1 /*
2 * Copyright (C) 2015 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 #ifndef ART_COMPILER_OPTIMIZING_COMMON_ARM64_H_
18 #define ART_COMPILER_OPTIMIZING_COMMON_ARM64_H_
19
20 #include "code_generator.h"
21 #include "instruction_simplifier_shared.h"
22 #include "locations.h"
23 #include "nodes.h"
24 #include "utils/arm64/assembler_arm64.h"
25
26 // TODO(VIXL): Make VIXL compile with -Wshadow.
27 #pragma GCC diagnostic push
28 #pragma GCC diagnostic ignored "-Wshadow"
29 #include "aarch64/disasm-aarch64.h"
30 #include "aarch64/macro-assembler-aarch64.h"
31 #include "aarch64/simulator-aarch64.h"
32 #pragma GCC diagnostic pop
33
34 namespace art {
35
36 using helpers::CanFitInShifterOperand;
37 using helpers::HasShifterOperand;
38
39 namespace arm64 {
40 namespace helpers {
41
42 // Convenience helpers to ease conversion to and from VIXL operands.
43 static_assert((SP == 31) && (WSP == 31) && (XZR == 32) && (WZR == 32),
44 "Unexpected values for register codes.");
45
VIXLRegCodeFromART(int code)46 inline int VIXLRegCodeFromART(int code) {
47 if (code == SP) {
48 return vixl::aarch64::kSPRegInternalCode;
49 }
50 if (code == XZR) {
51 return vixl::aarch64::kZeroRegCode;
52 }
53 return code;
54 }
55
ARTRegCodeFromVIXL(int code)56 inline int ARTRegCodeFromVIXL(int code) {
57 if (code == vixl::aarch64::kSPRegInternalCode) {
58 return SP;
59 }
60 if (code == vixl::aarch64::kZeroRegCode) {
61 return XZR;
62 }
63 return code;
64 }
65
XRegisterFrom(Location location)66 inline vixl::aarch64::Register XRegisterFrom(Location location) {
67 DCHECK(location.IsRegister()) << location;
68 return vixl::aarch64::XRegister(VIXLRegCodeFromART(location.reg()));
69 }
70
WRegisterFrom(Location location)71 inline vixl::aarch64::Register WRegisterFrom(Location location) {
72 DCHECK(location.IsRegister()) << location;
73 return vixl::aarch64::WRegister(VIXLRegCodeFromART(location.reg()));
74 }
75
RegisterFrom(Location location,DataType::Type type)76 inline vixl::aarch64::Register RegisterFrom(Location location, DataType::Type type) {
77 DCHECK(type != DataType::Type::kVoid && !DataType::IsFloatingPointType(type)) << type;
78 return type == DataType::Type::kInt64 ? XRegisterFrom(location) : WRegisterFrom(location);
79 }
80
OutputRegister(HInstruction * instr)81 inline vixl::aarch64::Register OutputRegister(HInstruction* instr) {
82 return RegisterFrom(instr->GetLocations()->Out(), instr->GetType());
83 }
84
InputRegisterAt(HInstruction * instr,int input_index)85 inline vixl::aarch64::Register InputRegisterAt(HInstruction* instr, int input_index) {
86 return RegisterFrom(instr->GetLocations()->InAt(input_index),
87 instr->InputAt(input_index)->GetType());
88 }
89
DRegisterFrom(Location location)90 inline vixl::aarch64::VRegister DRegisterFrom(Location location) {
91 DCHECK(location.IsFpuRegister()) << location;
92 return vixl::aarch64::DRegister(location.reg());
93 }
94
QRegisterFrom(Location location)95 inline vixl::aarch64::VRegister QRegisterFrom(Location location) {
96 DCHECK(location.IsFpuRegister()) << location;
97 return vixl::aarch64::QRegister(location.reg());
98 }
99
VRegisterFrom(Location location)100 inline vixl::aarch64::VRegister VRegisterFrom(Location location) {
101 DCHECK(location.IsFpuRegister()) << location;
102 return vixl::aarch64::VRegister(location.reg());
103 }
104
ZRegisterFrom(Location location)105 inline vixl::aarch64::ZRegister ZRegisterFrom(Location location) {
