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 1117 matching lines @@
       Src1 /= 3;
     } else if (Src1 % 2 == 0) {
       if (Count2 == 0)
         ++CountOps;
       ++Count2;
       Src1 /= 2;
     } else {
       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.
   constexpr uint32_t MaxOpsForOptimizedMul = 3;
   if (CountOps > MaxOpsForOptimizedMul)
     return false;
   _mov(T, Src0);
   Constant *Zero = Ctx->getConstantZero(IceType_i32);
   for (uint32_t i = 0; i < Count9; ++i) {
     constexpr uint16_t Shift = 3; // log2(9-1)
@@ skipping 4684 matching lines @@
                                                        int32_t RegNum) {
   return makeZeroedRegister(Ty, RegNum);
 }
 
 template <typename TraitsType>
 Variable *TargetX86Base<TraitsType>::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)
+    // 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;
 }
 
 template <typename TraitsType>
 Variable *TargetX86Base<TraitsType>::makeVectorOfOnes(Type Ty, int32_t RegNum) {
   Variable *Dest = makeVectorOfZeros(Ty, RegNum);
   Variable *MinusOne = makeVectorOfMinusOnes(Ty);
   _psub(Dest, MinusOne);
   return Dest;
 }
@@ skipping 700 matching lines @@
   }
   // the offset is not eligible for blinding or pooling, return the original
   // mem operand
   return MemOperand;
 }
 
 } // end of namespace X86NAMESPACE
 } // end of namespace Ice
 
 #endif // SUBZERO_SRC_ICETARGETLOWERINGX86BASEIMPL_H