 new (GetGraph()->GetAllocator()) LocationSummary(add, LocationSummary::kNoCall);
3743 switch (add->GetResultType()) {
3744 case DataType::Type::kInt32: {
3745 locations->SetInAt(0, Location::RequiresRegister());
3746 locations->SetInAt(1, Location::RegisterOrConstant(add->InputAt(1)));
3747 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
3748 break;
3749 }
3750
3751 case DataType::Type::kInt64: {
3752 locations->SetInAt(0, Location::RequiresRegister());
3753 locations->SetInAt(1, ArmEncodableConstantOrRegister(add->InputAt(1), ADD));
3754 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
3755 break;
3756 }
3757
3758 case DataType::Type::kFloat32:
3759 case DataType::Type::kFloat64: {
3760 locations->SetInAt(0, Location::RequiresFpuRegister());
3761 locations->SetInAt(1, Location::RequiresFpuRegister());
3762 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
3763 break;
3764 }
3765
3766 default:
3767 LOG(FATAL) << "Unexpected add type " << add->GetResultType();
3768 }
3769 }
3770
VisitAdd(HAdd * add)3771 void InstructionCodeGeneratorARMVIXL::VisitAdd(HAdd* add) {
3772 LocationSummary* locations = add->GetLocations();
3773 Location out = locations->Out();
3774 Location first = locations->InAt(0);
3775 Location second = locations->InAt(1);
3776
3777 switch (add->GetResultType()) {
3778 case DataType::Type::kInt32: {
3779 __ Add(OutputRegister(add), InputRegisterAt(add, 0), InputOperandAt(add, 1));
3780 }
3781 break;
3782
3783 case DataType::Type::kInt64: {
3784 if (second.IsConstant()) {
3785 uint64_t value = static_cast<uint64_t>(Int64FromConstant(second.GetConstant()));
3786 GenerateAddLongConst(out, first, value);
3787 } else {
3788 DCHECK(second.IsRegisterPair());
3789 __ Adds(LowRegisterFrom(out), LowRegisterFrom(first), LowRegisterFrom(second));
3790 __ Adc(HighRegisterFrom(out), HighRegisterFrom(first), HighRegisterFrom(second));
3791 }
3792 break;
3793 }
3794
3795 case DataType::Type::kFloat32:
3796 case DataType::Type::kFloat64:
3797 __ Vadd(OutputVRegister(add), InputVRegisterAt(add, 0), InputVRegisterAt(add, 1));
3798 break;
3799
3800 default:
3801 LOG(FATAL) << "Unexpected add type " << add->GetResultType();
3802 }
3803 }
3804
VisitSub(HSub * sub)3805 void LocationsBuilderARMVIXL::VisitSub(HSub* sub) {
3806 LocationSummary* locations =
3807 new (GetGraph()->GetAllocator()) LocationSummary(sub, LocationSummary::kNoCall);
3808 switch (sub->GetResultType()) {
3809 case DataType::Type::kInt32: {
3810 locations->SetInAt(0, Location::RequiresRegister());
3811 locations->SetInAt(1, Location::RegisterOrConstant(sub->InputAt(1)));
3812 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
3813 break;
3814 }
3815
3816 case DataType::Type::kInt64: {
3817 locations->SetInAt(0, Location::RequiresRegister());
3818 locations->SetInAt(1, ArmEncodableConstantOrRegister(sub->InputAt(1), SUB));
3819 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
3820 break;
3821 }
3822 case DataType::Type::kFloat32:
3823 case DataType::Type::kFloat64: {
3824 locations->SetInAt(0, Location::RequiresFpuRegister());
3825 locations->SetInAt(1, Location::RequiresFpuRegister());
3826 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
3827 break;
3828 }
3829 default:
3830 LOG(FATAL) << "Unexpected sub type " << sub->GetResultType();
3831 }
3832 }
3833
VisitSub(HSub * sub)3834 void InstructionCodeGeneratorARMVIXL::VisitSub(HSub* sub) {
3835 LocationSummary* locations = sub->GetLocations();
3836 Location out = locations->Out();
3837 Location first = locations->InAt(0);
3838 Location second = locations->InAt(1);
3839 switch (sub->GetResultType()) {
3840 case DataType::Type::kInt32: {
3841 __ Sub(OutputRegister(sub), InputRegisterAt(sub, 0), InputOperandAt(sub, 1));
3842 break;
3843 }
3844
3845 case DataType::Type::kInt64: {
3846 if (second.IsConstant()) {
3847 uint64_t value = static_cast<uint64_t>(Int64FromConstant(second.GetConstant()));
3848 GenerateAddLongConst(out, first, -value);
3849 } else {
3850 DCHECK(second.IsRegisterPair());
3851 __ Subs(LowRegisterFrom(out), LowRegisterFrom(first), LowRegisterFrom(second));
3852 __ Sbc(HighRegisterFrom(out), HighRegisterFrom(first), HighRegisterFrom(second));
3853 }
3854 break;
3855 }
3856
3857 case DataType::Type::kFloat32:
3858 case DataType::Type::kFloat64:
3859 __ Vsub(OutputVRegister(sub), InputVRegisterAt(sub, 0), InputVRegisterAt(sub, 1));
3860 break;
3861
3862 default:
3863 LOG(FATAL) << "Unexpected sub type " << sub->GetResultType();
3864 }
3865 }
3866
VisitMul(HMul * mul)3867 void LocationsBuilderARMVIXL::VisitMul(HMul* mul) {
3868 LocationSummary* locations =
3869 new (GetGraph()->GetAllocator()) LocationSummary(mul, LocationSummary::kNoCall);
3870 switch (mul->GetResultType()) {
3871 case DataType::Type::kInt32:
3872 case DataType::Type::kInt64: {
3873 locations->SetInAt(0, Location::RequiresRegister());
3874 locations->SetInAt(1, Location::RequiresRegister());
3875 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
3876 break;
3877 }
3878
3879 case DataType::Type::kFloat32:
3880 case DataType::Type::kFloat64: {
3881 locations->SetInAt(0, Location::RequiresFpuRegister());
3882 locations->SetInAt(1, Location::RequiresFpuRegister());
3883 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
3884 break;
3885 }
3886
3887 default:
3888 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType();
3889 }
3890 }
3891
VisitMul(HMul * mul)3892 void InstructionCodeGeneratorARMVIXL::VisitMul(HMul* mul) {
3893 LocationSummary* locations = mul->GetLocations();
3894 Location out = locations->Out();
3895 Location first = locations->InAt(0);
3896 Location second = locations->InAt(1);
3897 switch (mul->GetResultType()) {
3898 case DataType::Type::kInt32: {
3899 __ Mul(OutputRegister(mul), InputRegisterAt(mul, 0), InputRegisterAt(mul, 1));
3900 break;
3901 }
3902 case DataType::Type::kInt64: {
3903 vixl32::Register out_hi = HighRegisterFrom(out);
3904 vixl32::Register out_lo = LowRegisterFrom(out);
3905 vixl32::Register in1_hi = HighRegisterFrom(first);
3906 vixl32::Register in1_lo = LowRegisterFrom(first);
3907 vixl32::Register in2_hi = HighRegisterFrom(second);
3908 vixl32::Register in2_lo = LowRegisterFrom(second);
3909
3910 // Extra checks to protect caused by the existence of R1_R2.
3911 // The algorithm is wrong if out.hi is either in1.lo or in2.lo:
3912 // (e.g. in1=r0_r1, in2=r2_r3 and out=r1_r2);
3913 DCHECK(!out_hi.Is(in1_lo));
3914 DCHECK(!out_hi.Is(in2_lo));
3915
3916 // input: in1 - 64 bits, in2 - 64 bits
3917 // output: out
3918 // formula: out.hi : out.lo = (in1.lo * in2.hi + in1.hi * in2.lo)* 2^32 + in1.lo * in2.lo
3919 // parts: out.hi = in1.lo * in2.hi + in1.hi * in2.lo + (in1.lo * in2.lo)[63:32]
3920 // parts: out.lo = (in1.lo * in2.lo)[31:0]
3921
3922 UseScratchRegisterScope temps(GetVIXLAssembler());
3923 vixl32::Register temp = temps.Acquire();
3924 // temp <- in1.lo * in2.hi
3925 __ Mul(temp, in1_lo, in2_hi);
3926 // out.hi <- in1.lo * in2.hi + in1.hi * in2.lo
3927 __ Mla(out_hi, in1_hi, in2_lo, temp);
3928 // out.lo <- (in1.lo * in2.lo)[31:0];
3929 __ Umull(out_lo, temp, in1_lo, in2_lo);
3930 // out.hi <- in2.hi * in1.lo + in2.lo * in1.hi + (in1.lo * in2.lo)[63:32]
3931 __ Add(out_hi, out_hi, temp);
3932 break;
3933 }
3934
3935 case DataType::Type::kFloat32:
3936 case DataType::Type::kFloat64:
3937 __ Vmul(OutputVRegister(mul), InputVRegisterAt(mul, 0), InputVRegisterAt(mul, 1));
3938 break;
3939
3940 default:
3941 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType();
3942 }
3943 }
3944
DivRemOneOrMinusOne(HBinaryOperation * instruction)3945 void InstructionCodeGeneratorARMVIXL::DivRemOneOrMinusOne(HBinaryOperation* instruction) {
3946 DCHECK(instruction->IsDiv() || instruction->IsRem());
3947 DCHECK(instruction->GetResultType() == DataType::Type::kInt32);
3948
3949 Location second = instruction->GetLocations()->InAt(1);
3950 DCHECK(second.IsConstant());
3951
3952 vixl32::Register out = OutputRegister(instruction);
3953 vixl32::Register dividend = InputRegisterAt(instruction, 0);
3954 int32_t imm = Int32ConstantFrom(second);
3955 DCHECK(imm == 1 || imm == -1);
3956
3957 if (instruction->IsRem()) {
3958 __ Mov(out, 0);
3959 } else {
3960 if (imm == 1) {
3961 __ Mov(out, dividend);
3962 } else {
3963 __ Rsb(out, dividend, 0);
3964 }
3965 }
3966 }
3967
DivRemByPowerOfTwo(HBinaryOperation * instruction)3968 void InstructionCodeGeneratorARMVIXL::DivRemByPowerOfTwo(HBinaryOperation* instruction) {
3969 DCHECK(instruction->IsDiv() || instruction->IsRem());
3970 DCHECK(instruction->GetResultType() == DataType::Type::kInt32);
3971
3972 LocationSummary* locations = instruction->GetLocations();
3973 Location second = locations->InAt(1);
3974 DCHECK(second.IsConstant());
3975
3976 vixl32::Register out = OutputRegister(instruction);
3977 vixl32::Register dividend = InputRegisterAt(instruction, 0);
3978 vixl32::Register temp = RegisterFrom(locations->GetTemp(0));
3979 int32_t imm = Int32ConstantFrom(second);
3980 uint32_t abs_imm = static_cast<uint32_t>(AbsOrMin(imm));
3981 int ctz_imm = CTZ(abs_imm);
3982
3983 if (ctz_imm == 1) {
3984 __ Lsr(temp, dividend, 32 - ctz_imm);
3985 } else {
3986 __ Asr(temp, dividend, 31);
3987 __ Lsr(temp, temp, 32 - ctz_imm);
3988 }
3989 __ Add(out, temp, dividend);
3990
3991 if (instruction->IsDiv()) {
3992 __ Asr(out, out, ctz_imm);
3993 if (imm < 0) {
3994 __ Rsb(out, out, 0);
3995 }
3996 } else {
3997 __ Ubfx(out, out, 0, ctz_imm);
3998 __ Sub(out, out, temp);
3999 }
4000 }
4001
GenerateDivRemWithAnyConstant(HBinaryOperation * instruction)4002 void InstructionCodeGeneratorARMVIXL::GenerateDivRemWithAnyConstant(HBinaryOperation* instruction) {
4003 DCHECK(instruction->IsDiv() || instruction->IsRem());
4004 DCHECK(instruction->GetResultType() == DataType::Type::kInt32);
4005
4006 LocationSummary* locations = instruction->GetLocations();
4007 Location second = locations->InAt(1);
4008 DCHECK(second.IsConstant());
4009
4010 vixl32::Register out = OutputRegister(instruction);
4011 vixl32::Register dividend = InputRegisterAt(instruction, 0);
4012 vixl32::Register temp1 = RegisterFrom(locations->GetTemp(0));
4013 vixl32::Register temp2 = RegisterFrom(locations->GetTemp(1));
4014 int32_t imm = Int32ConstantFrom(second);
4015
4016 int64_t magic;
4017 int shift;
4018 CalculateMagicAndShiftForDivRem(imm, /* is_long= */ false, &magic, &shift);
4019
4020 // TODO(VIXL): Change the static cast to Operand::From() after VIXL is fixed.
4021 __ Mov(temp1, static_cast<int32_t>(magic));
4022 __ Smull(temp2, temp1, dividend, temp1);
4023
4024 if (imm > 0 && magic < 0) {
4025 __ Add(temp1, temp1, dividend);
4026 } else if (imm < 0 && magic > 0) {
4027 __ Sub(temp1, temp1, dividend);
4028 }
4029
4030 if (shift != 0) {
4031 __ Asr(temp1, temp1, shift);
4032 }
4033
4034 if (instruction->IsDiv()) {
4035 __ Sub(out, temp1, Operand(temp1, vixl32::Shift(ASR), 31));
4036 } else {
4037 __ Sub(temp1, temp1, Operand(temp1, vixl32::Shift(ASR), 31));
4038 // TODO: Strength reduction for mls.
4039 __ Mov(temp2, imm);
4040 __ Mls(out, temp1, temp2, dividend);
4041 }
4042 }
4043
GenerateDivRemConstantIntegral(HBinaryOperation * instruction)4044 void InstructionCodeGeneratorARMVIXL::GenerateDivRemConstantIntegral(
4045 HBinaryOperation* instruction) {
4046 DCHECK(instruction->IsDiv() || instruction->IsRem());
4047 DCHECK(instruction->GetResultType() == DataType::Type::kInt32);
4048
4049 Location second = instruction->GetLocations()->InAt(1);
4050 DCHECK(second.IsConstant());
4051
4052 int32_t imm = Int32ConstantFrom(second);
4053 if (imm == 0) {
4054 // Do not generate anything. DivZeroCheck would prevent any code to be executed.
4055 } else if (imm == 1 || imm == -1) {
4056 DivRemOneOrMinusOne(instruction);
4057 } else if (IsPowerOfTwo(AbsOrMin(imm))) {
4058 DivRemByPowerOfTwo(instruction);
4059 } else {
4060 DCHECK(imm <= -2 || imm >= 2);
4061 GenerateDivRemWithAnyConstant(instruction);
4062 }
4063 }
4064
VisitDiv(HDiv * div)4065 void LocationsBuilderARMVIXL::VisitDiv(HDiv* div) {
4066 LocationSummary::CallKind call_kind = LocationSummary::kNoCall;
4067 if (div->GetResultType() == DataType::Type::kInt64) {
4068 // pLdiv runtime call.
4069 call_kind = LocationSummary::kCallOnMainOnly;
4070 } else if (div->GetResultType() == DataType::Type::kInt32 && div->InputAt(1)->IsConstant()) {
4071 // sdiv will be replaced by other instruction sequence.
4072 } else if (div->GetResultType() == DataType::Type::kInt32 &&
4073 !codegen_->GetInstructionSetFeatures().HasDivideInstruction()) {
4074 // pIdivmod runtime call.
4075 call_kind = LocationSummary::kCallOnMainOnly;
4076 }
4077
4078 LocationSummary* locations = new (GetGraph()->GetAllocator()) LocationSummary(div, call_kind);
4079
4080 switch (div->GetResultType()) {
4081 case DataType::Type::kInt32: {
4082 if (div->InputAt(1)->IsConstant()) {
4083 locations->SetInAt(0, Location::RequiresRegister());
4084 locations->SetInAt(1, Location::ConstantLocation(div->InputAt(1)->AsConstant()));
4085 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
4086 int32_t value = Int32ConstantFrom(div->InputAt(1));
4087 if (value == 1 || value == 0 || value == -1) {
4088 // No temp register required.
4089 } else {
4090 locations->AddTemp(Location::RequiresRegister());
4091 if (!IsPowerOfTwo(AbsOrMin(value))) {
4092 locations->AddTemp(Location::RequiresRegister());
4093 }
4094 }
4095 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) {
4096 locations->SetInAt(0, Location::RequiresRegister());
4097 locations->SetInAt(1, Location::RequiresRegister());
4098 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
4099 } else {
4100 InvokeRuntimeCallingConventionARMVIXL calling_convention;
4101 locations->SetInAt(0, LocationFrom(calling_convention.GetRegisterAt(0)));
4102 locations->SetInAt(1, LocationFrom(calling_convention.GetRegisterAt(1)));
4103 // Note: divmod will compute both the quotient and the remainder as the pair R0 and R1, but
4104 // we only need the former.
4105 locations->SetOut(LocationFrom(r0));
4106 }
4107 break;
4108 }
4109 case DataType::Type::kInt64: {
4110 InvokeRuntimeCallingConventionARMVIXL calling_convention;
4111 locations->SetInAt(0, LocationFrom(
4112 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1)));
4113 locations->SetInAt(1, LocationFrom(
4114 calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3)));
4115 locations->SetOut(LocationFrom(r0, r1));
4116 break;
4117 }
4118 case DataType::Type::kFloat32:
4119 case DataType::Type::kFloat64: {
4120 locations->SetInAt(0, Location::RequiresFpuRegister());
4121 locations->SetInAt(1, Location::RequiresFpuRegister());
4122 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
4123 break;
4124 }
4125
4126 default:
4127 LOG(FATAL) << "Unexpected div type " << div->GetResultType();
4128 }
4129 }
4130
VisitDiv(HDiv * div)4131 void InstructionCodeGeneratorARMVIXL::VisitDiv(HDiv* div) {
4132 Location lhs = div->GetLocations()->InAt(0);
4133 Location rhs = div->GetLocations()->InAt(1);
4134
4135 switch (div->GetResultType()) {
4136 case DataType::Type::kInt32: {
4137 if (rhs.IsConstant()) {
4138 GenerateDivRemConstantIntegral(div);
4139 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) {
4140 __ Sdiv(OutputRegister(div), InputRegisterAt(div, 0), InputRegisterAt(div, 1));
4141 } else {
4142 InvokeRuntimeCallingConventionARMVIXL calling_convention;
4143 DCHECK(calling_convention.GetRegisterAt(0).Is(RegisterFrom(lhs)));
4144 DCHECK(calling_convention.GetRegisterAt(1).Is(RegisterFrom(rhs)));
4145 DCHECK(r0.Is(OutputRegister(div)));
4146
4147 codegen_->InvokeRuntime(kQuickIdivmod, div, div->GetDexPc());
4148 CheckEntrypointTypes<kQuickIdivmod, int32_t, int32_t, int32_t>();
4149 }
4150 break;
4151 }
4152
4153 case DataType::Type::kInt64: {
4154 InvokeRuntimeCallingConventionARMVIXL calling_convention;
4155 DCHECK(calling_convention.GetRegisterAt(0).Is(LowRegisterFrom(lhs)));
4156 DCHECK(calling_convention.GetRegisterAt(1).Is(HighRegisterFrom(lhs)));
4157 DCHECK(calling_convention.GetRegisterAt(2).Is(LowRegisterFrom(rhs)));
4158 DCHECK(calling_convention.GetRegisterAt(3).Is(HighRegisterFrom(rhs)));
4159 DCHECK(LowRegisterFrom(div->GetLocations()->Out()).Is(r0));
4160 DCHECK(HighRegisterFrom(div->GetLocations()->Out()).Is(r1));
4161
4162 codegen_->InvokeRuntime(kQuickLdiv, div, div->GetDexPc());
4163 CheckEntrypointTypes<kQuickLdiv, int64_t, int64_t, int64_t>();
4164 break;
4165 }
4166
4167 case DataType::Type::kFloat32:
4168 case DataType::Type::kFloat64:
4169 __ Vdiv(OutputVRegister(div), InputVRegisterAt(div, 0), InputVRegisterAt(div, 1));
4170 break;
4171
4172 default:
4173 LOG(FATAL) << "Unexpected div type " << div->GetResultType();
4174 }
4175 }
4176
VisitRem(HRem * rem)4177 void LocationsBuilderARMVIXL::VisitRem(HRem* rem) {
4178 DataType::Type type = rem->GetResultType();
4179
4180 // Most remainders are implemented in the runtime.
4181 LocationSummary::CallKind call_kind = LocationSummary::kCallOnMainOnly;
4182 if (rem->GetResultType() == DataType::Type::kInt32 && rem->InputAt(1)->IsConstant()) {
4183 // sdiv will be replaced by other instruction sequence.
4184 call_kind = LocationSummary::kNoCall;
4185 } else if ((rem->GetResultType() == DataType::Type::kInt32)
4186 && codegen_->GetInstructionSetFeatures().HasDivideInstruction()) {
4187 // Have hardware divide instruction for int, do it with three instructions.
4188 call_kind = LocationSummary::kNoCall;
4189 }
4190
4191 LocationSummary* locations = new (GetGraph()->GetAllocator()) LocationSummary(rem, call_kind);
4192
4193 switch (type) {
4194 case DataType::Type::kInt32: {
4195 if (rem->InputAt(1)->IsConstant()) {
4196 locations->SetInAt(0, Location::RequiresRegister());
4197 locations->SetInAt(1, Location::ConstantLocation(rem->InputAt(1)->AsConstant()));
4198 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
4199 int32_t value = Int32ConstantFrom(rem->InputAt(1));
4200 if (value == 1 || value == 0 || value == -1) {
4201 // No temp register required.
4202 } else {
4203 locations->AddTemp(Location::RequiresRegister());
4204 if (!IsPowerOfTwo(AbsOrMin(value))) {
4205 locations->AddTemp(Location::RequiresRegister());
4206 }
4207 }
4208 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) {
4209 locations->SetInAt(0, Location::RequiresRegister());
4210 locations->SetInAt(1, Location::RequiresRegister());
4211 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
4212 locations->AddTemp(Location::RequiresRegister());
4213 } else {
4214 InvokeRuntimeCallingConventionARMVIXL calling_convention;
4215 locations->SetInAt(0, LocationFrom(calling_convention.GetRegisterAt(0)));
4216 locations->SetInAt(1, LocationFrom(calling_convention.GetRegisterAt(1)));
4217 // Note: divmod will compute both the quotient and the remainder as the pair R0 and R1, but
4218 // we only need the latter.
4219 locations->SetOut(LocationFrom(r1));
4220 }
4221 break;
4222 }
4223 case DataType::Type::kInt64: {
4224 InvokeRuntimeCallingConventionARMVIXL calling_convention;
4225 locations->SetInAt(0, LocationFrom(
4226 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1)));
4227 locations->SetInAt(1, LocationFrom(
4228 calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3)));
4229 // The runtime helper puts the output in R2,R3.
4230 locations->SetOut(LocationFrom(r2, r3));
4231 break;
4232 }
4233 case DataType::Type::kFloat32: {
4234 InvokeRuntimeCallingConventionARMVIXL calling_convention;
4235 locations->SetInAt(0, LocationFrom(calling_convention.GetFpuRegisterAt(0)));
4236 locations->SetInAt(1, LocationFrom(calling_convention.GetFpuRegisterAt(1)));
4237 locations->SetOut(LocationFrom(s0));
4238 break;
4239 }
4240
4241 case DataType::Type::kFloat64: {
4242 InvokeRuntimeCallingConventionARMVIXL calling_convention;
4243 locations->SetInAt(0, LocationFrom(
4244 calling_convention.GetFpuRegisterAt(0), calling_convention.GetFpuRegisterAt(1)));
4245 locations->SetInAt(1, LocationFrom(
4246 calling_convention.GetFpuRegisterAt(2), calling_convention.GetFpuRegisterAt(3)));
4247 locations->SetOut(LocationFrom(s0, s1));
4248 break;
4249 }
4250
4251 default:
4252 LOG(FATAL) << "Unexpected rem type " << type;
4253 }
4254 }
4255
VisitRem(HRem * rem)4256 void InstructionCodeGeneratorARMVIXL::VisitRem(HRem* rem) {
4257 LocationSummary* locations = rem->GetLocations();
4258 Location second = locations->InAt(1);
4259
4260 DataType::Type type = rem->GetResultType();
4261 switch (type) {
4262 case DataType::Type::kInt32: {
4263 vixl32::Register reg1 = InputRegisterAt(rem, 0);
4264 vixl32::Register out_reg = OutputRegister(rem);
4265 if (second.IsConstant()) {
4266 GenerateDivRemConstantIntegral(rem);
4267 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) {
4268 vixl32::Register reg2 = RegisterFrom(second);
4269 vixl32::Register temp = RegisterFrom(locations->GetTemp(0));
4270
4271 // temp = reg1 / reg2 (integer division)
4272 // dest = reg1 - temp * reg2
4273 __ Sdiv(temp, reg1, reg2);
4274 __ Mls(out_reg, temp, reg2, reg1);
4275 } else {
4276 InvokeRuntimeCallingConventionARMVIXL calling_convention;
4277 DCHECK(reg1.Is(calling_convention.GetRegisterAt(0)));
4278 DCHECK(RegisterFrom(second).Is(calling_convention.GetRegisterAt(1)));
4279 DCHECK(out_reg.Is(r1));
4280
4281 codegen_->InvokeRuntime(kQuickIdivmod, rem, rem->GetDexPc());
4282 CheckEntrypointTypes<kQuickIdivmod, int32_t, int32_t, int32_t>();
4283 }
4284 break;
4285 }
4286
4287 case DataType::Type::kInt64: {
4288 codegen_->InvokeRuntime(kQuickLmod, rem, rem->GetDexPc());
4289 CheckEntrypointTypes<kQuickLmod, int64_t, int64_t, int64_t>();
4290 break;
4291 }
4292
4293 case DataType::Type::kFloat32: {
4294 codegen_->InvokeRuntime(kQuickFmodf, rem, rem->GetDexPc());
4295 CheckEntrypointTypes<kQuickFmodf, float, float, float>();
4296 break;
4297 }
4298
4299 case DataType::Type::kFloat64: {
4300 codegen_->InvokeRuntime(kQuickFmod, rem, rem->GetDexPc());
4301 CheckEntrypointTypes<kQuickFmod, double, double, double>();
4302 break;
4303 }
4304
4305 default:
4306 LOG(FATAL) << "Unexpected rem type " << type;
4307 }
4308 }
4309
CreateMinMaxLocations(ArenaAllocator * allocator,HBinaryOperation * minmax)4310 static void CreateMinMaxLocations(ArenaAllocator* allocator, HBinaryOperation* minmax) {
4311 LocationSummary* locations = new (allocator) LocationSummary(minmax);
4312 switch (minmax->GetResultType()) {
4313 case DataType::Type::kInt32:
4314 locations->SetInAt(0, Location::RequiresRegister());
4315 locations->SetInAt(1, Location::RequiresRegister());
4316 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
4317 break;
4318 case DataType::Type::kInt64:
4319 locations->SetInAt(0, Location::RequiresRegister());
4320 locations->SetInAt(1, Location::RequiresRegister());
4321 locations->SetOut(Location::SameAsFirstInput());
4322 break;
4323 case DataType::Type::kFloat32:
4324 locations->SetInAt(0, Location::RequiresFpuRegister());
4325 locations->SetInAt(1, Location::RequiresFpuRegister());
4326 locations->SetOut(Location::SameAsFirstInput());
4327 locations->AddTemp(Location::RequiresRegister());
4328 break;
4329 case DataType::Type::kFloat64:
4330 locations->SetInAt(0, Location::RequiresFpuRegister());
4331 locations->SetInAt(1, Location::RequiresFpuRegister());
4332 locations->SetOut(Location::SameAsFirstInput());
4333 break;
4334 default:
4335 LOG(FATAL) << "Unexpected type for HMinMax " << minmax->GetResultType();
4336 }
4337 }
4338
GenerateMinMaxInt(LocationSummary * locations,bool is_min)4339 void InstructionCodeGeneratorARMVIXL::GenerateMinMaxInt(LocationSummary* locations, bool is_min) {
4340 Location op1_loc = locations->InAt(0);
4341 Location op2_loc = locations->InAt(1);
4342 Location out_loc = locations->Out();
4343
4344 vixl32::Register op1 = RegisterFrom(op1_loc);
4345 vixl32::Register op2 = RegisterFrom(op2_loc);
4346 vixl32::Register out = RegisterFrom(out_loc);
4347
4348 __ Cmp(op1, op2);
4349
4350 {
4351 ExactAssemblyScope aas(GetVIXLAssembler(),
4352 3 * kMaxInstructionSizeInBytes,
4353 CodeBufferCheckScope::kMaximumSize);
4354
4355 __ ite(is_min ? lt : gt);
4356 __ mov(is_min ? lt : gt, out, op1);
4357 __ mov(is_min ? ge : le, out, op2);
4358 }
4359 }
4360
GenerateMinMaxLong(LocationSummary * locations,bool is_min)4361 void InstructionCodeGeneratorARMVIXL::GenerateMinMaxLong(LocationSummary* locations, bool is_min) {
4362 Location op1_loc = locations->InAt(0);
4363 Location op2_loc = locations->InAt(1);
4364 Location out_loc = locations->Out();
4365
4366 // Optimization: don't generate any code if inputs are the same.
4367 if (op1_loc.Equals(op2_loc)) {
4368 DCHECK(out_loc.Equals(op1_loc)); // out_loc is set as SameAsFirstInput() in location builder.
4369 return;
4370 }
4371
4372 vixl32::Register op1_lo = LowRegisterFrom(op1_loc);
4373 vixl32::Register op1_hi = HighRegisterFrom(op1_loc);
4374 vixl32::Register op2_lo = LowRegisterFrom(op2_loc);
4375 vixl32::Register op2_hi = HighRegisterFrom(op2_loc);
4376 vixl32::Register out_lo = LowRegisterFrom(out_loc);
4377 vixl32::Register out_hi = HighRegisterFrom(out_loc);
4378 UseScratchRegisterScope temps(GetVIXLAssembler());
4379 const vixl32::Register temp = temps.Acquire();
4380
4381 DCHECK(op1_lo.Is(out_lo));
4382 DCHECK(op1_hi.Is(out_hi));
4383
4384 // Compare op1 >= op2, or op1 < op2.
4385 __ Cmp(out_lo, op2_lo);
4386 __ Sbcs(temp, out_hi, op2_hi);
4387
4388 // Now GE/LT condition code is correct for the long comparison.
4389 {
4390 vixl32::ConditionType cond = is_min ? ge : lt;
4391 ExactAssemblyScope it_scope(GetVIXLAssembler(),
4392 3 * kMaxInstructionSizeInBytes,
4393 CodeBufferCheckScope::kMaximumSize);
4394 __ itt(cond);
4395 __ mov(cond, out_lo, op2_lo);
4396 __ mov(cond, out_hi, op2_hi);
4397 }
4398 }
4399
GenerateMinMaxFloat(HInstruction * minmax,bool is_min)4400 void InstructionCodeGeneratorARMVIXL::GenerateMinMaxFloat(HInstruction* minmax, bool is_min) {
4401 LocationSummary* locations = minmax->GetLocations();
4402 Location op1_loc = locations->InAt(0);
4403 Location op2_loc = locations->InAt(1);
4404 Location out_loc = locations->Out();
4405
4406 // Optimization: don't generate any code if inputs are the same.
4407 if (op1_loc.Equals(op2_loc)) {
4408 DCHECK(out_loc.Equals(op1_loc)); // out_loc is set as SameAsFirstInput() in location builder.
4409 return;
4410 }
4411
4412 vixl32::SRegister op1 = SRegisterFrom(op1_loc);
4413 vixl32::SRegister op2 = SRegisterFrom(op2_loc);
4414 vixl32::SRegister out = SRegisterFrom(out_loc);
4415
4416 UseScratchRegisterScope temps(GetVIXLAssembler());
4417 const vixl32::Register temp1 = temps.Acquire();
4418 vixl32::Register temp2 = RegisterFrom(locations->GetTemp(0));
4419 vixl32::Label nan, done;
4420 vixl32::Label* final_label = codegen_->GetFinalLabel(minmax, &done);
4421
4422 DCHECK(op1.Is(out));
4423
4424 __ Vcmp(op1, op2);
4425 __ Vmrs(RegisterOrAPSR_nzcv(kPcCode), FPSCR);
4426 __ B(vs, &nan, /* is_far_target= */ false); // if un-ordered, go to NaN handling.
4427
4428 // op1 <> op2
4429 vixl32::ConditionType cond = is_min ? gt : lt;
4430 {
4431 ExactAssemblyScope it_scope(GetVIXLAssembler(),
4432 2 * kMaxInstructionSizeInBytes,
4433 CodeBufferCheckScope::kMaximumSize);
4434 __ it(cond);
4435 __ vmov(cond, F32, out, op2);
4436 }
4437 // for <>(not equal), we've done min/max calculation.
4438 __ B(ne, final_label, /* is_far_target= */ false);
4439
4440 // handle op1 == op2, max(+0.0,-0.0), min(+0.0,-0.0).
4441 __ Vmov(temp1, op1);
4442 __ Vmov(temp2, op2);
4443 if (is_min) {
4444 __ Orr(temp1, temp1, temp2);
4445 } else {
4446 __ And(temp1, temp1, temp2);
4447 }
4448 __ Vmov(out, temp1);
4449 __ B(final_label);
4450
4451 // handle NaN input.
4452 __ Bind(&nan);
4453 __ Movt(temp1, High16Bits(kNanFloat)); // 0x7FC0xxxx is a NaN.
4454 __ Vmov(out, temp1);
4455
4456 if (done.IsReferenced()) {
4457 __ Bind(&done);
4458 }
4459 }
4460
GenerateMinMaxDouble(HInstruction * minmax,bool is_min)4461 void InstructionCodeGeneratorARMVIXL::GenerateMinMaxDouble(HInstruction* minmax, bool is_min) {
4462 LocationSummary* locations = minmax->GetLocations();
4463 Location op1_loc = locations->InAt(0);
4464 Location op2_loc = locations->InAt(1);
4465 Location out_loc = locations->Out();
4466
4467 // Optimization: don't generate any code if inputs are the same.
4468 if (op1_loc.Equals(op2_loc)) {
4469 DCHECK(out_loc.Equals(op1_loc)); // out_loc is set as SameAsFirstInput() in.
4470 return;
4471 }
4472
4473 vixl32::DRegister op1 = DRegisterFrom(op1_loc);
4474 vixl32::DRegister op2 = DRegisterFrom(op2_loc);
4475 vixl32::DRegister out = DRegisterFrom(out_loc);
4476 vixl32::Label handle_nan_eq, done;
4477 vixl32::Label* final_label = codegen_->GetFinalLabel(minmax, &done);
4478
4479 DCHECK(op1.Is(out));
4480
4481 __ Vcmp(op1, op2);
4482 __ Vmrs(RegisterOrAPSR_nzcv(kPcCode), FPSCR);
4483 __ B(vs, &handle_nan_eq, /* is_far_target= */ false); // if un-ordered, go to NaN handling.
4484
4485 // op1 <> op2
4486 vixl32::ConditionType cond = is_min ? gt : lt;
4487 {
4488 ExactAssemblyScope it_scope(GetVIXLAssembler(),
4489 2 * kMaxInstructionSizeInBytes,
4490 CodeBufferCheckScope::kMaximumSize);
4491 __ it(cond);
4492 __ vmov(cond, F64, out, op2);
4493 }
4494 // for <>(not equal), we've done min/max calculation.
4495 __ B(ne, final_label, /* is_far_target= */ false);
4496
4497 // handle op1 == op2, max(+0.0,-0.0).
4498 if (!is_min) {
4499 __ Vand(F64, out, op1, op2);
4500 __ B(final_label);
4501 }
4502
4503 // handle op1 == op2, min(+0.0,-0.0), NaN input.
4504 __ Bind(&handle_nan_eq);
4505 __ Vorr(F64, out, op1, op2); // assemble op1/-0.0/NaN.