106 DCHECK(location.IsFpuRegister()) << location;
107 return vixl::aarch64::ZRegister(location.reg());
108 }
109
SRegisterFrom(Location location)110 inline vixl::aarch64::VRegister SRegisterFrom(Location location) {
111 DCHECK(location.IsFpuRegister()) << location;
112 return vixl::aarch64::SRegister(location.reg());
113 }
114
HRegisterFrom(Location location)115 inline vixl::aarch64::VRegister HRegisterFrom(Location location) {
116 DCHECK(location.IsFpuRegister()) << location;
117 return vixl::aarch64::HRegister(location.reg());
118 }
119
FPRegisterFrom(Location location,DataType::Type type)120 inline vixl::aarch64::VRegister FPRegisterFrom(Location location, DataType::Type type) {
121 DCHECK(DataType::IsFloatingPointType(type)) << type;
122 return type == DataType::Type::kFloat64 ? DRegisterFrom(location) : SRegisterFrom(location);
123 }
124
OutputFPRegister(HInstruction * instr)125 inline vixl::aarch64::VRegister OutputFPRegister(HInstruction* instr) {
126 return FPRegisterFrom(instr->GetLocations()->Out(), instr->GetType());
127 }
128
InputFPRegisterAt(HInstruction * instr,int input_index)129 inline vixl::aarch64::VRegister InputFPRegisterAt(HInstruction* instr, int input_index) {
130 return FPRegisterFrom(instr->GetLocations()->InAt(input_index),
131 instr->InputAt(input_index)->GetType());
132 }
133
CPURegisterFrom(Location location,DataType::Type type)134 inline vixl::aarch64::CPURegister CPURegisterFrom(Location location, DataType::Type type) {
135 return DataType::IsFloatingPointType(type)
136 ? vixl::aarch64::CPURegister(FPRegisterFrom(location, type))
137 : vixl::aarch64::CPURegister(RegisterFrom(location, type));
138 }
139
OutputCPURegister(HInstruction * instr)140 inline vixl::aarch64::CPURegister OutputCPURegister(HInstruction* instr) {
141 return DataType::IsFloatingPointType(instr->GetType())
142 ? static_cast<vixl::aarch64::CPURegister>(OutputFPRegister(instr))
143 : static_cast<vixl::aarch64::CPURegister>(OutputRegister(instr));
144 }
145
InputCPURegisterAt(HInstruction * instr,int index)146 inline vixl::aarch64::CPURegister InputCPURegisterAt(HInstruction* instr, int index) {
147 return DataType::IsFloatingPointType(instr->InputAt(index)->GetType())
148 ? static_cast<vixl::aarch64::CPURegister>(InputFPRegisterAt(instr, index))
149 : static_cast<vixl::aarch64::CPURegister>(InputRegisterAt(instr, index));
150 }
151
InputCPURegisterOrZeroRegAt(HInstruction * instr,int index)152 inline vixl::aarch64::CPURegister InputCPURegisterOrZeroRegAt(HInstruction* instr,
153 int index) {
154 HInstruction* input = instr->InputAt(index);
155 DataType::Type input_type = input->GetType();
156 if (input->IsConstant() && input->AsConstant()->IsZeroBitPattern()) {
157 return (DataType::Size(input_type) >= vixl::aarch64::kXRegSizeInBytes)
158 ? vixl::aarch64::Register(vixl::aarch64::xzr)
159 : vixl::aarch64::Register(vixl::aarch64::wzr);
160 }
161 return InputCPURegisterAt(instr, index);
162 }
163
Int64FromLocation(Location location)164 inline int64_t Int64FromLocation(Location location) {
165 return Int64FromConstant(location.GetConstant());
166 }
167
OperandFrom(Location location,DataType::Type type)168 inline vixl::aarch64::Operand OperandFrom(Location location, DataType::Type type) {
169 if (location.IsRegister()) {
170 return vixl::aarch64::Operand(RegisterFrom(location, type));
171 } else {
172 return vixl::aarch64::Operand(Int64FromLocation(location));
