Ben Murdoch | da12d29 | 2016-06-02 14:46:10 +0100 | [diff] [blame] | 1 | // Copyright 2014 the V8 project authors. All rights reserved. |
| 2 | // Use of this source code is governed by a BSD-style license that can be |
| 3 | // found in the LICENSE file. |
| 4 | |
| 5 | #include "src/deoptimizer.h" |
| 6 | #include "src/codegen.h" |
| 7 | #include "src/full-codegen/full-codegen.h" |
| 8 | #include "src/register-configuration.h" |
| 9 | #include "src/safepoint-table.h" |
| 10 | |
| 11 | namespace v8 { |
| 12 | namespace internal { |
| 13 | |
| 14 | // LAY + LGHI/LHI + BRCL |
| 15 | const int Deoptimizer::table_entry_size_ = 16; |
| 16 | |
| 17 | int Deoptimizer::patch_size() { |
| 18 | #if V8_TARGET_ARCH_S390X |
| 19 | const int kCallInstructionSize = 16; |
| 20 | #else |
| 21 | const int kCallInstructionSize = 10; |
| 22 | #endif |
| 23 | return kCallInstructionSize; |
| 24 | } |
| 25 | |
| 26 | void Deoptimizer::EnsureRelocSpaceForLazyDeoptimization(Handle<Code> code) { |
| 27 | // Empty because there is no need for relocation information for the code |
| 28 | // patching in Deoptimizer::PatchCodeForDeoptimization below. |
| 29 | } |
| 30 | |
| 31 | void Deoptimizer::PatchCodeForDeoptimization(Isolate* isolate, Code* code) { |
| 32 | Address code_start_address = code->instruction_start(); |
| 33 | |
| 34 | // Invalidate the relocation information, as it will become invalid by the |
| 35 | // code patching below, and is not needed any more. |
| 36 | code->InvalidateRelocation(); |
| 37 | |
| 38 | if (FLAG_zap_code_space) { |
| 39 | // Fail hard and early if we enter this code object again. |
| 40 | byte* pointer = code->FindCodeAgeSequence(); |
| 41 | if (pointer != NULL) { |
| 42 | pointer += kNoCodeAgeSequenceLength; |
| 43 | } else { |
| 44 | pointer = code->instruction_start(); |
| 45 | } |
| 46 | CodePatcher patcher(isolate, pointer, 2); |
| 47 | patcher.masm()->bkpt(0); |
| 48 | |
| 49 | DeoptimizationInputData* data = |
| 50 | DeoptimizationInputData::cast(code->deoptimization_data()); |
| 51 | int osr_offset = data->OsrPcOffset()->value(); |
| 52 | if (osr_offset > 0) { |
| 53 | CodePatcher osr_patcher(isolate, code->instruction_start() + osr_offset, |
| 54 | 2); |
| 55 | osr_patcher.masm()->bkpt(0); |
| 56 | } |
| 57 | } |
| 58 | |
| 59 | DeoptimizationInputData* deopt_data = |
| 60 | DeoptimizationInputData::cast(code->deoptimization_data()); |
| 61 | #ifdef DEBUG |
| 62 | Address prev_call_address = NULL; |
| 63 | #endif |
| 64 | // For each LLazyBailout instruction insert a call to the corresponding |
| 65 | // deoptimization entry. |
| 66 | for (int i = 0; i < deopt_data->DeoptCount(); i++) { |
| 67 | if (deopt_data->Pc(i)->value() == -1) continue; |
| 68 | Address call_address = code_start_address + deopt_data->Pc(i)->value(); |
| 69 | Address deopt_entry = GetDeoptimizationEntry(isolate, i, LAZY); |
| 70 | // We need calls to have a predictable size in the unoptimized code, but |
| 71 | // this is optimized code, so we don't have to have a predictable size. |