4506
4507 if (done.IsReferenced()) {
4508 __ Bind(&done);
4509 }
4510 }
4511
GenerateMinMax(HBinaryOperation * minmax,bool is_min)4512 void InstructionCodeGeneratorARMVIXL::GenerateMinMax(HBinaryOperation* minmax, bool is_min) {
4513 DataType::Type type = minmax->GetResultType();
4514 switch (type) {
4515 case DataType::Type::kInt32:
4516 GenerateMinMaxInt(minmax->GetLocations(), is_min);
4517 break;
4518 case DataType::Type::kInt64:
4519 GenerateMinMaxLong(minmax->GetLocations(), is_min);
4520 break;
4521 case DataType::Type::kFloat32:
4522 GenerateMinMaxFloat(minmax, is_min);
4523 break;
4524 case DataType::Type::kFloat64:
4525 GenerateMinMaxDouble(minmax, is_min);
4526 break;
4527 default:
4528 LOG(FATAL) << "Unexpected type for HMinMax " << type;
4529 }
4530 }
4531
VisitMin(HMin * min)4532 void LocationsBuilderARMVIXL::VisitMin(HMin* min) {
4533 CreateMinMaxLocations(GetGraph()->GetAllocator(), min);
4534 }
4535
VisitMin(HMin * min)4536 void InstructionCodeGeneratorARMVIXL::VisitMin(HMin* min) {
4537 GenerateMinMax(min, /*is_min*/ true);
4538 }
4539
VisitMax(HMax * max)4540 void LocationsBuilderARMVIXL::VisitMax(HMax* max) {
4541 CreateMinMaxLocations(GetGraph()->GetAllocator(), max);
4542 }
4543
VisitMax(HMax * max)4544 void InstructionCodeGeneratorARMVIXL::VisitMax(HMax* max) {
4545 GenerateMinMax(max, /*is_min*/ false);
4546 }
4547
VisitAbs(HAbs * abs)4548 void LocationsBuilderARMVIXL::VisitAbs(HAbs* abs) {
4549 LocationSummary* locations = new (GetGraph()->GetAllocator()) LocationSummary(abs);
4550 switch (abs->GetResultType()) {
4551 case DataType::Type::kInt32:
4552 case DataType::Type::kInt64:
4553 locations->SetInAt(0, Location::RequiresRegister());
4554 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
4555 locations->AddTemp(Location::RequiresRegister());
4556 break;
4557 case DataType::Type::kFloat32:
4558 case DataType::Type::kFloat64:
4559 locations->SetInAt(0, Location::RequiresFpuRegister());
4560 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
4561 break;
4562 default:
4563 LOG(FATAL) << "Unexpected type for abs operation " << abs->GetResultType();
4564 }
4565 }
4566
VisitAbs(HAbs * abs)4567 void InstructionCodeGeneratorARMVIXL::VisitAbs(HAbs* abs) {
4568 LocationSummary* locations = abs->GetLocations();
4569 switch (abs->GetResultType()) {
4570 case DataType::Type::kInt32: {
4571 vixl32::Register in_reg = RegisterFrom(locations->InAt(0));
4572 vixl32::Register out_reg = RegisterFrom(locations->Out());
4573 vixl32::Register mask = RegisterFrom(locations->GetTemp(0));
4574 __ Asr(mask, in_reg, 31);
4575 __ Add(out_reg, in_reg, mask);
4576 __ Eor(out_reg, out_reg, mask);
4577 break;
4578 }
4579 case DataType::Type::kInt64: {
4580 Location in = locations->InAt(0);
4581 vixl32::Register in_reg_lo = LowRegisterFrom(in);
4582 vixl32::Register in_reg_hi = HighRegisterFrom(in);
4583 Location output = locations->Out();
4584 vixl32::Register out_reg_lo = LowRegisterFrom(output);
4585 vixl32::Register out_reg_hi = HighRegisterFrom(output);
4586 DCHECK(!out_reg_lo.Is(in_reg_hi)) << "Diagonal overlap unexpected.";
4587 vixl32::Register mask = RegisterFrom(locations->GetTemp(0));
4588 __ Asr(mask, in_reg_hi, 31);
4589 __ Adds(out_reg_lo, in_reg_lo, mask);
4590 __ Adc(out_reg_hi, in_reg_hi, mask);
4591 __ Eor(out_reg_lo, out_reg_lo, mask);
4592 __ Eor(out_reg_hi, out_reg_hi, mask);
4593 break;
4594 }
4595 case DataType::Type::kFloat32:
4596 case DataType::Type::kFloat64:
4597 __ Vabs(OutputVRegister(abs), InputVRegisterAt(abs, 0));
4598 break;
4599 default:
4600 LOG(FATAL) << "Unexpected type for abs operation " << abs->GetResultType();
4601 }
4602 }
4603
VisitDivZeroCheck(HDivZeroCheck * instruction)4604 void LocationsBuilderARMVIXL::VisitDivZeroCheck(HDivZeroCheck* instruction) {
4605 LocationSummary* locations = codegen_->CreateThrowingSlowPathLocations(instruction);
4606 locations->SetInAt(0, Location::RegisterOrConstant(instruction->InputAt(0)));
4607 }
4608
VisitDivZeroCheck(HDivZeroCheck * instruction)4609 void InstructionCodeGeneratorARMVIXL::VisitDivZeroCheck(HDivZeroCheck* instruction) {
4610 DivZeroCheckSlowPathARMVIXL* slow_path =
4611 new (codegen_->GetScopedAllocator()) DivZeroCheckSlowPathARMVIXL(instruction);
4612 codegen_->AddSlowPath(slow_path);
4613
4614 LocationSummary* locations = instruction->GetLocations();
4615 Location value = locations->InAt(0);
4616
4617 switch (instruction->GetType()) {
4618 case DataType::Type::kBool:
4619 case DataType::Type::kUint8:
4620 case DataType::Type::kInt8:
4621 case DataType::Type::kUint16:
4622 case DataType::Type::kInt16:
4623 case DataType::Type::kInt32: {
4624 if (value.IsRegister()) {
4625 __ CompareAndBranchIfZero(InputRegisterAt(instruction, 0), slow_path->GetEntryLabel());
4626 } else {
4627 DCHECK(value.IsConstant()) << value;
4628 if (Int32ConstantFrom(value) == 0) {
4629 __ B(slow_path->GetEntryLabel());
4630 }
4631 }
4632 break;
4633 }
4634 case DataType::Type::kInt64: {
4635 if (value.IsRegisterPair()) {
4636 UseScratchRegisterScope temps(GetVIXLAssembler());
4637 vixl32::Register temp = temps.Acquire();
4638 __ Orrs(temp, LowRegisterFrom(value), HighRegisterFrom(value));
4639 __ B(eq, slow_path->GetEntryLabel());
4640 } else {
4641 DCHECK(value.IsConstant()) << value;
4642 if (Int64ConstantFrom(value) == 0) {
4643 __ B(slow_path->GetEntryLabel());
4644 }
4645 }
4646 break;
4647 }
4648 default:
4649 LOG(FATAL) << "Unexpected type for HDivZeroCheck " << instruction->GetType();
4650 }
4651 }
4652
HandleIntegerRotate(HRor * ror)4653 void InstructionCodeGeneratorARMVIXL::HandleIntegerRotate(HRor* ror) {
4654 LocationSummary* locations = ror->GetLocations();
4655 vixl32::Register in = InputRegisterAt(ror, 0);
4656 Location rhs = locations->InAt(1);
4657 vixl32::Register out = OutputRegister(ror);
4658
4659 if (rhs.IsConstant()) {
4660 // Arm32 and Thumb2 assemblers require a rotation on the interval [1,31],
4661 // so map all rotations to a +ve. equivalent in that range.
4662 // (e.g. left *or* right by -2 bits == 30 bits in the same direction.)
4663 uint32_t rot = CodeGenerator::GetInt32ValueOf(rhs.GetConstant()) & 0x1F;
4664 if (rot) {
4665 // Rotate, mapping left rotations to right equivalents if necessary.
4666 // (e.g. left by 2 bits == right by 30.)
4667 __ Ror(out, in, rot);
4668 } else if (!out.Is(in)) {
4669 __ Mov(out, in);
4670 }
4671 } else {
4672 __ Ror(out, in, RegisterFrom(rhs));
4673 }
4674 }
4675
4676 // Gain some speed by mapping all Long rotates onto equivalent pairs of Integer
4677 // rotates by swapping input regs (effectively rotating by the first 32-bits of
4678 // a larger rotation) or flipping direction (thus treating larger right/left
4679 // rotations as sub-word sized rotations in the other direction) as appropriate.
HandleLongRotate(HRor * ror)4680 void InstructionCodeGeneratorARMVIXL::HandleLongRotate(HRor* ror) {
4681 LocationSummary* locations = ror->GetLocations();
4682 vixl32::Register in_reg_lo = LowRegisterFrom(locations->InAt(0));
4683 vixl32::Register in_reg_hi = HighRegisterFrom(locations->InAt(0));
4684 Location rhs = locations->InAt(1);
4685 vixl32::Register out_reg_lo = LowRegisterFrom(locations->Out());
4686 vixl32::Register out_reg_hi = HighRegisterFrom(locations->Out());
4687
4688 if (rhs.IsConstant()) {
4689 uint64_t rot = CodeGenerator::GetInt64ValueOf(rhs.GetConstant());
4690 // Map all rotations to +ve. equivalents on the interval [0,63].
4691 rot &= kMaxLongShiftDistance;
4692 // For rotates over a word in size, 'pre-rotate' by 32-bits to keep rotate
4693 // logic below to a simple pair of binary orr.
4694 // (e.g. 34 bits == in_reg swap + 2 bits right.)
4695 if (rot >= kArmBitsPerWord) {
4696 rot -= kArmBitsPerWord;
4697 std::swap(in_reg_hi, in_reg_lo);
4698 }
4699 // Rotate, or mov to out for zero or word size rotations.
4700 if (rot != 0u) {
4701 __ Lsr(out_reg_hi, in_reg_hi, Operand::From(rot));
4702 __ Orr(out_reg_hi, out_reg_hi, Operand(in_reg_lo, ShiftType::LSL, kArmBitsPerWord - rot));
4703 __ Lsr(out_reg_lo, in_reg_lo, Operand::From(rot));
4704 __ Orr(out_reg_lo, out_reg_lo, Operand(in_reg_hi, ShiftType::LSL, kArmBitsPerWord - rot));
4705 } else {
4706 __ Mov(out_reg_lo, in_reg_lo);
4707 __ Mov(out_reg_hi, in_reg_hi);
4708 }
4709 } else {
4710 vixl32::Register shift_right = RegisterFrom(locations->GetTemp(0));
4711 vixl32::Register shift_left = RegisterFrom(locations->GetTemp(1));
4712 vixl32::Label end;
4713 vixl32::Label shift_by_32_plus_shift_right;
4714 vixl32::Label* final_label = codegen_->GetFinalLabel(ror, &end);
4715
4716 __ And(shift_right, RegisterFrom(rhs), 0x1F);
4717 __ Lsrs(shift_left, RegisterFrom(rhs), 6);
4718 __ Rsb(LeaveFlags, shift_left, shift_right, Operand::From(kArmBitsPerWord));
4719 __ B(cc, &shift_by_32_plus_shift_right, /* is_far_target= */ false);
4720
4721 // out_reg_hi = (reg_hi << shift_left) | (reg_lo >> shift_right).
4722 // out_reg_lo = (reg_lo << shift_left) | (reg_hi >> shift_right).
4723 __ Lsl(out_reg_hi, in_reg_hi, shift_left);
4724 __ Lsr(out_reg_lo, in_reg_lo, shift_right);
4725 __ Add(out_reg_hi, out_reg_hi, out_reg_lo);
4726 __ Lsl(out_reg_lo, in_reg_lo, shift_left);
4727 __ Lsr(shift_left, in_reg_hi, shift_right);
4728 __ Add(out_reg_lo, out_reg_lo, shift_left);
4729 __ B(final_label);
4730
4731 __ Bind(&shift_by_32_plus_shift_right); // Shift by 32+shift_right.
4732 // out_reg_hi = (reg_hi >> shift_right) | (reg_lo << shift_left).
4733 // out_reg_lo = (reg_lo >> shift_right) | (reg_hi << shift_left).
4734 __ Lsr(out_reg_hi, in_reg_hi, shift_right);
4735 __ Lsl(out_reg_lo, in_reg_lo, shift_left);
4736 __ Add(out_reg_hi, out_reg_hi, out_reg_lo);
4737 __ Lsr(out_reg_lo, in_reg_lo, shift_right);
4738 __ Lsl(shift_right, in_reg_hi, shift_left);
4739 __ Add(out_reg_lo, out_reg_lo, shift_right);
4740
4741 if (end.IsReferenced()) {
4742 __ Bind(&end);
4743 }
4744 }
4745 }
4746
VisitRor(HRor * ror)4747 void LocationsBuilderARMVIXL::VisitRor(HRor* ror) {
4748 LocationSummary* locations =
4749 new (GetGraph()->GetAllocator()) LocationSummary(ror, LocationSummary::kNoCall);
4750 switch (ror->GetResultType()) {
4751 case DataType::Type::kInt32: {
4752 locations->SetInAt(0, Location::RequiresRegister());
4753 locations->SetInAt(1, Location::RegisterOrConstant(ror->InputAt(1)));
4754 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
4755 break;
4756 }
4757 case DataType::Type::kInt64: {
4758 locations->SetInAt(0, Location::RequiresRegister());
4759 if (ror->InputAt(1)->IsConstant()) {
4760 locations->SetInAt(1, Location::ConstantLocation(ror->InputAt(1)->AsConstant()));
4761 } else {
4762 locations->SetInAt(1, Location::RequiresRegister());
4763 locations->AddTemp(Location::RequiresRegister());
4764 locations->AddTemp(Location::RequiresRegister());
4765 }
4766 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
4767 break;
4768 }
4769 default:
4770 LOG(FATAL) << "Unexpected operation type " << ror->GetResultType();
4771 }
4772 }
4773
VisitRor(HRor * ror)4774 void InstructionCodeGeneratorARMVIXL::VisitRor(HRor* ror) {
4775 DataType::Type type = ror->GetResultType();
4776 switch (type) {
4777 case DataType::Type::kInt32: {
4778 HandleIntegerRotate(ror);
4779 break;
4780 }
4781 case DataType::Type::kInt64: {
4782 HandleLongRotate(ror);
4783 break;
4784 }
4785 default:
4786 LOG(FATAL) << "Unexpected operation type " << type;
4787 UNREACHABLE();
4788 }
4789 }
4790
HandleShift(HBinaryOperation * op)4791 void LocationsBuilderARMVIXL::HandleShift(HBinaryOperation* op) {
4792 DCHECK(op->IsShl() || op->IsShr() || op->IsUShr());
4793
4794 LocationSummary* locations =
4795 new (GetGraph()->GetAllocator()) LocationSummary(op, LocationSummary::kNoCall);
4796
4797 switch (op->GetResultType()) {
4798 case DataType::Type::kInt32: {
4799 locations->SetInAt(0, Location::RequiresRegister());
4800 if (op->InputAt(1)->IsConstant()) {
4801 locations->SetInAt(1, Location::ConstantLocation(op->InputAt(1)->AsConstant()));
4802 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
4803 } else {
4804 locations->SetInAt(1, Location::RequiresRegister());
4805 // Make the output overlap, as it will be used to hold the masked
4806 // second input.
4807 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
4808 }
4809 break;
4810 }
4811 case DataType::Type::kInt64: {
4812 locations->SetInAt(0, Location::RequiresRegister());
4813 if (op->InputAt(1)->IsConstant()) {
4814 locations->SetInAt(1, Location::ConstantLocation(op->InputAt(1)->AsConstant()));
4815 // For simplicity, use kOutputOverlap even though we only require that low registers
4816 // don't clash with high registers which the register allocator currently guarantees.
4817 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
4818 } else {
4819 locations->SetInAt(1, Location::RequiresRegister());
4820 locations->AddTemp(Location::RequiresRegister());
4821 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
4822 }
4823 break;
4824 }
4825 default:
4826 LOG(FATAL) << "Unexpected operation type " << op->GetResultType();
4827 }
4828 }
4829
HandleShift(HBinaryOperation * op)4830 void InstructionCodeGeneratorARMVIXL::HandleShift(HBinaryOperation* op) {
4831 DCHECK(op->IsShl() || op->IsShr() || op->IsUShr());
4832
4833 LocationSummary* locations = op->GetLocations();
4834 Location out = locations->Out();
4835 Location first = locations->InAt(0);
4836 Location second = locations->InAt(1);
4837
4838 DataType::Type type = op->GetResultType();
4839 switch (type) {
4840 case DataType::Type::kInt32: {
4841 vixl32::Register out_reg = OutputRegister(op);
4842 vixl32::Register first_reg = InputRegisterAt(op, 0);
4843 if (second.IsRegister()) {
4844 vixl32::Register second_reg = RegisterFrom(second);
4845 // ARM doesn't mask the shift count so we need to do it ourselves.
4846 __ And(out_reg, second_reg, kMaxIntShiftDistance);
4847 if (op->IsShl()) {
4848 __ Lsl(out_reg, first_reg, out_reg);
4849 } else if (op->IsShr()) {
4850 __ Asr(out_reg, first_reg, out_reg);
4851 } else {
4852 __ Lsr(out_reg, first_reg, out_reg);
4853 }
4854 } else {
4855 int32_t cst = Int32ConstantFrom(second);
4856 uint32_t shift_value = cst & kMaxIntShiftDistance;
4857 if (shift_value == 0) { // ARM does not support shifting with 0 immediate.
4858 __ Mov(out_reg, first_reg);
4859 } else if (op->IsShl()) {
4860 __ Lsl(out_reg, first_reg, shift_value);
4861 } else if (op->IsShr()) {
4862 __ Asr(out_reg, first_reg, shift_value);
4863 } else {
4864 __ Lsr(out_reg, first_reg, shift_value);
4865 }
4866 }
4867 break;
4868 }
4869 case DataType::Type::kInt64: {
4870 vixl32::Register o_h = HighRegisterFrom(out);
4871 vixl32::Register o_l = LowRegisterFrom(out);
4872
4873 vixl32::Register high = HighRegisterFrom(first);
4874 vixl32::Register low = LowRegisterFrom(first);
4875
4876 if (second.IsRegister()) {
4877 vixl32::Register temp = RegisterFrom(locations->GetTemp(0));
4878
4879 vixl32::Register second_reg = RegisterFrom(second);
4880
4881 if (op->IsShl()) {
4882 __ And(o_l, second_reg, kMaxLongShiftDistance);
4883 // Shift the high part
4884 __ Lsl(o_h, high, o_l);
4885 // Shift the low part and `or` what overflew on the high part
4886 __ Rsb(temp, o_l, Operand::From(kArmBitsPerWord));
4887 __ Lsr(temp, low, temp);
4888 __ Orr(o_h, o_h, temp);
4889 // If the shift is > 32 bits, override the high part
4890 __ Subs(temp, o_l, Operand::From(kArmBitsPerWord));
4891 {
4892 ExactAssemblyScope guard(GetVIXLAssembler(),
4893 2 * vixl32::kMaxInstructionSizeInBytes,
4894 CodeBufferCheckScope::kMaximumSize);
4895 __ it(pl);
4896 __ lsl(pl, o_h, low, temp);
4897 }
4898 // Shift the low part
4899 __ Lsl(o_l, low, o_l);
4900 } else if (op->IsShr()) {
4901 __ And(o_h, second_reg, kMaxLongShiftDistance);
4902 // Shift the low part
4903 __ Lsr(o_l, low, o_h);
4904 // Shift the high part and `or` what underflew on the low part
4905 __ Rsb(temp, o_h, Operand::From(kArmBitsPerWord));
4906 __ Lsl(temp, high, temp);
4907 __ Orr(o_l, o_l, temp);
4908 // If the shift is > 32 bits, override the low part
4909 __ Subs(temp, o_h, Operand::From(kArmBitsPerWord));
4910 {
4911 ExactAssemblyScope guard(GetVIXLAssembler(),
4912 2 * vixl32::kMaxInstructionSizeInBytes,
4913 CodeBufferCheckScope::kMaximumSize);
4914 __ it(pl);
4915 __ asr(pl, o_l, high, temp);
4916 }
4917 // Shift the high part
4918 __ Asr(o_h, high, o_h);
4919 } else {
4920 __ And(o_h, second_reg, kMaxLongShiftDistance);
4921 // same as Shr except we use `Lsr`s and not `Asr`s
4922 __ Lsr(o_l, low, o_h);
4923 __ Rsb(temp, o_h, Operand::From(kArmBitsPerWord));
4924 __ Lsl(temp, high, temp);
4925 __ Orr(o_l, o_l, temp);
4926 __ Subs(temp, o_h, Operand::From(kArmBitsPerWord));
4927 {
4928 ExactAssemblyScope guard(GetVIXLAssembler(),
4929 2 * vixl32::kMaxInstructionSizeInBytes,
4930 CodeBufferCheckScope::kMaximumSize);
4931 __ it(pl);
4932 __ lsr(pl, o_l, high, temp);
4933 }
4934 __ Lsr(o_h, high, o_h);
4935 }
4936 } else {
4937 // Register allocator doesn't create partial overlap.
4938 DCHECK(!o_l.Is(high));
4939 DCHECK(!o_h.Is(low));
4940 int32_t cst = Int32ConstantFrom(second);
4941 uint32_t shift_value = cst & kMaxLongShiftDistance;
4942 if (shift_value > 32) {
4943 if (op->IsShl()) {
4944 __ Lsl(o_h, low, shift_value - 32);
4945 __ Mov(o_l, 0);
4946 } else if (op->IsShr()) {
4947 __ Asr(o_l, high, shift_value - 32);
4948 __ Asr(o_h, high, 31);
4949 } else {
4950 __ Lsr(o_l, high, shift_value - 32);
4951 __ Mov(o_h, 0);
4952 }
4953 } else if (shift_value == 32) {
4954 if (op->IsShl()) {
4955 __ Mov(o_h, low);
4956 __ Mov(o_l, 0);
4957 } else if (op->IsShr()) {
4958 __ Mov(o_l, high);
4959 __ Asr(o_h, high, 31);
4960 } else {
4961 __ Mov(o_l, high);
4962 __ Mov(o_h, 0);
4963 }
4964 } else if (shift_value == 1) {
4965 if (op->IsShl()) {
4966 __ Lsls(o_l, low, 1);
4967 __ Adc(o_h, high, high);
4968 } else if (op->IsShr()) {
4969 __ Asrs(o_h, high, 1);
4970 __ Rrx(o_l, low);
4971 } else {
4972 __ Lsrs(o_h, high, 1);
4973 __ Rrx(o_l, low);
4974 }
4975 } else if (shift_value == 0) {
4976 __ Mov(o_l, low);
4977 __ Mov(o_h, high);
4978 } else {
4979 DCHECK(0 < shift_value && shift_value < 32) << shift_value;
4980 if (op->IsShl()) {
4981 __ Lsl(o_h, high, shift_value);
4982 __ Orr(o_h, o_h, Operand(low, ShiftType::LSR, 32 - shift_value));
4983 __ Lsl(o_l, low, shift_value);
4984 } else if (op->IsShr()) {
4985 __ Lsr(o_l, low, shift_value);
4986 __ Orr(o_l, o_l, Operand(high, ShiftType::LSL, 32 - shift_value));
4987 __ Asr(o_h, high, shift_value);
4988 } else {
4989 __ Lsr(o_l, low, shift_value);
4990 __ Orr(o_l, o_l, Operand(high, ShiftType::LSL, 32 - shift_value));
4991 __ Lsr(o_h, high, shift_value);
4992 }
4993 }
4994 }
4995 break;
4996 }
4997 default:
4998 LOG(FATAL) << "Unexpected operation type " << type;
4999 UNREACHABLE();
5000 }
5001 }
5002
VisitShl(HShl * shl)5003 void LocationsBuilderARMVIXL::VisitShl(HShl* shl) {
5004 HandleShift(shl);
5005 }
5006
VisitShl(HShl * shl)5007 void InstructionCodeGeneratorARMVIXL::VisitShl(HShl* shl) {
5008 HandleShift(shl);
5009 }
5010
VisitShr(HShr * shr)5011 void LocationsBuilderARMVIXL::VisitShr(HShr* shr) {
5012 HandleShift(shr);
5013 }
5014
VisitShr(HShr * shr)5015 void InstructionCodeGeneratorARMVIXL::VisitShr(HShr* shr) {
5016 HandleShift(shr);
5017 }
5018
VisitUShr(HUShr * ushr)5019 void LocationsBuilderARMVIXL::VisitUShr(HUShr* ushr) {
5020 HandleShift(ushr);
5021 }
5022
VisitUShr(HUShr * ushr)5023 void InstructionCodeGeneratorARMVIXL::VisitUShr(HUShr* ushr) {
5024 HandleShift(ushr);
5025 }
5026
VisitNewInstance(HNewInstance * instruction)5027 void LocationsBuilderARMVIXL::VisitNewInstance(HNewInstance* instruction) {
5028 LocationSummary* locations = new (GetGraph()->GetAllocator()) LocationSummary(
5029 instruction, LocationSummary::kCallOnMainOnly);
5030 InvokeRuntimeCallingConventionARMVIXL calling_convention;
5031 locations->SetInAt(0, LocationFrom(calling_convention.GetRegisterAt(0)));
5032 locations->SetOut(LocationFrom(r0));
5033 }
5034
VisitNewInstance(HNewInstance * instruction)5035 void InstructionCodeGeneratorARMVIXL::VisitNewInstance(HNewInstance* instruction) {
5036 codegen_->InvokeRuntime(instruction->GetEntrypoint(), instruction, instruction->GetDexPc());
5037 CheckEntrypointTypes<kQuickAllocObjectWithChecks, void*, mirror::Class*>();
5038 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 11);
5039 }
5040
VisitNewArray(HNewArray * instruction)5041 void LocationsBuilderARMVIXL::VisitNewArray(HNewArray* instruction) {
5042 LocationSummary* locations = new (GetGraph()->GetAllocator()) LocationSummary(
5043 instruction, LocationSummary::kCallOnMainOnly);
5044 InvokeRuntimeCallingConventionARMVIXL calling_convention;
5045 locations->SetOut(LocationFrom(r0));
5046 locations->SetInAt(0, LocationFrom(calling_convention.GetRegisterAt(0)));
5047 locations->SetInAt(1, LocationFrom(calling_convention.GetRegisterAt(1)));
5048 }
5049
VisitNewArray(HNewArray * instruction)5050 void InstructionCodeGeneratorARMVIXL::VisitNewArray(HNewArray* instruction) {
5051 // Note: if heap poisoning is enabled, the entry point takes care of poisoning the reference.
5052 QuickEntrypointEnum entrypoint = CodeGenerator::GetArrayAllocationEntrypoint(instruction);
5053 codegen_->InvokeRuntime(entrypoint, instruction, instruction->GetDexPc());
5054 CheckEntrypointTypes<kQuickAllocArrayResolved, void*, mirror::Class*, int32_t>();
5055 DCHECK(!codegen_->IsLeafMethod());
5056 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 12);
5057 }
5058
VisitParameterValue(HParameterValue * instruction)5059 void LocationsBuilderARMVIXL::VisitParameterValue(HParameterValue* instruction) {
5060 LocationSummary* locations =
5061 new (GetGraph()->GetAllocator()) LocationSummary(instruction, LocationSummary::kNoCall);
5062 Location location = parameter_visitor_.GetNextLocation(instruction->GetType());
5063 if (location.IsStackSlot()) {
5064 location = Location::StackSlot(location.GetStackIndex() + codegen_->GetFrameSize());
5065 } else if (location.IsDoubleStackSlot()) {
5066 location = Location::DoubleStackSlot(location.GetStackIndex() + codegen_->GetFrameSize());
5067 }
5068 locations->SetOut(location);
5069 }
5070
VisitParameterValue(HParameterValue * instruction ATTRIBUTE_UNUSED)5071 void InstructionCodeGeneratorARMVIXL::VisitParameterValue(
5072 HParameterValue* instruction ATTRIBUTE_UNUSED) {
5073 // Nothing to do, the parameter is already at its location.
5074 }
5075
VisitCurrentMethod(HCurrentMethod * instruction)5076 void LocationsBuilderARMVIXL::VisitCurrentMethod(HCurrentMethod* instruction) {
5077 LocationSummary* locations =
5078 new (GetGraph()->GetAllocator()) LocationSummary(instruction, LocationSummary::kNoCall);
5079 locations->SetOut(LocationFrom(kMethodRegister));
5080 }
5081
VisitCurrentMethod(HCurrentMethod * instruction ATTRIBUTE_UNUSED)5082 void InstructionCodeGeneratorARMVIXL::VisitCurrentMethod(
5083 HCurrentMethod* instruction ATTRIBUTE_UNUSED) {
5084 // Nothing to do, the method is already at its location.
5085 }
5086
VisitNot(HNot * not_)5087 void LocationsBuilderARMVIXL::VisitNot(HNot* not_) {
5088 LocationSummary* locations =
5089 new (GetGraph()->GetAllocator()) LocationSummary(not_, LocationSummary::kNoCall);
5090 locations->SetInAt(0, Location::RequiresRegister());
5091 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
5092 }
5093
VisitNot(HNot * not_)5094 void InstructionCodeGeneratorARMVIXL::VisitNot(HNot* not_) {
5095 LocationSummary* locations = not_->GetLocations();
5096 Location out = locations->Out();
5097 Location in = locations->InAt(0);
5098 switch (not_->GetResultType()) {
5099 case DataType::Type::kInt32:
5100 __ Mvn(OutputRegister(not_), InputRegisterAt(not_, 0));
5101 break;
5102
5103 case DataType::Type::kInt64:
5104 __ Mvn(LowRegisterFrom(out), LowRegisterFrom(in));
5105 __ Mvn(HighRegisterFrom(out), HighRegisterFrom(in));
5106 break;
5107
5108 default:
5109 LOG(FATAL) << "Unimplemented type for not operation " << not_->GetResultType();
5110 }
5111 }
5112
VisitBooleanNot(HBooleanNot * bool_not)5113 void LocationsBuilderARMVIXL::VisitBooleanNot(HBooleanNot* bool_not) {
5114 LocationSummary* locations =
5115 new (GetGraph()->GetAllocator()) LocationSummary(bool_not, LocationSummary::kNoCall);
5116 locations->SetInAt(0, Location::RequiresRegister());
5117 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
5118 }
5119
VisitBooleanNot(HBooleanNot * bool_not)5120 void InstructionCodeGeneratorARMVIXL::VisitBooleanNot(HBooleanNot* bool_not) {
5121 __ Eor(OutputRegister(bool_not), InputRegister(bool_not), 1);
5122 }
5123
VisitCompare(HCompare * compare)5124 void LocationsBuilderARMVIXL::VisitCompare(HCompare* compare) {
5125 LocationSummary* locations =
5126 new (GetGraph()->GetAllocator()) LocationSummary(compare, LocationSummary::kNoCall);
5127 switch (compare->InputAt(0)->GetType()) {
5128 case DataType::Type::kBool:
5129 case DataType::Type::kUint8:
5130 case DataType::Type::kInt8:
5131 case DataType::Type::kUint16:
5132 case DataType::Type::kInt16:
5133 case DataType::Type::kInt32:
5134 case DataType::Type::kInt64: {
5135 locations->SetInAt(0, Location::RequiresRegister());
5136 locations->SetInAt(1, Location::RequiresRegister());
5137 // Output overlaps because it is written before doing the low comparison.
5138 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
5139 break;
5140 }
5141 case DataType::Type::kFloat32:
5142 case DataType::Type::kFloat64: {
5143 locations->SetInAt(0, Location::RequiresFpuRegister());
5144 locations->SetInAt(1, ArithmeticZeroOrFpuRegister(compare->InputAt(1)));
5145 locations->SetOut(Location::RequiresRegister());
5146 break;
5147 }
5148 default:
5149 LOG(FATAL) << "Unexpected type for compare operation " << compare->InputAt(0)->GetType();
5150 }
5151 }
5152
VisitCompare(HCompare * compare)5153 void InstructionCodeGeneratorARMVIXL::VisitCompare(HCompare* compare) {
5154 LocationSummary* locations = compare->GetLocations();
5155 vixl32::Register out = OutputRegister(compare);
5156 Location left = locations->InAt(0);
5157 Location right = locations->InAt(1);
5158
5159 vixl32::Label less, greater, done;
5160 vixl32::Label* final_label = codegen_->GetFinalLabel(compare, &done);
5161 DataType::Type type = compare->InputAt(0)->GetType();
5162 vixl32::Condition less_cond = vixl32::Condition::None();
5163 switch (type) {
5164 case DataType::Type::kBool:
5165 case DataType::Type::kUint8:
5166 case DataType::Type::kInt8:
5167 case DataType::Type::kUint16:
5168 case DataType::Type::kInt16:
5169 case DataType::Type::kInt32: {
5170 // Emit move to `out` before the `Cmp`, as `Mov` might affect the status flags.
5171 __ Mov(out, 0);
5172 __ Cmp(RegisterFrom(left), RegisterFrom(right)); // Signed compare.
5173 less_cond = lt;
5174 break;
5175 }
5176 case DataType::Type::kInt64: {
5177 __ Cmp(HighRegisterFrom(left), HighRegisterFrom(right)); // Signed compare.
5178 __ B(lt, &less, /* is_far_target= */ false);
5179 __ B(gt, &greater, /* is_far_target= */ false);
5180 // Emit move to `out` before the last `Cmp`, as `Mov` might affect the status flags.
5181 __ Mov(out, 0);
5182 __ Cmp(LowRegisterFrom(left), LowRegisterFrom(right)); // Unsigned compare.
5183 less_cond = lo;
5184 break;
5185 }
5186 case DataType::Type::kFloat32:
5187 case DataType::Type::kFloat64: {
5188 __ Mov(out, 0);
5189 GenerateVcmp(compare, codegen_);
5190 // To branch on the FP compare result we transfer FPSCR to APSR (encoded as PC in VMRS).
5191 __ Vmrs(RegisterOrAPSR_nzcv(kPcCode), FPSCR);
5192 less_cond = ARMFPCondition(kCondLT, compare->IsGtBias());
5193 break;
5194 }
5195 default:
5196 LOG(FATAL) << "Unexpected compare type " << type;
5197 UNREACHABLE();
5198 }
5199
5200 __ B(eq, final_label, /* is_far_target= */ false);
5201 __ B(less_cond, &less, /* is_far_target= */ false);
5202
5203 __ Bind(&greater);
5204 __ Mov(out, 1);
5205 __ B(final_label);
5206
5207 __ Bind(&less);
5208 __ Mov(out, -1);
5209
5210 if (done.IsReferenced()) {
5211 __ Bind(&done);
5212 }
5213 }
5214
VisitPhi(HPhi * instruction)5215 void LocationsBuilderARMVIXL::VisitPhi(HPhi* instruction) {
5216 LocationSummary* locations =
5217 new (GetGraph()->GetAllocator()) LocationSummary(instruction, LocationSummary::kNoCall);
5218 for (size_t i = 0, e = locations->GetInputCount(); i < e; ++i) {
5219 locations->SetInAt(i, Location::Any());
5220 }
5221 locations->SetOut(Location::Any());
5222 }
5223
VisitPhi(HPhi * instruction ATTRIBUTE_UNUSED)5224 void InstructionCodeGeneratorARMVIXL::VisitPhi(HPhi* instruction ATTRIBUTE_UNUSED) {
5225 LOG(FATAL) << "Unreachable";
5226 }
5227
GenerateMemoryBarrier(MemBarrierKind kind)5228 void CodeGeneratorARMVIXL::GenerateMemoryBarrier(MemBarrierKind kind) {
5229 // TODO (ported from quick): revisit ARM barrier kinds.
5230 DmbOptions flavor = DmbOptions::ISH; // Quiet C++ warnings.
5231 switch (kind) {
5232 case MemBarrierKind::kAnyStore:
5233 case MemBarrierKind::kLoadAny:
5234 case MemBarrierKind::kAnyAny: {
5235 flavor = DmbOptions::ISH;
5236 break;
5237 }
5238 case MemBarrierKind::kStoreStore: {
5239 flavor = DmbOptions::ISHST;
5240 break;
5241 }
5242 default:
5243 LOG(FATAL) << "Unexpected memory barrier " << kind;
5244 }
5245 __ Dmb(flavor);
5246 }
5247
GenerateWideAtomicLoad(vixl32::Register addr,uint32_t offset,vixl32::Register out_lo,vixl32::Register out_hi)5248 void InstructionCodeGeneratorARMVIXL::GenerateWideAtomicLoad(vixl32::Register addr,
5249 uint32_t offset,
5250 vixl32::Register out_lo,
5251 vixl32::Register out_hi) {
5252 UseScratchRegisterScope temps(GetVIXLAssembler());
5253 if (offset != 0) {
5254 vixl32::Register temp = temps.Acquire();
5255 __ Add(temp, addr, offset);
5256 addr = temp;
5257 }
5258 __ Ldrexd(out_lo, out_hi, MemOperand(addr));
5259 }
5260
GenerateWideAtomicStore(vixl32::Register addr,uint32_t offset,vixl32::Register value_lo,vixl32::Register value_hi,vixl32::Register temp1,vixl32::Register temp2,HInstruction * instruction)5261 void InstructionCodeGeneratorARMVIXL::GenerateWideAtomicStore(vixl32::Register addr,
5262 uint32_t offset,
5263 vixl32::Register value_lo,
5264 vixl32::Register value_hi,
5265 vixl32::Register temp1,
5266 vixl32::Register temp2,
5267 HInstruction* instruction) {
5268 UseScratchRegisterScope temps(GetVIXLAssembler());
5269 vixl32::Label fail;
5270 if (offset != 0) {
5271 vixl32::Register temp = temps.Acquire();
5272 __ Add(temp, addr, offset);
5273 addr = temp;
5274 }
5275 __ Bind(&fail);
5276 {
5277 // Ensure the pc position is recorded immediately after the `ldrexd` instruction.
5278 ExactAssemblyScope aas(GetVIXLAssembler(),
5279 vixl32::kMaxInstructionSizeInBytes,
5280 CodeBufferCheckScope::kMaximumSize);
5281 // We need a load followed by store. (The address used in a STREX instruction must
5282 // be the same as the address in the most recently executed LDREX instruction.)
5283 __ ldrexd(temp1, temp2, MemOperand(addr));
5284 codegen_->MaybeRecordImplicitNullCheck(instruction);
5285 }
5286 __ Strexd(temp1, value_lo, value_hi, MemOperand(addr));
5287 __ CompareAndBranchIfNonZero(temp1, &fail);
5288 }
5289
HandleFieldSet(HInstruction * instruction,const FieldInfo & field_info)5290 void LocationsBuilderARMVIXL::HandleFieldSet(
5291 HInstruction* instruction, const FieldInfo& field_info) {
5292 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet());
5293
5294 LocationSummary* locations =
5295 new (GetGraph()->GetAllocator()) LocationSummary(instruction, LocationSummary::kNoCall);
5296 locations->SetInAt(0, Location::RequiresRegister());
5297
5298 DataType::Type field_type = field_info.GetFieldType();
5299 if (DataType::IsFloatingPointType(field_type)) {
5300 locations->SetInAt(1, Location::RequiresFpuRegister());
5301 } else {
5302 locations->SetInAt(1, Location::RequiresRegister());
5303 }
5304
5305 bool is_wide = field_type == DataType::Type::kInt64 || field_type == DataType::Type::kFloat64;
5306 bool generate_volatile = field_info.IsVolatile()
5307 && is_wide
5308 && !codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd();
5309 bool needs_write_barrier =
5310 CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1));
5311 // Temporary registers for the write barrier.
5312 // TODO: consider renaming StoreNeedsWriteBarrier to StoreNeedsGCMark.
5313 if (needs_write_barrier) {
5314 locations->AddTemp(Location::RequiresRegister()); // Possibly used for reference poisoning too.
