[Atomics] Make Atomics.compareExchange a builtin using TF

src/compiler/arm/code-generator-arm.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/arm/macro-assembler-arm.h"
 #include "src/compilation-info.h"
 #include "src/compiler/code-generator-impl.h"
 #include "src/compiler/gap-resolver.h"
@@ skipping 415 matching lines @@
     __ dmb(ISH);                                                              \
     __ bind(&exchange);                                                       \
     __ add(i.TempRegister(0), i.InputRegister(0), i.InputRegister(1));        \
     __ load_instr(i.OutputRegister(0), i.TempRegister(0));                    \
     __ store_instr(i.TempRegister(0), i.InputRegister(2), i.TempRegister(0)); \
     __ teq(i.TempRegister(0), Operand(0));                                    \
     __ b(ne, &exchange);                                                      \
     __ dmb(ISH);                                                              \
   } while (0)
 
+#define ASSEMBLE_ATOMIC_COMPARE_EXCHANGE_INTEGER(load_instr, store_instr,     \
+                                                 mask)                        \
+  do {                                                                        \
+    Label compareExchange;                                                    \
+    Label exit;                                                               \
+    __ dmb(ISH);                                                              \

[binji, 2017/02/28 23:28:53]

The optimized code is a bit hard to follow, but the idea is that you can do an
initial zero-extend on the new_value, then ldrexb (which will also zero-extend).
Then you can just do a normal cmp/bne without any of the masking.


  uxtb    r6, r2
  ldrexb  r0, [r1]
  cmp     r0, r6
  bne.n   fail
  dmb     ish
loop:
  strexb  r5, r3, [r1]
  cmp     r5, #0
  beq.w   ok
  ldrexb  r0, [r1]
  cmp     r0, r6
  beq.n   loop
ok:
  clrex
  movs    r1, #0
  dmb     ish
  b.n     done
  ...
fail:
  dmb ish
  movs    r1, #1
done:
  cmp     r1, #0
  it      ne
  movne   r0, r2
  sxtb    r0, r0

It's also worth noting that they only loop while the comparison is still valid.
This is a nice optimization, but I think we can do something simpler and just do
one iteration, rather than loop at all. Basically, something like this:

old_extend = uxt(old);
read = ldrex(addr)
if (read == old_extend) {
  strex(addr, new);
} else {
  clrex;
}
return sign_or_zero_extend(read);

+    __ bind(&compareExchange);                                                \
+    __ add(i.TempRegister(0), i.InputRegister(0), i.InputRegister(1));        \
+    __ load_instr(i.OutputRegister(0), i.TempRegister(0));                    \
+    __ eor(i.TempRegister(1), i.OutputRegister(0),                            \
+           Operand(i.InputRegister(2)));                                      \
+    __ And(i.TempRegister(1), i.TempRegister(1), Operand(mask));              \
+    __ teq(i.TempRegister(1), Operand(0));                                    \
+    __ b(ne, &exit);                                                          \
+    __ store_instr(i.TempRegister(0), i.InputRegister(3), i.TempRegister(0)); \
+    __ teq(i.TempRegister(0), Operand(0));                                    \
+    __ b(ne, &compareExchange);                                               \
+    __ bind(&exit);                                                           \
+    __ dmb(ISH);                                                              \
+  } while (0)
+
 #define ASSEMBLE_IEEE754_BINOP(name)                                           \
   do {                                                                         \
     /* TODO(bmeurer): We should really get rid of this special instruction, */ \
     /* and generate a CallAddress instruction instead. */                      \
     FrameScope scope(masm(), StackFrame::MANUAL);                              \
     __ PrepareCallCFunction(0, 2, kScratchReg);                                \
     __ MovToFloatParameters(i.InputDoubleRegister(0),                          \
                             i.InputDoubleRegister(1));                         \
     __ CallCFunction(ExternalReference::ieee754_##name##_function(isolate()),  \
                      0, 2);                                                    \
@@ skipping 1586 matching lines @@
       ASSEMBLE_ATOMIC_EXCHANGE_INTEGER(ldrexh, strexh);
       __ sxth(i.OutputRegister(0), i.OutputRegister(0));
       break;
     case kAtomicExchangeUint16:
       ASSEMBLE_ATOMIC_EXCHANGE_INTEGER(ldrexh, strexh);
       __ uxth(i.OutputRegister(0), i.OutputRegister(0));
       break;
     case kAtomicExchangeWord32:
       ASSEMBLE_ATOMIC_EXCHANGE_INTEGER(ldrex, strex);
       break;
+    case kAtomicCompareExchangeInt8:
+      ASSEMBLE_ATOMIC_COMPARE_EXCHANGE_INTEGER(ldrexb, strexb, 0x000000ff);
+      __ sxtb(i.OutputRegister(0), i.OutputRegister(0));
+      break;
+    case kAtomicCompareExchangeUint8:
+      ASSEMBLE_ATOMIC_COMPARE_EXCHANGE_INTEGER(ldrexb, strexb, 0x000000ff);
+      __ uxtb(i.OutputRegister(0), i.OutputRegister(0));
+      break;
+    case kAtomicCompareExchangeInt16:
+      ASSEMBLE_ATOMIC_COMPARE_EXCHANGE_INTEGER(ldrexh, strexh, 0x0000ffff);
+      __ sxth(i.OutputRegister(0), i.OutputRegister(0));
+      break;
+    case kAtomicCompareExchangeUint16:
+      ASSEMBLE_ATOMIC_COMPARE_EXCHANGE_INTEGER(ldrexh, strexh, 0x0000ffff);
+      __ uxth(i.OutputRegister(0), i.OutputRegister(0));
+      break;
+    case kAtomicCompareExchangeWord32:
+      ASSEMBLE_ATOMIC_COMPARE_EXCHANGE_INTEGER(ldrex, strex, 0xffffffff);
+      break;
   }
   return kSuccess;
 }  // NOLINT(readability/fn_size)
 
 
 // Assembles branches after an instruction.
 void CodeGenerator::AssembleArchBranch(Instruction* instr, BranchInfo* branch) {
   ArmOperandConverter i(this, instr);
   Label* tlabel = branch->true_label;
   Label* flabel = branch->false_label;
@@ skipping 584 matching lines @@
       padding_size -= v8::internal::Assembler::kInstrSize;
     }
   }
 }
 
 #undef __
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8