Implement vector constants which are splat of
integers with mov + vdup.  8003375.  This is
currently disabled by default because LICM will
not hoist a VDUP, so it pessimizes the code if
the construct occurs inside a loop (8248029).



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@109799 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 56d2440..69b24d0 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -57,6 +57,13 @@
   cl::desc("Generate tail calls (TEMPORARY OPTION)."),
   cl::init(true));
 
+// This option should go away when Machine LICM is smart enough to hoist a 
+// reg-to-reg VDUP.
+static cl::opt<bool>
+EnableARMVDUPsplat("arm-vdup-splat", cl::Hidden,
+  cl::desc("Generate VDUP for integer constant splats (TEMPORARY OPTION)."),
+  cl::init(false));
+
 static cl::opt<bool>
 EnableARMLongCalls("arm-long-calls", cl::Hidden,
   cl::desc("Generate calls via indirect call instructions"),
@@ -3257,9 +3264,30 @@
   return true;
 }
 
+// If N is an integer constant that can be moved into a register in one
+// instruction, return an SDValue of such a constant (will become a MOV
+// instruction).  Otherwise return null.
+static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
+                                     const ARMSubtarget *ST, DebugLoc dl) {
+  uint64_t Val;
+  if (!isa<ConstantSDNode>(N))
+    return SDValue();
+  Val = cast<ConstantSDNode>(N)->getZExtValue();
+
+  if (ST->isThumb1Only()) {
+    if (Val <= 255 || ~Val <= 255)
+      return DAG.getConstant(Val, MVT::i32);
+  } else {
+    if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
+      return DAG.getConstant(Val, MVT::i32);
+  }
+  return SDValue();
+}
+
 // If this is a case we can't handle, return null and let the default
 // expansion code take care of it.
-static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
+static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG, 
+                                 const ARMSubtarget *ST) {
   BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
   DebugLoc dl = Op.getDebugLoc();
   EVT VT = Op.getValueType();
@@ -3319,15 +3347,41 @@
   if (isOnlyLowElement)
     return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
 
-  // If all elements are constants, fall back to the default expansion, which
-  // will generate a load from the constant pool.
+  unsigned EltSize = VT.getVectorElementType().getSizeInBits();
+
+  if (EnableARMVDUPsplat) {
+    // Use VDUP for non-constant splats.  For f32 constant splats, reduce to
+    // i32 and try again.
+    if (usesOnlyOneValue && EltSize <= 32) {
+      if (!isConstant)
+        return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
+      if (VT.getVectorElementType().isFloatingPoint()) {
+        SmallVector<SDValue, 8> Ops;
+        for (unsigned i = 0; i < NumElts; ++i)
+          Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, 
+                                    Op.getOperand(i)));
+        SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &Ops[0],
+                                  NumElts);
+        return DAG.getNode(ISD::BIT_CONVERT, dl, VT, 
+                           LowerBUILD_VECTOR(Val, DAG, ST));
+      }
+      SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
+      if (Val.getNode())
+        return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
+    }
+  }
+
+  // If all elements are constants and the case above didn't get hit, fall back
+  // to the default expansion, which will generate a load from the constant
+  // pool.
   if (isConstant)
     return SDValue();
 
-  // Use VDUP for non-constant splats.
-  unsigned EltSize = VT.getVectorElementType().getSizeInBits();
-  if (usesOnlyOneValue && EltSize <= 32)
-    return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
+  if (!EnableARMVDUPsplat) {
+    // Use VDUP for non-constant splats.
+    if (usesOnlyOneValue && EltSize <= 32)
+      return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
+  }
 
   // Vectors with 32- or 64-bit elements can be built by directly assigning
   // the subregisters.  Lower it to an ARMISD::BUILD_VECTOR so the operands
@@ -3647,7 +3701,7 @@
   case ISD::SRA_PARTS:     return LowerShiftRightParts(Op, DAG);
   case ISD::CTTZ:          return LowerCTTZ(Op.getNode(), DAG, Subtarget);
   case ISD::VSETCC:        return LowerVSETCC(Op, DAG);
-  case ISD::BUILD_VECTOR:  return LowerBUILD_VECTOR(Op, DAG);
+  case ISD::BUILD_VECTOR:  return LowerBUILD_VECTOR(Op, DAG, Subtarget);
   case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
   case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
   case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);