5315 locations->AddTemp(Location::RequiresRegister());
5316 } else if (generate_volatile) {
5317 // ARM encoding have some additional constraints for ldrexd/strexd:
5318 // - registers need to be consecutive
5319 // - the first register should be even but not R14.
5320 // We don't test for ARM yet, and the assertion makes sure that we
5321 // revisit this if we ever enable ARM encoding.
5322 DCHECK_EQ(InstructionSet::kThumb2, codegen_->GetInstructionSet());
5323
5324 locations->AddTemp(Location::RequiresRegister());
5325 locations->AddTemp(Location::RequiresRegister());
5326 if (field_type == DataType::Type::kFloat64) {
5327 // For doubles we need two more registers to copy the value.
5328 locations->AddTemp(LocationFrom(r2));
5329 locations->AddTemp(LocationFrom(r3));
5330 }
5331 }
5332 }
5333
HandleFieldSet(HInstruction * instruction,const FieldInfo & field_info,bool value_can_be_null)5334 void InstructionCodeGeneratorARMVIXL::HandleFieldSet(HInstruction* instruction,
5335 const FieldInfo& field_info,
5336 bool value_can_be_null) {
5337 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet());
5338
5339 LocationSummary* locations = instruction->GetLocations();
5340 vixl32::Register base = InputRegisterAt(instruction, 0);
5341 Location value = locations->InAt(1);
5342
5343 bool is_volatile = field_info.IsVolatile();
5344 bool atomic_ldrd_strd = codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd();
5345 DataType::Type field_type = field_info.GetFieldType();
5346 uint32_t offset = field_info.GetFieldOffset().Uint32Value();
5347 bool needs_write_barrier =
5348 CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1));
5349
5350 if (is_volatile) {
5351 codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyStore);
5352 }
5353
5354 switch (field_type) {
5355 case DataType::Type::kBool:
5356 case DataType::Type::kUint8:
5357 case DataType::Type::kInt8:
5358 case DataType::Type::kUint16:
5359 case DataType::Type::kInt16:
5360 case DataType::Type::kInt32: {
5361 StoreOperandType operand_type = GetStoreOperandType(field_type);
5362 GetAssembler()->StoreToOffset(operand_type, RegisterFrom(value), base, offset);
5363 break;
5364 }
5365
5366 case DataType::Type::kReference: {
5367 if (kPoisonHeapReferences && needs_write_barrier) {
5368 // Note that in the case where `value` is a null reference,
5369 // we do not enter this block, as a null reference does not
5370 // need poisoning.
5371 DCHECK_EQ(field_type, DataType::Type::kReference);
5372 vixl32::Register temp = RegisterFrom(locations->GetTemp(0));
5373 __ Mov(temp, RegisterFrom(value));
5374 GetAssembler()->PoisonHeapReference(temp);
5375 GetAssembler()->StoreToOffset(kStoreWord, temp, base, offset);
5376 } else {
5377 GetAssembler()->StoreToOffset(kStoreWord, RegisterFrom(value), base, offset);
5378 }
5379 break;
5380 }
5381
5382 case DataType::Type::kInt64: {
5383 if (is_volatile && !atomic_ldrd_strd) {
5384 GenerateWideAtomicStore(base,
5385 offset,
5386 LowRegisterFrom(value),
5387 HighRegisterFrom(value),
5388 RegisterFrom(locations->GetTemp(0)),
5389 RegisterFrom(locations->GetTemp(1)),
5390 instruction);
5391 } else {
5392 GetAssembler()->StoreToOffset(kStoreWordPair, LowRegisterFrom(value), base, offset);
5393 codegen_->MaybeRecordImplicitNullCheck(instruction);
5394 }
5395 break;
5396 }
5397
5398 case DataType::Type::kFloat32: {
5399 GetAssembler()->StoreSToOffset(SRegisterFrom(value), base, offset);
5400 break;
5401 }
5402
5403 case DataType::Type::kFloat64: {
5404 vixl32::DRegister value_reg = DRegisterFrom(value);
5405 if (is_volatile && !atomic_ldrd_strd) {
5406 vixl32::Register value_reg_lo = RegisterFrom(locations->GetTemp(0));
5407 vixl32::Register value_reg_hi = RegisterFrom(locations->GetTemp(1));
5408
5409 __ Vmov(value_reg_lo, value_reg_hi, value_reg);
5410
5411 GenerateWideAtomicStore(base,
5412 offset,
5413 value_reg_lo,
5414 value_reg_hi,
5415 RegisterFrom(locations->GetTemp(2)),
5416 RegisterFrom(locations->GetTemp(3)),
5417 instruction);
5418 } else {
5419 GetAssembler()->StoreDToOffset(value_reg, base, offset);
5420 codegen_->MaybeRecordImplicitNullCheck(instruction);
5421 }
5422 break;
5423 }
5424
5425 case DataType::Type::kUint32:
5426 case DataType::Type::kUint64:
5427 case DataType::Type::kVoid:
5428 LOG(FATAL) << "Unreachable type " << field_type;
5429 UNREACHABLE();
5430 }
5431
5432 // Longs and doubles are handled in the switch.
5433 if (field_type != DataType::Type::kInt64 && field_type != DataType::Type::kFloat64) {
5434 // TODO(VIXL): Here and for other calls to `MaybeRecordImplicitNullCheck` in this method, we
5435 // should use a scope and the assembler to emit the store instruction to guarantee that we
5436 // record the pc at the correct position. But the `Assembler` does not automatically handle
5437 // unencodable offsets. Practically, everything is fine because the helper and VIXL, at the time
5438 // of writing, do generate the store instruction last.
5439 codegen_->MaybeRecordImplicitNullCheck(instruction);
5440 }
5441
5442 if (CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1))) {
5443 vixl32::Register temp = RegisterFrom(locations->GetTemp(0));
5444 vixl32::Register card = RegisterFrom(locations->GetTemp(1));
5445 codegen_->MarkGCCard(temp, card, base, RegisterFrom(value), value_can_be_null);
5446 }
5447
5448 if (is_volatile) {
5449 codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyAny);
5450 }
5451 }
5452
HandleFieldGet(HInstruction * instruction,const FieldInfo & field_info)5453 void LocationsBuilderARMVIXL::HandleFieldGet(HInstruction* instruction,
5454 const FieldInfo& field_info) {
5455 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet());
5456
5457 bool object_field_get_with_read_barrier =
5458 kEmitCompilerReadBarrier && (field_info.GetFieldType() == DataType::Type::kReference);
5459 LocationSummary* locations =
5460 new (GetGraph()->GetAllocator()) LocationSummary(instruction,
5461 object_field_get_with_read_barrier
5462 ? LocationSummary::kCallOnSlowPath
5463 : LocationSummary::kNoCall);
5464 if (object_field_get_with_read_barrier && kUseBakerReadBarrier) {
5465 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers.
5466 }
5467 locations->SetInAt(0, Location::RequiresRegister());
5468
5469 bool volatile_for_double = field_info.IsVolatile()
5470 && (field_info.GetFieldType() == DataType::Type::kFloat64)
5471 && !codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd();
5472 // The output overlaps in case of volatile long: we don't want the
5473 // code generated by GenerateWideAtomicLoad to overwrite the
5474 // object's location. Likewise, in the case of an object field get
5475 // with read barriers enabled, we do not want the load to overwrite
5476 // the object's location, as we need it to emit the read barrier.
5477 bool overlap =
5478 (field_info.IsVolatile() && (field_info.GetFieldType() == DataType::Type::kInt64)) ||
5479 object_field_get_with_read_barrier;
5480
5481 if (DataType::IsFloatingPointType(instruction->GetType())) {
5482 locations->SetOut(Location::RequiresFpuRegister());
5483 } else {
5484 locations->SetOut(Location::RequiresRegister(),
5485 (overlap ? Location::kOutputOverlap : Location::kNoOutputOverlap));
5486 }
5487 if (volatile_for_double) {
5488 // ARM encoding have some additional constraints for ldrexd/strexd:
5489 // - registers need to be consecutive
5490 // - the first register should be even but not R14.
5491 // We don't test for ARM yet, and the assertion makes sure that we
5492 // revisit this if we ever enable ARM encoding.
5493 DCHECK_EQ(InstructionSet::kThumb2, codegen_->GetInstructionSet());
5494 locations->AddTemp(Location::RequiresRegister());
5495 locations->AddTemp(Location::RequiresRegister());
5496 } else if (object_field_get_with_read_barrier && kUseBakerReadBarrier) {
5497 // We need a temporary register for the read barrier load in
5498 // CodeGeneratorARMVIXL::GenerateFieldLoadWithBakerReadBarrier()
5499 // only if the offset is too big.
5500 if (field_info.GetFieldOffset().Uint32Value() >= kReferenceLoadMinFarOffset) {
5501 locations->AddTemp(Location::RequiresRegister());
5502 }
5503 }
5504 }
5505
ArithmeticZeroOrFpuRegister(HInstruction * input)5506 Location LocationsBuilderARMVIXL::ArithmeticZeroOrFpuRegister(HInstruction* input) {
5507 DCHECK(DataType::IsFloatingPointType(input->GetType())) << input->GetType();
5508 if ((input->IsFloatConstant() && (input->AsFloatConstant()->IsArithmeticZero())) ||
5509 (input->IsDoubleConstant() && (input->AsDoubleConstant()->IsArithmeticZero()))) {
5510 return Location::ConstantLocation(input->AsConstant());
5511 } else {
5512 return Location::RequiresFpuRegister();
5513 }
5514 }
5515
ArmEncodableConstantOrRegister(HInstruction * constant,Opcode opcode)5516 Location LocationsBuilderARMVIXL::ArmEncodableConstantOrRegister(HInstruction* constant,
5517 Opcode opcode) {
5518 DCHECK(!DataType::IsFloatingPointType(constant->GetType()));
5519 if (constant->IsConstant() &&
5520 CanEncodeConstantAsImmediate(constant->AsConstant(), opcode)) {
5521 return Location::ConstantLocation(constant->AsConstant());
5522 }
5523 return Location::RequiresRegister();
5524 }
5525
CanEncode32BitConstantAsImmediate(CodeGeneratorARMVIXL * codegen,uint32_t value,Opcode opcode,vixl32::FlagsUpdate flags_update=vixl32::FlagsUpdate::DontCare)5526 static bool CanEncode32BitConstantAsImmediate(
5527 CodeGeneratorARMVIXL* codegen,
5528 uint32_t value,
5529 Opcode opcode,
5530 vixl32::FlagsUpdate flags_update = vixl32::FlagsUpdate::DontCare) {
5531 ArmVIXLAssembler* assembler = codegen->GetAssembler();
5532 if (assembler->ShifterOperandCanHold(opcode, value, flags_update)) {
5533 return true;
5534 }
5535 Opcode neg_opcode = kNoOperand;
5536 uint32_t neg_value = 0;
5537 switch (opcode) {
5538 case AND: neg_opcode = BIC; neg_value = ~value; break;
5539 case ORR: neg_opcode = ORN; neg_value = ~value; break;
5540 case ADD: neg_opcode = SUB; neg_value = -value; break;
5541 case ADC: neg_opcode = SBC; neg_value = ~value; break;
5542 case SUB: neg_opcode = ADD; neg_value = -value; break;
5543 case SBC: neg_opcode = ADC; neg_value = ~value; break;
5544 case MOV: neg_opcode = MVN; neg_value = ~value; break;
5545 default:
5546 return false;
5547 }
5548
5549 if (assembler->ShifterOperandCanHold(neg_opcode, neg_value, flags_update)) {
5550 return true;
5551 }
5552
5553 return opcode == AND && IsPowerOfTwo(value + 1);
5554 }
5555
CanEncodeConstantAsImmediate(HConstant * input_cst,Opcode opcode)5556 bool LocationsBuilderARMVIXL::CanEncodeConstantAsImmediate(HConstant* input_cst, Opcode opcode) {
5557 uint64_t value = static_cast<uint64_t>(Int64FromConstant(input_cst));
5558 if (DataType::Is64BitType(input_cst->GetType())) {
5559 Opcode high_opcode = opcode;
5560 vixl32::FlagsUpdate low_flags_update = vixl32::FlagsUpdate::DontCare;
5561 switch (opcode) {
5562 case SUB:
5563 // Flip the operation to an ADD.
5564 value = -value;
5565 opcode = ADD;
5566 FALLTHROUGH_INTENDED;
5567 case ADD:
5568 if (Low32Bits(value) == 0u) {
5569 return CanEncode32BitConstantAsImmediate(codegen_, High32Bits(value), opcode);
5570 }
5571 high_opcode = ADC;
5572 low_flags_update = vixl32::FlagsUpdate::SetFlags;
5573 break;
5574 default:
5575 break;
5576 }
5577 return CanEncode32BitConstantAsImmediate(codegen_, High32Bits(value), high_opcode) &&
5578 CanEncode32BitConstantAsImmediate(codegen_, Low32Bits(value), opcode, low_flags_update);
5579 } else {
5580 return CanEncode32BitConstantAsImmediate(codegen_, Low32Bits(value), opcode);
5581 }
5582 }
5583
HandleFieldGet(HInstruction * instruction,const FieldInfo & field_info)5584 void InstructionCodeGeneratorARMVIXL::HandleFieldGet(HInstruction* instruction,
5585 const FieldInfo& field_info) {
5586 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet());
5587
5588 LocationSummary* locations = instruction->GetLocations();
5589 vixl32::Register base = InputRegisterAt(instruction, 0);
5590 Location out = locations->Out();
5591 bool is_volatile = field_info.IsVolatile();
5592 bool atomic_ldrd_strd = codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd();
5593 DCHECK_EQ(DataType::Size(field_info.GetFieldType()), DataType::Size(instruction->GetType()));
5594 DataType::Type load_type = instruction->GetType();
5595 uint32_t offset = field_info.GetFieldOffset().Uint32Value();
5596
5597 switch (load_type) {
5598 case DataType::Type::kBool:
5599 case DataType::Type::kUint8:
5600 case DataType::Type::kInt8:
5601 case DataType::Type::kUint16:
5602 case DataType::Type::kInt16:
5603 case DataType::Type::kInt32: {
5604 LoadOperandType operand_type = GetLoadOperandType(load_type);
5605 GetAssembler()->LoadFromOffset(operand_type, RegisterFrom(out), base, offset);
5606 break;
5607 }
5608
5609 case DataType::Type::kReference: {
5610 // /* HeapReference<Object> */ out = *(base + offset)
5611 if (kEmitCompilerReadBarrier && kUseBakerReadBarrier) {
5612 Location maybe_temp = (locations->GetTempCount() != 0) ? locations->GetTemp(0) : Location();
5613 // Note that a potential implicit null check is handled in this
5614 // CodeGeneratorARMVIXL::GenerateFieldLoadWithBakerReadBarrier call.
5615 codegen_->GenerateFieldLoadWithBakerReadBarrier(
5616 instruction, out, base, offset, maybe_temp, /* needs_null_check= */ true);
5617 if (is_volatile) {
5618 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny);
5619 }
5620 } else {
5621 GetAssembler()->LoadFromOffset(kLoadWord, RegisterFrom(out), base, offset);
5622 codegen_->MaybeRecordImplicitNullCheck(instruction);
5623 if (is_volatile) {
5624 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny);
5625 }
5626 // If read barriers are enabled, emit read barriers other than
5627 // Baker's using a slow path (and also unpoison the loaded
5628 // reference, if heap poisoning is enabled).
5629 codegen_->MaybeGenerateReadBarrierSlow(instruction, out, out, locations->InAt(0), offset);
5630 }
5631 break;
5632 }
5633
5634 case DataType::Type::kInt64:
5635 if (is_volatile && !atomic_ldrd_strd) {
5636 GenerateWideAtomicLoad(base, offset, LowRegisterFrom(out), HighRegisterFrom(out));
5637 } else {
5638 GetAssembler()->LoadFromOffset(kLoadWordPair, LowRegisterFrom(out), base, offset);
5639 }
5640 break;
5641
5642 case DataType::Type::kFloat32:
5643 GetAssembler()->LoadSFromOffset(SRegisterFrom(out), base, offset);
5644 break;
5645
5646 case DataType::Type::kFloat64: {
5647 vixl32::DRegister out_dreg = DRegisterFrom(out);
5648 if (is_volatile && !atomic_ldrd_strd) {
5649 vixl32::Register lo = RegisterFrom(locations->GetTemp(0));
5650 vixl32::Register hi = RegisterFrom(locations->GetTemp(1));
5651 GenerateWideAtomicLoad(base, offset, lo, hi);
5652 // TODO(VIXL): Do we need to be immediately after the ldrexd instruction? If so we need a
5653 // scope.
5654 codegen_->MaybeRecordImplicitNullCheck(instruction);
5655 __ Vmov(out_dreg, lo, hi);
5656 } else {
5657 GetAssembler()->LoadDFromOffset(out_dreg, base, offset);
5658 codegen_->MaybeRecordImplicitNullCheck(instruction);
5659 }
5660 break;
5661 }
5662
5663 case DataType::Type::kUint32:
5664 case DataType::Type::kUint64:
5665 case DataType::Type::kVoid:
5666 LOG(FATAL) << "Unreachable type " << load_type;
5667 UNREACHABLE();
5668 }
5669
5670 if (load_type == DataType::Type::kReference || load_type == DataType::Type::kFloat64) {
5671 // Potential implicit null checks, in the case of reference or
5672 // double fields, are handled in the previous switch statement.
5673 } else {
5674 // Address cases other than reference and double that may require an implicit null check.
5675 // TODO(VIXL): Here and for other calls to `MaybeRecordImplicitNullCheck` in this method, we
5676 // should use a scope and the assembler to emit the load instruction to guarantee that we
5677 // record the pc at the correct position. But the `Assembler` does not automatically handle
5678 // unencodable offsets. Practically, everything is fine because the helper and VIXL, at the time
5679 // of writing, do generate the store instruction last.
5680 codegen_->MaybeRecordImplicitNullCheck(instruction);
5681 }
5682
5683 if (is_volatile) {
5684 if (load_type == DataType::Type::kReference) {
5685 // Memory barriers, in the case of references, are also handled
5686 // in the previous switch statement.
5687 } else {
5688 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny);
5689 }
5690 }
5691 }
5692
VisitInstanceFieldSet(HInstanceFieldSet * instruction)5693 void LocationsBuilderARMVIXL::VisitInstanceFieldSet(HInstanceFieldSet* instruction) {
5694 HandleFieldSet(instruction, instruction->GetFieldInfo());
5695 }
5696
VisitInstanceFieldSet(HInstanceFieldSet * instruction)5697 void InstructionCodeGeneratorARMVIXL::VisitInstanceFieldSet(HInstanceFieldSet* instruction) {
5698 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull());
5699 }
5700
VisitInstanceFieldGet(HInstanceFieldGet * instruction)5701 void LocationsBuilderARMVIXL::VisitInstanceFieldGet(HInstanceFieldGet* instruction) {
5702 HandleFieldGet(instruction, instruction->GetFieldInfo());
5703 }
5704
VisitInstanceFieldGet(HInstanceFieldGet * instruction)5705 void InstructionCodeGeneratorARMVIXL::VisitInstanceFieldGet(HInstanceFieldGet* instruction) {
5706 HandleFieldGet(instruction, instruction->GetFieldInfo());
5707 }
5708
VisitStaticFieldGet(HStaticFieldGet * instruction)5709 void LocationsBuilderARMVIXL::VisitStaticFieldGet(HStaticFieldGet* instruction) {
5710 HandleFieldGet(instruction, instruction->GetFieldInfo());
5711 }
5712
VisitStaticFieldGet(HStaticFieldGet * instruction)5713 void InstructionCodeGeneratorARMVIXL::VisitStaticFieldGet(HStaticFieldGet* instruction) {
5714 HandleFieldGet(instruction, instruction->GetFieldInfo());
5715 }
5716
VisitStaticFieldSet(HStaticFieldSet * instruction)5717 void LocationsBuilderARMVIXL::VisitStaticFieldSet(HStaticFieldSet* instruction) {
5718 HandleFieldSet(instruction, instruction->GetFieldInfo());
5719 }
5720
VisitStaticFieldSet(HStaticFieldSet * instruction)5721 void InstructionCodeGeneratorARMVIXL::VisitStaticFieldSet(HStaticFieldSet* instruction) {
5722 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull());
5723 }
5724
VisitUnresolvedInstanceFieldGet(HUnresolvedInstanceFieldGet * instruction)5725 void LocationsBuilderARMVIXL::VisitUnresolvedInstanceFieldGet(
5726 HUnresolvedInstanceFieldGet* instruction) {
5727 FieldAccessCallingConventionARMVIXL calling_convention;
5728 codegen_->CreateUnresolvedFieldLocationSummary(
5729 instruction, instruction->GetFieldType(), calling_convention);
5730 }
5731
VisitUnresolvedInstanceFieldGet(HUnresolvedInstanceFieldGet * instruction)5732 void InstructionCodeGeneratorARMVIXL::VisitUnresolvedInstanceFieldGet(
5733 HUnresolvedInstanceFieldGet* instruction) {
5734 FieldAccessCallingConventionARMVIXL calling_convention;
5735 codegen_->GenerateUnresolvedFieldAccess(instruction,
5736 instruction->GetFieldType(),
5737 instruction->GetFieldIndex(),
5738 instruction->GetDexPc(),
5739 calling_convention);
5740 }
5741
VisitUnresolvedInstanceFieldSet(HUnresolvedInstanceFieldSet * instruction)5742 void LocationsBuilderARMVIXL::VisitUnresolvedInstanceFieldSet(
5743 HUnresolvedInstanceFieldSet* instruction) {
5744 FieldAccessCallingConventionARMVIXL calling_convention;
5745 codegen_->CreateUnresolvedFieldLocationSummary(
5746 instruction, instruction->GetFieldType(), calling_convention);
5747 }
5748
VisitUnresolvedInstanceFieldSet(HUnresolvedInstanceFieldSet * instruction)5749 void InstructionCodeGeneratorARMVIXL::VisitUnresolvedInstanceFieldSet(
5750 HUnresolvedInstanceFieldSet* instruction) {
5751 FieldAccessCallingConventionARMVIXL calling_convention;
5752 codegen_->GenerateUnresolvedFieldAccess(instruction,
5753 instruction->GetFieldType(),
5754 instruction->GetFieldIndex(),
5755 instruction->GetDexPc(),
5756 calling_convention);
5757 }
5758
VisitUnresolvedStaticFieldGet(HUnresolvedStaticFieldGet * instruction)5759 void LocationsBuilderARMVIXL::VisitUnresolvedStaticFieldGet(
5760 HUnresolvedStaticFieldGet* instruction) {
5761 FieldAccessCallingConventionARMVIXL calling_convention;
5762 codegen_->CreateUnresolvedFieldLocationSummary(
5763 instruction, instruction->GetFieldType(), calling_convention);
5764 }
5765
VisitUnresolvedStaticFieldGet(HUnresolvedStaticFieldGet * instruction)5766 void InstructionCodeGeneratorARMVIXL::VisitUnresolvedStaticFieldGet(
5767 HUnresolvedStaticFieldGet* instruction) {
5768 FieldAccessCallingConventionARMVIXL calling_convention;
5769 codegen_->GenerateUnresolvedFieldAccess(instruction,
5770 instruction->GetFieldType(),
5771 instruction->GetFieldIndex(),
5772 instruction->GetDexPc(),
5773 calling_convention);
5774 }
5775
VisitUnresolvedStaticFieldSet(HUnresolvedStaticFieldSet * instruction)5776 void LocationsBuilderARMVIXL::VisitUnresolvedStaticFieldSet(
5777 HUnresolvedStaticFieldSet* instruction) {
5778 FieldAccessCallingConventionARMVIXL calling_convention;
5779 codegen_->CreateUnresolvedFieldLocationSummary(
5780 instruction, instruction->GetFieldType(), calling_convention);
5781 }
5782
VisitUnresolvedStaticFieldSet(HUnresolvedStaticFieldSet * instruction)5783 void InstructionCodeGeneratorARMVIXL::VisitUnresolvedStaticFieldSet(
5784 HUnresolvedStaticFieldSet* instruction) {
5785 FieldAccessCallingConventionARMVIXL calling_convention;
5786 codegen_->GenerateUnresolvedFieldAccess(instruction,
5787 instruction->GetFieldType(),
5788 instruction->GetFieldIndex(),
5789 instruction->GetDexPc(),
5790 calling_convention);
5791 }
5792
VisitNullCheck(HNullCheck * instruction)5793 void LocationsBuilderARMVIXL::VisitNullCheck(HNullCheck* instruction) {
5794 LocationSummary* locations = codegen_->CreateThrowingSlowPathLocations(instruction);
5795 locations->SetInAt(0, Location::RequiresRegister());
5796 }
5797
GenerateImplicitNullCheck(HNullCheck * instruction)5798 void CodeGeneratorARMVIXL::GenerateImplicitNullCheck(HNullCheck* instruction) {
5799 if (CanMoveNullCheckToUser(instruction)) {
5800 return;
5801 }
5802
5803 UseScratchRegisterScope temps(GetVIXLAssembler());
5804 // Ensure the pc position is recorded immediately after the `ldr` instruction.
5805 ExactAssemblyScope aas(GetVIXLAssembler(),
5806 vixl32::kMaxInstructionSizeInBytes,
5807 CodeBufferCheckScope::kMaximumSize);
5808 __ ldr(temps.Acquire(), MemOperand(InputRegisterAt(instruction, 0)));
5809 RecordPcInfo(instruction, instruction->GetDexPc());
5810 }
5811
GenerateExplicitNullCheck(HNullCheck * instruction)5812 void CodeGeneratorARMVIXL::GenerateExplicitNullCheck(HNullCheck* instruction) {
5813 NullCheckSlowPathARMVIXL* slow_path =
5814 new (GetScopedAllocator()) NullCheckSlowPathARMVIXL(instruction);
5815 AddSlowPath(slow_path);
5816 __ CompareAndBranchIfZero(InputRegisterAt(instruction, 0), slow_path->GetEntryLabel());
5817 }
5818
VisitNullCheck(HNullCheck * instruction)5819 void InstructionCodeGeneratorARMVIXL::VisitNullCheck(HNullCheck* instruction) {
5820 codegen_->GenerateNullCheck(instruction);
5821 }
5822
LoadFromShiftedRegOffset(DataType::Type type,Location out_loc,vixl32::Register base,vixl32::Register reg_index,vixl32::Condition cond)5823 void CodeGeneratorARMVIXL::LoadFromShiftedRegOffset(DataType::Type type,
5824 Location out_loc,
5825 vixl32::Register base,
5826 vixl32::Register reg_index,
5827 vixl32::Condition cond) {
5828 uint32_t shift_count = DataType::SizeShift(type);
5829 MemOperand mem_address(base, reg_index, vixl32::LSL, shift_count);
5830
5831 switch (type) {
5832 case DataType::Type::kBool:
5833 case DataType::Type::kUint8:
5834 __ Ldrb(cond, RegisterFrom(out_loc), mem_address);
5835 break;
5836 case DataType::Type::kInt8:
5837 __ Ldrsb(cond, RegisterFrom(out_loc), mem_address);
5838 break;
5839 case DataType::Type::kUint16:
5840 __ Ldrh(cond, RegisterFrom(out_loc), mem_address);
5841 break;
5842 case DataType::Type::kInt16:
5843 __ Ldrsh(cond, RegisterFrom(out_loc), mem_address);
5844 break;
5845 case DataType::Type::kReference:
5846 case DataType::Type::kInt32:
5847 __ Ldr(cond, RegisterFrom(out_loc), mem_address);
5848 break;
5849 // T32 doesn't support LoadFromShiftedRegOffset mem address mode for these types.
5850 case DataType::Type::kInt64:
5851 case DataType::Type::kFloat32:
5852 case DataType::Type::kFloat64:
5853 default:
5854 LOG(FATAL) << "Unreachable type " << type;
5855 UNREACHABLE();
5856 }
5857 }
5858
StoreToShiftedRegOffset(DataType::Type type,Location loc,vixl32::Register base,vixl32::Register reg_index,vixl32::Condition cond)5859 void CodeGeneratorARMVIXL::StoreToShiftedRegOffset(DataType::Type type,
5860 Location loc,
5861 vixl32::Register base,
5862 vixl32::Register reg_index,
5863 vixl32::Condition cond) {
5864 uint32_t shift_count = DataType::SizeShift(type);
5865 MemOperand mem_address(base, reg_index, vixl32::LSL, shift_count);
5866
5867 switch (type) {
5868 case DataType::Type::kBool:
5869 case DataType::Type::kUint8:
5870 case DataType::Type::kInt8:
5871 __ Strb(cond, RegisterFrom(loc), mem_address);
5872 break;
5873 case DataType::Type::kUint16:
5874 case DataType::Type::kInt16:
5875 __ Strh(cond, RegisterFrom(loc), mem_address);
5876 break;
5877 case DataType::Type::kReference:
5878 case DataType::Type::kInt32:
5879 __ Str(cond, RegisterFrom(loc), mem_address);
5880 break;
5881 // T32 doesn't support StoreToShiftedRegOffset mem address mode for these types.
5882 case DataType::Type::kInt64:
5883 case DataType::Type::kFloat32:
5884 case DataType::Type::kFloat64:
5885 default:
5886 LOG(FATAL) << "Unreachable type " << type;
5887 UNREACHABLE();
5888 }
5889 }
5890
VisitArrayGet(HArrayGet * instruction)5891 void LocationsBuilderARMVIXL::VisitArrayGet(HArrayGet* instruction) {
5892 bool object_array_get_with_read_barrier =
5893 kEmitCompilerReadBarrier && (instruction->GetType() == DataType::Type::kReference);
5894 LocationSummary* locations =
5895 new (GetGraph()->GetAllocator()) LocationSummary(instruction,
5896 object_array_get_with_read_barrier
5897 ? LocationSummary::kCallOnSlowPath
5898 : LocationSummary::kNoCall);
5899 if (object_array_get_with_read_barrier && kUseBakerReadBarrier) {
5900 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers.
5901 }
5902 locations->SetInAt(0, Location::RequiresRegister());
5903 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1)));
5904 if (DataType::IsFloatingPointType(instruction->GetType())) {
5905 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
5906 } else {
5907 // The output overlaps in the case of an object array get with
5908 // read barriers enabled: we do not want the move to overwrite the
5909 // array's location, as we need it to emit the read barrier.
5910 locations->SetOut(
5911 Location::RequiresRegister(),
5912 object_array_get_with_read_barrier ? Location::kOutputOverlap : Location::kNoOutputOverlap);
5913 }
5914 if (object_array_get_with_read_barrier && kUseBakerReadBarrier) {
5915 if (instruction->GetIndex()->IsConstant()) {
5916 // Array loads with constant index are treated as field loads.
5917 // We need a temporary register for the read barrier load in
5918 // CodeGeneratorARMVIXL::GenerateFieldLoadWithBakerReadBarrier()
5919 // only if the offset is too big.
5920 uint32_t offset = CodeGenerator::GetArrayDataOffset(instruction);
5921 uint32_t index = instruction->GetIndex()->AsIntConstant()->GetValue();
5922 offset += index << DataType::SizeShift(DataType::Type::kReference);
5923 if (offset >= kReferenceLoadMinFarOffset) {
5924 locations->AddTemp(Location::RequiresRegister());
5925 }
5926 } else {
5927 // We need a non-scratch temporary for the array data pointer in
5928 // CodeGeneratorARMVIXL::GenerateArrayLoadWithBakerReadBarrier().
5929 locations->AddTemp(Location::RequiresRegister());
5930 }
5931 } else if (mirror::kUseStringCompression && instruction->IsStringCharAt()) {
5932 // Also need a temporary for String compression feature.
5933 locations->AddTemp(Location::RequiresRegister());
5934 }
5935 }
5936
VisitArrayGet(HArrayGet * instruction)5937 void InstructionCodeGeneratorARMVIXL::VisitArrayGet(HArrayGet* instruction) {
5938 LocationSummary* locations = instruction->GetLocations();
5939 Location obj_loc = locations->InAt(0);
5940 vixl32::Register obj = InputRegisterAt(instruction, 0);
5941 Location index = locations->InAt(1);
5942 Location out_loc = locations->Out();
5943 uint32_t data_offset = CodeGenerator::GetArrayDataOffset(instruction);
5944 DataType::Type type = instruction->GetType();
5945 const bool maybe_compressed_char_at = mirror::kUseStringCompression &&
5946 instruction->IsStringCharAt();
5947 HInstruction* array_instr = instruction->GetArray();
5948 bool has_intermediate_address = array_instr->IsIntermediateAddress();
5949
5950 switch (type) {
5951 case DataType::Type::kBool:
5952 case DataType::Type::kUint8:
5953 case DataType::Type::kInt8:
5954 case DataType::Type::kUint16:
5955 case DataType::Type::kInt16:
5956 case DataType::Type::kInt32: {
5957 vixl32::Register length;
5958 if (maybe_compressed_char_at) {
5959 length = RegisterFrom(locations->GetTemp(0));
5960 uint32_t count_offset = mirror::String::CountOffset().Uint32Value();
5961 GetAssembler()->LoadFromOffset(kLoadWord, length, obj, count_offset);
5962 codegen_->MaybeRecordImplicitNullCheck(instruction);
5963 }
5964 if (index.IsConstant()) {
5965 int32_t const_index = Int32ConstantFrom(index);
5966 if (maybe_compressed_char_at) {
5967 vixl32::Label uncompressed_load, done;
5968 vixl32::Label* final_label = codegen_->GetFinalLabel(instruction, &done);
5969 __ Lsrs(length, length, 1u); // LSRS has a 16-bit encoding, TST (immediate) does not.
5970 static_assert(static_cast<uint32_t>(mirror::StringCompressionFlag::kCompressed) == 0u,
5971 "Expecting 0=compressed, 1=uncompressed");
5972 __ B(cs, &uncompressed_load, /* is_far_target= */ false);
5973 GetAssembler()->LoadFromOffset(kLoadUnsignedByte,
5974 RegisterFrom(out_loc),
5975 obj,
5976 data_offset + const_index);
5977 __ B(final_label);
5978 __ Bind(&uncompressed_load);
5979 GetAssembler()->LoadFromOffset(GetLoadOperandType(DataType::Type::kUint16),
5980 RegisterFrom(out_loc),
5981 obj,
5982 data_offset + (const_index << 1));
5983 if (done.IsReferenced()) {
5984 __ Bind(&done);
5985 }
5986 } else {
5987 uint32_t full_offset = data_offset + (const_index << DataType::SizeShift(type));
5988
5989 LoadOperandType load_type = GetLoadOperandType(type);
5990 GetAssembler()->LoadFromOffset(load_type, RegisterFrom(out_loc), obj, full_offset);
5991 }
5992 } else {
5993 UseScratchRegisterScope temps(GetVIXLAssembler());
5994 vixl32::Register temp = temps.Acquire();
5995
5996 if (has_intermediate_address) {
5997 // We do not need to compute the intermediate address from the array: the
5998 // input instruction has done it already. See the comment in
5999 // `TryExtractArrayAccessAddress()`.
6000 if (kIsDebugBuild) {
6001 HIntermediateAddress* tmp = array_instr->AsIntermediateAddress();
6002 DCHECK_EQ(Uint64ConstantFrom(tmp->GetOffset()), data_offset);
6003 }
6004 temp = obj;
6005 } else {
6006 __ Add(temp, obj, data_offset);
6007 }
6008 if (maybe_compressed_char_at) {
6009 vixl32::Label uncompressed_load, done;
6010 vixl32::Label* final_label = codegen_->GetFinalLabel(instruction, &done);
6011 __ Lsrs(length, length, 1u); // LSRS has a 16-bit encoding, TST (immediate) does not.
6012 static_assert(static_cast<uint32_t>(mirror::StringCompressionFlag::kCompressed) == 0u,
6013 "Expecting 0=compressed, 1=uncompressed");
6014 __ B(cs, &uncompressed_load, /* is_far_target= */ false);
6015 __ Ldrb(RegisterFrom(out_loc), MemOperand(temp, RegisterFrom(index), vixl32::LSL, 0));
6016 __ B(final_label);
6017 __ Bind(&uncompressed_load);
6018 __ Ldrh(RegisterFrom(out_loc), MemOperand(temp, RegisterFrom(index), vixl32::LSL, 1));
6019 if (done.IsReferenced()) {
6020 __ Bind(&done);
6021 }
6022 } else {
6023 codegen_->LoadFromShiftedRegOffset(type, out_loc, temp, RegisterFrom(index));
6024 }
6025 }
6026 break;
6027 }
6028
6029 case DataType::Type::kReference: {
6030 // The read barrier instrumentation of object ArrayGet
6031 // instructions does not support the HIntermediateAddress
6032 // instruction.
6033 DCHECK(!(has_intermediate_address && kEmitCompilerReadBarrier));
6034
6035 static_assert(
6036 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t),
6037 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes.");
6038 // /* HeapReference<Object> */ out =
6039 // *(obj + data_offset + index * sizeof(HeapReference<Object>))
6040 if (kEmitCompilerReadBarrier && kUseBakerReadBarrier) {
6041 // Note that a potential implicit null check is handled in this
6042 // CodeGeneratorARMVIXL::GenerateArrayLoadWithBakerReadBarrier call.
6043 DCHECK(!instruction->CanDoImplicitNullCheckOn(instruction->InputAt(0)));
6044 if (index.IsConstant()) {
6045 // Array load with a constant index can be treated as a field load.