| 72 | int call_size_in_bytes = MacroAssembler::CallSizeNotPredictableCodeSize( |
| 73 | deopt_entry, kRelocInfo_NONEPTR); |
| 74 | DCHECK(call_size_in_bytes <= patch_size()); |
| 75 | CodePatcher patcher(isolate, call_address, call_size_in_bytes); |
| 76 | patcher.masm()->Call(deopt_entry, kRelocInfo_NONEPTR); |
| 77 | DCHECK(prev_call_address == NULL || |
| 78 | call_address >= prev_call_address + patch_size()); |
| 79 | DCHECK(call_address + patch_size() <= code->instruction_end()); |
| 80 | #ifdef DEBUG |
| 81 | prev_call_address = call_address; |
| 82 | #endif |
| 83 | } |
| 84 | } |
| 85 | |
| 86 | void Deoptimizer::SetPlatformCompiledStubRegisters( |
| 87 | FrameDescription* output_frame, CodeStubDescriptor* descriptor) { |
| 88 | ApiFunction function(descriptor->deoptimization_handler()); |
| 89 | ExternalReference xref(&function, ExternalReference::BUILTIN_CALL, isolate_); |
| 90 | intptr_t handler = reinterpret_cast<intptr_t>(xref.address()); |
| 91 | int params = descriptor->GetHandlerParameterCount(); |
| 92 | output_frame->SetRegister(r2.code(), params); |
| 93 | output_frame->SetRegister(r3.code(), handler); |
| 94 | } |
| 95 | |
| 96 | void Deoptimizer::CopyDoubleRegisters(FrameDescription* output_frame) { |
| 97 | for (int i = 0; i < DoubleRegister::kNumRegisters; ++i) { |
| 98 | double double_value = input_->GetDoubleRegister(i); |
| 99 | output_frame->SetDoubleRegister(i, double_value); |
| 100 | } |
| 101 | } |
| 102 | |
| 103 | #define __ masm()-> |
| 104 | |
| 105 | // This code tries to be close to ia32 code so that any changes can be |
| 106 | // easily ported. |
| 107 | void Deoptimizer::TableEntryGenerator::Generate() { |
| 108 | GeneratePrologue(); |
| 109 | |
| 110 | // Save all the registers onto the stack |
| 111 | const int kNumberOfRegisters = Register::kNumRegisters; |
| 112 | |
| 113 | RegList restored_regs = kJSCallerSaved | kCalleeSaved; |
| 114 | |
| 115 | const int kDoubleRegsSize = kDoubleSize * DoubleRegister::kNumRegisters; |
| 116 | |
| 117 | // Save all double registers before messing with them. |
| 118 | __ lay(sp, MemOperand(sp, -kDoubleRegsSize)); |
Ben Murdoch | 61f157c | 2016-09-16 13:49:30 +0100 | [diff] [blame] | 119 | const RegisterConfiguration* config = RegisterConfiguration::Crankshaft(); |
Ben Murdoch | da12d29 | 2016-06-02 14:46:10 +0100 | [diff] [blame] | 120 | for (int i = 0; i < config->num_allocatable_double_registers(); ++i) { |
| 121 | int code = config->GetAllocatableDoubleCode(i); |
| 122 | const DoubleRegister dreg = DoubleRegister::from_code(code); |
| 123 | int offset = code * kDoubleSize; |
| 124 | __ StoreDouble(dreg, MemOperand(sp, offset)); |
| 125 | } |
| 126 | |
| 127 | // Push all GPRs onto the stack |
| 128 | __ lay(sp, MemOperand(sp, -kNumberOfRegisters * kPointerSize)); |
| 129 | __ StoreMultipleP(r0, sp, MemOperand(sp)); // Save all 16 registers |
| 130 | |
| 131 | __ mov(ip, Operand(ExternalReference(Isolate::kCEntryFPAddress, isolate()))); |
| 132 | __ StoreP(fp, MemOperand(ip)); |
| 133 | |
| 134 | const int kSavedRegistersAreaSize = |