173 }
174 }
175
InputOperandAt(HInstruction * instr,int input_index)176 inline vixl::aarch64::Operand InputOperandAt(HInstruction* instr, int input_index) {
177 return OperandFrom(instr->GetLocations()->InAt(input_index),
178 instr->InputAt(input_index)->GetType());
179 }
180
StackOperandFrom(Location location)181 inline vixl::aarch64::MemOperand StackOperandFrom(Location location) {
182 return vixl::aarch64::MemOperand(vixl::aarch64::sp, location.GetStackIndex());
183 }
184
SveStackOperandFrom(Location location)185 inline vixl::aarch64::SVEMemOperand SveStackOperandFrom(Location location) {
186 return vixl::aarch64::SVEMemOperand(vixl::aarch64::sp, location.GetStackIndex());
187 }
188
189 inline vixl::aarch64::MemOperand HeapOperand(const vixl::aarch64::Register& base,
190 size_t offset = 0) {
191 // A heap reference must be 32bit, so fit in a W register.
192 DCHECK(base.IsW());
193 return vixl::aarch64::MemOperand(base.X(), offset);
194 }
195
196 inline vixl::aarch64::MemOperand HeapOperand(const vixl::aarch64::Register& base,
197 const vixl::aarch64::Register& regoffset,
198 vixl::aarch64::Shift shift = vixl::aarch64::LSL,
199 unsigned shift_amount = 0) {
200 // A heap reference must be 32bit, so fit in a W register.
201 DCHECK(base.IsW());
202 return vixl::aarch64::MemOperand(base.X(), regoffset, shift, shift_amount);
203 }
204
HeapOperand(const vixl::aarch64::Register & base,Offset offset)205 inline vixl::aarch64::MemOperand HeapOperand(const vixl::aarch64::Register& base,
206 Offset offset) {
207 return HeapOperand(base, offset.SizeValue());
208 }
209
HeapOperandFrom(Location location,Offset offset)210 inline vixl::aarch64::MemOperand HeapOperandFrom(Location location, Offset offset) {
211 return HeapOperand(RegisterFrom(location, DataType::Type::kReference), offset);
212 }
213
LocationFrom(const vixl::aarch64::Register & reg)214 inline Location LocationFrom(const vixl::aarch64::Register& reg) {
215 return Location::RegisterLocation(ARTRegCodeFromVIXL(reg.GetCode()));
216 }
217
LocationFrom(const vixl::aarch64::VRegister & fpreg)218 inline Location LocationFrom(const vixl::aarch64::VRegister& fpreg) {
219 return Location::FpuRegisterLocation(fpreg.GetCode());
220 }
221
LocationFrom(const vixl::aarch64::ZRegister & zreg)222 inline Location LocationFrom(const vixl::aarch64::ZRegister& zreg) {
223 return Location::FpuRegisterLocation(zreg.GetCode());
224 }
225
OperandFromMemOperand(const vixl::aarch64::MemOperand & mem_op)226 inline vixl::aarch64::Operand OperandFromMemOperand(
227 const vixl::aarch64::MemOperand& mem_op) {
228 if (mem_op.IsImmediateOffset()) {
229 return vixl::aarch64::Operand(mem_op.GetOffset());
230 } else {
231 DCHECK(mem_op.IsRegisterOffset());
232 if (mem_op.GetExtend() != vixl::aarch64::NO_EXTEND) {
233 return vixl::aarch64::Operand(mem_op.GetRegisterOffset(),
234 mem_op.GetExtend(),
235 mem_op.GetShiftAmount());
236 } else if (mem_op.GetShift() != vixl::aarch64::NO_SHIFT) {
237 return vixl::aarch64::Operand(mem_op.GetRegisterOffset(),
238 mem_op.GetShift(),
239 mem_op.GetShiftAmount());
240 } else {
241 LOG(FATAL) << "Should not reach here";
242 UNREACHABLE();
243 }
244 }
245 }
246
AddSubCanEncodeAsImmediate(int64_t value)247 inline bool AddSubCanEncodeAsImmediate(int64_t value) {
248 // If `value` does not fit but `-value` does, VIXL will automatically use
249 // the 'opposite' instruction.