6046 Location maybe_temp =
6047 (locations->GetTempCount() != 0) ? locations->GetTemp(0) : Location();
6048 data_offset += Int32ConstantFrom(index) << DataType::SizeShift(type);
6049 codegen_->GenerateFieldLoadWithBakerReadBarrier(instruction,
6050 out_loc,
6051 obj,
6052 data_offset,
6053 maybe_temp,
6054 /* needs_null_check= */ false);
6055 } else {
6056 Location temp = locations->GetTemp(0);
6057 codegen_->GenerateArrayLoadWithBakerReadBarrier(
6058 out_loc, obj, data_offset, index, temp, /* needs_null_check= */ false);
6059 }
6060 } else {
6061 vixl32::Register out = OutputRegister(instruction);
6062 if (index.IsConstant()) {
6063 size_t offset =
6064 (Int32ConstantFrom(index) << TIMES_4) + data_offset;
6065 GetAssembler()->LoadFromOffset(kLoadWord, out, obj, offset);
6066 // TODO(VIXL): Here and for other calls to `MaybeRecordImplicitNullCheck` in this method,
6067 // we should use a scope and the assembler to emit the load instruction to guarantee that
6068 // we record the pc at the correct position. But the `Assembler` does not automatically
6069 // handle unencodable offsets. Practically, everything is fine because the helper and
6070 // VIXL, at the time of writing, do generate the store instruction last.
6071 codegen_->MaybeRecordImplicitNullCheck(instruction);
6072 // If read barriers are enabled, emit read barriers other than
6073 // Baker's using a slow path (and also unpoison the loaded
6074 // reference, if heap poisoning is enabled).
6075 codegen_->MaybeGenerateReadBarrierSlow(instruction, out_loc, out_loc, obj_loc, offset);
6076 } else {
6077 UseScratchRegisterScope temps(GetVIXLAssembler());
6078 vixl32::Register temp = temps.Acquire();
6079
6080 if (has_intermediate_address) {
6081 // We do not need to compute the intermediate address from the array: the
6082 // input instruction has done it already. See the comment in
6083 // `TryExtractArrayAccessAddress()`.
6084 if (kIsDebugBuild) {
6085 HIntermediateAddress* tmp = array_instr->AsIntermediateAddress();
6086 DCHECK_EQ(Uint64ConstantFrom(tmp->GetOffset()), data_offset);
6087 }
6088 temp = obj;
6089 } else {
6090 __ Add(temp, obj, data_offset);
6091 }
6092 codegen_->LoadFromShiftedRegOffset(type, out_loc, temp, RegisterFrom(index));
6093 temps.Close();
6094 // TODO(VIXL): Use a scope to ensure that we record the pc position immediately after the
6095 // load instruction. Practically, everything is fine because the helper and VIXL, at the
6096 // time of writing, do generate the store instruction last.
6097 codegen_->MaybeRecordImplicitNullCheck(instruction);
6098 // If read barriers are enabled, emit read barriers other than
6099 // Baker's using a slow path (and also unpoison the loaded
6100 // reference, if heap poisoning is enabled).
6101 codegen_->MaybeGenerateReadBarrierSlow(
6102 instruction, out_loc, out_loc, obj_loc, data_offset, index);
6103 }
6104 }
6105 break;
6106 }
6107
6108 case DataType::Type::kInt64: {
6109 if (index.IsConstant()) {
6110 size_t offset =
6111 (Int32ConstantFrom(index) << TIMES_8) + data_offset;
6112 GetAssembler()->LoadFromOffset(kLoadWordPair, LowRegisterFrom(out_loc), obj, offset);
6113 } else {
6114 UseScratchRegisterScope temps(GetVIXLAssembler());
6115 vixl32::Register temp = temps.Acquire();
6116 __ Add(temp, obj, Operand(RegisterFrom(index), vixl32::LSL, TIMES_8));
6117 GetAssembler()->LoadFromOffset(kLoadWordPair, LowRegisterFrom(out_loc), temp, data_offset);
6118 }
6119 break;
6120 }
6121
6122 case DataType::Type::kFloat32: {
6123 vixl32::SRegister out = SRegisterFrom(out_loc);
6124 if (index.IsConstant()) {
6125 size_t offset = (Int32ConstantFrom(index) << TIMES_4) + data_offset;
6126 GetAssembler()->LoadSFromOffset(out, obj, offset);
6127 } else {
6128 UseScratchRegisterScope temps(GetVIXLAssembler());
6129 vixl32::Register temp = temps.Acquire();
6130 __ Add(temp, obj, Operand(RegisterFrom(index), vixl32::LSL, TIMES_4));
6131 GetAssembler()->LoadSFromOffset(out, temp, data_offset);
6132 }
6133 break;
6134 }
6135
6136 case DataType::Type::kFloat64: {
6137 if (index.IsConstant()) {
6138 size_t offset = (Int32ConstantFrom(index) << TIMES_8) + data_offset;
6139 GetAssembler()->LoadDFromOffset(DRegisterFrom(out_loc), obj, offset);
6140 } else {
6141 UseScratchRegisterScope temps(GetVIXLAssembler());
6142 vixl32::Register temp = temps.Acquire();
6143 __ Add(temp, obj, Operand(RegisterFrom(index), vixl32::LSL, TIMES_8));
6144 GetAssembler()->LoadDFromOffset(DRegisterFrom(out_loc), temp, data_offset);
6145 }
6146 break;
6147 }
6148
6149 case DataType::Type::kUint32:
6150 case DataType::Type::kUint64:
6151 case DataType::Type::kVoid:
6152 LOG(FATAL) << "Unreachable type " << type;
6153 UNREACHABLE();
6154 }
6155
6156 if (type == DataType::Type::kReference) {
6157 // Potential implicit null checks, in the case of reference
6158 // arrays, are handled in the previous switch statement.
6159 } else if (!maybe_compressed_char_at) {
6160 // TODO(VIXL): Use a scope to ensure we record the pc info immediately after
6161 // the preceding load instruction.
6162 codegen_->MaybeRecordImplicitNullCheck(instruction);
6163 }
6164 }
6165
VisitArraySet(HArraySet * instruction)6166 void LocationsBuilderARMVIXL::VisitArraySet(HArraySet* instruction) {
6167 DataType::Type value_type = instruction->GetComponentType();
6168
6169 bool needs_write_barrier =
6170 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue());
6171 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck();
6172
6173 LocationSummary* locations = new (GetGraph()->GetAllocator()) LocationSummary(
6174 instruction,
6175 may_need_runtime_call_for_type_check ?
6176 LocationSummary::kCallOnSlowPath :
6177 LocationSummary::kNoCall);
6178
6179 locations->SetInAt(0, Location::RequiresRegister());
6180 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1)));
6181 if (DataType::IsFloatingPointType(value_type)) {
6182 locations->SetInAt(2, Location::RequiresFpuRegister());
6183 } else {
6184 locations->SetInAt(2, Location::RequiresRegister());
6185 }
6186 if (needs_write_barrier) {
6187 // Temporary registers for the write barrier.
6188 locations->AddTemp(Location::RequiresRegister()); // Possibly used for ref. poisoning too.
6189 locations->AddTemp(Location::RequiresRegister());
6190 }
6191 }
6192
VisitArraySet(HArraySet * instruction)6193 void InstructionCodeGeneratorARMVIXL::VisitArraySet(HArraySet* instruction) {
6194 LocationSummary* locations = instruction->GetLocations();
6195 vixl32::Register array = InputRegisterAt(instruction, 0);
6196 Location index = locations->InAt(1);
6197 DataType::Type value_type = instruction->GetComponentType();
6198 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck();
6199 bool needs_write_barrier =
6200 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue());
6201 uint32_t data_offset =
6202 mirror::Array::DataOffset(DataType::Size(value_type)).Uint32Value();
6203 Location value_loc = locations->InAt(2);
6204 HInstruction* array_instr = instruction->GetArray();
6205 bool has_intermediate_address = array_instr->IsIntermediateAddress();
6206
6207 switch (value_type) {
6208 case DataType::Type::kBool:
6209 case DataType::Type::kUint8:
6210 case DataType::Type::kInt8:
6211 case DataType::Type::kUint16:
6212 case DataType::Type::kInt16:
6213 case DataType::Type::kInt32: {
6214 if (index.IsConstant()) {
6215 int32_t const_index = Int32ConstantFrom(index);
6216 uint32_t full_offset =
6217 data_offset + (const_index << DataType::SizeShift(value_type));
6218 StoreOperandType store_type = GetStoreOperandType(value_type);
6219 GetAssembler()->StoreToOffset(store_type, RegisterFrom(value_loc), array, full_offset);
6220 } else {
6221 UseScratchRegisterScope temps(GetVIXLAssembler());
6222 vixl32::Register temp = temps.Acquire();
6223
6224 if (has_intermediate_address) {
6225 // We do not need to compute the intermediate address from the array: the
6226 // input instruction has done it already. See the comment in
6227 // `TryExtractArrayAccessAddress()`.
6228 if (kIsDebugBuild) {
6229 HIntermediateAddress* tmp = array_instr->AsIntermediateAddress();
6230 DCHECK_EQ(Uint64ConstantFrom(tmp->GetOffset()), data_offset);
6231 }
6232 temp = array;
6233 } else {
6234 __ Add(temp, array, data_offset);
6235 }
6236 codegen_->StoreToShiftedRegOffset(value_type, value_loc, temp, RegisterFrom(index));
6237 }
6238 break;
6239 }
6240
6241 case DataType::Type::kReference: {
6242 vixl32::Register value = RegisterFrom(value_loc);
6243 // TryExtractArrayAccessAddress optimization is never applied for non-primitive ArraySet.
6244 // See the comment in instruction_simplifier_shared.cc.
6245 DCHECK(!has_intermediate_address);
6246
6247 if (instruction->InputAt(2)->IsNullConstant()) {
6248 // Just setting null.
6249 if (index.IsConstant()) {
6250 size_t offset =
6251 (Int32ConstantFrom(index) << TIMES_4) + data_offset;
6252 GetAssembler()->StoreToOffset(kStoreWord, value, array, offset);
6253 } else {
6254 DCHECK(index.IsRegister()) << index;
6255 UseScratchRegisterScope temps(GetVIXLAssembler());
6256 vixl32::Register temp = temps.Acquire();
6257 __ Add(temp, array, data_offset);
6258 codegen_->StoreToShiftedRegOffset(value_type, value_loc, temp, RegisterFrom(index));
6259 }
6260 // TODO(VIXL): Use a scope to ensure we record the pc info immediately after the preceding
6261 // store instruction.
6262 codegen_->MaybeRecordImplicitNullCheck(instruction);
6263 DCHECK(!needs_write_barrier);
6264 DCHECK(!may_need_runtime_call_for_type_check);
6265 break;
6266 }
6267
6268 DCHECK(needs_write_barrier);
6269 Location temp1_loc = locations->GetTemp(0);
6270 vixl32::Register temp1 = RegisterFrom(temp1_loc);
6271 Location temp2_loc = locations->GetTemp(1);
6272 vixl32::Register temp2 = RegisterFrom(temp2_loc);
6273 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
6274 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value();
6275 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value();
6276 vixl32::Label done;
6277 vixl32::Label* final_label = codegen_->GetFinalLabel(instruction, &done);
6278 SlowPathCodeARMVIXL* slow_path = nullptr;
6279
6280 if (may_need_runtime_call_for_type_check) {
6281 slow_path = new (codegen_->GetScopedAllocator()) ArraySetSlowPathARMVIXL(instruction);
6282 codegen_->AddSlowPath(slow_path);
6283 if (instruction->GetValueCanBeNull()) {
6284 vixl32::Label non_zero;
6285 __ CompareAndBranchIfNonZero(value, &non_zero);
6286 if (index.IsConstant()) {
6287 size_t offset =
6288 (Int32ConstantFrom(index) << TIMES_4) + data_offset;
6289 GetAssembler()->StoreToOffset(kStoreWord, value, array, offset);
6290 } else {
6291 DCHECK(index.IsRegister()) << index;
6292 UseScratchRegisterScope temps(GetVIXLAssembler());
6293 vixl32::Register temp = temps.Acquire();
6294 __ Add(temp, array, data_offset);
6295 codegen_->StoreToShiftedRegOffset(value_type, value_loc, temp, RegisterFrom(index));
6296 }
6297 // TODO(VIXL): Use a scope to ensure we record the pc info immediately after the preceding
6298 // store instruction.
6299 codegen_->MaybeRecordImplicitNullCheck(instruction);
6300 __ B(final_label);
6301 __ Bind(&non_zero);
6302 }
6303
6304 // Note that when read barriers are enabled, the type checks
6305 // are performed without read barriers. This is fine, even in
6306 // the case where a class object is in the from-space after
6307 // the flip, as a comparison involving such a type would not
6308 // produce a false positive; it may of course produce a false
6309 // negative, in which case we would take the ArraySet slow
6310 // path.
6311
6312 {
6313 // Ensure we record the pc position immediately after the `ldr` instruction.
6314 ExactAssemblyScope aas(GetVIXLAssembler(),
6315 vixl32::kMaxInstructionSizeInBytes,
6316 CodeBufferCheckScope::kMaximumSize);
6317 // /* HeapReference<Class> */ temp1 = array->klass_
6318 __ ldr(temp1, MemOperand(array, class_offset));
6319 codegen_->MaybeRecordImplicitNullCheck(instruction);
6320 }
6321 GetAssembler()->MaybeUnpoisonHeapReference(temp1);
6322
6323 // /* HeapReference<Class> */ temp1 = temp1->component_type_
6324 GetAssembler()->LoadFromOffset(kLoadWord, temp1, temp1, component_offset);
6325 // /* HeapReference<Class> */ temp2 = value->klass_
6326 GetAssembler()->LoadFromOffset(kLoadWord, temp2, value, class_offset);
6327 // If heap poisoning is enabled, no need to unpoison `temp1`
6328 // nor `temp2`, as we are comparing two poisoned references.
6329 __ Cmp(temp1, temp2);
6330
6331 if (instruction->StaticTypeOfArrayIsObjectArray()) {
6332 vixl32::Label do_put;
6333 __ B(eq, &do_put, /* is_far_target= */ false);
6334 // If heap poisoning is enabled, the `temp1` reference has
6335 // not been unpoisoned yet; unpoison it now.
6336 GetAssembler()->MaybeUnpoisonHeapReference(temp1);
6337
6338 // /* HeapReference<Class> */ temp1 = temp1->super_class_
6339 GetAssembler()->LoadFromOffset(kLoadWord, temp1, temp1, super_offset);
6340 // If heap poisoning is enabled, no need to unpoison
6341 // `temp1`, as we are comparing against null below.
6342 __ CompareAndBranchIfNonZero(temp1, slow_path->GetEntryLabel());
6343 __ Bind(&do_put);
6344 } else {
6345 __ B(ne, slow_path->GetEntryLabel());
6346 }
6347 }
6348
6349 vixl32::Register source = value;
6350 if (kPoisonHeapReferences) {
6351 // Note that in the case where `value` is a null reference,
6352 // we do not enter this block, as a null reference does not
6353 // need poisoning.
6354 DCHECK_EQ(value_type, DataType::Type::kReference);
6355 __ Mov(temp1, value);
6356 GetAssembler()->PoisonHeapReference(temp1);
6357 source = temp1;
6358 }
6359
6360 if (index.IsConstant()) {
6361 size_t offset =
6362 (Int32ConstantFrom(index) << TIMES_4) + data_offset;
6363 GetAssembler()->StoreToOffset(kStoreWord, source, array, offset);
6364 } else {
6365 DCHECK(index.IsRegister()) << index;
6366
6367 UseScratchRegisterScope temps(GetVIXLAssembler());
6368 vixl32::Register temp = temps.Acquire();
6369 __ Add(temp, array, data_offset);
6370 codegen_->StoreToShiftedRegOffset(value_type,
6371 LocationFrom(source),
6372 temp,
6373 RegisterFrom(index));
6374 }
6375
6376 if (!may_need_runtime_call_for_type_check) {
6377 // TODO(VIXL): Ensure we record the pc position immediately after the preceding store
6378 // instruction.
6379 codegen_->MaybeRecordImplicitNullCheck(instruction);
6380 }
6381
6382 codegen_->MarkGCCard(temp1, temp2, array, value, instruction->GetValueCanBeNull());
6383
6384 if (done.IsReferenced()) {
6385 __ Bind(&done);
6386 }
6387
6388 if (slow_path != nullptr) {
6389 __ Bind(slow_path->GetExitLabel());
6390 }
6391
6392 break;
6393 }
6394
6395 case DataType::Type::kInt64: {
6396 Location value = locations->InAt(2);
6397 if (index.IsConstant()) {
6398 size_t offset =
6399 (Int32ConstantFrom(index) << TIMES_8) + data_offset;
6400 GetAssembler()->StoreToOffset(kStoreWordPair, LowRegisterFrom(value), array, offset);
6401 } else {
6402 UseScratchRegisterScope temps(GetVIXLAssembler());
6403 vixl32::Register temp = temps.Acquire();
6404 __ Add(temp, array, Operand(RegisterFrom(index), vixl32::LSL, TIMES_8));
6405 GetAssembler()->StoreToOffset(kStoreWordPair, LowRegisterFrom(value), temp, data_offset);
6406 }
6407 break;
6408 }
6409
6410 case DataType::Type::kFloat32: {
6411 Location value = locations->InAt(2);
6412 DCHECK(value.IsFpuRegister());
6413 if (index.IsConstant()) {
6414 size_t offset = (Int32ConstantFrom(index) << TIMES_4) + data_offset;
6415 GetAssembler()->StoreSToOffset(SRegisterFrom(value), array, offset);
6416 } else {
6417 UseScratchRegisterScope temps(GetVIXLAssembler());
6418 vixl32::Register temp = temps.Acquire();
6419 __ Add(temp, array, Operand(RegisterFrom(index), vixl32::LSL, TIMES_4));
6420 GetAssembler()->StoreSToOffset(SRegisterFrom(value), temp, data_offset);
6421 }
6422 break;
6423 }
6424
6425 case DataType::Type::kFloat64: {
6426 Location value = locations->InAt(2);
6427 DCHECK(value.IsFpuRegisterPair());
6428 if (index.IsConstant()) {
6429 size_t offset = (Int32ConstantFrom(index) << TIMES_8) + data_offset;
6430 GetAssembler()->StoreDToOffset(DRegisterFrom(value), array, offset);
6431 } else {
6432 UseScratchRegisterScope temps(GetVIXLAssembler());
6433 vixl32::Register temp = temps.Acquire();
6434 __ Add(temp, array, Operand(RegisterFrom(index), vixl32::LSL, TIMES_8));
6435 GetAssembler()->StoreDToOffset(DRegisterFrom(value), temp, data_offset);
6436 }
6437 break;
6438 }
6439
6440 case DataType::Type::kUint32:
6441 case DataType::Type::kUint64:
6442 case DataType::Type::kVoid:
6443 LOG(FATAL) << "Unreachable type " << value_type;
6444 UNREACHABLE();
6445 }
6446
6447 // Objects are handled in the switch.
6448 if (value_type != DataType::Type::kReference) {
6449 // TODO(VIXL): Ensure we record the pc position immediately after the preceding store
6450 // instruction.
6451 codegen_->MaybeRecordImplicitNullCheck(instruction);
6452 }
6453 }
6454
VisitArrayLength(HArrayLength * instruction)6455 void LocationsBuilderARMVIXL::VisitArrayLength(HArrayLength* instruction) {
6456 LocationSummary* locations =
6457 new (GetGraph()->GetAllocator()) LocationSummary(instruction, LocationSummary::kNoCall);
6458 locations->SetInAt(0, Location::RequiresRegister());
6459 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
6460 }
6461
VisitArrayLength(HArrayLength * instruction)6462 void InstructionCodeGeneratorARMVIXL::VisitArrayLength(HArrayLength* instruction) {
6463 uint32_t offset = CodeGenerator::GetArrayLengthOffset(instruction);
6464 vixl32::Register obj = InputRegisterAt(instruction, 0);
6465 vixl32::Register out = OutputRegister(instruction);
6466 {
6467 ExactAssemblyScope aas(GetVIXLAssembler(),
6468 vixl32::kMaxInstructionSizeInBytes,
6469 CodeBufferCheckScope::kMaximumSize);
6470 __ ldr(out, MemOperand(obj, offset));
6471 codegen_->MaybeRecordImplicitNullCheck(instruction);
6472 }
6473 // Mask out compression flag from String's array length.
6474 if (mirror::kUseStringCompression && instruction->IsStringLength()) {
6475 __ Lsr(out, out, 1u);
6476 }
6477 }
6478
VisitIntermediateAddress(HIntermediateAddress * instruction)6479 void LocationsBuilderARMVIXL::VisitIntermediateAddress(HIntermediateAddress* instruction) {
6480 LocationSummary* locations =
6481 new (GetGraph()->GetAllocator()) LocationSummary(instruction, LocationSummary::kNoCall);
6482
6483 locations->SetInAt(0, Location::RequiresRegister());
6484 locations->SetInAt(1, Location::RegisterOrConstant(instruction->GetOffset()));
6485 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
6486 }
6487
VisitIntermediateAddress(HIntermediateAddress * instruction)6488 void InstructionCodeGeneratorARMVIXL::VisitIntermediateAddress(HIntermediateAddress* instruction) {
6489 vixl32::Register out = OutputRegister(instruction);
6490 vixl32::Register first = InputRegisterAt(instruction, 0);
6491 Location second = instruction->GetLocations()->InAt(1);
6492
6493 if (second.IsRegister()) {
6494 __ Add(out, first, RegisterFrom(second));
6495 } else {
6496 __ Add(out, first, Int32ConstantFrom(second));
6497 }
6498 }
6499
VisitIntermediateAddressIndex(HIntermediateAddressIndex * instruction)6500 void LocationsBuilderARMVIXL::VisitIntermediateAddressIndex(
6501 HIntermediateAddressIndex* instruction) {
6502 LOG(FATAL) << "Unreachable " << instruction->GetId();
6503 }
6504
VisitIntermediateAddressIndex(HIntermediateAddressIndex * instruction)6505 void InstructionCodeGeneratorARMVIXL::VisitIntermediateAddressIndex(
6506 HIntermediateAddressIndex* instruction) {
6507 LOG(FATAL) << "Unreachable " << instruction->GetId();
6508 }
6509
VisitBoundsCheck(HBoundsCheck * instruction)6510 void LocationsBuilderARMVIXL::VisitBoundsCheck(HBoundsCheck* instruction) {
6511 RegisterSet caller_saves = RegisterSet::Empty();
6512 InvokeRuntimeCallingConventionARMVIXL calling_convention;
6513 caller_saves.Add(LocationFrom(calling_convention.GetRegisterAt(0)));
6514 caller_saves.Add(LocationFrom(calling_convention.GetRegisterAt(1)));
6515 LocationSummary* locations = codegen_->CreateThrowingSlowPathLocations(instruction, caller_saves);
6516
6517 HInstruction* index = instruction->InputAt(0);
6518 HInstruction* length = instruction->InputAt(1);
6519 // If both index and length are constants we can statically check the bounds. But if at least one
6520 // of them is not encodable ArmEncodableConstantOrRegister will create
6521 // Location::RequiresRegister() which is not desired to happen. Instead we create constant
6522 // locations.
6523 bool both_const = index->IsConstant() && length->IsConstant();
6524 locations->SetInAt(0, both_const
6525 ? Location::ConstantLocation(index->AsConstant())
6526 : ArmEncodableConstantOrRegister(index, CMP));
6527 locations->SetInAt(1, both_const
6528 ? Location::ConstantLocation(length->AsConstant())
6529 : ArmEncodableConstantOrRegister(length, CMP));
6530 }
6531
VisitBoundsCheck(HBoundsCheck * instruction)6532 void InstructionCodeGeneratorARMVIXL::VisitBoundsCheck(HBoundsCheck* instruction) {
6533 LocationSummary* locations = instruction->GetLocations();
6534 Location index_loc = locations->InAt(0);
6535 Location length_loc = locations->InAt(1);
6536
6537 if (length_loc.IsConstant()) {
6538 int32_t length = Int32ConstantFrom(length_loc);
6539 if (index_loc.IsConstant()) {
6540 // BCE will remove the bounds check if we are guaranteed to pass.
6541 int32_t index = Int32ConstantFrom(index_loc);
6542 if (index < 0 || index >= length) {
6543 SlowPathCodeARMVIXL* slow_path =
6544 new (codegen_->GetScopedAllocator()) BoundsCheckSlowPathARMVIXL(instruction);
6545 codegen_->AddSlowPath(slow_path);
6546 __ B(slow_path->GetEntryLabel());
6547 } else {
6548 // Some optimization after BCE may have generated this, and we should not
6549 // generate a bounds check if it is a valid range.
6550 }
6551 return;
6552 }
6553
6554 SlowPathCodeARMVIXL* slow_path =
6555 new (codegen_->GetScopedAllocator()) BoundsCheckSlowPathARMVIXL(instruction);
6556 __ Cmp(RegisterFrom(index_loc), length);
6557 codegen_->AddSlowPath(slow_path);
6558 __ B(hs, slow_path->GetEntryLabel());
6559 } else {
6560 SlowPathCodeARMVIXL* slow_path =
6561 new (codegen_->GetScopedAllocator()) BoundsCheckSlowPathARMVIXL(instruction);
6562 __ Cmp(RegisterFrom(length_loc), InputOperandAt(instruction, 0));
6563 codegen_->AddSlowPath(slow_path);
6564 __ B(ls, slow_path->GetEntryLabel());
6565 }
6566 }
6567
MarkGCCard(vixl32::Register temp,vixl32::Register card,vixl32::Register object,vixl32::Register value,bool can_be_null)6568 void CodeGeneratorARMVIXL::MarkGCCard(vixl32::Register temp,
6569 vixl32::Register card,
6570 vixl32::Register object,
6571 vixl32::Register value,
6572 bool can_be_null) {
6573 vixl32::Label is_null;
6574 if (can_be_null) {
6575 __ CompareAndBranchIfZero(value, &is_null);
6576 }
6577 // Load the address of the card table into `card`.
6578 GetAssembler()->LoadFromOffset(
6579 kLoadWord, card, tr, Thread::CardTableOffset<kArmPointerSize>().Int32Value());
6580 // Calculate the offset (in the card table) of the card corresponding to
6581 // `object`.
6582 __ Lsr(temp, object, Operand::From(gc::accounting::CardTable::kCardShift));
6583 // Write the `art::gc::accounting::CardTable::kCardDirty` value into the
6584 // `object`'s card.
6585 //
6586 // Register `card` contains the address of the card table. Note that the card
6587 // table's base is biased during its creation so that it always starts at an
6588 // address whose least-significant byte is equal to `kCardDirty` (see
6589 // art::gc::accounting::CardTable::Create). Therefore the STRB instruction
6590 // below writes the `kCardDirty` (byte) value into the `object`'s card
6591 // (located at `card + object >> kCardShift`).
6592 //
6593 // This dual use of the value in register `card` (1. to calculate the location
6594 // of the card to mark; and 2. to load the `kCardDirty` value) saves a load
6595 // (no need to explicitly load `kCardDirty` as an immediate value).
6596 __ Strb(card, MemOperand(card, temp));
6597 if (can_be_null) {
6598 __ Bind(&is_null);
6599 }
6600 }
6601
VisitParallelMove(HParallelMove * instruction ATTRIBUTE_UNUSED)6602 void LocationsBuilderARMVIXL::VisitParallelMove(HParallelMove* instruction ATTRIBUTE_UNUSED) {
6603 LOG(FATAL) << "Unreachable";
6604 }
6605
VisitParallelMove(HParallelMove * instruction)6606 void InstructionCodeGeneratorARMVIXL::VisitParallelMove(HParallelMove* instruction) {
6607 if (instruction->GetNext()->IsSuspendCheck() &&
6608 instruction->GetBlock()->GetLoopInformation() != nullptr) {
6609 HSuspendCheck* suspend_check = instruction->GetNext()->AsSuspendCheck();
6610 // The back edge will generate the suspend check.
6611 codegen_->ClearSpillSlotsFromLoopPhisInStackMap(suspend_check, instruction);
6612 }
6613
6614 codegen_->GetMoveResolver()->EmitNativeCode(instruction);
6615 }
6616
VisitSuspendCheck(HSuspendCheck * instruction)6617 void LocationsBuilderARMVIXL::VisitSuspendCheck(HSuspendCheck* instruction) {
6618 LocationSummary* locations = new (GetGraph()->GetAllocator()) LocationSummary(
6619 instruction, LocationSummary::kCallOnSlowPath);
6620 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers.
6621 }
6622
VisitSuspendCheck(HSuspendCheck * instruction)6623 void InstructionCodeGeneratorARMVIXL::VisitSuspendCheck(HSuspendCheck* instruction) {
6624 HBasicBlock* block = instruction->GetBlock();
6625 if (block->GetLoopInformation() != nullptr) {
6626 DCHECK(block->GetLoopInformation()->GetSuspendCheck() == instruction);
6627 // The back edge will generate the suspend check.
6628 return;
6629 }
6630 if (block->IsEntryBlock() && instruction->GetNext()->IsGoto()) {
6631 // The goto will generate the suspend check.
6632 return;
6633 }
6634 GenerateSuspendCheck(instruction, nullptr);
6635 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 13);
6636 }
6637
GenerateSuspendCheck(HSuspendCheck * instruction,HBasicBlock * successor)6638 void InstructionCodeGeneratorARMVIXL::GenerateSuspendCheck(HSuspendCheck* instruction,
6639 HBasicBlock* successor) {
6640 SuspendCheckSlowPathARMVIXL* slow_path =
6641 down_cast<SuspendCheckSlowPathARMVIXL*>(instruction->GetSlowPath());
6642 if (slow_path == nullptr) {
6643 slow_path =
6644 new (codegen_->GetScopedAllocator()) SuspendCheckSlowPathARMVIXL(instruction, successor);
6645 instruction->SetSlowPath(slow_path);
6646 codegen_->AddSlowPath(slow_path);
6647 if (successor != nullptr) {
6648 DCHECK(successor->IsLoopHeader());
6649 }
6650 } else {
6651 DCHECK_EQ(slow_path->GetSuccessor(), successor);
6652 }
6653
6654 UseScratchRegisterScope temps(GetVIXLAssembler());
6655 vixl32::Register temp = temps.Acquire();
6656 GetAssembler()->LoadFromOffset(
6657 kLoadUnsignedHalfword, temp, tr, Thread::ThreadFlagsOffset<kArmPointerSize>().Int32Value());
6658 if (successor == nullptr) {
6659 __ CompareAndBranchIfNonZero(temp, slow_path->GetEntryLabel());
6660 __ Bind(slow_path->GetReturnLabel());
6661 } else {
6662 __ CompareAndBranchIfZero(temp, codegen_->GetLabelOf(successor));
6663 __ B(slow_path->GetEntryLabel());
6664 }
6665 }
6666
GetAssembler() const6667 ArmVIXLAssembler* ParallelMoveResolverARMVIXL::GetAssembler() const {
6668 return codegen_->GetAssembler();
6669 }
6670
EmitMove(size_t index)6671 void ParallelMoveResolverARMVIXL::EmitMove(size_t index) {
6672 UseScratchRegisterScope temps(GetAssembler()->GetVIXLAssembler());
6673 MoveOperands* move = moves_[index];
6674 Location source = move->GetSource();
6675 Location destination = move->GetDestination();
6676
6677 if (source.IsRegister()) {
6678 if (destination.IsRegister()) {
6679 __ Mov(RegisterFrom(destination), RegisterFrom(source));
6680 } else if (destination.IsFpuRegister()) {
6681 __ Vmov(SRegisterFrom(destination), RegisterFrom(source));
6682 } else {
6683 DCHECK(destination.IsStackSlot());
6684 GetAssembler()->StoreToOffset(kStoreWord,
6685 RegisterFrom(source),
6686 sp,
6687 destination.GetStackIndex());
6688 }
6689 } else if (source.IsStackSlot()) {
6690 if (destination.IsRegister()) {
6691 GetAssembler()->LoadFromOffset(kLoadWord,
6692 RegisterFrom(destination),
6693 sp,
6694 source.GetStackIndex());
6695 } else if (destination.IsFpuRegister()) {
6696 GetAssembler()->LoadSFromOffset(SRegisterFrom(destination), sp, source.GetStackIndex());
6697 } else {
6698 DCHECK(destination.IsStackSlot());
6699 vixl32::Register temp = temps.Acquire();
6700 GetAssembler()->LoadFromOffset(kLoadWord, temp, sp, source.GetStackIndex());
6701 GetAssembler()->StoreToOffset(kStoreWord, temp, sp, destination.GetStackIndex());
6702 }
6703 } else if (source.IsFpuRegister()) {
6704 if (destination.IsRegister()) {
6705 __ Vmov(RegisterFrom(destination), SRegisterFrom(source));
6706 } else if (destination.IsFpuRegister()) {
6707 __ Vmov(SRegisterFrom(destination), SRegisterFrom(source));
6708 } else {
6709 DCHECK(destination.IsStackSlot());
6710 GetAssembler()->StoreSToOffset(SRegisterFrom(source), sp, destination.GetStackIndex());
6711 }
6712 } else if (source.IsDoubleStackSlot()) {
6713 if (destination.IsDoubleStackSlot()) {
6714 vixl32::DRegister temp = temps.AcquireD();
6715 GetAssembler()->LoadDFromOffset(temp, sp, source.GetStackIndex());
6716 GetAssembler()->StoreDToOffset(temp, sp, destination.GetStackIndex());
6717 } else if (destination.IsRegisterPair()) {
6718 DCHECK(ExpectedPairLayout(destination));
6719 GetAssembler()->LoadFromOffset(
6720 kLoadWordPair, LowRegisterFrom(destination), sp, source.GetStackIndex());
6721 } else {
6722 DCHECK(destination.IsFpuRegisterPair()) << destination;
6723 GetAssembler()->LoadDFromOffset(DRegisterFrom(destination), sp, source.GetStackIndex());
6724 }
6725 } else if (source.IsRegisterPair()) {
6726 if (destination.IsRegisterPair()) {
6727 __ Mov(LowRegisterFrom(destination), LowRegisterFrom(source));
6728 __ Mov(HighRegisterFrom(destination), HighRegisterFrom(source));
6729 } else if (destination.IsFpuRegisterPair()) {
6730 __ Vmov(DRegisterFrom(destination), LowRegisterFrom(source), HighRegisterFrom(source));
6731 } else {
6732 DCHECK(destination.IsDoubleStackSlot()) << destination;
6733 DCHECK(ExpectedPairLayout(source));
6734 GetAssembler()->StoreToOffset(kStoreWordPair,
6735 LowRegisterFrom(source),
6736 sp,
6737 destination.GetStackIndex());
6738 }
6739 } else if (source.IsFpuRegisterPair()) {
6740 if (destination.IsRegisterPair()) {
6741 __ Vmov(LowRegisterFrom(destination), HighRegisterFrom(destination), DRegisterFrom(source));
6742 } else if (destination.IsFpuRegisterPair()) {
6743 __ Vmov(DRegisterFrom(destination), DRegisterFrom(source));
6744 } else {
6745 DCHECK(destination.IsDoubleStackSlot()) << destination;
6746 GetAssembler()->StoreDToOffset(DRegisterFrom(source), sp, destination.GetStackIndex());
6747 }
6748 } else {
6749 DCHECK(source.IsConstant()) << source;
6750 HConstant* constant = source.GetConstant();
6751 if (constant->IsIntConstant() || constant->IsNullConstant()) {
6752 int32_t value = CodeGenerator::GetInt32ValueOf(constant);
6753 if (destination.IsRegister()) {
6754 __ Mov(RegisterFrom(destination), value);
6755 } else {
6756 DCHECK(destination.IsStackSlot());
6757 vixl32::Register temp = temps.Acquire();
6758 __ Mov(temp, value);
6759 GetAssembler()->StoreToOffset(kStoreWord, temp, sp, destination.GetStackIndex());
6760 }
6761 } else if (constant->IsLongConstant()) {
6762 int64_t value = Int64ConstantFrom(source);
6763 if (destination.IsRegisterPair()) {
6764 __ Mov(LowRegisterFrom(destination), Low32Bits(value));
6765 __ Mov(HighRegisterFrom(destination), High32Bits(value));
6766 } else {
6767 DCHECK(destination.IsDoubleStackSlot()) << destination;
6768 vixl32::Register temp = temps.Acquire();
6769 __ Mov(temp, Low32Bits(value));
6770 GetAssembler()->StoreToOffset(kStoreWord, temp, sp, destination.GetStackIndex());
6771 __ Mov(temp, High32Bits(value));
6772 GetAssembler()->StoreToOffset(kStoreWord,
6773 temp,
6774 sp,
6775 destination.GetHighStackIndex(kArmWordSize));
6776 }
6777 } else if (constant->IsDoubleConstant()) {
6778 double value = constant->AsDoubleConstant()->GetValue();
6779 if (destination.IsFpuRegisterPair()) {
6780 __ Vmov(DRegisterFrom(destination), value);
6781 } else {
6782 DCHECK(destination.IsDoubleStackSlot()) << destination;
6783 uint64_t int_value = bit_cast<uint64_t, double>(value);
6784 vixl32::Register temp = temps.Acquire();
6785 __ Mov(temp, Low32Bits(int_value));
6786 GetAssembler()->StoreToOffset(kStoreWord, temp, sp, destination.GetStackIndex());
6787 __ Mov(temp, High32Bits(int_value));
6788 GetAssembler()->StoreToOffset(kStoreWord,
6789 temp,
6790 sp,
6791 destination.GetHighStackIndex(kArmWordSize));
6792 }
6793 } else {
6794 DCHECK(constant->IsFloatConstant()) << constant->DebugName();
6795 float value = constant->AsFloatConstant()->GetValue();
6796 if (destination.IsFpuRegister()) {
6797 __ Vmov(SRegisterFrom(destination), value);
6798 } else {
6799 DCHECK(destination.IsStackSlot());
6800 vixl32::Register temp = temps.Acquire();
6801 __ Mov(temp, bit_cast<int32_t, float>(value));
6802 GetAssembler()->StoreToOffset(kStoreWord, temp, sp, destination.GetStackIndex());
6803 }
6804 }
6805 }
6806 }
6807
Exchange(vixl32::Register reg,int mem)6808 void ParallelMoveResolverARMVIXL::Exchange(vixl32::Register reg, int mem) {
6809 UseScratchRegisterScope temps(GetAssembler()->GetVIXLAssembler());
6810 vixl32::Register temp = temps.Acquire();
6811 __ Mov(temp, reg);
6812 GetAssembler()->LoadFromOffset(kLoadWord, reg, sp, mem);
6813 GetAssembler()->StoreToOffset(kStoreWord, temp, sp, mem);
6814 }
6815
Exchange(int mem1,int mem2)6816 void ParallelMoveResolverARMVIXL::Exchange(int mem1, int mem2) {
6817 // TODO(VIXL32): Double check the performance of this implementation.
