Add option to force filetype=asm for testing

src/IceTargetLoweringX86BaseImpl.h

Index: src/IceTargetLoweringX86BaseImpl.h
diff --git a/src/IceTargetLoweringX86BaseImpl.h b/src/IceTargetLoweringX86BaseImpl.h
index 5127768437978d0623cfa2e52d4ea0e0767bfee5..08dc25c446463dfaa11c08d5e9d5709f0dd3907a 100644
--- a/src/IceTargetLoweringX86BaseImpl.h
+++ b/src/IceTargetLoweringX86BaseImpl.h
@@ -1135,8 +1135,8 @@ bool TargetX86Base<TraitsType>::optimizeScalarMul(Variable *Dest, Operand *Src0,
       return false;
     }
   }
-  // Lea optimization only works for i16 and i32 types, not i8.
-  if (Ty != IceType_i16 && Ty != IceType_i32 && (Count3 || Count5 || Count9))
+  // Lea optimization only works for i32 types, not i8 or i16.

[Jim Stichnoth, 2016/01/10 03:08:56]

i32 type

[sehr, 2016/01/11 21:49:48]

Done.

+  if (Ty != IceType_i32 && (Count3 || Count5 || Count9))
     return false;
   // Limit the number of lea/shl operations for a single multiply, to a
   // somewhat arbitrary choice of 3.
@@ -5841,7 +5841,14 @@ Variable *TargetX86Base<TraitsType>::makeVectorOfMinusOnes(Type Ty,
   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)
+    // Making a vector of minus ones of type f64 is currently only used for the
+    // fabs intrinsic.  To use the f64 type to create this mask with pcmpeqq
+    // requires SSE 4.1.  Since we're just creating a mask, pcmpeqd does the
+    // same job and only requires SSE2.
+    _pcmpeq(MinusOnes, MinusOnes, IceType_f32);
+  else
+    _pcmpeq(MinusOnes, MinusOnes);
   return MinusOnes;
 }