| 135 | (kNumberOfRegisters * kPointerSize) + kDoubleRegsSize; |
| 136 | |
| 137 | // Get the bailout id from the stack. |
| 138 | __ LoadP(r4, MemOperand(sp, kSavedRegistersAreaSize)); |
| 139 | |
| 140 | // Cleanse the Return address for 31-bit |
| 141 | __ CleanseP(r14); |
| 142 | |
| 143 | // Get the address of the location in the code object (r5)(return |
| 144 | // address for lazy deoptimization) and compute the fp-to-sp delta in |
| 145 | // register r6. |
| 146 | __ LoadRR(r5, r14); |
| 147 | __ la(r6, MemOperand(sp, kSavedRegistersAreaSize + (1 * kPointerSize))); |
| 148 | __ SubP(r6, fp, r6); |
| 149 | |
| 150 | // Allocate a new deoptimizer object. |
| 151 | // Pass six arguments in r2 to r7. |
| 152 | __ PrepareCallCFunction(6, r7); |
| 153 | __ LoadImmP(r2, Operand::Zero()); |
| 154 | Label context_check; |
| 155 | __ LoadP(r3, MemOperand(fp, CommonFrameConstants::kContextOrFrameTypeOffset)); |
| 156 | __ JumpIfSmi(r3, &context_check); |
| 157 | __ LoadP(r2, MemOperand(fp, JavaScriptFrameConstants::kFunctionOffset)); |
| 158 | __ bind(&context_check); |
| 159 | __ LoadImmP(r3, Operand(type())); // bailout type, |
| 160 | // r4: bailout id already loaded. |
| 161 | // r5: code address or 0 already loaded. |
| 162 | // r6: Fp-to-sp delta. |
| 163 | // Parm6: isolate is passed on the stack. |
| 164 | __ mov(r7, Operand(ExternalReference::isolate_address(isolate()))); |
| 165 | __ StoreP(r7, MemOperand(sp, kStackFrameExtraParamSlot * kPointerSize)); |
| 166 | |
| 167 | // Call Deoptimizer::New(). |
| 168 | { |
| 169 | AllowExternalCallThatCantCauseGC scope(masm()); |
| 170 | __ CallCFunction(ExternalReference::new_deoptimizer_function(isolate()), 6); |
| 171 | } |
| 172 | |
| 173 | // Preserve "deoptimizer" object in register r2 and get the input |
| 174 | // frame descriptor pointer to r3 (deoptimizer->input_); |
| 175 | __ LoadP(r3, MemOperand(r2, Deoptimizer::input_offset())); |
| 176 | |
| 177 | // Copy core registers into FrameDescription::registers_[kNumRegisters]. |
| 178 | // DCHECK(Register::kNumRegisters == kNumberOfRegisters); |
| 179 | // __ mvc(MemOperand(r3, FrameDescription::registers_offset()), |
| 180 | // MemOperand(sp), kNumberOfRegisters * kPointerSize); |
| 181 | // Copy core registers into FrameDescription::registers_[kNumRegisters]. |
| 182 | // TODO(john.yan): optimize the following code by using mvc instruction |
| 183 | DCHECK(Register::kNumRegisters == kNumberOfRegisters); |
| 184 | for (int i = 0; i < kNumberOfRegisters; i++) { |
| 185 | int offset = (i * kPointerSize) + FrameDescription::registers_offset(); |
| 186 | __ LoadP(r4, MemOperand(sp, i * kPointerSize)); |
| 187 | __ StoreP(r4, MemOperand(r3, offset)); |
| 188 | } |
| 189 | |
| 190 | int double_regs_offset = FrameDescription::double_registers_offset(); |
| 191 | // Copy double registers to |
| 192 | // double_registers_[DoubleRegister::kNumRegisters] |
| 193 | for (int i = 0; i < config->num_allocatable_double_registers(); ++i) { |
| 194 | int code = config->GetAllocatableDoubleCode(i); |