6818 UseScratchRegisterScope temps(GetAssembler()->GetVIXLAssembler());
6819 vixl32::Register temp1 = temps.Acquire();
6820 ScratchRegisterScope ensure_scratch(
6821 this, temp1.GetCode(), r0.GetCode(), codegen_->GetNumberOfCoreRegisters());
6822 vixl32::Register temp2(ensure_scratch.GetRegister());
6823
6824 int stack_offset = ensure_scratch.IsSpilled() ? kArmWordSize : 0;
6825 GetAssembler()->LoadFromOffset(kLoadWord, temp1, sp, mem1 + stack_offset);
6826 GetAssembler()->LoadFromOffset(kLoadWord, temp2, sp, mem2 + stack_offset);
6827 GetAssembler()->StoreToOffset(kStoreWord, temp1, sp, mem2 + stack_offset);
6828 GetAssembler()->StoreToOffset(kStoreWord, temp2, sp, mem1 + stack_offset);
6829 }
6830
EmitSwap(size_t index)6831 void ParallelMoveResolverARMVIXL::EmitSwap(size_t index) {
6832 MoveOperands* move = moves_[index];
6833 Location source = move->GetSource();
6834 Location destination = move->GetDestination();
6835 UseScratchRegisterScope temps(GetAssembler()->GetVIXLAssembler());
6836
6837 if (source.IsRegister() && destination.IsRegister()) {
6838 vixl32::Register temp = temps.Acquire();
6839 DCHECK(!RegisterFrom(source).Is(temp));
6840 DCHECK(!RegisterFrom(destination).Is(temp));
6841 __ Mov(temp, RegisterFrom(destination));
6842 __ Mov(RegisterFrom(destination), RegisterFrom(source));
6843 __ Mov(RegisterFrom(source), temp);
6844 } else if (source.IsRegister() && destination.IsStackSlot()) {
6845 Exchange(RegisterFrom(source), destination.GetStackIndex());
6846 } else if (source.IsStackSlot() && destination.IsRegister()) {
6847 Exchange(RegisterFrom(destination), source.GetStackIndex());
6848 } else if (source.IsStackSlot() && destination.IsStackSlot()) {
6849 Exchange(source.GetStackIndex(), destination.GetStackIndex());
6850 } else if (source.IsFpuRegister() && destination.IsFpuRegister()) {
6851 vixl32::Register temp = temps.Acquire();
6852 __ Vmov(temp, SRegisterFrom(source));
6853 __ Vmov(SRegisterFrom(source), SRegisterFrom(destination));
6854 __ Vmov(SRegisterFrom(destination), temp);
6855 } else if (source.IsRegisterPair() && destination.IsRegisterPair()) {
6856 vixl32::DRegister temp = temps.AcquireD();
6857 __ Vmov(temp, LowRegisterFrom(source), HighRegisterFrom(source));
6858 __ Mov(LowRegisterFrom(source), LowRegisterFrom(destination));
6859 __ Mov(HighRegisterFrom(source), HighRegisterFrom(destination));
6860 __ Vmov(LowRegisterFrom(destination), HighRegisterFrom(destination), temp);
6861 } else if (source.IsRegisterPair() || destination.IsRegisterPair()) {
6862 vixl32::Register low_reg = LowRegisterFrom(source.IsRegisterPair() ? source : destination);
6863 int mem = source.IsRegisterPair() ? destination.GetStackIndex() : source.GetStackIndex();
6864 DCHECK(ExpectedPairLayout(source.IsRegisterPair() ? source : destination));
6865 vixl32::DRegister temp = temps.AcquireD();
6866 __ Vmov(temp, low_reg, vixl32::Register(low_reg.GetCode() + 1));
6867 GetAssembler()->LoadFromOffset(kLoadWordPair, low_reg, sp, mem);
6868 GetAssembler()->StoreDToOffset(temp, sp, mem);
6869 } else if (source.IsFpuRegisterPair() && destination.IsFpuRegisterPair()) {
6870 vixl32::DRegister first = DRegisterFrom(source);
6871 vixl32::DRegister second = DRegisterFrom(destination);
6872 vixl32::DRegister temp = temps.AcquireD();
6873 __ Vmov(temp, first);
6874 __ Vmov(first, second);
6875 __ Vmov(second, temp);
6876 } else if (source.IsFpuRegisterPair() || destination.IsFpuRegisterPair()) {
6877 vixl32::DRegister reg = source.IsFpuRegisterPair()
6878 ? DRegisterFrom(source)
6879 : DRegisterFrom(destination);
6880 int mem = source.IsFpuRegisterPair()
6881 ? destination.GetStackIndex()
6882 : source.GetStackIndex();
6883 vixl32::DRegister temp = temps.AcquireD();
6884 __ Vmov(temp, reg);
6885 GetAssembler()->LoadDFromOffset(reg, sp, mem);
6886 GetAssembler()->StoreDToOffset(temp, sp, mem);
6887 } else if (source.IsFpuRegister() || destination.IsFpuRegister()) {
6888 vixl32::SRegister reg = source.IsFpuRegister()
6889 ? SRegisterFrom(source)
6890 : SRegisterFrom(destination);
6891 int mem = source.IsFpuRegister()
6892 ? destination.GetStackIndex()
6893 : source.GetStackIndex();
6894 vixl32::Register temp = temps.Acquire();
6895 __ Vmov(temp, reg);
6896 GetAssembler()->LoadSFromOffset(reg, sp, mem);
6897 GetAssembler()->StoreToOffset(kStoreWord, temp, sp, mem);
6898 } else if (source.IsDoubleStackSlot() && destination.IsDoubleStackSlot()) {
6899 vixl32::DRegister temp1 = temps.AcquireD();
6900 vixl32::DRegister temp2 = temps.AcquireD();
6901 __ Vldr(temp1, MemOperand(sp, source.GetStackIndex()));
6902 __ Vldr(temp2, MemOperand(sp, destination.GetStackIndex()));
6903 __ Vstr(temp1, MemOperand(sp, destination.GetStackIndex()));
6904 __ Vstr(temp2, MemOperand(sp, source.GetStackIndex()));
6905 } else {
6906 LOG(FATAL) << "Unimplemented" << source << " <-> " << destination;
6907 }
6908 }
6909
SpillScratch(int reg)6910 void ParallelMoveResolverARMVIXL::SpillScratch(int reg) {
6911 __ Push(vixl32::Register(reg));
6912 }
6913
RestoreScratch(int reg)6914 void ParallelMoveResolverARMVIXL::RestoreScratch(int reg) {
6915 __ Pop(vixl32::Register(reg));
6916 }
6917
GetSupportedLoadClassKind(HLoadClass::LoadKind desired_class_load_kind)6918 HLoadClass::LoadKind CodeGeneratorARMVIXL::GetSupportedLoadClassKind(
6919 HLoadClass::LoadKind desired_class_load_kind) {
6920 switch (desired_class_load_kind) {
6921 case HLoadClass::LoadKind::kInvalid:
6922 LOG(FATAL) << "UNREACHABLE";
6923 UNREACHABLE();
6924 case HLoadClass::LoadKind::kReferrersClass:
6925 break;
6926 case HLoadClass::LoadKind::kBootImageLinkTimePcRelative:
6927 case HLoadClass::LoadKind::kBootImageRelRo:
6928 case HLoadClass::LoadKind::kBssEntry:
6929 DCHECK(!Runtime::Current()->UseJitCompilation());
6930 break;
6931 case HLoadClass::LoadKind::kJitBootImageAddress:
6932 case HLoadClass::LoadKind::kJitTableAddress:
6933 DCHECK(Runtime::Current()->UseJitCompilation());
6934 break;
6935 case HLoadClass::LoadKind::kRuntimeCall:
6936 break;
6937 }
6938 return desired_class_load_kind;
6939 }
6940
VisitLoadClass(HLoadClass * cls)6941 void LocationsBuilderARMVIXL::VisitLoadClass(HLoadClass* cls) {
6942 HLoadClass::LoadKind load_kind = cls->GetLoadKind();
6943 if (load_kind == HLoadClass::LoadKind::kRuntimeCall) {
6944 InvokeRuntimeCallingConventionARMVIXL calling_convention;
6945 CodeGenerator::CreateLoadClassRuntimeCallLocationSummary(
6946 cls,
6947 LocationFrom(calling_convention.GetRegisterAt(0)),
6948 LocationFrom(r0));
6949 DCHECK(calling_convention.GetRegisterAt(0).Is(r0));
6950 return;
6951 }
6952 DCHECK(!cls->NeedsAccessCheck());
6953
6954 const bool requires_read_barrier = kEmitCompilerReadBarrier && !cls->IsInBootImage();
6955 LocationSummary::CallKind call_kind = (cls->NeedsEnvironment() || requires_read_barrier)
6956 ? LocationSummary::kCallOnSlowPath
6957 : LocationSummary::kNoCall;
6958 LocationSummary* locations = new (GetGraph()->GetAllocator()) LocationSummary(cls, call_kind);
6959 if (kUseBakerReadBarrier && requires_read_barrier && !cls->NeedsEnvironment()) {
6960 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers.
6961 }
6962
6963 if (load_kind == HLoadClass::LoadKind::kReferrersClass) {
6964 locations->SetInAt(0, Location::RequiresRegister());
6965 }
6966 locations->SetOut(Location::RequiresRegister());
6967 if (load_kind == HLoadClass::LoadKind::kBssEntry) {
6968 if (!kUseReadBarrier || kUseBakerReadBarrier) {
6969 // Rely on the type resolution or initialization and marking to save everything we need.
6970 locations->SetCustomSlowPathCallerSaves(OneRegInReferenceOutSaveEverythingCallerSaves());
6971 } else {
6972 // For non-Baker read barrier we have a temp-clobbering call.
6973 }
6974 }
6975 }
6976
6977 // NO_THREAD_SAFETY_ANALYSIS as we manipulate handles whose internal object we know does not
6978 // move.
VisitLoadClass(HLoadClass * cls)6979 void InstructionCodeGeneratorARMVIXL::VisitLoadClass(HLoadClass* cls) NO_THREAD_SAFETY_ANALYSIS {
6980 HLoadClass::LoadKind load_kind = cls->GetLoadKind();
6981 if (load_kind == HLoadClass::LoadKind::kRuntimeCall) {
6982 codegen_->GenerateLoadClassRuntimeCall(cls);
6983 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 14);
6984 return;
6985 }
6986 DCHECK(!cls->NeedsAccessCheck());
6987
6988 LocationSummary* locations = cls->GetLocations();
6989 Location out_loc = locations->Out();
6990 vixl32::Register out = OutputRegister(cls);
6991
6992 const ReadBarrierOption read_barrier_option = cls->IsInBootImage()
6993 ? kWithoutReadBarrier
6994 : kCompilerReadBarrierOption;
6995 bool generate_null_check = false;
6996 switch (load_kind) {
6997 case HLoadClass::LoadKind::kReferrersClass: {
6998 DCHECK(!cls->CanCallRuntime());
6999 DCHECK(!cls->MustGenerateClinitCheck());
7000 // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_
7001 vixl32::Register current_method = InputRegisterAt(cls, 0);
7002 codegen_->GenerateGcRootFieldLoad(cls,
7003 out_loc,
7004 current_method,
7005 ArtMethod::DeclaringClassOffset().Int32Value(),
7006 read_barrier_option);
7007 break;
7008 }
7009 case HLoadClass::LoadKind::kBootImageLinkTimePcRelative: {
7010 DCHECK(codegen_->GetCompilerOptions().IsBootImage());
7011 DCHECK_EQ(read_barrier_option, kWithoutReadBarrier);
7012 CodeGeneratorARMVIXL::PcRelativePatchInfo* labels =
7013 codegen_->NewBootImageTypePatch(cls->GetDexFile(), cls->GetTypeIndex());
7014 codegen_->EmitMovwMovtPlaceholder(labels, out);
7015 break;
7016 }
7017 case HLoadClass::LoadKind::kBootImageRelRo: {
7018 DCHECK(!codegen_->GetCompilerOptions().IsBootImage());
7019 CodeGeneratorARMVIXL::PcRelativePatchInfo* labels =
7020 codegen_->NewBootImageRelRoPatch(codegen_->GetBootImageOffset(cls));
7021 codegen_->EmitMovwMovtPlaceholder(labels, out);
7022 __ Ldr(out, MemOperand(out, /* offset= */ 0));
7023 break;
7024 }
7025 case HLoadClass::LoadKind::kBssEntry: {
7026 CodeGeneratorARMVIXL::PcRelativePatchInfo* labels =
7027 codegen_->NewTypeBssEntryPatch(cls->GetDexFile(), cls->GetTypeIndex());
7028 codegen_->EmitMovwMovtPlaceholder(labels, out);
7029 codegen_->GenerateGcRootFieldLoad(cls, out_loc, out, /* offset= */ 0, read_barrier_option);
7030 generate_null_check = true;
7031 break;
7032 }
7033 case HLoadClass::LoadKind::kJitBootImageAddress: {
7034 DCHECK_EQ(read_barrier_option, kWithoutReadBarrier);
7035 uint32_t address = reinterpret_cast32<uint32_t>(cls->GetClass().Get());
7036 DCHECK_NE(address, 0u);
7037 __ Ldr(out, codegen_->DeduplicateBootImageAddressLiteral(address));
7038 break;
7039 }
7040 case HLoadClass::LoadKind::kJitTableAddress: {
7041 __ Ldr(out, codegen_->DeduplicateJitClassLiteral(cls->GetDexFile(),
7042 cls->GetTypeIndex(),
7043 cls->GetClass()));
7044 // /* GcRoot<mirror::Class> */ out = *out
7045 codegen_->GenerateGcRootFieldLoad(cls, out_loc, out, /* offset= */ 0, read_barrier_option);
7046 break;
7047 }
7048 case HLoadClass::LoadKind::kRuntimeCall:
7049 case HLoadClass::LoadKind::kInvalid:
7050 LOG(FATAL) << "UNREACHABLE";
7051 UNREACHABLE();
7052 }
7053
7054 if (generate_null_check || cls->MustGenerateClinitCheck()) {
7055 DCHECK(cls->CanCallRuntime());
7056 LoadClassSlowPathARMVIXL* slow_path =
7057 new (codegen_->GetScopedAllocator()) LoadClassSlowPathARMVIXL(cls, cls);
7058 codegen_->AddSlowPath(slow_path);
7059 if (generate_null_check) {
7060 __ CompareAndBranchIfZero(out, slow_path->GetEntryLabel());
7061 }
7062 if (cls->MustGenerateClinitCheck()) {
7063 GenerateClassInitializationCheck(slow_path, out);
7064 } else {
7065 __ Bind(slow_path->GetExitLabel());
7066 }
7067 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 15);
7068 }
7069 }
7070
VisitLoadMethodHandle(HLoadMethodHandle * load)7071 void LocationsBuilderARMVIXL::VisitLoadMethodHandle(HLoadMethodHandle* load) {
7072 InvokeRuntimeCallingConventionARMVIXL calling_convention;
7073 Location location = LocationFrom(calling_convention.GetRegisterAt(0));
7074 CodeGenerator::CreateLoadMethodHandleRuntimeCallLocationSummary(load, location, location);
7075 }
7076
VisitLoadMethodHandle(HLoadMethodHandle * load)7077 void InstructionCodeGeneratorARMVIXL::VisitLoadMethodHandle(HLoadMethodHandle* load) {
7078 codegen_->GenerateLoadMethodHandleRuntimeCall(load);
7079 }
7080
VisitLoadMethodType(HLoadMethodType * load)7081 void LocationsBuilderARMVIXL::VisitLoadMethodType(HLoadMethodType* load) {
7082 InvokeRuntimeCallingConventionARMVIXL calling_convention;
7083 Location location = LocationFrom(calling_convention.GetRegisterAt(0));
7084 CodeGenerator::CreateLoadMethodTypeRuntimeCallLocationSummary(load, location, location);
7085 }
7086
VisitLoadMethodType(HLoadMethodType * load)7087 void InstructionCodeGeneratorARMVIXL::VisitLoadMethodType(HLoadMethodType* load) {
7088 codegen_->GenerateLoadMethodTypeRuntimeCall(load);
7089 }
7090
VisitClinitCheck(HClinitCheck * check)7091 void LocationsBuilderARMVIXL::VisitClinitCheck(HClinitCheck* check) {
7092 LocationSummary* locations =
7093 new (GetGraph()->GetAllocator()) LocationSummary(check, LocationSummary::kCallOnSlowPath);
7094 locations->SetInAt(0, Location::RequiresRegister());
7095 if (check->HasUses()) {
7096 locations->SetOut(Location::SameAsFirstInput());
7097 }
7098 // Rely on the type initialization to save everything we need.
7099 locations->SetCustomSlowPathCallerSaves(OneRegInReferenceOutSaveEverythingCallerSaves());
7100 }
7101
VisitClinitCheck(HClinitCheck * check)7102 void InstructionCodeGeneratorARMVIXL::VisitClinitCheck(HClinitCheck* check) {
7103 // We assume the class is not null.
7104 LoadClassSlowPathARMVIXL* slow_path =
7105 new (codegen_->GetScopedAllocator()) LoadClassSlowPathARMVIXL(check->GetLoadClass(), check);
7106 codegen_->AddSlowPath(slow_path);
7107 GenerateClassInitializationCheck(slow_path, InputRegisterAt(check, 0));
7108 }
7109
GenerateClassInitializationCheck(LoadClassSlowPathARMVIXL * slow_path,vixl32::Register class_reg)7110 void InstructionCodeGeneratorARMVIXL::GenerateClassInitializationCheck(
7111 LoadClassSlowPathARMVIXL* slow_path, vixl32::Register class_reg) {
7112 UseScratchRegisterScope temps(GetVIXLAssembler());
7113 vixl32::Register temp = temps.Acquire();
7114 constexpr size_t status_lsb_position = SubtypeCheckBits::BitStructSizeOf();
7115 const size_t status_byte_offset =
7116 mirror::Class::StatusOffset().SizeValue() + (status_lsb_position / kBitsPerByte);
7117 constexpr uint32_t shifted_initialized_value =
7118 enum_cast<uint32_t>(ClassStatus::kInitialized) << (status_lsb_position % kBitsPerByte);
7119
7120 GetAssembler()->LoadFromOffset(kLoadUnsignedByte, temp, class_reg, status_byte_offset);
7121 __ Cmp(temp, shifted_initialized_value);
7122 __ B(lo, slow_path->GetEntryLabel());
7123 // Even if the initialized flag is set, we may be in a situation where caches are not synced
7124 // properly. Therefore, we do a memory fence.
7125 __ Dmb(ISH);
7126 __ Bind(slow_path->GetExitLabel());
7127 }
7128
GenerateBitstringTypeCheckCompare(HTypeCheckInstruction * check,vixl32::Register temp,vixl32::FlagsUpdate flags_update)7129 void InstructionCodeGeneratorARMVIXL::GenerateBitstringTypeCheckCompare(
7130 HTypeCheckInstruction* check,
7131 vixl32::Register temp,
7132 vixl32::FlagsUpdate flags_update) {
7133 uint32_t path_to_root = check->GetBitstringPathToRoot();
7134 uint32_t mask = check->GetBitstringMask();
7135 DCHECK(IsPowerOfTwo(mask + 1));
7136 size_t mask_bits = WhichPowerOf2(mask + 1);
7137
7138 // Note that HInstanceOf shall check for zero value in `temp` but HCheckCast needs
7139 // the Z flag for BNE. This is indicated by the `flags_update` parameter.
7140 if (mask_bits == 16u) {
7141 // Load only the bitstring part of the status word.
7142 __ Ldrh(temp, MemOperand(temp, mirror::Class::StatusOffset().Int32Value()));
7143 // Check if the bitstring bits are equal to `path_to_root`.
7144 if (flags_update == SetFlags) {
7145 __ Cmp(temp, path_to_root);
7146 } else {
7147 __ Sub(temp, temp, path_to_root);
7148 }
7149 } else {
7150 // /* uint32_t */ temp = temp->status_
7151 __ Ldr(temp, MemOperand(temp, mirror::Class::StatusOffset().Int32Value()));
7152 if (GetAssembler()->ShifterOperandCanHold(SUB, path_to_root)) {
7153 // Compare the bitstring bits using SUB.
7154 __ Sub(temp, temp, path_to_root);
7155 // Shift out bits that do not contribute to the comparison.
7156 __ Lsl(flags_update, temp, temp, dchecked_integral_cast<uint32_t>(32u - mask_bits));
7157 } else if (IsUint<16>(path_to_root)) {
7158 if (temp.IsLow()) {
7159 // Note: Optimized for size but contains one more dependent instruction than necessary.
7160 // MOVW+SUB(register) would be 8 bytes unless we find a low-reg temporary but the
7161 // macro assembler would use the high reg IP for the constant by default.
7162 // Compare the bitstring bits using SUB.
7163 __ Sub(temp, temp, path_to_root & 0x00ffu); // 16-bit SUB (immediate) T2
7164 __ Sub(temp, temp, path_to_root & 0xff00u); // 32-bit SUB (immediate) T3
7165 // Shift out bits that do not contribute to the comparison.
7166 __ Lsl(flags_update, temp, temp, dchecked_integral_cast<uint32_t>(32u - mask_bits));
7167 } else {
7168 // Extract the bitstring bits.
7169 __ Ubfx(temp, temp, 0, mask_bits);
7170 // Check if the bitstring bits are equal to `path_to_root`.
7171 if (flags_update == SetFlags) {
7172 __ Cmp(temp, path_to_root);
7173 } else {
7174 __ Sub(temp, temp, path_to_root);
7175 }
7176 }
7177 } else {
7178 // Shift out bits that do not contribute to the comparison.
7179 __ Lsl(temp, temp, dchecked_integral_cast<uint32_t>(32u - mask_bits));
7180 // Check if the shifted bitstring bits are equal to `path_to_root << (32u - mask_bits)`.
7181 if (flags_update == SetFlags) {
7182 __ Cmp(temp, path_to_root << (32u - mask_bits));
7183 } else {
7184 __ Sub(temp, temp, path_to_root << (32u - mask_bits));
7185 }
7186 }
7187 }
7188 }
7189
GetSupportedLoadStringKind(HLoadString::LoadKind desired_string_load_kind)7190 HLoadString::LoadKind CodeGeneratorARMVIXL::GetSupportedLoadStringKind(
7191 HLoadString::LoadKind desired_string_load_kind) {
7192 switch (desired_string_load_kind) {
7193 case HLoadString::LoadKind::kBootImageLinkTimePcRelative:
7194 case HLoadString::LoadKind::kBootImageRelRo:
7195 case HLoadString::LoadKind::kBssEntry:
7196 DCHECK(!Runtime::Current()->UseJitCompilation());
7197 break;
7198 case HLoadString::LoadKind::kJitBootImageAddress:
7199 case HLoadString::LoadKind::kJitTableAddress:
7200 DCHECK(Runtime::Current()->UseJitCompilation());
7201 break;
7202 case HLoadString::LoadKind::kRuntimeCall:
7203 break;
7204 }
7205 return desired_string_load_kind;
7206 }
7207
VisitLoadString(HLoadString * load)7208 void LocationsBuilderARMVIXL::VisitLoadString(HLoadString* load) {
7209 LocationSummary::CallKind call_kind = CodeGenerator::GetLoadStringCallKind(load);
7210 LocationSummary* locations = new (GetGraph()->GetAllocator()) LocationSummary(load, call_kind);
7211 HLoadString::LoadKind load_kind = load->GetLoadKind();
7212 if (load_kind == HLoadString::LoadKind::kRuntimeCall) {
7213 locations->SetOut(LocationFrom(r0));
7214 } else {
7215 locations->SetOut(Location::RequiresRegister());
7216 if (load_kind == HLoadString::LoadKind::kBssEntry) {
7217 if (!kUseReadBarrier || kUseBakerReadBarrier) {
7218 // Rely on the pResolveString and marking to save everything we need, including temps.
7219 locations->SetCustomSlowPathCallerSaves(OneRegInReferenceOutSaveEverythingCallerSaves());
7220 } else {
7221 // For non-Baker read barrier we have a temp-clobbering call.
7222 }
7223 }
7224 }
7225 }
7226
7227 // NO_THREAD_SAFETY_ANALYSIS as we manipulate handles whose internal object we know does not
7228 // move.
VisitLoadString(HLoadString * load)7229 void InstructionCodeGeneratorARMVIXL::VisitLoadString(HLoadString* load) NO_THREAD_SAFETY_ANALYSIS {
7230 LocationSummary* locations = load->GetLocations();
7231 Location out_loc = locations->Out();
7232 vixl32::Register out = OutputRegister(load);
7233 HLoadString::LoadKind load_kind = load->GetLoadKind();
7234
7235 switch (load_kind) {
7236 case HLoadString::LoadKind::kBootImageLinkTimePcRelative: {
7237 DCHECK(codegen_->GetCompilerOptions().IsBootImage());
7238 CodeGeneratorARMVIXL::PcRelativePatchInfo* labels =
7239 codegen_->NewBootImageStringPatch(load->GetDexFile(), load->GetStringIndex());
7240 codegen_->EmitMovwMovtPlaceholder(labels, out);
7241 return;
7242 }
7243 case HLoadString::LoadKind::kBootImageRelRo: {
7244 DCHECK(!codegen_->GetCompilerOptions().IsBootImage());
7245 CodeGeneratorARMVIXL::PcRelativePatchInfo* labels =
7246 codegen_->NewBootImageRelRoPatch(codegen_->GetBootImageOffset(load));
7247 codegen_->EmitMovwMovtPlaceholder(labels, out);
7248 __ Ldr(out, MemOperand(out, /* offset= */ 0));
7249 return;
7250 }
7251 case HLoadString::LoadKind::kBssEntry: {
7252 CodeGeneratorARMVIXL::PcRelativePatchInfo* labels =
7253 codegen_->NewStringBssEntryPatch(load->GetDexFile(), load->GetStringIndex());
7254 codegen_->EmitMovwMovtPlaceholder(labels, out);
7255 codegen_->GenerateGcRootFieldLoad(
7256 load, out_loc, out, /* offset= */ 0, kCompilerReadBarrierOption);
7257 LoadStringSlowPathARMVIXL* slow_path =
7258 new (codegen_->GetScopedAllocator()) LoadStringSlowPathARMVIXL(load);
7259 codegen_->AddSlowPath(slow_path);
7260 __ CompareAndBranchIfZero(out, slow_path->GetEntryLabel());
7261 __ Bind(slow_path->GetExitLabel());
7262 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 16);
7263 return;
7264 }
7265 case HLoadString::LoadKind::kJitBootImageAddress: {
7266 uint32_t address = reinterpret_cast32<uint32_t>(load->GetString().Get());
7267 DCHECK_NE(address, 0u);
7268 __ Ldr(out, codegen_->DeduplicateBootImageAddressLiteral(address));
7269 return;
7270 }
7271 case HLoadString::LoadKind::kJitTableAddress: {
7272 __ Ldr(out, codegen_->DeduplicateJitStringLiteral(load->GetDexFile(),
7273 load->GetStringIndex(),
7274 load->GetString()));
7275 // /* GcRoot<mirror::String> */ out = *out
7276 codegen_->GenerateGcRootFieldLoad(
7277 load, out_loc, out, /* offset= */ 0, kCompilerReadBarrierOption);
7278 return;
7279 }
7280 default:
7281 break;
7282 }
7283
7284 // TODO: Re-add the compiler code to do string dex cache lookup again.
7285 DCHECK_EQ(load->GetLoadKind(), HLoadString::LoadKind::kRuntimeCall);
7286 InvokeRuntimeCallingConventionARMVIXL calling_convention;
7287 __ Mov(calling_convention.GetRegisterAt(0), load->GetStringIndex().index_);
7288 codegen_->InvokeRuntime(kQuickResolveString, load, load->GetDexPc());
7289 CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>();
7290 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 17);
7291 }
7292
GetExceptionTlsOffset()7293 static int32_t GetExceptionTlsOffset() {
7294 return Thread::ExceptionOffset<kArmPointerSize>().Int32Value();
7295 }
7296
VisitLoadException(HLoadException * load)7297 void LocationsBuilderARMVIXL::VisitLoadException(HLoadException* load) {
7298 LocationSummary* locations =
7299 new (GetGraph()->GetAllocator()) LocationSummary(load, LocationSummary::kNoCall);
7300 locations->SetOut(Location::RequiresRegister());
7301 }
7302
VisitLoadException(HLoadException * load)7303 void InstructionCodeGeneratorARMVIXL::VisitLoadException(HLoadException* load) {
7304 vixl32::Register out = OutputRegister(load);
7305 GetAssembler()->LoadFromOffset(kLoadWord, out, tr, GetExceptionTlsOffset());
7306 }
7307
7308
VisitClearException(HClearException * clear)7309 void LocationsBuilderARMVIXL::VisitClearException(HClearException* clear) {
7310 new (GetGraph()->GetAllocator()) LocationSummary(clear, LocationSummary::kNoCall);
7311 }
7312
VisitClearException(HClearException * clear ATTRIBUTE_UNUSED)7313 void InstructionCodeGeneratorARMVIXL::VisitClearException(HClearException* clear ATTRIBUTE_UNUSED) {
7314 UseScratchRegisterScope temps(GetVIXLAssembler());
7315 vixl32::Register temp = temps.Acquire();
7316 __ Mov(temp, 0);
7317 GetAssembler()->StoreToOffset(kStoreWord, temp, tr, GetExceptionTlsOffset());
7318 }
7319
VisitThrow(HThrow * instruction)7320 void LocationsBuilderARMVIXL::VisitThrow(HThrow* instruction) {
7321 LocationSummary* locations = new (GetGraph()->GetAllocator()) LocationSummary(
7322 instruction, LocationSummary::kCallOnMainOnly);
7323 InvokeRuntimeCallingConventionARMVIXL calling_convention;
7324 locations->SetInAt(0, LocationFrom(calling_convention.GetRegisterAt(0)));
7325 }
7326
VisitThrow(HThrow * instruction)7327 void InstructionCodeGeneratorARMVIXL::VisitThrow(HThrow* instruction) {
7328 codegen_->InvokeRuntime(kQuickDeliverException, instruction, instruction->GetDexPc());
7329 CheckEntrypointTypes<kQuickDeliverException, void, mirror::Object*>();
7330 }
7331
7332 // Temp is used for read barrier.
NumberOfInstanceOfTemps(TypeCheckKind type_check_kind)7333 static size_t NumberOfInstanceOfTemps(TypeCheckKind type_check_kind) {
7334 if (kEmitCompilerReadBarrier &&
7335 (kUseBakerReadBarrier ||
7336 type_check_kind == TypeCheckKind::kAbstractClassCheck ||
7337 type_check_kind == TypeCheckKind::kClassHierarchyCheck ||
7338 type_check_kind == TypeCheckKind::kArrayObjectCheck)) {
7339 return 1;
7340 }
7341 return 0;
7342 }
7343
7344 // Interface case has 3 temps, one for holding the number of interfaces, one for the current
7345 // interface pointer, one for loading the current interface.
7346 // The other checks have one temp for loading the object's class.
NumberOfCheckCastTemps(TypeCheckKind type_check_kind)7347 static size_t NumberOfCheckCastTemps(TypeCheckKind type_check_kind) {
7348 if (type_check_kind == TypeCheckKind::kInterfaceCheck) {
7349 return 3;
7350 }
7351 return 1 + NumberOfInstanceOfTemps(type_check_kind);
7352 }
7353
VisitInstanceOf(HInstanceOf * instruction)7354 void LocationsBuilderARMVIXL::VisitInstanceOf(HInstanceOf* instruction) {
7355 LocationSummary::CallKind call_kind = LocationSummary::kNoCall;
7356 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind();
7357 bool baker_read_barrier_slow_path = false;
7358 switch (type_check_kind) {
7359 case TypeCheckKind::kExactCheck:
7360 case TypeCheckKind::kAbstractClassCheck:
7361 case TypeCheckKind::kClassHierarchyCheck:
7362 case TypeCheckKind::kArrayObjectCheck: {
7363 bool needs_read_barrier = CodeGenerator::InstanceOfNeedsReadBarrier(instruction);
7364 call_kind = needs_read_barrier ? LocationSummary::kCallOnSlowPath : LocationSummary::kNoCall;
7365 baker_read_barrier_slow_path = kUseBakerReadBarrier && needs_read_barrier;
7366 break;
7367 }
7368 case TypeCheckKind::kArrayCheck:
7369 case TypeCheckKind::kUnresolvedCheck:
7370 case TypeCheckKind::kInterfaceCheck:
7371 call_kind = LocationSummary::kCallOnSlowPath;
7372 break;
7373 case TypeCheckKind::kBitstringCheck:
7374 break;
7375 }
7376
7377 LocationSummary* locations =
7378 new (GetGraph()->GetAllocator()) LocationSummary(instruction, call_kind);
7379 if (baker_read_barrier_slow_path) {
7380 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers.
7381 }
7382 locations->SetInAt(0, Location::RequiresRegister());
7383 if (type_check_kind == TypeCheckKind::kBitstringCheck) {
7384 locations->SetInAt(1, Location::ConstantLocation(instruction->InputAt(1)->AsConstant()));
7385 locations->SetInAt(2, Location::ConstantLocation(instruction->InputAt(2)->AsConstant()));
7386 locations->SetInAt(3, Location::ConstantLocation(instruction->InputAt(3)->AsConstant()));
7387 } else {
7388 locations->SetInAt(1, Location::RequiresRegister());
7389 }
7390 // The "out" register is used as a temporary, so it overlaps with the inputs.
7391 // Note that TypeCheckSlowPathARM uses this register too.
7392 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
7393 locations->AddRegisterTemps(NumberOfInstanceOfTemps(type_check_kind));
7394 }
7395
VisitInstanceOf(HInstanceOf * instruction)7396 void InstructionCodeGeneratorARMVIXL::VisitInstanceOf(HInstanceOf* instruction) {
7397 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind();
7398 LocationSummary* locations = instruction->GetLocations();
7399 Location obj_loc = locations->InAt(0);
7400 vixl32::Register obj = InputRegisterAt(instruction, 0);
7401 vixl32::Register cls = (type_check_kind == TypeCheckKind::kBitstringCheck)
7402 ? vixl32::Register()
7403 : InputRegisterAt(instruction, 1);
7404 Location out_loc = locations->Out();
7405 vixl32::Register out = OutputRegister(instruction);
7406 const size_t num_temps = NumberOfInstanceOfTemps(type_check_kind);
7407 DCHECK_LE(num_temps, 1u);
7408 Location maybe_temp_loc = (num_temps >= 1) ? locations->GetTemp(0) : Location::NoLocation();
7409 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
7410 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value();
7411 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value();
7412 uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value();
7413 vixl32::Label done;
7414 vixl32::Label* const final_label = codegen_->GetFinalLabel(instruction, &done);
7415 SlowPathCodeARMVIXL* slow_path = nullptr;
7416
7417 // Return 0 if `obj` is null.
7418 // avoid null check if we know obj is not null.
7419 if (instruction->MustDoNullCheck()) {
7420 DCHECK(!out.Is(obj));
7421 __ Mov(out, 0);
7422 __ CompareAndBranchIfZero(obj, final_label, /* is_far_target= */ false);
7423 }
7424
7425 switch (type_check_kind) {
7426 case TypeCheckKind::kExactCheck: {
7427 ReadBarrierOption read_barrier_option =
7428 CodeGenerator::ReadBarrierOptionForInstanceOf(instruction);
7429 // /* HeapReference<Class> */ out = obj->klass_
7430 GenerateReferenceLoadTwoRegisters(instruction,
7431 out_loc,
7432 obj_loc,
7433 class_offset,
7434 maybe_temp_loc,
7435 read_barrier_option);
7436 // Classes must be equal for the instanceof to succeed.
7437 __ Cmp(out, cls);
7438 // We speculatively set the result to false without changing the condition
7439 // flags, which allows us to avoid some branching later.
7440 __ Mov(LeaveFlags, out, 0);
7441
7442 // Since IT blocks longer than a 16-bit instruction are deprecated by ARMv8,
7443 // we check that the output is in a low register, so that a 16-bit MOV
7444 // encoding can be used.
7445 if (out.IsLow()) {
7446 // We use the scope because of the IT block that follows.
7447 ExactAssemblyScope guard(GetVIXLAssembler(),
7448 2 * vixl32::k16BitT32InstructionSizeInBytes,
7449 CodeBufferCheckScope::kExactSize);
7450
7451 __ it(eq);
7452 __ mov(eq, out, 1);
7453 } else {
7454 __ B(ne, final_label, /* is_far_target= */ false);
7455 __ Mov(out, 1);
7456 }
7457
7458 break;
7459 }
7460
7461 case TypeCheckKind::kAbstractClassCheck: {
7462 ReadBarrierOption read_barrier_option =
7463 CodeGenerator::ReadBarrierOptionForInstanceOf(instruction);
7464 // /* HeapReference<Class> */ out = obj->klass_
7465 GenerateReferenceLoadTwoRegisters(instruction,
7466 out_loc,
7467 obj_loc,
7468 class_offset,
7469 maybe_temp_loc,
7470 read_barrier_option);
7471 // If the class is abstract, we eagerly fetch the super class of the
7472 // object to avoid doing a comparison we know will fail.
7473 vixl32::Label loop;
7474 __ Bind(&loop);
7475 // /* HeapReference<Class> */ out = out->super_class_
7476 GenerateReferenceLoadOneRegister(instruction,
7477 out_loc,
7478 super_offset,
7479 maybe_temp_loc,
7480 read_barrier_option);
7481 // If `out` is null, we use it for the result, and jump to the final label.
