[Atomics] Make Atomics.compareExchange a builtin using TF

src/compiler/arm64/code-generator-arm64.cc

 // Copyright 2014 the V8 project authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "src/compiler/code-generator.h"
 
 #include "src/arm64/frames-arm64.h"
 #include "src/arm64/macro-assembler-arm64.h"
 #include "src/compilation-info.h"
 #include "src/compiler/code-generator-impl.h"
@@ skipping 81 matching lines @@
   Operand InputOperand64(size_t index) { return InputOperand(index); }
 
   Operand InputOperand32(size_t index) {
     return ToOperand32(instr_->InputAt(index));
   }
 
   Register OutputRegister64() { return OutputRegister(); }
 
   Register OutputRegister32() { return ToRegister(instr_->Output()).W(); }
 
-  Register TempRegister32() { return ToRegister(instr_->TempAt(0)).W(); }
+  Register TempRegister32(size_t index) {
+    return ToRegister(instr_->TempAt(index)).W();
+  }
 
   Operand InputOperand2_32(size_t index) {
     switch (AddressingModeField::decode(instr_->opcode())) {
       case kMode_None:
         return InputOperand32(index);
       case kMode_Operand2_R_LSL_I:
         return Operand(InputRegister32(index), LSL, InputInt5(index + 1));
       case kMode_Operand2_R_LSR_I:
         return Operand(InputRegister32(index), LSR, InputInt5(index + 1));
       case kMode_Operand2_R_ASR_I:
@@ skipping 415 matching lines @@
                  MemOperand(i.InputRegister(0), i.InputRegister(1))); \
     __ Dmb(InnerShareable, BarrierAll);                               \
   } while (0)
 
 #define ASSEMBLE_ATOMIC_EXCHANGE_INTEGER(load_instr, store_instr)      \
   do {                                                                 \
     Label exchange;                                                    \
     __ bind(&exchange);                                                \
     __ Add(i.TempRegister(0), i.InputRegister(0), i.InputRegister(1)); \
     __ load_instr(i.OutputRegister32(), i.TempRegister(0));            \
-    __ store_instr(i.TempRegister32(), i.InputRegister32(2),           \
+    __ store_instr(i.TempRegister32(0), i.InputRegister32(2),          \
                    i.TempRegister(0));                                 \
-    __ cbnz(i.TempRegister32(), &exchange);                            \
+    __ cbnz(i.TempRegister32(0), &exchange);                           \
+  } while (0)
+
+#define ASSEMBLE_ATOMIC_COMPARE_EXCHANGE_INTEGER(load_instr, store_instr) \
+  do {                                                                    \
+    Label compareExchange;                                                \
+    Label exit;                                                           \
+    __ bind(&compareExchange);                                            \
+    __ Add(i.TempRegister(0), i.InputRegister(0), i.InputRegister(1));    \
+    __ load_instr(i.OutputRegister32(), i.TempRegister(0));               \
+    __ cmp(i.TempRegister32(1), i.OutputRegister32());                    \
+    __ B(ne, &exit);                                                      \
+    __ store_instr(i.TempRegister32(0), i.InputRegister32(3),             \
+                   i.TempRegister(0));                                    \
+    __ cbnz(i.TempRegister32(0), &compareExchange);                       \
+    __ bind(&exit);                                                       \
   } while (0)
 
 #define ASSEMBLE_IEEE754_BINOP(name)                                          \
   do {                                                                        \
     FrameScope scope(masm(), StackFrame::MANUAL);                             \
     __ CallCFunction(ExternalReference::ieee754_##name##_function(isolate()), \
                      0, 2);                                                   \
   } while (0)
 
 #define ASSEMBLE_IEEE754_UNOP(name)                                           \
@@ skipping 1103 matching lines @@
     case kAtomicExchangeInt16:
       ASSEMBLE_ATOMIC_EXCHANGE_INTEGER(ldaxrh, stlxrh);
       __ Sxth(i.OutputRegister(0), i.OutputRegister(0));
       break;
     case kAtomicExchangeUint16:
       ASSEMBLE_ATOMIC_EXCHANGE_INTEGER(ldaxrh, stlxrh);
       break;
     case kAtomicExchangeWord32:
       ASSEMBLE_ATOMIC_EXCHANGE_INTEGER(ldaxr, stlxr);
       break;
+    case kAtomicCompareExchangeInt8:
+      __ Uxtb(i.TempRegister(1), i.InputRegister(2));
+      ASSEMBLE_ATOMIC_COMPARE_EXCHANGE_INTEGER(ldaxrb, stlxrb);
+      __ Sxtb(i.OutputRegister(0), i.OutputRegister(0));
+      break;
+    case kAtomicCompareExchangeUint8:
+      __ Uxtb(i.TempRegister(1), i.InputRegister(2));
+      ASSEMBLE_ATOMIC_COMPARE_EXCHANGE_INTEGER(ldaxrb, stlxrb);

[binji, 2017/03/10 23:08:02]

Is it necessary to zero-extend the output register here, or does the ldaxrb do
that already?

[aseemgarg, 2017/03/10 23:14:32]

ldaxrb/ldaxrh zero extends always. So, we need to take care for signed cases but
not unsigned cases.

[aseemgarg, 2017/03/11 00:45:43]

It is verified that nodes that require 64bit input do not directly use 32bit
output from another node. So, ldaxrb will correctly zero extend to 32 bit and if
that value is needed by a 64 bit operator, it will have to go through a casting
operation.

+      break;
+    case kAtomicCompareExchangeInt16:
+      __ Uxth(i.TempRegister(1), i.InputRegister(2));
+      ASSEMBLE_ATOMIC_COMPARE_EXCHANGE_INTEGER(ldaxrh, stlxrh);
+      __ Sxth(i.OutputRegister(0), i.OutputRegister(0));
+      break;
+    case kAtomicCompareExchangeUint16:
+      __ Uxth(i.TempRegister(1), i.InputRegister(2));
+      ASSEMBLE_ATOMIC_COMPARE_EXCHANGE_INTEGER(ldaxrh, stlxrh);
+      break;
+    case kAtomicCompareExchangeWord32:
+      __ mov(i.TempRegister(1), i.InputRegister(2));
+      ASSEMBLE_ATOMIC_COMPARE_EXCHANGE_INTEGER(ldaxr, stlxr);
+      break;
   }
   return kSuccess;
 }  // NOLINT(readability/fn_size)
 
 
 // Assemble branches after this instruction.
 void CodeGenerator::AssembleArchBranch(Instruction* instr, BranchInfo* branch) {
   Arm64OperandConverter i(this, instr);
   Label* tlabel = branch->true_label;
   Label* flabel = branch->false_label;
@@ skipping 533 matching lines @@
       padding_size -= kInstructionSize;
     }
   }
 }
 
 #undef __
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8