Add option to force filetype=asm for testing

src/IceTargetLoweringX86BaseImpl.h

 //===- subzero/src/IceTargetLoweringX86BaseImpl.h - x86 lowering -*- C++ -*-==//
 //
 //                        The Subzero Code Generator
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 ///
 /// \file
@@ skipping 5730 matching lines @@
 Variable *TargetX86Base<Machine>::makeVectorOfZeros(Type Ty, int32_t RegNum) {
   return makeZeroedRegister(Ty, RegNum);
 }
 
 template <class Machine>
 Variable *TargetX86Base<Machine>::makeVectorOfMinusOnes(Type Ty,
                                                         int32_t RegNum) {
   Variable *MinusOnes = makeReg(Ty, RegNum);
   // Insert a FakeDef so the live range of MinusOnes is not overestimated.
   Context.insert<InstFakeDef>(MinusOnes);
-  _pcmpeq(MinusOnes, MinusOnes);
+  if (Ty == IceType_f64)
+    // pcmpeqq requires SSE 4.1, while pcmpeqd does the same job and only

[Jim Stichnoth, 2015/12/20 18:42:23]

Could you also note that f64 is (currently) used only for fabs intrinsic
lowering?  Otherwise I was having a hard time figuring out how this code path
was taken.

[sehr, 2016/01/07 18:53:12]

Done.

+    // requires SSE2.
+    _pcmpeq(MinusOnes, MinusOnes, IceType_f32);
+  else
+    _pcmpeq(MinusOnes, MinusOnes);
   return MinusOnes;
 }
 
 template <class Machine>
 Variable *TargetX86Base<Machine>::makeVectorOfOnes(Type Ty, int32_t RegNum) {
   Variable *Dest = makeVectorOfZeros(Ty, RegNum);
   Variable *MinusOne = makeVectorOfMinusOnes(Ty);
   _psub(Dest, MinusOne);
   return Dest;
 }
@@ skipping 641 matching lines @@
   }
   // the offset is not eligible for blinding or pooling, return the original
   // mem operand
   return MemOperand;
 }
 
 } // end of namespace X86Internal
 } // end of namespace Ice
 
 #endif // SUBZERO_SRC_ICETARGETLOWERINGX86BASEIMPL_H