7482 __ CompareAndBranchIfZero(out, final_label, /* is_far_target= */ false);
7483 __ Cmp(out, cls);
7484 __ B(ne, &loop, /* is_far_target= */ false);
7485 __ Mov(out, 1);
7486 break;
7487 }
7488
7489 case TypeCheckKind::kClassHierarchyCheck: {
7490 ReadBarrierOption read_barrier_option =
7491 CodeGenerator::ReadBarrierOptionForInstanceOf(instruction);
7492 // /* HeapReference<Class> */ out = obj->klass_
7493 GenerateReferenceLoadTwoRegisters(instruction,
7494 out_loc,
7495 obj_loc,
7496 class_offset,
7497 maybe_temp_loc,
7498 read_barrier_option);
7499 // Walk over the class hierarchy to find a match.
7500 vixl32::Label loop, success;
7501 __ Bind(&loop);
7502 __ Cmp(out, cls);
7503 __ B(eq, &success, /* is_far_target= */ false);
7504 // /* HeapReference<Class> */ out = out->super_class_
7505 GenerateReferenceLoadOneRegister(instruction,
7506 out_loc,
7507 super_offset,
7508 maybe_temp_loc,
7509 read_barrier_option);
7510 // This is essentially a null check, but it sets the condition flags to the
7511 // proper value for the code that follows the loop, i.e. not `eq`.
7512 __ Cmp(out, 1);
7513 __ B(hs, &loop, /* is_far_target= */ false);
7514
7515 // Since IT blocks longer than a 16-bit instruction are deprecated by ARMv8,
7516 // we check that the output is in a low register, so that a 16-bit MOV
7517 // encoding can be used.
7518 if (out.IsLow()) {
7519 // If `out` is null, we use it for the result, and the condition flags
7520 // have already been set to `ne`, so the IT block that comes afterwards
7521 // (and which handles the successful case) turns into a NOP (instead of
7522 // overwriting `out`).
7523 __ Bind(&success);
7524
7525 // We use the scope because of the IT block that follows.
7526 ExactAssemblyScope guard(GetVIXLAssembler(),
7527 2 * vixl32::k16BitT32InstructionSizeInBytes,
7528 CodeBufferCheckScope::kExactSize);
7529
7530 // There is only one branch to the `success` label (which is bound to this
7531 // IT block), and it has the same condition, `eq`, so in that case the MOV
7532 // is executed.
7533 __ it(eq);
7534 __ mov(eq, out, 1);
7535 } else {
7536 // If `out` is null, we use it for the result, and jump to the final label.
7537 __ B(final_label);
7538 __ Bind(&success);
7539 __ Mov(out, 1);
7540 }
7541
7542 break;
7543 }
7544
7545 case TypeCheckKind::kArrayObjectCheck: {
7546 ReadBarrierOption read_barrier_option =
7547 CodeGenerator::ReadBarrierOptionForInstanceOf(instruction);
7548 // /* HeapReference<Class> */ out = obj->klass_
7549 GenerateReferenceLoadTwoRegisters(instruction,
7550 out_loc,
7551 obj_loc,
7552 class_offset,
7553 maybe_temp_loc,
7554 read_barrier_option);
7555 // Do an exact check.
7556 vixl32::Label exact_check;
7557 __ Cmp(out, cls);
7558 __ B(eq, &exact_check, /* is_far_target= */ false);
7559 // Otherwise, we need to check that the object's class is a non-primitive array.
7560 // /* HeapReference<Class> */ out = out->component_type_
7561 GenerateReferenceLoadOneRegister(instruction,
7562 out_loc,
7563 component_offset,
7564 maybe_temp_loc,
7565 read_barrier_option);
7566 // If `out` is null, we use it for the result, and jump to the final label.
7567 __ CompareAndBranchIfZero(out, final_label, /* is_far_target= */ false);
7568 GetAssembler()->LoadFromOffset(kLoadUnsignedHalfword, out, out, primitive_offset);
7569 static_assert(Primitive::kPrimNot == 0, "Expected 0 for kPrimNot");
7570 __ Cmp(out, 0);
7571 // We speculatively set the result to false without changing the condition
7572 // flags, which allows us to avoid some branching later.
7573 __ Mov(LeaveFlags, out, 0);
7574
7575 // Since IT blocks longer than a 16-bit instruction are deprecated by ARMv8,
7576 // we check that the output is in a low register, so that a 16-bit MOV
7577 // encoding can be used.
7578 if (out.IsLow()) {
7579 __ Bind(&exact_check);
7580
7581 // We use the scope because of the IT block that follows.
7582 ExactAssemblyScope guard(GetVIXLAssembler(),
7583 2 * vixl32::k16BitT32InstructionSizeInBytes,
7584 CodeBufferCheckScope::kExactSize);
7585
7586 __ it(eq);
7587 __ mov(eq, out, 1);
7588 } else {
7589 __ B(ne, final_label, /* is_far_target= */ false);
7590 __ Bind(&exact_check);
7591 __ Mov(out, 1);
7592 }
7593
7594 break;
7595 }
7596
7597 case TypeCheckKind::kArrayCheck: {
7598 // No read barrier since the slow path will retry upon failure.
7599 // /* HeapReference<Class> */ out = obj->klass_
7600 GenerateReferenceLoadTwoRegisters(instruction,
7601 out_loc,
7602 obj_loc,
7603 class_offset,
7604 maybe_temp_loc,
7605 kWithoutReadBarrier);
7606 __ Cmp(out, cls);
7607 DCHECK(locations->OnlyCallsOnSlowPath());
7608 slow_path = new (codegen_->GetScopedAllocator()) TypeCheckSlowPathARMVIXL(
7609 instruction, /* is_fatal= */ false);
7610 codegen_->AddSlowPath(slow_path);
7611 __ B(ne, slow_path->GetEntryLabel());
7612 __ Mov(out, 1);
7613 break;
7614 }
7615
7616 case TypeCheckKind::kUnresolvedCheck:
7617 case TypeCheckKind::kInterfaceCheck: {
7618 // Note that we indeed only call on slow path, but we always go
7619 // into the slow path for the unresolved and interface check
7620 // cases.
7621 //
7622 // We cannot directly call the InstanceofNonTrivial runtime
7623 // entry point without resorting to a type checking slow path
7624 // here (i.e. by calling InvokeRuntime directly), as it would
7625 // require to assign fixed registers for the inputs of this
7626 // HInstanceOf instruction (following the runtime calling
7627 // convention), which might be cluttered by the potential first
7628 // read barrier emission at the beginning of this method.
7629 //
7630 // TODO: Introduce a new runtime entry point taking the object
7631 // to test (instead of its class) as argument, and let it deal
7632 // with the read barrier issues. This will let us refactor this
7633 // case of the `switch` code as it was previously (with a direct
7634 // call to the runtime not using a type checking slow path).
7635 // This should also be beneficial for the other cases above.
7636 DCHECK(locations->OnlyCallsOnSlowPath());
7637 slow_path = new (codegen_->GetScopedAllocator()) TypeCheckSlowPathARMVIXL(
7638 instruction, /* is_fatal= */ false);
7639 codegen_->AddSlowPath(slow_path);
7640 __ B(slow_path->GetEntryLabel());
7641 break;
7642 }
7643
7644 case TypeCheckKind::kBitstringCheck: {
7645 // /* HeapReference<Class> */ temp = obj->klass_
7646 GenerateReferenceLoadTwoRegisters(instruction,
7647 out_loc,
7648 obj_loc,
7649 class_offset,
7650 maybe_temp_loc,
7651 kWithoutReadBarrier);
7652
7653 GenerateBitstringTypeCheckCompare(instruction, out, DontCare);
7654 // If `out` is a low reg and we would have another low reg temp, we could
7655 // optimize this as RSBS+ADC, see GenerateConditionWithZero().
7656 //
7657 // Also, in some cases when `out` is a low reg and we're loading a constant to IP
7658 // it would make sense to use CMP+MOV+IT+MOV instead of SUB+CLZ+LSR as the code size
7659 // would be the same and we would have fewer direct data dependencies.
7660 codegen_->GenerateConditionWithZero(kCondEQ, out, out); // CLZ+LSR
7661 break;
7662 }
7663 }
7664
7665 if (done.IsReferenced()) {
7666 __ Bind(&done);
7667 }
7668
7669 if (slow_path != nullptr) {
7670 __ Bind(slow_path->GetExitLabel());
7671 }
7672 }
7673
VisitCheckCast(HCheckCast * instruction)7674 void LocationsBuilderARMVIXL::VisitCheckCast(HCheckCast* instruction) {
7675 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind();
7676 LocationSummary::CallKind call_kind = CodeGenerator::GetCheckCastCallKind(instruction);
7677 LocationSummary* locations =
7678 new (GetGraph()->GetAllocator()) LocationSummary(instruction, call_kind);
7679 locations->SetInAt(0, Location::RequiresRegister());
7680 if (type_check_kind == TypeCheckKind::kBitstringCheck) {
7681 locations->SetInAt(1, Location::ConstantLocation(instruction->InputAt(1)->AsConstant()));
7682 locations->SetInAt(2, Location::ConstantLocation(instruction->InputAt(2)->AsConstant()));
7683 locations->SetInAt(3, Location::ConstantLocation(instruction->InputAt(3)->AsConstant()));
7684 } else {
7685 locations->SetInAt(1, Location::RequiresRegister());
7686 }
7687 locations->AddRegisterTemps(NumberOfCheckCastTemps(type_check_kind));
7688 }
7689
VisitCheckCast(HCheckCast * instruction)7690 void InstructionCodeGeneratorARMVIXL::VisitCheckCast(HCheckCast* instruction) {
7691 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind();
7692 LocationSummary* locations = instruction->GetLocations();
7693 Location obj_loc = locations->InAt(0);
7694 vixl32::Register obj = InputRegisterAt(instruction, 0);
7695 vixl32::Register cls = (type_check_kind == TypeCheckKind::kBitstringCheck)
7696 ? vixl32::Register()
7697 : InputRegisterAt(instruction, 1);
7698 Location temp_loc = locations->GetTemp(0);
7699 vixl32::Register temp = RegisterFrom(temp_loc);
7700 const size_t num_temps = NumberOfCheckCastTemps(type_check_kind);
7701 DCHECK_LE(num_temps, 3u);
7702 Location maybe_temp2_loc = (num_temps >= 2) ? locations->GetTemp(1) : Location::NoLocation();
7703 Location maybe_temp3_loc = (num_temps >= 3) ? locations->GetTemp(2) : Location::NoLocation();
7704 const uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
7705 const uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value();
7706 const uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value();
7707 const uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value();
7708 const uint32_t iftable_offset = mirror::Class::IfTableOffset().Uint32Value();
7709 const uint32_t array_length_offset = mirror::Array::LengthOffset().Uint32Value();
7710 const uint32_t object_array_data_offset =
7711 mirror::Array::DataOffset(kHeapReferenceSize).Uint32Value();
7712
7713 bool is_type_check_slow_path_fatal = CodeGenerator::IsTypeCheckSlowPathFatal(instruction);
7714 SlowPathCodeARMVIXL* type_check_slow_path =
7715 new (codegen_->GetScopedAllocator()) TypeCheckSlowPathARMVIXL(
7716 instruction, is_type_check_slow_path_fatal);
7717 codegen_->AddSlowPath(type_check_slow_path);
7718
7719 vixl32::Label done;
7720 vixl32::Label* final_label = codegen_->GetFinalLabel(instruction, &done);
7721 // Avoid null check if we know obj is not null.
7722 if (instruction->MustDoNullCheck()) {
7723 __ CompareAndBranchIfZero(obj, final_label, /* is_far_target= */ false);
7724 }
7725
7726 switch (type_check_kind) {
7727 case TypeCheckKind::kExactCheck:
7728 case TypeCheckKind::kArrayCheck: {
7729 // /* HeapReference<Class> */ temp = obj->klass_
7730 GenerateReferenceLoadTwoRegisters(instruction,
7731 temp_loc,
7732 obj_loc,
7733 class_offset,
7734 maybe_temp2_loc,
7735 kWithoutReadBarrier);
7736
7737 __ Cmp(temp, cls);
7738 // Jump to slow path for throwing the exception or doing a
7739 // more involved array check.
7740 __ B(ne, type_check_slow_path->GetEntryLabel());
7741 break;
7742 }
7743
7744 case TypeCheckKind::kAbstractClassCheck: {
7745 // /* HeapReference<Class> */ temp = obj->klass_
7746 GenerateReferenceLoadTwoRegisters(instruction,
7747 temp_loc,
7748 obj_loc,
7749 class_offset,
7750 maybe_temp2_loc,
7751 kWithoutReadBarrier);
7752
7753 // If the class is abstract, we eagerly fetch the super class of the
7754 // object to avoid doing a comparison we know will fail.
7755 vixl32::Label loop;
7756 __ Bind(&loop);
7757 // /* HeapReference<Class> */ temp = temp->super_class_
7758 GenerateReferenceLoadOneRegister(instruction,
7759 temp_loc,
7760 super_offset,
7761 maybe_temp2_loc,
7762 kWithoutReadBarrier);
7763
7764 // If the class reference currently in `temp` is null, jump to the slow path to throw the
7765 // exception.
7766 __ CompareAndBranchIfZero(temp, type_check_slow_path->GetEntryLabel());
7767
7768 // Otherwise, compare the classes.
7769 __ Cmp(temp, cls);
7770 __ B(ne, &loop, /* is_far_target= */ false);
7771 break;
7772 }
7773
7774 case TypeCheckKind::kClassHierarchyCheck: {
7775 // /* HeapReference<Class> */ temp = obj->klass_
7776 GenerateReferenceLoadTwoRegisters(instruction,
7777 temp_loc,
7778 obj_loc,
7779 class_offset,
7780 maybe_temp2_loc,
7781 kWithoutReadBarrier);
7782
7783 // Walk over the class hierarchy to find a match.
7784 vixl32::Label loop;
7785 __ Bind(&loop);
7786 __ Cmp(temp, cls);
7787 __ B(eq, final_label, /* is_far_target= */ false);
7788
7789 // /* HeapReference<Class> */ temp = temp->super_class_
7790 GenerateReferenceLoadOneRegister(instruction,
7791 temp_loc,
7792 super_offset,
7793 maybe_temp2_loc,
7794 kWithoutReadBarrier);
7795
7796 // If the class reference currently in `temp` is null, jump to the slow path to throw the
7797 // exception.
7798 __ CompareAndBranchIfZero(temp, type_check_slow_path->GetEntryLabel());
7799 // Otherwise, jump to the beginning of the loop.
7800 __ B(&loop);
7801 break;
7802 }
7803
7804 case TypeCheckKind::kArrayObjectCheck: {
7805 // /* HeapReference<Class> */ temp = obj->klass_
7806 GenerateReferenceLoadTwoRegisters(instruction,
7807 temp_loc,
7808 obj_loc,
7809 class_offset,
7810 maybe_temp2_loc,
7811 kWithoutReadBarrier);
7812
7813 // Do an exact check.
7814 __ Cmp(temp, cls);
7815 __ B(eq, final_label, /* is_far_target= */ false);
7816
7817 // Otherwise, we need to check that the object's class is a non-primitive array.
7818 // /* HeapReference<Class> */ temp = temp->component_type_
7819 GenerateReferenceLoadOneRegister(instruction,
7820 temp_loc,
7821 component_offset,
7822 maybe_temp2_loc,
7823 kWithoutReadBarrier);
7824 // If the component type is null, jump to the slow path to throw the exception.
7825 __ CompareAndBranchIfZero(temp, type_check_slow_path->GetEntryLabel());
7826 // Otherwise,the object is indeed an array, jump to label `check_non_primitive_component_type`
7827 // to further check that this component type is not a primitive type.
7828 GetAssembler()->LoadFromOffset(kLoadUnsignedHalfword, temp, temp, primitive_offset);
7829 static_assert(Primitive::kPrimNot == 0, "Expected 0 for kPrimNot");
7830 __ CompareAndBranchIfNonZero(temp, type_check_slow_path->GetEntryLabel());
7831 break;
7832 }
7833
7834 case TypeCheckKind::kUnresolvedCheck:
7835 // We always go into the type check slow path for the unresolved check case.
7836 // We cannot directly call the CheckCast runtime entry point
7837 // without resorting to a type checking slow path here (i.e. by
7838 // calling InvokeRuntime directly), as it would require to
7839 // assign fixed registers for the inputs of this HInstanceOf
7840 // instruction (following the runtime calling convention), which
7841 // might be cluttered by the potential first read barrier
7842 // emission at the beginning of this method.
7843
7844 __ B(type_check_slow_path->GetEntryLabel());
7845 break;
7846
7847 case TypeCheckKind::kInterfaceCheck: {
7848 // Avoid read barriers to improve performance of the fast path. We can not get false
7849 // positives by doing this.
7850 // /* HeapReference<Class> */ temp = obj->klass_
7851 GenerateReferenceLoadTwoRegisters(instruction,
7852 temp_loc,
7853 obj_loc,
7854 class_offset,
7855 maybe_temp2_loc,
7856 kWithoutReadBarrier);
7857
7858 // /* HeapReference<Class> */ temp = temp->iftable_
7859 GenerateReferenceLoadTwoRegisters(instruction,
7860 temp_loc,
7861 temp_loc,
7862 iftable_offset,
7863 maybe_temp2_loc,
7864 kWithoutReadBarrier);
7865 // Iftable is never null.
7866 __ Ldr(RegisterFrom(maybe_temp2_loc), MemOperand(temp, array_length_offset));
7867 // Loop through the iftable and check if any class matches.
7868 vixl32::Label start_loop;
7869 __ Bind(&start_loop);
7870 __ CompareAndBranchIfZero(RegisterFrom(maybe_temp2_loc),
7871 type_check_slow_path->GetEntryLabel());
7872 __ Ldr(RegisterFrom(maybe_temp3_loc), MemOperand(temp, object_array_data_offset));
7873 GetAssembler()->MaybeUnpoisonHeapReference(RegisterFrom(maybe_temp3_loc));
7874 // Go to next interface.
7875 __ Add(temp, temp, Operand::From(2 * kHeapReferenceSize));
7876 __ Sub(RegisterFrom(maybe_temp2_loc), RegisterFrom(maybe_temp2_loc), 2);
7877 // Compare the classes and continue the loop if they do not match.
7878 __ Cmp(cls, RegisterFrom(maybe_temp3_loc));
7879 __ B(ne, &start_loop, /* is_far_target= */ false);
7880 break;
7881 }
7882
7883 case TypeCheckKind::kBitstringCheck: {
7884 // /* HeapReference<Class> */ temp = obj->klass_
7885 GenerateReferenceLoadTwoRegisters(instruction,
7886 temp_loc,
7887 obj_loc,
7888 class_offset,
7889 maybe_temp2_loc,
7890 kWithoutReadBarrier);
7891
7892 GenerateBitstringTypeCheckCompare(instruction, temp, SetFlags);
7893 __ B(ne, type_check_slow_path->GetEntryLabel());
7894 break;
7895 }
7896 }
7897 if (done.IsReferenced()) {
7898 __ Bind(&done);
7899 }
7900
7901 __ Bind(type_check_slow_path->GetExitLabel());
7902 }
7903
VisitMonitorOperation(HMonitorOperation * instruction)7904 void LocationsBuilderARMVIXL::VisitMonitorOperation(HMonitorOperation* instruction) {
7905 LocationSummary* locations = new (GetGraph()->GetAllocator()) LocationSummary(
7906 instruction, LocationSummary::kCallOnMainOnly);
7907 InvokeRuntimeCallingConventionARMVIXL calling_convention;
7908 locations->SetInAt(0, LocationFrom(calling_convention.GetRegisterAt(0)));
7909 }
7910
VisitMonitorOperation(HMonitorOperation * instruction)7911 void InstructionCodeGeneratorARMVIXL::VisitMonitorOperation(HMonitorOperation* instruction) {
7912 codegen_->InvokeRuntime(instruction->IsEnter() ? kQuickLockObject : kQuickUnlockObject,
7913 instruction,
7914 instruction->GetDexPc());
7915 if (instruction->IsEnter()) {
7916 CheckEntrypointTypes<kQuickLockObject, void, mirror::Object*>();
7917 } else {
7918 CheckEntrypointTypes<kQuickUnlockObject, void, mirror::Object*>();
7919 }
7920 codegen_->MaybeGenerateMarkingRegisterCheck(/* code= */ 18);
7921 }
7922
VisitAnd(HAnd * instruction)7923 void LocationsBuilderARMVIXL::VisitAnd(HAnd* instruction) {
7924 HandleBitwiseOperation(instruction, AND);
7925 }
7926
VisitOr(HOr * instruction)7927 void LocationsBuilderARMVIXL::VisitOr(HOr* instruction) {
7928 HandleBitwiseOperation(instruction, ORR);
7929 }
7930
VisitXor(HXor * instruction)7931 void LocationsBuilderARMVIXL::VisitXor(HXor* instruction) {
7932 HandleBitwiseOperation(instruction, EOR);
7933 }
7934
HandleBitwiseOperation(HBinaryOperation * instruction,Opcode opcode)7935 void LocationsBuilderARMVIXL::HandleBitwiseOperation(HBinaryOperation* instruction, Opcode opcode) {
7936 LocationSummary* locations =
7937 new (GetGraph()->GetAllocator()) LocationSummary(instruction, LocationSummary::kNoCall);
7938 DCHECK(instruction->GetResultType() == DataType::Type::kInt32
7939 || instruction->GetResultType() == DataType::Type::kInt64);
7940 // Note: GVN reorders commutative operations to have the constant on the right hand side.
7941 locations->SetInAt(0, Location::RequiresRegister());
7942 locations->SetInAt(1, ArmEncodableConstantOrRegister(instruction->InputAt(1), opcode));
7943 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
7944 }
7945
VisitAnd(HAnd * instruction)7946 void InstructionCodeGeneratorARMVIXL::VisitAnd(HAnd* instruction) {
7947 HandleBitwiseOperation(instruction);
7948 }
7949
VisitOr(HOr * instruction)7950 void InstructionCodeGeneratorARMVIXL::VisitOr(HOr* instruction) {
7951 HandleBitwiseOperation(instruction);
7952 }
7953
VisitXor(HXor * instruction)7954 void InstructionCodeGeneratorARMVIXL::VisitXor(HXor* instruction) {
7955 HandleBitwiseOperation(instruction);
7956 }
7957
VisitBitwiseNegatedRight(HBitwiseNegatedRight * instruction)7958 void LocationsBuilderARMVIXL::VisitBitwiseNegatedRight(HBitwiseNegatedRight* instruction) {
7959 LocationSummary* locations =
7960 new (GetGraph()->GetAllocator()) LocationSummary(instruction, LocationSummary::kNoCall);
7961 DCHECK(instruction->GetResultType() == DataType::Type::kInt32
7962 || instruction->GetResultType() == DataType::Type::kInt64);
7963
7964 locations->SetInAt(0, Location::RequiresRegister());
7965 locations->SetInAt(1, Location::RequiresRegister());
7966 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
7967 }
7968
VisitBitwiseNegatedRight(HBitwiseNegatedRight * instruction)7969 void InstructionCodeGeneratorARMVIXL::VisitBitwiseNegatedRight(HBitwiseNegatedRight* instruction) {
7970 LocationSummary* locations = instruction->GetLocations();
7971 Location first = locations->InAt(0);
7972 Location second = locations->InAt(1);
7973 Location out = locations->Out();
7974
7975 if (instruction->GetResultType() == DataType::Type::kInt32) {
7976 vixl32::Register first_reg = RegisterFrom(first);
7977 vixl32::Register second_reg = RegisterFrom(second);
7978 vixl32::Register out_reg = RegisterFrom(out);
7979
7980 switch (instruction->GetOpKind()) {
7981 case HInstruction::kAnd:
7982 __ Bic(out_reg, first_reg, second_reg);
7983 break;
7984 case HInstruction::kOr:
7985 __ Orn(out_reg, first_reg, second_reg);
7986 break;
7987 // There is no EON on arm.
7988 case HInstruction::kXor:
7989 default:
7990 LOG(FATAL) << "Unexpected instruction " << instruction->DebugName();
7991 UNREACHABLE();
7992 }
7993 return;
7994
7995 } else {
7996 DCHECK_EQ(instruction->GetResultType(), DataType::Type::kInt64);
7997 vixl32::Register first_low = LowRegisterFrom(first);
7998 vixl32::Register first_high = HighRegisterFrom(first);
7999 vixl32::Register second_low = LowRegisterFrom(second);
8000 vixl32::Register second_high = HighRegisterFrom(second);
8001 vixl32::Register out_low = LowRegisterFrom(out);
8002 vixl32::Register out_high = HighRegisterFrom(out);
8003
8004 switch (instruction->GetOpKind()) {
8005 case HInstruction::kAnd:
8006 __ Bic(out_low, first_low, second_low);
8007 __ Bic(out_high, first_high, second_high);
8008 break;
8009 case HInstruction::kOr:
8010 __ Orn(out_low, first_low, second_low);
8011 __ Orn(out_high, first_high, second_high);
8012 break;
8013 // There is no EON on arm.
8014 case HInstruction::kXor:
8015 default:
8016 LOG(FATAL) << "Unexpected instruction " << instruction->DebugName();
8017 UNREACHABLE();
8018 }
8019 }
8020 }
8021
VisitDataProcWithShifterOp(HDataProcWithShifterOp * instruction)8022 void LocationsBuilderARMVIXL::VisitDataProcWithShifterOp(
8023 HDataProcWithShifterOp* instruction) {
8024 DCHECK(instruction->GetType() == DataType::Type::kInt32 ||
8025 instruction->GetType() == DataType::Type::kInt64);
8026 LocationSummary* locations =
8027 new (GetGraph()->GetAllocator()) LocationSummary(instruction, LocationSummary::kNoCall);
8028 const bool overlap = instruction->GetType() == DataType::Type::kInt64 &&
8029 HDataProcWithShifterOp::IsExtensionOp(instruction->GetOpKind());
8030
8031 locations->SetInAt(0, Location::RequiresRegister());
8032 locations->SetInAt(1, Location::RequiresRegister());
8033 locations->SetOut(Location::RequiresRegister(),
8034 overlap ? Location::kOutputOverlap : Location::kNoOutputOverlap);
8035 }
8036
VisitDataProcWithShifterOp(HDataProcWithShifterOp * instruction)8037 void InstructionCodeGeneratorARMVIXL::VisitDataProcWithShifterOp(
8038 HDataProcWithShifterOp* instruction) {
8039 const LocationSummary* const locations = instruction->GetLocations();
8040 const HInstruction::InstructionKind kind = instruction->GetInstrKind();
8041 const HDataProcWithShifterOp::OpKind op_kind = instruction->GetOpKind();
8042
8043 if (instruction->GetType() == DataType::Type::kInt32) {
8044 const vixl32::Register first = InputRegisterAt(instruction, 0);
8045 const vixl32::Register output = OutputRegister(instruction);
8046 const vixl32::Register second = instruction->InputAt(1)->GetType() == DataType::Type::kInt64
8047 ? LowRegisterFrom(locations->InAt(1))
8048 : InputRegisterAt(instruction, 1);
8049
8050 if (HDataProcWithShifterOp::IsExtensionOp(op_kind)) {
8051 DCHECK_EQ(kind, HInstruction::kAdd);
8052
8053 switch (op_kind) {
8054 case HDataProcWithShifterOp::kUXTB:
8055 __ Uxtab(output, first, second);
8056 break;
8057 case HDataProcWithShifterOp::kUXTH:
8058 __ Uxtah(output, first, second);
8059 break;
8060 case HDataProcWithShifterOp::kSXTB:
8061 __ Sxtab(output, first, second);
8062 break;
8063 case HDataProcWithShifterOp::kSXTH:
8064 __ Sxtah(output, first, second);
8065 break;
8066 default:
8067 LOG(FATAL) << "Unexpected operation kind: " << op_kind;
8068 UNREACHABLE();
8069 }
8070 } else {
8071 GenerateDataProcInstruction(kind,
8072 output,
8073 first,
8074 Operand(second,
8075 ShiftFromOpKind(op_kind),
8076 instruction->GetShiftAmount()),
8077 codegen_);
8078 }
8079 } else {
8080 DCHECK_EQ(instruction->GetType(), DataType::Type::kInt64);
8081
8082 if (HDataProcWithShifterOp::IsExtensionOp(op_kind)) {
8083 const vixl32::Register second = InputRegisterAt(instruction, 1);
8084
8085 DCHECK(!LowRegisterFrom(locations->Out()).Is(second));
8086 GenerateDataProc(kind,
8087 locations->Out(),
8088 locations->InAt(0),
8089 second,
8090 Operand(second, ShiftType::ASR, 31),
8091 codegen_);
8092 } else {
8093 GenerateLongDataProc(instruction, codegen_);
8094 }
8095 }
8096 }
8097
8098 // TODO(VIXL): Remove optimizations in the helper when they are implemented in vixl.
GenerateAndConst(vixl32::Register out,vixl32::Register first,uint32_t value)8099 void InstructionCodeGeneratorARMVIXL::GenerateAndConst(vixl32::Register out,
8100 vixl32::Register first,
8101 uint32_t value) {
8102 // Optimize special cases for individual halfs of `and-long` (`and` is simplified earlier).
8103 if (value == 0xffffffffu) {
8104 if (!out.Is(first)) {
8105 __ Mov(out, first);
8106 }
8107 return;
8108 }
8109 if (value == 0u) {
8110 __ Mov(out, 0);
8111 return;
8112 }
8113 if (GetAssembler()->ShifterOperandCanHold(AND, value)) {
8114 __ And(out, first, value);
8115 } else if (GetAssembler()->ShifterOperandCanHold(BIC, ~value)) {
8116 __ Bic(out, first, ~value);
8117 } else {
8118 DCHECK(IsPowerOfTwo(value + 1));
8119 __ Ubfx(out, first, 0, WhichPowerOf2(value + 1));
8120 }
8121 }
8122
8123 // TODO(VIXL): Remove optimizations in the helper when they are implemented in vixl.
GenerateOrrConst(vixl32::Register out,vixl32::Register first,uint32_t value)8124 void InstructionCodeGeneratorARMVIXL::GenerateOrrConst(vixl32::Register out,
8125 vixl32::Register first,
8126 uint32_t value) {
8127 // Optimize special cases for individual halfs of `or-long` (`or` is simplified earlier).
8128 if (value == 0u) {
8129 if (!out.Is(first)) {
8130 __ Mov(out, first);
8131 }
8132 return;
8133 }
8134 if (value == 0xffffffffu) {
8135 __ Mvn(out, 0);
8136 return;
8137 }
8138 if (GetAssembler()->ShifterOperandCanHold(ORR, value)) {
8139 __ Orr(out, first, value);
8140 } else {
8141 DCHECK(GetAssembler()->ShifterOperandCanHold(ORN, ~value));
8142 __ Orn(out, first, ~value);
8143 }
8144 }
8145
8146 // TODO(VIXL): Remove optimizations in the helper when they are implemented in vixl.
GenerateEorConst(vixl32::Register out,vixl32::Register first,uint32_t value)8147 void InstructionCodeGeneratorARMVIXL::GenerateEorConst(vixl32::Register out,
8148 vixl32::Register first,
8149 uint32_t value) {
8150 // Optimize special case for individual halfs of `xor-long` (`xor` is simplified earlier).
8151 if (value == 0u) {
8152 if (!out.Is(first)) {
8153 __ Mov(out, first);
8154 }
8155 return;
8156 }
8157 __ Eor(out, first, value);
8158 }
8159
GenerateAddLongConst(Location out,Location first,uint64_t value)8160 void InstructionCodeGeneratorARMVIXL::GenerateAddLongConst(Location out,
8161 Location first,
8162 uint64_t value) {
8163 vixl32::Register out_low = LowRegisterFrom(out);
8164 vixl32::Register out_high = HighRegisterFrom(out);
8165 vixl32::Register first_low = LowRegisterFrom(first);
8166 vixl32::Register first_high = HighRegisterFrom(first);
8167 uint32_t value_low = Low32Bits(value);
8168 uint32_t value_high = High32Bits(value);
8169 if (value_low == 0u) {
8170 if (!out_low.Is(first_low)) {
8171 __ Mov(out_low, first_low);
8172 }
8173 __ Add(out_high, first_high, value_high);
8174 return;
8175 }
8176 __ Adds(out_low, first_low, value_low);
8177 if (GetAssembler()->ShifterOperandCanHold(ADC, value_high)) {
8178 __ Adc(out_high, first_high, value_high);
8179 } else {
8180 DCHECK(GetAssembler()->ShifterOperandCanHold(SBC, ~value_high));
8181 __ Sbc(out_high, first_high, ~value_high);
8182 }
8183 }
8184
HandleBitwiseOperation(HBinaryOperation * instruction)8185 void InstructionCodeGeneratorARMVIXL::HandleBitwiseOperation(HBinaryOperation* instruction) {
8186 LocationSummary* locations = instruction->GetLocations();
8187 Location first = locations->InAt(0);
8188 Location second = locations->InAt(1);
8189 Location out = locations->Out();
8190
8191 if (second.IsConstant()) {
8192 uint64_t value = static_cast<uint64_t>(Int64FromConstant(second.GetConstant()));
8193 uint32_t value_low = Low32Bits(value);
8194 if (instruction->GetResultType() == DataType::Type::kInt32) {
8195 vixl32::Register first_reg = InputRegisterAt(instruction, 0);
8196 vixl32::Register out_reg = OutputRegister(instruction);
8197 if (instruction->IsAnd()) {
8198 GenerateAndConst(out_reg, first_reg, value_low);
8199 } else if (instruction->IsOr()) {
8200 GenerateOrrConst(out_reg, first_reg, value_low);
8201 } else {
8202 DCHECK(instruction->IsXor());
8203 GenerateEorConst(out_reg, first_reg, value_low);
8204 }
8205 } else {
8206 DCHECK_EQ(instruction->GetResultType(), DataType::Type::kInt64);
8207 uint32_t value_high = High32Bits(value);
8208 vixl32::Register first_low = LowRegisterFrom(first);
8209 vixl32::Register first_high = HighRegisterFrom(first);
8210 vixl32::Register out_low = LowRegisterFrom(out);
8211 vixl32::Register out_high = HighRegisterFrom(out);
8212 if (instruction->IsAnd()) {
8213 GenerateAndConst(out_low, first_low, value_low);
8214 GenerateAndConst(out_high, first_high, value_high);
8215 } else if (instruction->IsOr()) {
8216 GenerateOrrConst(out_low, first_low, value_low);
8217 GenerateOrrConst(out_high, first_high, value_high);
8218 } else {
8219 DCHECK(instruction->IsXor());
8220 GenerateEorConst(out_low, first_low, value_low);
8221 GenerateEorConst(out_high, first_high, value_high);
8222 }
8223 }
8224 return;
8225 }
8226
8227 if (instruction->GetResultType() == DataType::Type::kInt32) {
8228 vixl32::Register first_reg = InputRegisterAt(instruction, 0);
8229 vixl32::Register second_reg = InputRegisterAt(instruction, 1);
8230 vixl32::Register out_reg = OutputRegister(instruction);
8231 if (instruction->IsAnd()) {
8232 __ And(out_reg, first_reg, second_reg);
8233 } else if (instruction->IsOr()) {
8234 __ Orr(out_reg, first_reg, second_reg);
8235 } else {
8236 DCHECK(instruction->IsXor());
8237 __ Eor(out_reg, first_reg, second_reg);
8238 }
8239 } else {
8240 DCHECK_EQ(instruction->GetResultType(), DataType::Type::kInt64);
8241 vixl32::Register first_low = LowRegisterFrom(first);
8242 vixl32::Register first_high = HighRegisterFrom(first);
8243 vixl32::Register second_low = LowRegisterFrom(second);
8244 vixl32::Register second_high = HighRegisterFrom(second);
8245 vixl32::Register out_low = LowRegisterFrom(out);
8246 vixl32::Register out_high = HighRegisterFrom(out);
8247 if (instruction->IsAnd()) {
8248 __ And(out_low, first_low, second_low);
8249 __ And(out_high, first_high, second_high);
8250 } else if (instruction->IsOr()) {
8251 __ Orr(out_low, first_low, second_low);
8252 __ Orr(out_high, first_high, second_high);
8253 } else {
8254 DCHECK(instruction->IsXor());
8255 __ Eor(out_low, first_low, second_low);
8256 __ Eor(out_high, first_high, second_high);
8257 }
8258 }
8259 }
8260
GenerateReferenceLoadOneRegister(HInstruction * instruction,Location out,uint32_t offset,Location maybe_temp,ReadBarrierOption read_barrier_option)8261 void InstructionCodeGeneratorARMVIXL::GenerateReferenceLoadOneRegister(
8262 HInstruction* instruction,
8263 Location out,
8264 uint32_t offset,
8265 Location maybe_temp,
8266 ReadBarrierOption read_barrier_option) {
8267 vixl32::Register out_reg = RegisterFrom(out);
8268 if (read_barrier_option == kWithReadBarrier) {
8269 CHECK(kEmitCompilerReadBarrier);
8270 DCHECK(maybe_temp.IsRegister()) << maybe_temp;
8271 if (kUseBakerReadBarrier) {
8272 // Load with fast path based Baker's read barrier.
8273 // /* HeapReference<Object> */ out = *(out + offset)
8274 codegen_->GenerateFieldLoadWithBakerReadBarrier(
8275 instruction, out, out_reg, offset, maybe_temp, /* needs_null_check= */ false);
8276 } else {
8277 // Load with slow path based read barrier.
8278 // Save the value of `out` into `maybe_temp` before overwriting it
8279 // in the following move operation, as we will need it for the
8280 // read barrier below.
8281 __ Mov(RegisterFrom(maybe_temp), out_reg);
8282 // /* HeapReference<Object> */ out = *(out + offset)
8283 GetAssembler()->LoadFromOffset(kLoadWord, out_reg, out_reg, offset);
8284 codegen_->GenerateReadBarrierSlow(instruction, out, out, maybe_temp, offset);
8285 }
8286 } else {
8287 // Plain load with no read barrier.