250 return vixl::aarch64::Assembler::IsImmAddSub(value)
251 || vixl::aarch64::Assembler::IsImmAddSub(-value);
252 }
253
Arm64CanEncodeConstantAsImmediate(HConstant * constant,HInstruction * instr)254 inline bool Arm64CanEncodeConstantAsImmediate(HConstant* constant, HInstruction* instr) {
255 int64_t value = CodeGenerator::GetInt64ValueOf(constant);
256
257 // TODO: Improve this when IsSIMDConstantEncodable method is implemented in VIXL.
258 if (instr->IsVecReplicateScalar()) {
259 if (constant->IsLongConstant()) {
260 return false;
261 } else if (constant->IsFloatConstant()) {
262 return vixl::aarch64::Assembler::IsImmFP32(constant->AsFloatConstant()->GetValue());
263 } else if (constant->IsDoubleConstant()) {
264 return vixl::aarch64::Assembler::IsImmFP64(constant->AsDoubleConstant()->GetValue());
265 }
266 return IsUint<8>(value);
267 }
268
269 // Code generation for Min/Max:
270 // Cmp left_op, right_op
271 // Csel dst, left_op, right_op, cond
272 if (instr->IsMin() || instr->IsMax()) {
273 if (constant->GetUses().HasExactlyOneElement()) {
274 // If value can be encoded as immediate for the Cmp, then let VIXL handle
275 // the constant generation for the Csel.
276 return AddSubCanEncodeAsImmediate(value);
277 }
278 // These values are encodable as immediates for Cmp and VIXL will use csinc and csinv
279 // with the zr register as right_op, hence no constant generation is required.
280 return constant->IsZeroBitPattern() || constant->IsOne() || constant->IsMinusOne();
281 }
282
283 // For single uses we let VIXL handle the constant generation since it will
284 // use registers that are not managed by the register allocator (wip0, wip1).
285 if (constant->GetUses().HasExactlyOneElement()) {
286 return true;
287 }
288
289 // Our code generator ensures shift distances are within an encodable range.
290 if (instr->IsRor()) {
291 return true;
292 }
293
294 if (instr->IsAnd() || instr->IsOr() || instr->IsXor()) {
295 // Uses logical operations.
296 return vixl::aarch64::Assembler::IsImmLogical(value, vixl::aarch64::kXRegSize);
297 } else if (instr->IsNeg()) {
298 // Uses mov -immediate.
299 return vixl::aarch64::Assembler::IsImmMovn(value, vixl::aarch64::kXRegSize);
300 } else {
301 DCHECK(instr->IsAdd() ||
302 instr->IsIntermediateAddress() ||
303 instr->IsBoundsCheck() ||
304 instr->IsCompare() ||
305 instr->IsCondition() ||
306 instr->IsSub())
307 << instr->DebugName();
308 // Uses aliases of ADD/SUB instructions.
309 return AddSubCanEncodeAsImmediate(value);
310 }
311 }
312
ARM64EncodableConstantOrRegister(HInstruction * constant,HInstruction * instr)313 inline Location ARM64EncodableConstantOrRegister(HInstruction* constant,
314 HInstruction* instr) {
315 if (constant->IsConstant()
316 && Arm64CanEncodeConstantAsImmediate(constant->AsConstant(), instr)) {
317 return Location::ConstantLocation(constant->AsConstant());
318 }
319
320 return Location::RequiresRegister();
321 }
322
323 // Check if registers in art register set have the same register code in vixl. If the register
324 // codes are same, we can initialize vixl register list simply by the register masks. Currently,
325 // only SP/WSP and ZXR/WZR codes are different between art and vixl.
