PPC: [turbofan] Introduce integer multiplication with overflow.

src/compiler/ppc/code-generator-ppc.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/ast/scopes.h"
 #include "src/compiler/code-generator-impl.h"
 #include "src/compiler/gap-resolver.h"
 #include "src/compiler/node-matchers.h"
@@ skipping 957 matching lines @@
     case kArchPrepareCallCFunction: {
       int const num_parameters = MiscField::decode(instr->opcode());
       __ PrepareCallCFunction(num_parameters, kScratchReg);
       // Frame alignment requires using FP-relative frame addressing.
       frame_access_state()->SetFrameAccessToFP();
       break;
     }
     case kArchPrepareTailCall:
       AssemblePrepareTailCall();
       break;
+    case kArchComment: {
+      Address comment_string = i.InputExternalReference(0).address();
+      __ RecordComment(reinterpret_cast<const char*>(comment_string));
+      break;
+    }
     case kArchCallCFunction: {
       int const num_parameters = MiscField::decode(instr->opcode());
       if (instr->InputAt(0)->IsImmediate()) {
         ExternalReference ref = i.InputExternalReference(0);
         __ CallCFunction(ref, num_parameters);
       } else {
         Register func = i.InputRegister(0);
         __ CallCFunction(func, num_parameters);
       }
       frame_access_state()->SetFrameAccessToDefault();
@@ skipping 305 matching lines @@
     case kPPC_Mul32:
       __ mullw(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1),
                LeaveOE, i.OutputRCBit());
       break;
 #if V8_TARGET_ARCH_PPC64
     case kPPC_Mul64:
       __ mulld(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1),
                LeaveOE, i.OutputRCBit());
       break;
 #endif
+    case kPPC_Mul32WithHigh32:

[john.yan, 2016/07/16 03:57:48]

The reason why this fails is that output reg can be one of the input reg.
Therefore, mullw could be modifying one of the input reg, which causes mulhw to
produce incorrect output.

+      __ mullw(i.OutputRegister(0),
+               i.InputRegister(0), i.InputRegister(1));  // low
+      __ mulhw(i.OutputRegister(1),
+               i.InputRegister(0), i.InputRegister(1));  // high
+      break;

[john.yan, 2016/07/16 03:54:45]

Try this sequence here:
 +    case kPPC_Mul32WithHigh32:
 +      if (i.OutputRegister(0).is(i.InputRegister(0)) ||
 +          i.OutputRegister(0).is(i.InputRegister(1)) ||
 +          i.OutputRegister(1).is(i.InputRegister(0)) ||
 +          i.OutputRegister(1).is(i.InputRegister(1))) {
 +        __ mullw(kScratchReg,
 +                 i.InputRegister(0), i.InputRegister(1));  // low
 +        __ mulhw(i.OutputRegister(1),
 +                 i.InputRegister(0), i.InputRegister(1));  // high
 +        __ mr(i.OutputRegister(0), kScratchReg);
 +      } else {
 +        __ mullw(i.OutputRegister(0),
 +                 i.InputRegister(0), i.InputRegister(1));  // low
 +        __ mulhw(i.OutputRegister(1),
 +                 i.InputRegister(0), i.InputRegister(1));  // high
 +      }
 +      break;

     case kPPC_MulHigh32:
       __ mulhw(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1),
                i.OutputRCBit());
       break;
     case kPPC_MulHighU32:
       __ mulhwu(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1),
                 i.OutputRCBit());
       break;
     case kPPC_MulDouble:
       ASSEMBLE_FLOAT_BINOP_RC(fmul, MiscField::decode(instr->opcode()));
@@ skipping 1031 matching lines @@
       padding_size -= v8::internal::Assembler::kInstrSize;
     }
   }
 }
 
 #undef __
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8