8288 // /* HeapReference<Object> */ out = *(out + offset)
8289 GetAssembler()->LoadFromOffset(kLoadWord, out_reg, out_reg, offset);
8290 GetAssembler()->MaybeUnpoisonHeapReference(out_reg);
8291 }
8292 }
8293
GenerateReferenceLoadTwoRegisters(HInstruction * instruction,Location out,Location obj,uint32_t offset,Location maybe_temp,ReadBarrierOption read_barrier_option)8294 void InstructionCodeGeneratorARMVIXL::GenerateReferenceLoadTwoRegisters(
8295 HInstruction* instruction,
8296 Location out,
8297 Location obj,
8298 uint32_t offset,
8299 Location maybe_temp,
8300 ReadBarrierOption read_barrier_option) {
8301 vixl32::Register out_reg = RegisterFrom(out);
8302 vixl32::Register obj_reg = RegisterFrom(obj);
8303 if (read_barrier_option == kWithReadBarrier) {
8304 CHECK(kEmitCompilerReadBarrier);
8305 if (kUseBakerReadBarrier) {
8306 DCHECK(maybe_temp.IsRegister()) << maybe_temp;
8307 // Load with fast path based Baker's read barrier.
8308 // /* HeapReference<Object> */ out = *(obj + offset)
8309 codegen_->GenerateFieldLoadWithBakerReadBarrier(
8310 instruction, out, obj_reg, offset, maybe_temp, /* needs_null_check= */ false);
8311 } else {
8312 // Load with slow path based read barrier.
8313 // /* HeapReference<Object> */ out = *(obj + offset)
8314 GetAssembler()->LoadFromOffset(kLoadWord, out_reg, obj_reg, offset);
8315 codegen_->GenerateReadBarrierSlow(instruction, out, out, obj, offset);
8316 }
8317 } else {
8318 // Plain load with no read barrier.
8319 // /* HeapReference<Object> */ out = *(obj + offset)
8320 GetAssembler()->LoadFromOffset(kLoadWord, out_reg, obj_reg, offset);
8321 GetAssembler()->MaybeUnpoisonHeapReference(out_reg);
8322 }
8323 }
8324
GenerateGcRootFieldLoad(HInstruction * instruction,Location root,vixl32::Register obj,uint32_t offset,ReadBarrierOption read_barrier_option)8325 void CodeGeneratorARMVIXL::GenerateGcRootFieldLoad(
8326 HInstruction* instruction,
8327 Location root,
8328 vixl32::Register obj,
8329 uint32_t offset,
8330 ReadBarrierOption read_barrier_option) {
8331 vixl32::Register root_reg = RegisterFrom(root);
8332 if (read_barrier_option == kWithReadBarrier) {
8333 DCHECK(kEmitCompilerReadBarrier);
8334 if (kUseBakerReadBarrier) {
8335 // Fast path implementation of art::ReadBarrier::BarrierForRoot when
8336 // Baker's read barrier are used.
8337
8338 // Query `art::Thread::Current()->GetIsGcMarking()` (stored in
8339 // the Marking Register) to decide whether we need to enter
8340 // the slow path to mark the GC root.
8341 //
8342 // We use shared thunks for the slow path; shared within the method
8343 // for JIT, across methods for AOT. That thunk checks the reference
8344 // and jumps to the entrypoint if needed.
8345 //
8346 // lr = &return_address;
8347 // GcRoot<mirror::Object> root = *(obj+offset); // Original reference load.
8348 // if (mr) { // Thread::Current()->GetIsGcMarking()
8349 // goto gc_root_thunk<root_reg>(lr)
8350 // }
8351 // return_address:
8352
8353 UseScratchRegisterScope temps(GetVIXLAssembler());
8354 temps.Exclude(ip);
8355 bool narrow = CanEmitNarrowLdr(root_reg, obj, offset);
8356 uint32_t custom_data = EncodeBakerReadBarrierGcRootData(root_reg.GetCode(), narrow);
8357
8358 size_t narrow_instructions = /* CMP */ (mr.IsLow() ? 1u : 0u) + /* LDR */ (narrow ? 1u : 0u);
8359 size_t wide_instructions = /* ADR+CMP+LDR+BNE */ 4u - narrow_instructions;
8360 size_t exact_size = wide_instructions * vixl32::k32BitT32InstructionSizeInBytes +
8361 narrow_instructions * vixl32::k16BitT32InstructionSizeInBytes;
8362 ExactAssemblyScope guard(GetVIXLAssembler(), exact_size);
8363 vixl32::Label return_address;
8364 EmitAdrCode adr(GetVIXLAssembler(), lr, &return_address);
8365 __ cmp(mr, Operand(0));
8366 // Currently the offset is always within range. If that changes,
8367 // we shall have to split the load the same way as for fields.
8368 DCHECK_LT(offset, kReferenceLoadMinFarOffset);
8369 ptrdiff_t old_offset = GetVIXLAssembler()->GetBuffer()->GetCursorOffset();
8370 __ ldr(EncodingSize(narrow ? Narrow : Wide), root_reg, MemOperand(obj, offset));
8371 EmitBakerReadBarrierBne(custom_data);
8372 __ bind(&return_address);
8373 DCHECK_EQ(old_offset - GetVIXLAssembler()->GetBuffer()->GetCursorOffset(),
8374 narrow ? BAKER_MARK_INTROSPECTION_GC_ROOT_LDR_NARROW_OFFSET
8375 : BAKER_MARK_INTROSPECTION_GC_ROOT_LDR_WIDE_OFFSET);
8376 } else {
8377 // GC root loaded through a slow path for read barriers other
8378 // than Baker's.
8379 // /* GcRoot<mirror::Object>* */ root = obj + offset
8380 __ Add(root_reg, obj, offset);
8381 // /* mirror::Object* */ root = root->Read()
8382 GenerateReadBarrierForRootSlow(instruction, root, root);
8383 }
8384 } else {
8385 // Plain GC root load with no read barrier.
8386 // /* GcRoot<mirror::Object> */ root = *(obj + offset)
8387 GetAssembler()->LoadFromOffset(kLoadWord, root_reg, obj, offset);
8388 // Note that GC roots are not affected by heap poisoning, thus we
8389 // do not have to unpoison `root_reg` here.
8390 }
8391 MaybeGenerateMarkingRegisterCheck(/* code= */ 19);
8392 }
8393
GenerateUnsafeCasOldValueAddWithBakerReadBarrier(vixl::aarch32::Register old_value,vixl::aarch32::Register adjusted_old_value,vixl::aarch32::Register expected)8394 void CodeGeneratorARMVIXL::GenerateUnsafeCasOldValueAddWithBakerReadBarrier(
8395 vixl::aarch32::Register old_value,
8396 vixl::aarch32::Register adjusted_old_value,
8397 vixl::aarch32::Register expected) {
8398 DCHECK(kEmitCompilerReadBarrier);
8399 DCHECK(kUseBakerReadBarrier);
8400
8401 // Similar to the Baker RB path in GenerateGcRootFieldLoad(), with an ADD instead of LDR.
8402 uint32_t custom_data = EncodeBakerReadBarrierUnsafeCasData(old_value.GetCode());
8403
8404 size_t narrow_instructions = /* CMP */ (mr.IsLow() ? 1u : 0u);
8405 size_t wide_instructions = /* ADR+CMP+ADD+BNE */ 4u - narrow_instructions;
8406 size_t exact_size = wide_instructions * vixl32::k32BitT32InstructionSizeInBytes +
8407 narrow_instructions * vixl32::k16BitT32InstructionSizeInBytes;
8408 ExactAssemblyScope guard(GetVIXLAssembler(), exact_size);
8409 vixl32::Label return_address;
8410 EmitAdrCode adr(GetVIXLAssembler(), lr, &return_address);
8411 __ cmp(mr, Operand(0));
8412 ptrdiff_t old_offset = GetVIXLAssembler()->GetBuffer()->GetCursorOffset();
8413 __ add(EncodingSize(Wide), old_value, adjusted_old_value, Operand(expected)); // Preserves flags.
8414 EmitBakerReadBarrierBne(custom_data);
8415 __ bind(&return_address);
8416 DCHECK_EQ(old_offset - GetVIXLAssembler()->GetBuffer()->GetCursorOffset(),
8417 BAKER_MARK_INTROSPECTION_UNSAFE_CAS_ADD_OFFSET);
8418 }
8419
GenerateFieldLoadWithBakerReadBarrier(HInstruction * instruction,Location ref,vixl32::Register obj,const vixl32::MemOperand & src,bool needs_null_check)8420 void CodeGeneratorARMVIXL::GenerateFieldLoadWithBakerReadBarrier(HInstruction* instruction,
8421 Location ref,
8422 vixl32::Register obj,
8423 const vixl32::MemOperand& src,
8424 bool needs_null_check) {
8425 DCHECK(kEmitCompilerReadBarrier);
8426 DCHECK(kUseBakerReadBarrier);
8427
8428 // Query `art::Thread::Current()->GetIsGcMarking()` (stored in the
8429 // Marking Register) to decide whether we need to enter the slow
8430 // path to mark the reference. Then, in the slow path, check the
8431 // gray bit in the lock word of the reference's holder (`obj`) to
8432 // decide whether to mark `ref` or not.
8433 //
8434 // We use shared thunks for the slow path; shared within the method
8435 // for JIT, across methods for AOT. That thunk checks the holder
8436 // and jumps to the entrypoint if needed. If the holder is not gray,
8437 // it creates a fake dependency and returns to the LDR instruction.
8438 //
8439 // lr = &gray_return_address;
8440 // if (mr) { // Thread::Current()->GetIsGcMarking()
8441 // goto field_thunk<holder_reg, base_reg>(lr)
8442 // }
8443 // not_gray_return_address:
8444 // // Original reference load. If the offset is too large to fit
8445 // // into LDR, we use an adjusted base register here.
8446 // HeapReference<mirror::Object> reference = *(obj+offset);
8447 // gray_return_address:
8448
8449 DCHECK(src.GetAddrMode() == vixl32::Offset);
8450 DCHECK_ALIGNED(src.GetOffsetImmediate(), sizeof(mirror::HeapReference<mirror::Object>));
8451 vixl32::Register ref_reg = RegisterFrom(ref, DataType::Type::kReference);
8452 bool narrow = CanEmitNarrowLdr(ref_reg, src.GetBaseRegister(), src.GetOffsetImmediate());
8453
8454 UseScratchRegisterScope temps(GetVIXLAssembler());
8455 temps.Exclude(ip);
8456 uint32_t custom_data =
8457 EncodeBakerReadBarrierFieldData(src.GetBaseRegister().GetCode(), obj.GetCode(), narrow);
8458
8459 {
8460 size_t narrow_instructions =
8461 /* CMP */ (mr.IsLow() ? 1u : 0u) +
8462 /* LDR+unpoison? */ (narrow ? (kPoisonHeapReferences ? 2u : 1u) : 0u);
8463 size_t wide_instructions =
8464 /* ADR+CMP+LDR+BNE+unpoison? */ (kPoisonHeapReferences ? 5u : 4u) - narrow_instructions;
8465 size_t exact_size = wide_instructions * vixl32::k32BitT32InstructionSizeInBytes +
8466 narrow_instructions * vixl32::k16BitT32InstructionSizeInBytes;
8467 ExactAssemblyScope guard(GetVIXLAssembler(), exact_size);
8468 vixl32::Label return_address;
8469 EmitAdrCode adr(GetVIXLAssembler(), lr, &return_address);
8470 __ cmp(mr, Operand(0));
8471 EmitBakerReadBarrierBne(custom_data);
8472 ptrdiff_t old_offset = GetVIXLAssembler()->GetBuffer()->GetCursorOffset();
8473 __ ldr(EncodingSize(narrow ? Narrow : Wide), ref_reg, src);
8474 if (needs_null_check) {
8475 MaybeRecordImplicitNullCheck(instruction);
8476 }
8477 // Note: We need a specific width for the unpoisoning NEG.
8478 if (kPoisonHeapReferences) {
8479 if (narrow) {
8480 // The only 16-bit encoding is T1 which sets flags outside IT block (i.e. RSBS, not RSB).
8481 __ rsbs(EncodingSize(Narrow), ref_reg, ref_reg, Operand(0));
8482 } else {
8483 __ rsb(EncodingSize(Wide), ref_reg, ref_reg, Operand(0));
8484 }
8485 }
8486 __ bind(&return_address);
8487 DCHECK_EQ(old_offset - GetVIXLAssembler()->GetBuffer()->GetCursorOffset(),
8488 narrow ? BAKER_MARK_INTROSPECTION_FIELD_LDR_NARROW_OFFSET
8489 : BAKER_MARK_INTROSPECTION_FIELD_LDR_WIDE_OFFSET);
8490 }
8491 MaybeGenerateMarkingRegisterCheck(/* code= */ 20, /* temp_loc= */ LocationFrom(ip));
8492 }
8493
GenerateFieldLoadWithBakerReadBarrier(HInstruction * instruction,Location ref,vixl32::Register obj,uint32_t offset,Location temp,bool needs_null_check)8494 void CodeGeneratorARMVIXL::GenerateFieldLoadWithBakerReadBarrier(HInstruction* instruction,
8495 Location ref,
8496 vixl32::Register obj,
8497 uint32_t offset,
8498 Location temp,
8499 bool needs_null_check) {
8500 DCHECK_ALIGNED(offset, sizeof(mirror::HeapReference<mirror::Object>));
8501 vixl32::Register base = obj;
8502 if (offset >= kReferenceLoadMinFarOffset) {
8503 base = RegisterFrom(temp);
8504 static_assert(IsPowerOfTwo(kReferenceLoadMinFarOffset), "Expecting a power of 2.");
8505 __ Add(base, obj, Operand(offset & ~(kReferenceLoadMinFarOffset - 1u)));
8506 offset &= (kReferenceLoadMinFarOffset - 1u);
8507 }
8508 GenerateFieldLoadWithBakerReadBarrier(
8509 instruction, ref, obj, MemOperand(base, offset), needs_null_check);
8510 }
8511
GenerateArrayLoadWithBakerReadBarrier(Location ref,vixl32::Register obj,uint32_t data_offset,Location index,Location temp,bool needs_null_check)8512 void CodeGeneratorARMVIXL::GenerateArrayLoadWithBakerReadBarrier(Location ref,
8513 vixl32::Register obj,
8514 uint32_t data_offset,
8515 Location index,
8516 Location temp,
8517 bool needs_null_check) {
8518 DCHECK(kEmitCompilerReadBarrier);
8519 DCHECK(kUseBakerReadBarrier);
8520
8521 static_assert(
8522 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t),
8523 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes.");
8524 ScaleFactor scale_factor = TIMES_4;
8525
8526 // Query `art::Thread::Current()->GetIsGcMarking()` (stored in the
8527 // Marking Register) to decide whether we need to enter the slow
8528 // path to mark the reference. Then, in the slow path, check the
8529 // gray bit in the lock word of the reference's holder (`obj`) to
8530 // decide whether to mark `ref` or not.
8531 //
8532 // We use shared thunks for the slow path; shared within the method
8533 // for JIT, across methods for AOT. That thunk checks the holder
8534 // and jumps to the entrypoint if needed. If the holder is not gray,
8535 // it creates a fake dependency and returns to the LDR instruction.
8536 //
8537 // lr = &gray_return_address;
8538 // if (mr) { // Thread::Current()->GetIsGcMarking()
8539 // goto array_thunk<base_reg>(lr)
8540 // }
8541 // not_gray_return_address:
8542 // // Original reference load. If the offset is too large to fit
8543 // // into LDR, we use an adjusted base register here.
8544 // HeapReference<mirror::Object> reference = data[index];
8545 // gray_return_address:
8546
8547 DCHECK(index.IsValid());
8548 vixl32::Register index_reg = RegisterFrom(index, DataType::Type::kInt32);
8549 vixl32::Register ref_reg = RegisterFrom(ref, DataType::Type::kReference);
8550 vixl32::Register data_reg = RegisterFrom(temp, DataType::Type::kInt32); // Raw pointer.
8551
8552 UseScratchRegisterScope temps(GetVIXLAssembler());
8553 temps.Exclude(ip);
8554 uint32_t custom_data = EncodeBakerReadBarrierArrayData(data_reg.GetCode());
8555
8556 __ Add(data_reg, obj, Operand(data_offset));
8557 {
8558 size_t narrow_instructions = /* CMP */ (mr.IsLow() ? 1u : 0u);
8559 size_t wide_instructions =
8560 /* ADR+CMP+BNE+LDR+unpoison? */ (kPoisonHeapReferences ? 5u : 4u) - narrow_instructions;
8561 size_t exact_size = wide_instructions * vixl32::k32BitT32InstructionSizeInBytes +
8562 narrow_instructions * vixl32::k16BitT32InstructionSizeInBytes;
8563 ExactAssemblyScope guard(GetVIXLAssembler(), exact_size);
8564 vixl32::Label return_address;
8565 EmitAdrCode adr(GetVIXLAssembler(), lr, &return_address);
8566 __ cmp(mr, Operand(0));
8567 EmitBakerReadBarrierBne(custom_data);
8568 ptrdiff_t old_offset = GetVIXLAssembler()->GetBuffer()->GetCursorOffset();
8569 __ ldr(ref_reg, MemOperand(data_reg, index_reg, vixl32::LSL, scale_factor));
8570 DCHECK(!needs_null_check); // The thunk cannot handle the null check.
8571 // Note: We need a Wide NEG for the unpoisoning.
8572 if (kPoisonHeapReferences) {
8573 __ rsb(EncodingSize(Wide), ref_reg, ref_reg, Operand(0));
8574 }
8575 __ bind(&return_address);
8576 DCHECK_EQ(old_offset - GetVIXLAssembler()->GetBuffer()->GetCursorOffset(),
8577 BAKER_MARK_INTROSPECTION_ARRAY_LDR_OFFSET);
8578 }
8579 MaybeGenerateMarkingRegisterCheck(/* code= */ 21, /* temp_loc= */ LocationFrom(ip));
8580 }
8581
MaybeGenerateMarkingRegisterCheck(int code,Location temp_loc)8582 void CodeGeneratorARMVIXL::MaybeGenerateMarkingRegisterCheck(int code, Location temp_loc) {
8583 // The following condition is a compile-time one, so it does not have a run-time cost.
8584 if (kEmitCompilerReadBarrier && kUseBakerReadBarrier && kIsDebugBuild) {
8585 // The following condition is a run-time one; it is executed after the
8586 // previous compile-time test, to avoid penalizing non-debug builds.
8587 if (GetCompilerOptions().EmitRunTimeChecksInDebugMode()) {
8588 UseScratchRegisterScope temps(GetVIXLAssembler());
8589 vixl32::Register temp = temp_loc.IsValid() ? RegisterFrom(temp_loc) : temps.Acquire();
8590 GetAssembler()->GenerateMarkingRegisterCheck(temp,
8591 kMarkingRegisterCheckBreakCodeBaseCode + code);
8592 }
8593 }
8594 }
8595
GenerateReadBarrierSlow(HInstruction * instruction,Location out,Location ref,Location obj,uint32_t offset,Location index)8596 void CodeGeneratorARMVIXL::GenerateReadBarrierSlow(HInstruction* instruction,
8597 Location out,
8598 Location ref,
8599 Location obj,
8600 uint32_t offset,
8601 Location index) {
8602 DCHECK(kEmitCompilerReadBarrier);
8603
8604 // Insert a slow path based read barrier *after* the reference load.
8605 //
8606 // If heap poisoning is enabled, the unpoisoning of the loaded
8607 // reference will be carried out by the runtime within the slow
8608 // path.
8609 //
8610 // Note that `ref` currently does not get unpoisoned (when heap
8611 // poisoning is enabled), which is alright as the `ref` argument is
8612 // not used by the artReadBarrierSlow entry point.
8613 //
8614 // TODO: Unpoison `ref` when it is used by artReadBarrierSlow.
8615 SlowPathCodeARMVIXL* slow_path = new (GetScopedAllocator())
8616 ReadBarrierForHeapReferenceSlowPathARMVIXL(instruction, out, ref, obj, offset, index);
8617 AddSlowPath(slow_path);
8618
8619 __ B(slow_path->GetEntryLabel());
8620 __ Bind(slow_path->GetExitLabel());
8621 }
8622
MaybeGenerateReadBarrierSlow(HInstruction * instruction,Location out,Location ref,Location obj,uint32_t offset,Location index)8623 void CodeGeneratorARMVIXL::MaybeGenerateReadBarrierSlow(HInstruction* instruction,
8624 Location out,
8625 Location ref,
8626 Location obj,
8627 uint32_t offset,
8628 Location index) {
8629 if (kEmitCompilerReadBarrier) {
8630 // Baker's read barriers shall be handled by the fast path
8631 // (CodeGeneratorARMVIXL::GenerateReferenceLoadWithBakerReadBarrier).
8632 DCHECK(!kUseBakerReadBarrier);
8633 // If heap poisoning is enabled, unpoisoning will be taken care of
8634 // by the runtime within the slow path.
8635 GenerateReadBarrierSlow(instruction, out, ref, obj, offset, index);
8636 } else if (kPoisonHeapReferences) {
8637 GetAssembler()->UnpoisonHeapReference(RegisterFrom(out));
8638 }
8639 }
8640
GenerateReadBarrierForRootSlow(HInstruction * instruction,Location out,Location root)8641 void CodeGeneratorARMVIXL::GenerateReadBarrierForRootSlow(HInstruction* instruction,
8642 Location out,
8643 Location root) {
8644 DCHECK(kEmitCompilerReadBarrier);
8645
8646 // Insert a slow path based read barrier *after* the GC root load.
8647 //
8648 // Note that GC roots are not affected by heap poisoning, so we do
8649 // not need to do anything special for this here.
8650 SlowPathCodeARMVIXL* slow_path =
8651 new (GetScopedAllocator()) ReadBarrierForRootSlowPathARMVIXL(instruction, out, root);
8652 AddSlowPath(slow_path);
8653
8654 __ B(slow_path->GetEntryLabel());
8655 __ Bind(slow_path->GetExitLabel());
8656 }
8657
8658 // Check if the desired_dispatch_info is supported. If it is, return it,
8659 // otherwise return a fall-back info that should be used instead.
GetSupportedInvokeStaticOrDirectDispatch(const HInvokeStaticOrDirect::DispatchInfo & desired_dispatch_info,ArtMethod * method ATTRIBUTE_UNUSED)8660 HInvokeStaticOrDirect::DispatchInfo CodeGeneratorARMVIXL::GetSupportedInvokeStaticOrDirectDispatch(
8661 const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
8662 ArtMethod* method ATTRIBUTE_UNUSED) {
8663 return desired_dispatch_info;
8664 }
8665
GetInvokeStaticOrDirectExtraParameter(HInvokeStaticOrDirect * invoke,vixl32::Register temp)8666 vixl32::Register CodeGeneratorARMVIXL::GetInvokeStaticOrDirectExtraParameter(
8667 HInvokeStaticOrDirect* invoke, vixl32::Register temp) {
8668 DCHECK_EQ(invoke->InputCount(), invoke->GetNumberOfArguments() + 1u);
8669 Location location = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex());
8670 if (!invoke->GetLocations()->Intrinsified()) {
8671 return RegisterFrom(location);
8672 }
8673 // For intrinsics we allow any location, so it may be on the stack.
8674 if (!location.IsRegister()) {
8675 GetAssembler()->LoadFromOffset(kLoadWord, temp, sp, location.GetStackIndex());
8676 return temp;
8677 }
8678 // For register locations, check if the register was saved. If so, get it from the stack.
8679 // Note: There is a chance that the register was saved but not overwritten, so we could
8680 // save one load. However, since this is just an intrinsic slow path we prefer this
8681 // simple and more robust approach rather that trying to determine if that's the case.
8682 SlowPathCode* slow_path = GetCurrentSlowPath();
8683 if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(RegisterFrom(location).GetCode())) {
8684 int stack_offset = slow_path->GetStackOffsetOfCoreRegister(RegisterFrom(location).GetCode());
8685 GetAssembler()->LoadFromOffset(kLoadWord, temp, sp, stack_offset);
8686 return temp;
8687 }
8688 return RegisterFrom(location);
8689 }
8690
GenerateStaticOrDirectCall(HInvokeStaticOrDirect * invoke,Location temp,SlowPathCode * slow_path)8691 void CodeGeneratorARMVIXL::GenerateStaticOrDirectCall(
8692 HInvokeStaticOrDirect* invoke, Location temp, SlowPathCode* slow_path) {
8693 Location callee_method = temp; // For all kinds except kRecursive, callee will be in temp.
8694 switch (invoke->GetMethodLoadKind()) {
8695 case HInvokeStaticOrDirect::MethodLoadKind::kStringInit: {
8696 uint32_t offset =
8697 GetThreadOffset<kArmPointerSize>(invoke->GetStringInitEntryPoint()).Int32Value();
8698 // temp = thread->string_init_entrypoint
8699 GetAssembler()->LoadFromOffset(kLoadWord, RegisterFrom(temp), tr, offset);
8700 break;
8701 }
8702 case HInvokeStaticOrDirect::MethodLoadKind::kRecursive:
8703 callee_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex());
8704 break;
8705 case HInvokeStaticOrDirect::MethodLoadKind::kBootImageLinkTimePcRelative: {
8706 DCHECK(GetCompilerOptions().IsBootImage());
8707 PcRelativePatchInfo* labels = NewBootImageMethodPatch(invoke->GetTargetMethod());
8708 vixl32::Register temp_reg = RegisterFrom(temp);
8709 EmitMovwMovtPlaceholder(labels, temp_reg);
8710 break;
8711 }
8712 case HInvokeStaticOrDirect::MethodLoadKind::kBootImageRelRo: {
8713 uint32_t boot_image_offset = GetBootImageOffset(invoke);
8714 PcRelativePatchInfo* labels = NewBootImageRelRoPatch(boot_image_offset);
8715 vixl32::Register temp_reg = RegisterFrom(temp);
8716 EmitMovwMovtPlaceholder(labels, temp_reg);
8717 GetAssembler()->LoadFromOffset(kLoadWord, temp_reg, temp_reg, /* offset*/ 0);
8718 break;
8719 }
8720 case HInvokeStaticOrDirect::MethodLoadKind::kBssEntry: {
8721 PcRelativePatchInfo* labels = NewMethodBssEntryPatch(
8722 MethodReference(&GetGraph()->GetDexFile(), invoke->GetDexMethodIndex()));
8723 vixl32::Register temp_reg = RegisterFrom(temp);
8724 EmitMovwMovtPlaceholder(labels, temp_reg);
8725 GetAssembler()->LoadFromOffset(kLoadWord, temp_reg, temp_reg, /* offset*/ 0);
8726 break;
8727 }
8728 case HInvokeStaticOrDirect::MethodLoadKind::kJitDirectAddress:
8729 __ Mov(RegisterFrom(temp), Operand::From(invoke->GetMethodAddress()));
8730 break;
8731 case HInvokeStaticOrDirect::MethodLoadKind::kRuntimeCall: {
8732 GenerateInvokeStaticOrDirectRuntimeCall(invoke, temp, slow_path);
8733 return; // No code pointer retrieval; the runtime performs the call directly.
8734 }
8735 }
8736
8737 switch (invoke->GetCodePtrLocation()) {
8738 case HInvokeStaticOrDirect::CodePtrLocation::kCallSelf:
8739 {
8740 // Use a scope to help guarantee that `RecordPcInfo()` records the correct pc.
8741 ExactAssemblyScope aas(GetVIXLAssembler(),
8742 vixl32::k32BitT32InstructionSizeInBytes,
8743 CodeBufferCheckScope::kMaximumSize);
8744 __ bl(GetFrameEntryLabel());
8745 RecordPcInfo(invoke, invoke->GetDexPc(), slow_path);
8746 }
8747 break;
8748 case HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod:
8749 // LR = callee_method->entry_point_from_quick_compiled_code_
8750 GetAssembler()->LoadFromOffset(
8751 kLoadWord,
8752 lr,
8753 RegisterFrom(callee_method),
8754 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmPointerSize).Int32Value());
8755 {
8756 // Use a scope to help guarantee that `RecordPcInfo()` records the correct pc.
8757 // blx in T32 has only 16bit encoding that's why a stricter check for the scope is used.
8758 ExactAssemblyScope aas(GetVIXLAssembler(),
8759 vixl32::k16BitT32InstructionSizeInBytes,
8760 CodeBufferCheckScope::kExactSize);
8761 // LR()
8762 __ blx(lr);
8763 RecordPcInfo(invoke, invoke->GetDexPc(), slow_path);
8764 }
8765 break;
8766 }
8767
8768 DCHECK(!IsLeafMethod());
8769 }
8770
GenerateVirtualCall(HInvokeVirtual * invoke,Location temp_location,SlowPathCode * slow_path)8771 void CodeGeneratorARMVIXL::GenerateVirtualCall(
8772 HInvokeVirtual* invoke, Location temp_location, SlowPathCode* slow_path) {
8773 vixl32::Register temp = RegisterFrom(temp_location);
8774 uint32_t method_offset = mirror::Class::EmbeddedVTableEntryOffset(
8775 invoke->GetVTableIndex(), kArmPointerSize).Uint32Value();
8776
8777 // Use the calling convention instead of the location of the receiver, as
8778 // intrinsics may have put the receiver in a different register. In the intrinsics
8779 // slow path, the arguments have been moved to the right place, so here we are
8780 // guaranteed that the receiver is the first register of the calling convention.
8781 InvokeDexCallingConventionARMVIXL calling_convention;
8782 vixl32::Register receiver = calling_convention.GetRegisterAt(0);
8783 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
8784 {
8785 // Make sure the pc is recorded immediately after the `ldr` instruction.
8786 ExactAssemblyScope aas(GetVIXLAssembler(),
8787 vixl32::kMaxInstructionSizeInBytes,
8788 CodeBufferCheckScope::kMaximumSize);
8789 // /* HeapReference<Class> */ temp = receiver->klass_
8790 __ ldr(temp, MemOperand(receiver, class_offset));
8791 MaybeRecordImplicitNullCheck(invoke);
8792 }
8793 // Instead of simply (possibly) unpoisoning `temp` here, we should
8794 // emit a read barrier for the previous class reference load.
8795 // However this is not required in practice, as this is an
8796 // intermediate/temporary reference and because the current
8797 // concurrent copying collector keeps the from-space memory
8798 // intact/accessible until the end of the marking phase (the
8799 // concurrent copying collector may not in the future).
8800 GetAssembler()->MaybeUnpoisonHeapReference(temp);
8801
8802 // temp = temp->GetMethodAt(method_offset);
8803 uint32_t entry_point = ArtMethod::EntryPointFromQuickCompiledCodeOffset(
8804 kArmPointerSize).Int32Value();
8805 GetAssembler()->LoadFromOffset(kLoadWord, temp, temp, method_offset);
8806 // LR = temp->GetEntryPoint();
8807 GetAssembler()->LoadFromOffset(kLoadWord, lr, temp, entry_point);
8808 {
8809 // Use a scope to help guarantee that `RecordPcInfo()` records the correct pc.
8810 // blx in T32 has only 16bit encoding that's why a stricter check for the scope is used.
8811 ExactAssemblyScope aas(GetVIXLAssembler(),
8812 vixl32::k16BitT32InstructionSizeInBytes,
8813 CodeBufferCheckScope::kExactSize);
8814 // LR();
8815 __ blx(lr);
8816 RecordPcInfo(invoke, invoke->GetDexPc(), slow_path);
8817 }
8818 }
8819
NewBootImageIntrinsicPatch(uint32_t intrinsic_data)8820 CodeGeneratorARMVIXL::PcRelativePatchInfo* CodeGeneratorARMVIXL::NewBootImageIntrinsicPatch(
8821 uint32_t intrinsic_data) {
8822 return NewPcRelativePatch(/* dex_file= */ nullptr, intrinsic_data, &boot_image_intrinsic_patches_);
8823 }
8824
NewBootImageRelRoPatch(uint32_t boot_image_offset)8825 CodeGeneratorARMVIXL::PcRelativePatchInfo* CodeGeneratorARMVIXL::NewBootImageRelRoPatch(
8826 uint32_t boot_image_offset) {
8827 return NewPcRelativePatch(/* dex_file= */ nullptr,
8828 boot_image_offset,
8829 &boot_image_method_patches_);
8830 }
8831
NewBootImageMethodPatch(MethodReference target_method)8832 CodeGeneratorARMVIXL::PcRelativePatchInfo* CodeGeneratorARMVIXL::NewBootImageMethodPatch(
8833 MethodReference target_method) {
8834 return NewPcRelativePatch(
8835 target_method.dex_file, target_method.index, &boot_image_method_patches_);
8836 }
8837
NewMethodBssEntryPatch(MethodReference target_method)8838 CodeGeneratorARMVIXL::PcRelativePatchInfo* CodeGeneratorARMVIXL::NewMethodBssEntryPatch(
8839 MethodReference target_method) {
8840 return NewPcRelativePatch(
8841 target_method.dex_file, target_method.index, &method_bss_entry_patches_);
8842 }
8843
NewBootImageTypePatch(const DexFile & dex_file,dex::TypeIndex type_index)8844 CodeGeneratorARMVIXL::PcRelativePatchInfo* CodeGeneratorARMVIXL::NewBootImageTypePatch(
8845 const DexFile& dex_file, dex::TypeIndex type_index) {
8846 return NewPcRelativePatch(&dex_file, type_index.index_, &boot_image_type_patches_);
8847 }
8848
NewTypeBssEntryPatch(const DexFile & dex_file,dex::TypeIndex type_index)8849 CodeGeneratorARMVIXL::PcRelativePatchInfo* CodeGeneratorARMVIXL::NewTypeBssEntryPatch(
8850 const DexFile& dex_file, dex::TypeIndex type_index) {
8851 return NewPcRelativePatch(&dex_file, type_index.index_, &type_bss_entry_patches_);
8852 }
8853
NewBootImageStringPatch(const DexFile & dex_file,dex::StringIndex string_index)8854 CodeGeneratorARMVIXL::PcRelativePatchInfo* CodeGeneratorARMVIXL::NewBootImageStringPatch(
8855 const DexFile& dex_file, dex::StringIndex string_index) {
8856 return NewPcRelativePatch(&dex_file, string_index.index_, &boot_image_string_patches_);
8857 }
8858
NewStringBssEntryPatch(const DexFile & dex_file,dex::StringIndex string_index)8859 CodeGeneratorARMVIXL::PcRelativePatchInfo* CodeGeneratorARMVIXL::NewStringBssEntryPatch(
8860 const DexFile& dex_file, dex::StringIndex string_index) {
8861 return NewPcRelativePatch(&dex_file, string_index.index_, &string_bss_entry_patches_);
8862 }
8863
NewPcRelativePatch(const DexFile * dex_file,uint32_t offset_or_index,ArenaDeque<PcRelativePatchInfo> * patches)8864 CodeGeneratorARMVIXL::PcRelativePatchInfo* CodeGeneratorARMVIXL::NewPcRelativePatch(
8865 const DexFile* dex_file, uint32_t offset_or_index, ArenaDeque<PcRelativePatchInfo>* patches) {
8866 patches->emplace_back(dex_file, offset_or_index);
8867 return &patches->back();
8868 }
8869
EmitBakerReadBarrierBne(uint32_t custom_data)8870 void CodeGeneratorARMVIXL::EmitBakerReadBarrierBne(uint32_t custom_data) {
8871 DCHECK(!__ AllowMacroInstructions()); // In ExactAssemblyScope.
8872 if (Runtime::Current()->UseJitCompilation()) {
8873 auto it = jit_baker_read_barrier_slow_paths_.FindOrAdd(custom_data);
8874 vixl::aarch32::Label* slow_path_entry = &it->second.label;
8875 __ b(ne, EncodingSize(Wide), slow_path_entry);
8876 } else {
8877 baker_read_barrier_patches_.emplace_back(custom_data);
8878 vixl::aarch32::Label* patch_label = &baker_read_barrier_patches_.back().label;
8879 __ bind(patch_label);
8880 vixl32::Label placeholder_label;
8881 __ b(ne, EncodingSize(Wide), &placeholder_label); // Placeholder, patched at link-time.