326 // Note: This function is only used for debug checks.
ArtVixlRegCodeCoherentForRegSet(uint32_t art_core_registers,size_t num_core,uint32_t art_fpu_registers,size_t num_fpu)327 inline bool ArtVixlRegCodeCoherentForRegSet(uint32_t art_core_registers,
328 size_t num_core,
329 uint32_t art_fpu_registers,
330 size_t num_fpu) {
331 // The register masks won't work if the number of register is larger than 32.
332 DCHECK_GE(sizeof(art_core_registers) * 8, num_core);
333 DCHECK_GE(sizeof(art_fpu_registers) * 8, num_fpu);
334 for (size_t art_reg_code = 0; art_reg_code < num_core; ++art_reg_code) {
335 if (RegisterSet::Contains(art_core_registers, art_reg_code)) {
336 if (art_reg_code != static_cast<size_t>(VIXLRegCodeFromART(art_reg_code))) {
337 return false;
338 }
339 }
340 }
341 // There is no register code translation for float registers.
342 return true;
343 }
344
ShiftFromOpKind(HDataProcWithShifterOp::OpKind op_kind)345 inline vixl::aarch64::Shift ShiftFromOpKind(HDataProcWithShifterOp::OpKind op_kind) {
346 switch (op_kind) {
347 case HDataProcWithShifterOp::kASR: return vixl::aarch64::ASR;
348 case HDataProcWithShifterOp::kLSL: return vixl::aarch64::LSL;
349 case HDataProcWithShifterOp::kLSR: return vixl::aarch64::LSR;
350 default:
351 LOG(FATAL) << "Unexpected op kind " << op_kind;
352 UNREACHABLE();
353 return vixl::aarch64::NO_SHIFT;
354 }
355 }
356
ExtendFromOpKind(HDataProcWithShifterOp::OpKind op_kind)357 inline vixl::aarch64::Extend ExtendFromOpKind(HDataProcWithShifterOp::OpKind op_kind) {
358 switch (op_kind) {
359 case HDataProcWithShifterOp::kUXTB: return vixl::aarch64::UXTB;
360 case HDataProcWithShifterOp::kUXTH: return vixl::aarch64::UXTH;
361 case HDataProcWithShifterOp::kUXTW: return vixl::aarch64::UXTW;
362 case HDataProcWithShifterOp::kSXTB: return vixl::aarch64::SXTB;
363 case HDataProcWithShifterOp::kSXTH: return vixl::aarch64::SXTH;
364 case HDataProcWithShifterOp::kSXTW: return vixl::aarch64::SXTW;
365 default:
366 LOG(FATAL) << "Unexpected op kind " << op_kind;
367 UNREACHABLE();
368 return vixl::aarch64::NO_EXTEND;
369 }
370 }
371
ShifterOperandSupportsExtension(HInstruction * instruction)372 inline bool ShifterOperandSupportsExtension(HInstruction* instruction) {
373 DCHECK(HasShifterOperand(instruction, InstructionSet::kArm64));
374 // Although the `neg` instruction is an alias of the `sub` instruction, `HNeg`
375 // does *not* support extension. This is because the `extended register` form
376 // of the `sub` instruction interprets the left register with code 31 as the
377 // stack pointer and not the zero register. (So does the `immediate` form.) In
378 // the other form `shifted register, the register with code 31 is interpreted
379 // as the zero register.
380 return instruction->IsAdd() || instruction->IsSub();
381 }
382
IsConstantZeroBitPattern(const HInstruction * instruction)383 inline bool IsConstantZeroBitPattern(const HInstruction* instruction) {
384 return instruction->IsConstant() && instruction->AsConstant()->IsZeroBitPattern();
385 }
386
387 } // namespace helpers
388 } // namespace arm64
389 } // namespace art
390
391 #endif // ART_COMPILER_OPTIMIZING_COMMON_ARM64_H_
392