| 195 | int dst_offset = code * kDoubleSize + double_regs_offset; |
| 196 | int src_offset = code * kDoubleSize + kNumberOfRegisters * kPointerSize; |
| 197 | // TODO(joransiu): MVC opportunity |
| 198 | __ LoadDouble(d0, MemOperand(sp, src_offset)); |
| 199 | __ StoreDouble(d0, MemOperand(r3, dst_offset)); |
| 200 | } |
| 201 | |
| 202 | // Remove the bailout id and the saved registers from the stack. |
| 203 | __ la(sp, MemOperand(sp, kSavedRegistersAreaSize + (1 * kPointerSize))); |
| 204 | |
| 205 | // Compute a pointer to the unwinding limit in register r4; that is |
| 206 | // the first stack slot not part of the input frame. |
| 207 | __ LoadP(r4, MemOperand(r3, FrameDescription::frame_size_offset())); |
| 208 | __ AddP(r4, sp); |
| 209 | |
| 210 | // Unwind the stack down to - but not including - the unwinding |
| 211 | // limit and copy the contents of the activation frame to the input |
| 212 | // frame description. |
| 213 | __ la(r5, MemOperand(r3, FrameDescription::frame_content_offset())); |
| 214 | Label pop_loop; |
| 215 | Label pop_loop_header; |
| 216 | __ b(&pop_loop_header, Label::kNear); |
| 217 | __ bind(&pop_loop); |
| 218 | __ pop(r6); |
| 219 | __ StoreP(r6, MemOperand(r5, 0)); |
| 220 | __ la(r5, MemOperand(r5, kPointerSize)); |
| 221 | __ bind(&pop_loop_header); |
| 222 | __ CmpP(r4, sp); |
| 223 | __ bne(&pop_loop); |
| 224 | |
| 225 | // Compute the output frame in the deoptimizer. |
| 226 | __ push(r2); // Preserve deoptimizer object across call. |
| 227 | // r2: deoptimizer object; r3: scratch. |
| 228 | __ PrepareCallCFunction(1, r3); |
| 229 | // Call Deoptimizer::ComputeOutputFrames(). |
| 230 | { |
| 231 | AllowExternalCallThatCantCauseGC scope(masm()); |
| 232 | __ CallCFunction( |
| 233 | ExternalReference::compute_output_frames_function(isolate()), 1); |
| 234 | } |
| 235 | __ pop(r2); // Restore deoptimizer object (class Deoptimizer). |
| 236 | |
| 237 | __ LoadP(sp, MemOperand(r2, Deoptimizer::caller_frame_top_offset())); |
| 238 | |
| 239 | // Replace the current (input) frame with the output frames. |
| 240 | Label outer_push_loop, inner_push_loop, outer_loop_header, inner_loop_header; |
| 241 | // Outer loop state: r6 = current "FrameDescription** output_", |
| 242 | // r3 = one past the last FrameDescription**. |
| 243 | __ LoadlW(r3, MemOperand(r2, Deoptimizer::output_count_offset())); |
| 244 | __ LoadP(r6, MemOperand(r2, Deoptimizer::output_offset())); // r6 is output_. |
| 245 | __ ShiftLeftP(r3, r3, Operand(kPointerSizeLog2)); |
| 246 | __ AddP(r3, r6, r3); |
| 247 | __ b(&outer_loop_header, Label::kNear); |
| 248 | |
| 249 | __ bind(&outer_push_loop); |
| 250 | // Inner loop state: r4 = current FrameDescription*, r5 = loop index. |
| 251 | __ LoadP(r4, MemOperand(r6, 0)); // output_[ix] |
| 252 | __ LoadP(r5, MemOperand(r4, FrameDescription::frame_size_offset())); |
| 253 | __ b(&inner_loop_header, Label::kNear); |
| 254 | |
| 255 | __ bind(&inner_push_loop); |
| 256 | __ AddP(r5, Operand(-sizeof(intptr_t))); |
| 257 | __ AddP(r8, r4, r5); |