8882 __ bind(&placeholder_label);
8883 }
8884 }
8885
DeduplicateBootImageAddressLiteral(uint32_t address)8886 VIXLUInt32Literal* CodeGeneratorARMVIXL::DeduplicateBootImageAddressLiteral(uint32_t address) {
8887 return DeduplicateUint32Literal(address, &uint32_literals_);
8888 }
8889
DeduplicateJitStringLiteral(const DexFile & dex_file,dex::StringIndex string_index,Handle<mirror::String> handle)8890 VIXLUInt32Literal* CodeGeneratorARMVIXL::DeduplicateJitStringLiteral(
8891 const DexFile& dex_file,
8892 dex::StringIndex string_index,
8893 Handle<mirror::String> handle) {
8894 ReserveJitStringRoot(StringReference(&dex_file, string_index), handle);
8895 return jit_string_patches_.GetOrCreate(
8896 StringReference(&dex_file, string_index),
8897 [this]() {
8898 return GetAssembler()->CreateLiteralDestroyedWithPool<uint32_t>(/* value= */ 0u);
8899 });
8900 }
8901
DeduplicateJitClassLiteral(const DexFile & dex_file,dex::TypeIndex type_index,Handle<mirror::Class> handle)8902 VIXLUInt32Literal* CodeGeneratorARMVIXL::DeduplicateJitClassLiteral(const DexFile& dex_file,
8903 dex::TypeIndex type_index,
8904 Handle<mirror::Class> handle) {
8905 ReserveJitClassRoot(TypeReference(&dex_file, type_index), handle);
8906 return jit_class_patches_.GetOrCreate(
8907 TypeReference(&dex_file, type_index),
8908 [this]() {
8909 return GetAssembler()->CreateLiteralDestroyedWithPool<uint32_t>(/* value= */ 0u);
8910 });
8911 }
8912
LoadBootImageAddress(vixl32::Register reg,uint32_t boot_image_reference)8913 void CodeGeneratorARMVIXL::LoadBootImageAddress(vixl32::Register reg,
8914 uint32_t boot_image_reference) {
8915 if (GetCompilerOptions().IsBootImage()) {
8916 CodeGeneratorARMVIXL::PcRelativePatchInfo* labels =
8917 NewBootImageIntrinsicPatch(boot_image_reference);
8918 EmitMovwMovtPlaceholder(labels, reg);
8919 } else if (GetCompilerOptions().GetCompilePic()) {
8920 CodeGeneratorARMVIXL::PcRelativePatchInfo* labels =
8921 NewBootImageRelRoPatch(boot_image_reference);
8922 EmitMovwMovtPlaceholder(labels, reg);
8923 __ Ldr(reg, MemOperand(reg, /* offset= */ 0));
8924 } else {
8925 DCHECK(Runtime::Current()->UseJitCompilation());
8926 gc::Heap* heap = Runtime::Current()->GetHeap();
8927 DCHECK(!heap->GetBootImageSpaces().empty());
8928 uintptr_t address =
8929 reinterpret_cast<uintptr_t>(heap->GetBootImageSpaces()[0]->Begin() + boot_image_reference);
8930 __ Ldr(reg, DeduplicateBootImageAddressLiteral(dchecked_integral_cast<uint32_t>(address)));
8931 }
8932 }
8933
AllocateInstanceForIntrinsic(HInvokeStaticOrDirect * invoke,uint32_t boot_image_offset)8934 void CodeGeneratorARMVIXL::AllocateInstanceForIntrinsic(HInvokeStaticOrDirect* invoke,
8935 uint32_t boot_image_offset) {
8936 DCHECK(invoke->IsStatic());
8937 InvokeRuntimeCallingConventionARMVIXL calling_convention;
8938 vixl32::Register argument = calling_convention.GetRegisterAt(0);
8939 if (GetCompilerOptions().IsBootImage()) {
8940 DCHECK_EQ(boot_image_offset, IntrinsicVisitor::IntegerValueOfInfo::kInvalidReference);
8941 // Load the class the same way as for HLoadClass::LoadKind::kBootImageLinkTimePcRelative.
8942 MethodReference target_method = invoke->GetTargetMethod();
8943 dex::TypeIndex type_idx = target_method.dex_file->GetMethodId(target_method.index).class_idx_;
8944 PcRelativePatchInfo* labels = NewBootImageTypePatch(*target_method.dex_file, type_idx);
8945 EmitMovwMovtPlaceholder(labels, argument);
8946 } else {
8947 LoadBootImageAddress(argument, boot_image_offset);
8948 }
8949 InvokeRuntime(kQuickAllocObjectInitialized, invoke, invoke->GetDexPc());
8950 CheckEntrypointTypes<kQuickAllocObjectWithChecks, void*, mirror::Class*>();
8951 }
8952
8953 template <linker::LinkerPatch (*Factory)(size_t, const DexFile*, uint32_t, uint32_t)>
EmitPcRelativeLinkerPatches(const ArenaDeque<PcRelativePatchInfo> & infos,ArenaVector<linker::LinkerPatch> * linker_patches)8954 inline void CodeGeneratorARMVIXL::EmitPcRelativeLinkerPatches(
8955 const ArenaDeque<PcRelativePatchInfo>& infos,
8956 ArenaVector<linker::LinkerPatch>* linker_patches) {
8957 for (const PcRelativePatchInfo& info : infos) {
8958 const DexFile* dex_file = info.target_dex_file;
8959 size_t offset_or_index = info.offset_or_index;
8960 DCHECK(info.add_pc_label.IsBound());
8961 uint32_t add_pc_offset = dchecked_integral_cast<uint32_t>(info.add_pc_label.GetLocation());
8962 // Add MOVW patch.
8963 DCHECK(info.movw_label.IsBound());
8964 uint32_t movw_offset = dchecked_integral_cast<uint32_t>(info.movw_label.GetLocation());
8965 linker_patches->push_back(Factory(movw_offset, dex_file, add_pc_offset, offset_or_index));
8966 // Add MOVT patch.
8967 DCHECK(info.movt_label.IsBound());
8968 uint32_t movt_offset = dchecked_integral_cast<uint32_t>(info.movt_label.GetLocation());
8969 linker_patches->push_back(Factory(movt_offset, dex_file, add_pc_offset, offset_or_index));
8970 }
8971 }
8972
8973 template <linker::LinkerPatch (*Factory)(size_t, uint32_t, uint32_t)>
NoDexFileAdapter(size_t literal_offset,const DexFile * target_dex_file,uint32_t pc_insn_offset,uint32_t boot_image_offset)8974 linker::LinkerPatch NoDexFileAdapter(size_t literal_offset,
8975 const DexFile* target_dex_file,
8976 uint32_t pc_insn_offset,
8977 uint32_t boot_image_offset) {
8978 DCHECK(target_dex_file == nullptr); // Unused for these patches, should be null.
8979 return Factory(literal_offset, pc_insn_offset, boot_image_offset);
8980 }
8981
EmitLinkerPatches(ArenaVector<linker::LinkerPatch> * linker_patches)8982 void CodeGeneratorARMVIXL::EmitLinkerPatches(ArenaVector<linker::LinkerPatch>* linker_patches) {
8983 DCHECK(linker_patches->empty());
8984 size_t size =
8985 /* MOVW+MOVT for each entry */ 2u * boot_image_method_patches_.size() +
8986 /* MOVW+MOVT for each entry */ 2u * method_bss_entry_patches_.size() +
8987 /* MOVW+MOVT for each entry */ 2u * boot_image_type_patches_.size() +
8988 /* MOVW+MOVT for each entry */ 2u * type_bss_entry_patches_.size() +
8989 /* MOVW+MOVT for each entry */ 2u * boot_image_string_patches_.size() +
8990 /* MOVW+MOVT for each entry */ 2u * string_bss_entry_patches_.size() +
8991 /* MOVW+MOVT for each entry */ 2u * boot_image_intrinsic_patches_.size() +
8992 baker_read_barrier_patches_.size();
8993 linker_patches->reserve(size);
8994 if (GetCompilerOptions().IsBootImage()) {
8995 EmitPcRelativeLinkerPatches<linker::LinkerPatch::RelativeMethodPatch>(
8996 boot_image_method_patches_, linker_patches);
8997 EmitPcRelativeLinkerPatches<linker::LinkerPatch::RelativeTypePatch>(
8998 boot_image_type_patches_, linker_patches);
8999 EmitPcRelativeLinkerPatches<linker::LinkerPatch::RelativeStringPatch>(
9000 boot_image_string_patches_, linker_patches);
9001 EmitPcRelativeLinkerPatches<NoDexFileAdapter<linker::LinkerPatch::IntrinsicReferencePatch>>(
9002 boot_image_intrinsic_patches_, linker_patches);
9003 } else {
9004 EmitPcRelativeLinkerPatches<NoDexFileAdapter<linker::LinkerPatch::DataBimgRelRoPatch>>(
9005 boot_image_method_patches_, linker_patches);
9006 DCHECK(boot_image_type_patches_.empty());
9007 DCHECK(boot_image_string_patches_.empty());
9008 DCHECK(boot_image_intrinsic_patches_.empty());
9009 }
9010 EmitPcRelativeLinkerPatches<linker::LinkerPatch::MethodBssEntryPatch>(
9011 method_bss_entry_patches_, linker_patches);
9012 EmitPcRelativeLinkerPatches<linker::LinkerPatch::TypeBssEntryPatch>(
9013 type_bss_entry_patches_, linker_patches);
9014 EmitPcRelativeLinkerPatches<linker::LinkerPatch::StringBssEntryPatch>(
9015 string_bss_entry_patches_, linker_patches);
9016 for (const BakerReadBarrierPatchInfo& info : baker_read_barrier_patches_) {
9017 linker_patches->push_back(linker::LinkerPatch::BakerReadBarrierBranchPatch(
9018 info.label.GetLocation(), info.custom_data));
9019 }
9020 DCHECK_EQ(size, linker_patches->size());
9021 }
9022
NeedsThunkCode(const linker::LinkerPatch & patch) const9023 bool CodeGeneratorARMVIXL::NeedsThunkCode(const linker::LinkerPatch& patch) const {
9024 return patch.GetType() == linker::LinkerPatch::Type::kBakerReadBarrierBranch ||
9025 patch.GetType() == linker::LinkerPatch::Type::kCallRelative;
9026 }
9027
EmitThunkCode(const linker::LinkerPatch & patch,ArenaVector<uint8_t> * code,std::string * debug_name)9028 void CodeGeneratorARMVIXL::EmitThunkCode(const linker::LinkerPatch& patch,
9029 /*out*/ ArenaVector<uint8_t>* code,
9030 /*out*/ std::string* debug_name) {
9031 arm::ArmVIXLAssembler assembler(GetGraph()->GetAllocator());
9032 switch (patch.GetType()) {
9033 case linker::LinkerPatch::Type::kCallRelative:
9034 // The thunk just uses the entry point in the ArtMethod. This works even for calls
9035 // to the generic JNI and interpreter trampolines.
9036 assembler.LoadFromOffset(
9037 arm::kLoadWord,
9038 vixl32::pc,
9039 vixl32::r0,
9040 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmPointerSize).Int32Value());
9041 assembler.GetVIXLAssembler()->Bkpt(0);
9042 if (GetCompilerOptions().GenerateAnyDebugInfo()) {
9043 *debug_name = "MethodCallThunk";
9044 }
9045 break;
9046 case linker::LinkerPatch::Type::kBakerReadBarrierBranch:
9047 DCHECK_EQ(patch.GetBakerCustomValue2(), 0u);
9048 CompileBakerReadBarrierThunk(assembler, patch.GetBakerCustomValue1(), debug_name);
9049 break;
9050 default:
9051 LOG(FATAL) << "Unexpected patch type " << patch.GetType();
9052 UNREACHABLE();
9053 }
9054
9055 // Ensure we emit the literal pool if any.
9056 assembler.FinalizeCode();
9057 code->resize(assembler.CodeSize());
9058 MemoryRegion code_region(code->data(), code->size());
9059 assembler.FinalizeInstructions(code_region);
9060 }
9061
DeduplicateUint32Literal(uint32_t value,Uint32ToLiteralMap * map)9062 VIXLUInt32Literal* CodeGeneratorARMVIXL::DeduplicateUint32Literal(
9063 uint32_t value,
9064 Uint32ToLiteralMap* map) {
9065 return map->GetOrCreate(
9066 value,
9067 [this, value]() {
9068 return GetAssembler()->CreateLiteralDestroyedWithPool<uint32_t>(/* value= */ value);
9069 });
9070 }
9071
VisitMultiplyAccumulate(HMultiplyAccumulate * instr)9072 void LocationsBuilderARMVIXL::VisitMultiplyAccumulate(HMultiplyAccumulate* instr) {
9073 LocationSummary* locations =
9074 new (GetGraph()->GetAllocator()) LocationSummary(instr, LocationSummary::kNoCall);
9075 locations->SetInAt(HMultiplyAccumulate::kInputAccumulatorIndex,
9076 Location::RequiresRegister());
9077 locations->SetInAt(HMultiplyAccumulate::kInputMulLeftIndex, Location::RequiresRegister());
9078 locations->SetInAt(HMultiplyAccumulate::kInputMulRightIndex, Location::RequiresRegister());
9079 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
9080 }
9081
VisitMultiplyAccumulate(HMultiplyAccumulate * instr)9082 void InstructionCodeGeneratorARMVIXL::VisitMultiplyAccumulate(HMultiplyAccumulate* instr) {
9083 vixl32::Register res = OutputRegister(instr);
9084 vixl32::Register accumulator =
9085 InputRegisterAt(instr, HMultiplyAccumulate::kInputAccumulatorIndex);
9086 vixl32::Register mul_left =
9087 InputRegisterAt(instr, HMultiplyAccumulate::kInputMulLeftIndex);
9088 vixl32::Register mul_right =
9089 InputRegisterAt(instr, HMultiplyAccumulate::kInputMulRightIndex);
9090
9091 if (instr->GetOpKind() == HInstruction::kAdd) {
9092 __ Mla(res, mul_left, mul_right, accumulator);
9093 } else {
9094 __ Mls(res, mul_left, mul_right, accumulator);
9095 }
9096 }
9097
VisitBoundType(HBoundType * instruction ATTRIBUTE_UNUSED)9098 void LocationsBuilderARMVIXL::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) {
9099 // Nothing to do, this should be removed during prepare for register allocator.
9100 LOG(FATAL) << "Unreachable";
9101 }
9102
VisitBoundType(HBoundType * instruction ATTRIBUTE_UNUSED)9103 void InstructionCodeGeneratorARMVIXL::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) {
9104 // Nothing to do, this should be removed during prepare for register allocator.
9105 LOG(FATAL) << "Unreachable";
9106 }
9107
9108 // Simple implementation of packed switch - generate cascaded compare/jumps.
VisitPackedSwitch(HPackedSwitch * switch_instr)9109 void LocationsBuilderARMVIXL::VisitPackedSwitch(HPackedSwitch* switch_instr) {
9110 LocationSummary* locations =
9111 new (GetGraph()->GetAllocator()) LocationSummary(switch_instr, LocationSummary::kNoCall);
9112 locations->SetInAt(0, Location::RequiresRegister());
9113 if (switch_instr->GetNumEntries() > kPackedSwitchCompareJumpThreshold &&
9114 codegen_->GetAssembler()->GetVIXLAssembler()->IsUsingT32()) {
9115 locations->AddTemp(Location::RequiresRegister()); // We need a temp for the table base.
9116 if (switch_instr->GetStartValue() != 0) {
9117 locations->AddTemp(Location::RequiresRegister()); // We need a temp for the bias.
9118 }
9119 }
9120 }
9121
9122 // TODO(VIXL): Investigate and reach the parity with old arm codegen.
VisitPackedSwitch(HPackedSwitch * switch_instr)9123 void InstructionCodeGeneratorARMVIXL::VisitPackedSwitch(HPackedSwitch* switch_instr) {
9124 int32_t lower_bound = switch_instr->GetStartValue();
9125 uint32_t num_entries = switch_instr->GetNumEntries();
9126 LocationSummary* locations = switch_instr->GetLocations();
9127 vixl32::Register value_reg = InputRegisterAt(switch_instr, 0);
9128 HBasicBlock* default_block = switch_instr->GetDefaultBlock();
9129
9130 if (num_entries <= kPackedSwitchCompareJumpThreshold ||
9131 !codegen_->GetAssembler()->GetVIXLAssembler()->IsUsingT32()) {
9132 // Create a series of compare/jumps.
9133 UseScratchRegisterScope temps(GetVIXLAssembler());
9134 vixl32::Register temp_reg = temps.Acquire();
9135 // Note: It is fine for the below AddConstantSetFlags() using IP register to temporarily store
9136 // the immediate, because IP is used as the destination register. For the other
9137 // AddConstantSetFlags() and GenerateCompareWithImmediate(), the immediate values are constant,
9138 // and they can be encoded in the instruction without making use of IP register.
9139 __ Adds(temp_reg, value_reg, -lower_bound);
9140
9141 const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors();
9142 // Jump to successors[0] if value == lower_bound.
9143 __ B(eq, codegen_->GetLabelOf(successors[0]));
9144 int32_t last_index = 0;
9145 for (; num_entries - last_index > 2; last_index += 2) {
9146 __ Adds(temp_reg, temp_reg, -2);
9147 // Jump to successors[last_index + 1] if value < case_value[last_index + 2].
9148 __ B(lo, codegen_->GetLabelOf(successors[last_index + 1]));
9149 // Jump to successors[last_index + 2] if value == case_value[last_index + 2].
9150 __ B(eq, codegen_->GetLabelOf(successors[last_index + 2]));
9151 }
9152 if (num_entries - last_index == 2) {
9153 // The last missing case_value.
9154 __ Cmp(temp_reg, 1);
9155 __ B(eq, codegen_->GetLabelOf(successors[last_index + 1]));
9156 }
9157
9158 // And the default for any other value.
9159 if (!codegen_->GoesToNextBlock(switch_instr->GetBlock(), default_block)) {
9160 __ B(codegen_->GetLabelOf(default_block));
9161 }
9162 } else {
9163 // Create a table lookup.
9164 vixl32::Register table_base = RegisterFrom(locations->GetTemp(0));
9165
9166 JumpTableARMVIXL* jump_table = codegen_->CreateJumpTable(switch_instr);
9167
9168 // Remove the bias.
9169 vixl32::Register key_reg;
9170 if (lower_bound != 0) {
9171 key_reg = RegisterFrom(locations->GetTemp(1));
9172 __ Sub(key_reg, value_reg, lower_bound);
9173 } else {
9174 key_reg = value_reg;
9175 }
9176
9177 // Check whether the value is in the table, jump to default block if not.
9178 __ Cmp(key_reg, num_entries - 1);
9179 __ B(hi, codegen_->GetLabelOf(default_block));
9180
9181 UseScratchRegisterScope temps(GetVIXLAssembler());
9182 vixl32::Register jump_offset = temps.Acquire();
9183
9184 // Load jump offset from the table.
9185 {
9186 const size_t jump_size = switch_instr->GetNumEntries() * sizeof(int32_t);
9187 ExactAssemblyScope aas(GetVIXLAssembler(),
9188 (vixl32::kMaxInstructionSizeInBytes * 4) + jump_size,
9189 CodeBufferCheckScope::kMaximumSize);
9190 __ adr(table_base, jump_table->GetTableStartLabel());
9191 __ ldr(jump_offset, MemOperand(table_base, key_reg, vixl32::LSL, 2));
9192
9193 // Jump to target block by branching to table_base(pc related) + offset.
9194 vixl32::Register target_address = table_base;
9195 __ add(target_address, table_base, jump_offset);
9196 __ bx(target_address);
9197
9198 jump_table->EmitTable(codegen_);
9199 }
9200 }
9201 }
9202
9203 // Copy the result of a call into the given target.
MoveFromReturnRegister(Location trg,DataType::Type type)9204 void CodeGeneratorARMVIXL::MoveFromReturnRegister(Location trg, DataType::Type type) {
9205 if (!trg.IsValid()) {
9206 DCHECK_EQ(type, DataType::Type::kVoid);
9207 return;
9208 }
9209
9210 DCHECK_NE(type, DataType::Type::kVoid);
9211
9212 Location return_loc = InvokeDexCallingConventionVisitorARMVIXL().GetReturnLocation(type);
9213 if (return_loc.Equals(trg)) {
9214 return;
9215 }
9216
9217 // TODO: Consider pairs in the parallel move resolver, then this could be nicely merged
9218 // with the last branch.
9219 if (type == DataType::Type::kInt64) {
9220 TODO_VIXL32(FATAL);
9221 } else if (type == DataType::Type::kFloat64) {
9222 TODO_VIXL32(FATAL);
9223 } else {
9224 // Let the parallel move resolver take care of all of this.
9225 HParallelMove parallel_move(GetGraph()->GetAllocator());
9226 parallel_move.AddMove(return_loc, trg, type, nullptr);
9227 GetMoveResolver()->EmitNativeCode(¶llel_move);
9228 }
9229 }
9230
VisitClassTableGet(HClassTableGet * instruction)9231 void LocationsBuilderARMVIXL::VisitClassTableGet(HClassTableGet* instruction) {
9232 LocationSummary* locations =
9233 new (GetGraph()->GetAllocator()) LocationSummary(instruction, LocationSummary::kNoCall);
9234 locations->SetInAt(0, Location::RequiresRegister());
9235 locations->SetOut(Location::RequiresRegister());
9236 }
9237
VisitClassTableGet(HClassTableGet * instruction)9238 void InstructionCodeGeneratorARMVIXL::VisitClassTableGet(HClassTableGet* instruction) {
9239 if (instruction->GetTableKind() == HClassTableGet::TableKind::kVTable) {
9240 uint32_t method_offset = mirror::Class::EmbeddedVTableEntryOffset(
9241 instruction->GetIndex(), kArmPointerSize).SizeValue();
9242 GetAssembler()->LoadFromOffset(kLoadWord,
9243 OutputRegister(instruction),
9244 InputRegisterAt(instruction, 0),
9245 method_offset);
9246 } else {
9247 uint32_t method_offset = static_cast<uint32_t>(ImTable::OffsetOfElement(
9248 instruction->GetIndex(), kArmPointerSize));
9249 GetAssembler()->LoadFromOffset(kLoadWord,
9250 OutputRegister(instruction),
9251 InputRegisterAt(instruction, 0),
9252 mirror::Class::ImtPtrOffset(kArmPointerSize).Uint32Value());
9253 GetAssembler()->LoadFromOffset(kLoadWord,
9254 OutputRegister(instruction),
9255 OutputRegister(instruction),
9256 method_offset);
9257 }
9258 }
9259
PatchJitRootUse(uint8_t * code,const uint8_t * roots_data,VIXLUInt32Literal * literal,uint64_t index_in_table)9260 static void PatchJitRootUse(uint8_t* code,
9261 const uint8_t* roots_data,
9262 VIXLUInt32Literal* literal,
9263 uint64_t index_in_table) {
9264 DCHECK(literal->IsBound());
9265 uint32_t literal_offset = literal->GetLocation();
9266 uintptr_t address =
9267 reinterpret_cast<uintptr_t>(roots_data) + index_in_table * sizeof(GcRoot<mirror::Object>);
9268 uint8_t* data = code + literal_offset;
9269 reinterpret_cast<uint32_t*>(data)[0] = dchecked_integral_cast<uint32_t>(address);
9270 }
9271
EmitJitRootPatches(uint8_t * code,const uint8_t * roots_data)9272 void CodeGeneratorARMVIXL::EmitJitRootPatches(uint8_t* code, const uint8_t* roots_data) {
9273 for (const auto& entry : jit_string_patches_) {
9274 const StringReference& string_reference = entry.first;
9275 VIXLUInt32Literal* table_entry_literal = entry.second;
9276 uint64_t index_in_table = GetJitStringRootIndex(string_reference);
9277 PatchJitRootUse(code, roots_data, table_entry_literal, index_in_table);
9278 }
9279 for (const auto& entry : jit_class_patches_) {
9280 const TypeReference& type_reference = entry.first;
9281 VIXLUInt32Literal* table_entry_literal = entry.second;
9282 uint64_t index_in_table = GetJitClassRootIndex(type_reference);
9283 PatchJitRootUse(code, roots_data, table_entry_literal, index_in_table);
9284 }
9285 }
9286
EmitMovwMovtPlaceholder(CodeGeneratorARMVIXL::PcRelativePatchInfo * labels,vixl32::Register out)9287 void CodeGeneratorARMVIXL::EmitMovwMovtPlaceholder(
9288 CodeGeneratorARMVIXL::PcRelativePatchInfo* labels,
9289 vixl32::Register out) {
9290 ExactAssemblyScope aas(GetVIXLAssembler(),
9291 3 * vixl32::kMaxInstructionSizeInBytes,
9292 CodeBufferCheckScope::kMaximumSize);
9293 // TODO(VIXL): Think about using mov instead of movw.
9294 __ bind(&labels->movw_label);
9295 __ movw(out, /* operand= */ 0u);
9296 __ bind(&labels->movt_label);
9297 __ movt(out, /* operand= */ 0u);
9298 __ bind(&labels->add_pc_label);
9299 __ add(out, out, pc);
9300 }
9301
9302 #undef __
9303 #undef QUICK_ENTRY_POINT
9304 #undef TODO_VIXL32
9305
9306 #define __ assembler.GetVIXLAssembler()->
9307
EmitGrayCheckAndFastPath(ArmVIXLAssembler & assembler,vixl32::Register base_reg,vixl32::MemOperand & lock_word,vixl32::Label * slow_path,int32_t raw_ldr_offset,vixl32::Label * throw_npe=nullptr)9308 static void EmitGrayCheckAndFastPath(ArmVIXLAssembler& assembler,
9309 vixl32::Register base_reg,
9310 vixl32::MemOperand& lock_word,
9311 vixl32::Label* slow_path,
9312 int32_t raw_ldr_offset,
9313 vixl32::Label* throw_npe = nullptr) {
9314 // Load the lock word containing the rb_state.
9315 __ Ldr(ip, lock_word);
9316 // Given the numeric representation, it's enough to check the low bit of the rb_state.
9317 static_assert(ReadBarrier::NonGrayState() == 0, "Expecting non-gray to have value 0");
9318 static_assert(ReadBarrier::GrayState() == 1, "Expecting gray to have value 1");
9319 __ Tst(ip, Operand(LockWord::kReadBarrierStateMaskShifted));
9320 __ B(ne, slow_path, /* is_far_target= */ false);
9321 // To throw NPE, we return to the fast path; the artificial dependence below does not matter.
9322 if (throw_npe != nullptr) {
9323 __ Bind(throw_npe);
9324 }
9325 __ Add(lr, lr, raw_ldr_offset);
9326 // Introduce a dependency on the lock_word including rb_state,
9327 // to prevent load-load reordering, and without using
9328 // a memory barrier (which would be more expensive).
9329 __ Add(base_reg, base_reg, Operand(ip, LSR, 32));
9330 __ Bx(lr); // And return back to the function.
9331 // Note: The fake dependency is unnecessary for the slow path.
9332 }
9333
9334 // Load the read barrier introspection entrypoint in register `entrypoint`
LoadReadBarrierMarkIntrospectionEntrypoint(ArmVIXLAssembler & assembler)9335 static vixl32::Register LoadReadBarrierMarkIntrospectionEntrypoint(ArmVIXLAssembler& assembler) {
9336 // The register where the read barrier introspection entrypoint is loaded
9337 // is the marking register. We clobber it here and the entrypoint restores it to 1.
9338 vixl32::Register entrypoint = mr;
9339 // entrypoint = Thread::Current()->pReadBarrierMarkReg12, i.e. pReadBarrierMarkIntrospection.
9340 DCHECK_EQ(ip.GetCode(), 12u);
9341 const int32_t entry_point_offset =
9342 Thread::ReadBarrierMarkEntryPointsOffset<kArmPointerSize>(ip.GetCode());
9343 __ Ldr(entrypoint, MemOperand(tr, entry_point_offset));
9344 return entrypoint;
9345 }
9346
CompileBakerReadBarrierThunk(ArmVIXLAssembler & assembler,uint32_t encoded_data,std::string * debug_name)9347 void CodeGeneratorARMVIXL::CompileBakerReadBarrierThunk(ArmVIXLAssembler& assembler,
9348 uint32_t encoded_data,
9349 /*out*/ std::string* debug_name) {
9350 BakerReadBarrierKind kind = BakerReadBarrierKindField::Decode(encoded_data);
9351 switch (kind) {
9352 case BakerReadBarrierKind::kField: {
9353 vixl32::Register base_reg(BakerReadBarrierFirstRegField::Decode(encoded_data));
9354 CheckValidReg(base_reg.GetCode());
9355 vixl32::Register holder_reg(BakerReadBarrierSecondRegField::Decode(encoded_data));
9356 CheckValidReg(holder_reg.GetCode());
9357 BakerReadBarrierWidth width = BakerReadBarrierWidthField::Decode(encoded_data);
9358 UseScratchRegisterScope temps(assembler.GetVIXLAssembler());
9359 temps.Exclude(ip);
9360 // If base_reg differs from holder_reg, the offset was too large and we must have emitted
9361 // an explicit null check before the load. Otherwise, for implicit null checks, we need to
9362 // null-check the holder as we do not necessarily do that check before going to the thunk.
9363 vixl32::Label throw_npe_label;
9364 vixl32::Label* throw_npe = nullptr;
9365 if (GetCompilerOptions().GetImplicitNullChecks() && holder_reg.Is(base_reg)) {
9366 throw_npe = &throw_npe_label;
9367 __ CompareAndBranchIfZero(holder_reg, throw_npe, /* is_far_target= */ false);
9368 }
9369 // Check if the holder is gray and, if not, add fake dependency to the base register
9370 // and return to the LDR instruction to load the reference. Otherwise, use introspection
9371 // to load the reference and call the entrypoint that performs further checks on the
9372 // reference and marks it if needed.
9373 vixl32::Label slow_path;
9374 MemOperand lock_word(holder_reg, mirror::Object::MonitorOffset().Int32Value());
9375 const int32_t raw_ldr_offset = (width == BakerReadBarrierWidth::kWide)
9376 ? BAKER_MARK_INTROSPECTION_FIELD_LDR_WIDE_OFFSET
9377 : BAKER_MARK_INTROSPECTION_FIELD_LDR_NARROW_OFFSET;
9378 EmitGrayCheckAndFastPath(
9379 assembler, base_reg, lock_word, &slow_path, raw_ldr_offset, throw_npe);
9380 __ Bind(&slow_path);
9381 const int32_t ldr_offset = /* Thumb state adjustment (LR contains Thumb state). */ -1 +
9382 raw_ldr_offset;
9383 vixl32::Register ep_reg = LoadReadBarrierMarkIntrospectionEntrypoint(assembler);
9384 if (width == BakerReadBarrierWidth::kWide) {
9385 MemOperand ldr_half_address(lr, ldr_offset + 2);
9386 __ Ldrh(ip, ldr_half_address); // Load the LDR immediate half-word with "Rt | imm12".
9387 __ Ubfx(ip, ip, 0, 12); // Extract the offset imm12.
9388 __ Ldr(ip, MemOperand(base_reg, ip)); // Load the reference.
9389 } else {
9390 MemOperand ldr_address(lr, ldr_offset);
9391 __ Ldrh(ip, ldr_address); // Load the LDR immediate, encoding T1.
9392 __ Add(ep_reg, // Adjust the entrypoint address to the entrypoint
9393 ep_reg, // for narrow LDR.
9394 Operand(BAKER_MARK_INTROSPECTION_FIELD_LDR_NARROW_ENTRYPOINT_OFFSET));
9395 __ Ubfx(ip, ip, 6, 5); // Extract the imm5, i.e. offset / 4.
9396 __ Ldr(ip, MemOperand(base_reg, ip, LSL, 2)); // Load the reference.
9397 }
9398 // Do not unpoison. With heap poisoning enabled, the entrypoint expects a poisoned reference.
9399 __ Bx(ep_reg); // Jump to the entrypoint.
9400 break;
9401 }
9402 case BakerReadBarrierKind::kArray: {
9403 vixl32::Register base_reg(BakerReadBarrierFirstRegField::Decode(encoded_data));
9404 CheckValidReg(base_reg.GetCode());
9405 DCHECK_EQ(kBakerReadBarrierInvalidEncodedReg,
9406 BakerReadBarrierSecondRegField::Decode(encoded_data));
9407 DCHECK(BakerReadBarrierWidthField::Decode(encoded_data) == BakerReadBarrierWidth::kWide);
9408 UseScratchRegisterScope temps(assembler.GetVIXLAssembler());
9409 temps.Exclude(ip);
9410 vixl32::Label slow_path;
9411 int32_t data_offset =
9412 mirror::Array::DataOffset(Primitive::ComponentSize(Primitive::kPrimNot)).Int32Value();
9413 MemOperand lock_word(base_reg, mirror::Object::MonitorOffset().Int32Value() - data_offset);
9414 DCHECK_LT(lock_word.GetOffsetImmediate(), 0);
9415 const int32_t raw_ldr_offset = BAKER_MARK_INTROSPECTION_ARRAY_LDR_OFFSET;
9416 EmitGrayCheckAndFastPath(assembler, base_reg, lock_word, &slow_path, raw_ldr_offset);
9417 __ Bind(&slow_path);
9418 const int32_t ldr_offset = /* Thumb state adjustment (LR contains Thumb state). */ -1 +
9419 raw_ldr_offset;
9420 MemOperand ldr_address(lr, ldr_offset + 2);
9421 __ Ldrb(ip, ldr_address); // Load the LDR (register) byte with "00 | imm2 | Rm",
9422 // i.e. Rm+32 because the scale in imm2 is 2.
9423 vixl32::Register ep_reg = LoadReadBarrierMarkIntrospectionEntrypoint(assembler);
9424 __ Bfi(ep_reg, ip, 3, 6); // Insert ip to the entrypoint address to create
9425 // a switch case target based on the index register.
9426 __ Mov(ip, base_reg); // Move the base register to ip0.
9427 __ Bx(ep_reg); // Jump to the entrypoint's array switch case.
9428 break;
9429 }
9430 case BakerReadBarrierKind::kGcRoot:
9431 case BakerReadBarrierKind::kUnsafeCas: {
9432 // Check if the reference needs to be marked and if so (i.e. not null, not marked yet
9433 // and it does not have a forwarding address), call the correct introspection entrypoint;
9434 // otherwise return the reference (or the extracted forwarding address).
9435 // There is no gray bit check for GC roots.
9436 vixl32::Register root_reg(BakerReadBarrierFirstRegField::Decode(encoded_data));
9437 CheckValidReg(root_reg.GetCode());
9438 DCHECK_EQ(kBakerReadBarrierInvalidEncodedReg,
9439 BakerReadBarrierSecondRegField::Decode(encoded_data));
9440 BakerReadBarrierWidth width = BakerReadBarrierWidthField::Decode(encoded_data);
9441 UseScratchRegisterScope temps(assembler.GetVIXLAssembler());
9442 temps.Exclude(ip);
9443 vixl32::Label return_label, not_marked, forwarding_address;
9444 __ CompareAndBranchIfZero(root_reg, &return_label, /* is_far_target= */ false);
9445 MemOperand lock_word(root_reg, mirror::Object::MonitorOffset().Int32Value());
9446 __ Ldr(ip, lock_word);
9447 __ Tst(ip, LockWord::kMarkBitStateMaskShifted);
9448 __ B(eq, ¬_marked);
9449 __ Bind(&return_label);
9450 __ Bx(lr);
9451 __ Bind(¬_marked);
9452 static_assert(LockWord::kStateShift == 30 && LockWord::kStateForwardingAddress == 3,
9453 "To use 'CMP ip, #modified-immediate; BHS', we need the lock word state in "
9454 " the highest bits and the 'forwarding address' state to have all bits set");
9455 __ Cmp(ip, Operand(0xc0000000));
9456 __ B(hs, &forwarding_address);
9457 vixl32::Register ep_reg = LoadReadBarrierMarkIntrospectionEntrypoint(assembler);
9458 // Adjust the art_quick_read_barrier_mark_introspection address in kBakerCcEntrypointRegister
9459 // to one of art_quick_read_barrier_mark_introspection_{gc_roots_{wide,narrow},unsafe_cas}.
9460 DCHECK(kind != BakerReadBarrierKind::kUnsafeCas || width == BakerReadBarrierWidth::kWide);
9461 int32_t entrypoint_offset =
9462 (kind == BakerReadBarrierKind::kGcRoot)
9463 ? (width == BakerReadBarrierWidth::kWide)
9464 ? BAKER_MARK_INTROSPECTION_GC_ROOT_LDR_WIDE_ENTRYPOINT_OFFSET
9465 : BAKER_MARK_INTROSPECTION_GC_ROOT_LDR_NARROW_ENTRYPOINT_OFFSET
9466 : BAKER_MARK_INTROSPECTION_UNSAFE_CAS_ENTRYPOINT_OFFSET;
9467 __ Add(ep_reg, ep_reg, Operand(entrypoint_offset));
9468 __ Mov(ip, root_reg);
9469 __ Bx(ep_reg);
9470 __ Bind(&forwarding_address);
9471 __ Lsl(root_reg, ip, LockWord::kForwardingAddressShift);
9472 __ Bx(lr);
9473 break;
9474 }
9475 default:
9476 LOG(FATAL) << "Unexpected kind: " << static_cast<uint32_t>(kind);
9477 UNREACHABLE();
9478 }
9479
9480 // For JIT, the slow path is considered part of the compiled method,
9481 // so JIT should pass null as `debug_name`. Tests may not have a runtime.
9482 DCHECK(Runtime::Current() == nullptr ||
9483 !Runtime::Current()->UseJitCompilation() ||
9484 debug_name == nullptr);
9485 if (debug_name != nullptr && GetCompilerOptions().GenerateAnyDebugInfo()) {
9486 std::ostringstream oss;
9487 oss << "BakerReadBarrierThunk";
9488 switch (kind) {
9489 case BakerReadBarrierKind::kField:
9490 oss << "Field";
9491 if (BakerReadBarrierWidthField::Decode(encoded_data) == BakerReadBarrierWidth::kWide) {
9492 oss << "Wide";
9493 }
9494 oss << "_r" << BakerReadBarrierFirstRegField::Decode(encoded_data)
9495 << "_r" << BakerReadBarrierSecondRegField::Decode(encoded_data);
9496 break;
9497 case BakerReadBarrierKind::kArray:
9498 oss << "Array_r" << BakerReadBarrierFirstRegField::Decode(encoded_data);
9499 DCHECK_EQ(kBakerReadBarrierInvalidEncodedReg,
9500 BakerReadBarrierSecondRegField::Decode(encoded_data));
9501 DCHECK(BakerReadBarrierWidthField::Decode(encoded_data) == BakerReadBarrierWidth::kWide);
9502 break;
9503 case BakerReadBarrierKind::kGcRoot:
9504 oss << "GcRoot";
9505 if (BakerReadBarrierWidthField::Decode(encoded_data) == BakerReadBarrierWidth::kWide) {
9506 oss << "Wide";
9507 }
9508 oss << "_r" << BakerReadBarrierFirstRegField::Decode(encoded_data);
9509 DCHECK_EQ(kBakerReadBarrierInvalidEncodedReg,
9510 BakerReadBarrierSecondRegField::Decode(encoded_data));
9511 break;
9512 case BakerReadBarrierKind::kUnsafeCas:
9513 oss << "UnsafeCas_r" << BakerReadBarrierFirstRegField::Decode(encoded_data);
9514 DCHECK_EQ(kBakerReadBarrierInvalidEncodedReg,
9515 BakerReadBarrierSecondRegField::Decode(encoded_data));
9516 DCHECK(BakerReadBarrierWidthField::Decode(encoded_data) == BakerReadBarrierWidth::kWide);
9517 break;
9518 }
9519 *debug_name = oss.str();
9520 }
9521 }
9522
9523 #undef __
9524
9525 } // namespace arm
9526 } // namespace art
9527