| 258 | __ LoadP(r8, MemOperand(r8, FrameDescription::frame_content_offset())); |
| 259 | __ push(r8); |
| 260 | |
| 261 | __ bind(&inner_loop_header); |
| 262 | __ CmpP(r5, Operand::Zero()); |
| 263 | __ bne(&inner_push_loop); // test for gt? |
| 264 | |
| 265 | __ AddP(r6, r6, Operand(kPointerSize)); |
| 266 | __ bind(&outer_loop_header); |
| 267 | __ CmpP(r6, r3); |
| 268 | __ blt(&outer_push_loop); |
| 269 | |
| 270 | __ LoadP(r3, MemOperand(r2, Deoptimizer::input_offset())); |
| 271 | for (int i = 0; i < config->num_allocatable_double_registers(); ++i) { |
| 272 | int code = config->GetAllocatableDoubleCode(i); |
| 273 | const DoubleRegister dreg = DoubleRegister::from_code(code); |
| 274 | int src_offset = code * kDoubleSize + double_regs_offset; |
| 275 | __ ld(dreg, MemOperand(r3, src_offset)); |
| 276 | } |
| 277 | |
| 278 | // Push state, pc, and continuation from the last output frame. |
| 279 | __ LoadP(r8, MemOperand(r4, FrameDescription::state_offset())); |
| 280 | __ push(r8); |
| 281 | __ LoadP(r8, MemOperand(r4, FrameDescription::pc_offset())); |
| 282 | __ push(r8); |
| 283 | __ LoadP(r8, MemOperand(r4, FrameDescription::continuation_offset())); |
| 284 | __ push(r8); |
| 285 | |
| 286 | // Restore the registers from the last output frame. |
| 287 | __ LoadRR(r1, r4); |
| 288 | for (int i = kNumberOfRegisters - 1; i > 0; i--) { |
| 289 | int offset = (i * kPointerSize) + FrameDescription::registers_offset(); |
| 290 | if ((restored_regs & (1 << i)) != 0) { |
| 291 | __ LoadP(ToRegister(i), MemOperand(r1, offset)); |
| 292 | } |
| 293 | } |
| 294 | |
| 295 | __ InitializeRootRegister(); |
| 296 | |
| 297 | __ pop(ip); // get continuation, leave pc on stack |
| 298 | __ pop(r14); |
| 299 | __ Jump(ip); |
| 300 | __ stop("Unreachable."); |
| 301 | } |
| 302 | |
| 303 | void Deoptimizer::TableEntryGenerator::GeneratePrologue() { |
| 304 | // Create a sequence of deoptimization entries. Note that any |
| 305 | // registers may be still live. |
| 306 | Label done; |
| 307 | for (int i = 0; i < count(); i++) { |
| 308 | int start = masm()->pc_offset(); |
| 309 | USE(start); |
| 310 | __ lay(sp, MemOperand(sp, -kPointerSize)); |
| 311 | __ LoadImmP(ip, Operand(i)); |
| 312 | __ b(&done); |
| 313 | int end = masm()->pc_offset(); |
| 314 | USE(end); |
| 315 | DCHECK(masm()->pc_offset() - start == table_entry_size_); |
| 316 | } |
| 317 | __ bind(&done); |
| 318 | __ StoreP(ip, MemOperand(sp)); |
| 319 | } |
| 320 | |
| 321 | void FrameDescription::SetCallerPc(unsigned offset, intptr_t value) { |
| 322 | SetFrameSlot(offset, value); |
| 323 | } |
| 324 | |
| 325 | void FrameDescription::SetCallerFp(unsigned offset, intptr_t value) { |
| 326 | SetFrameSlot(offset, value); |
| 327 | } |
| 328 | |
| 329 | void FrameDescription::SetCallerConstantPool(unsigned offset, intptr_t value) { |
| 330 | // No out-of-line constant pool support. |
| 331 | UNREACHABLE(); |
| 332 | } |
| 333 | |
| 334 | #undef __ |
| 335 | |
| 336 | } // namespace internal |
| 337 | } // namespace v8 |