[ARM] FP16: support vector INT_TO_FP and FP_TO_INT

This adds codegen support for the different vcvt_f16 variants.

Differential Revision: https://reviews.llvm.org/D50393

llvm-svn: 339227

1 files changed, 35 insertions, 7 deletions

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 63d6b35cc81..da99058b91a 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -651,9 +651,13 @@ ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
     // it have a FP_TO_[SU]INT instruction with a narrower destination than
     // source.
     setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
+    setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Custom);
     setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
+    setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
     setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
+    setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Custom);
     setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
+    setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Custom);
 
     setOperationAction(ISD::FP_ROUND,   MVT::v2f32, Expand);
     setOperationAction(ISD::FP_EXTEND,  MVT::v2f64, Expand);
@@ -4844,12 +4848,24 @@ static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
     return DAG.UnrollVectorOp(Op.getNode());
   }
 
-  assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
-         "Invalid type for custom lowering!");
-  if (VT != MVT::v4i16)
+  const bool HasFullFP16 =
+    static_cast<const ARMSubtarget&>(DAG.getSubtarget()).hasFullFP16();
+
+  EVT NewTy;
+  const EVT OpTy = Op.getOperand(0).getValueType();
+  if (OpTy == MVT::v4f32)
+    NewTy = MVT::v4i32;
+  else if (OpTy == MVT::v4f16 && HasFullFP16)
+    NewTy = MVT::v4i16;
+  else if (OpTy == MVT::v8f16 && HasFullFP16)
+    NewTy = MVT::v8i16;
+  else
+    llvm_unreachable("Invalid type for custom lowering!");
+
+  if (VT != MVT::v4i16 && VT != MVT::v8i16)
     return DAG.UnrollVectorOp(Op.getNode());
 
-  Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
+  Op = DAG.getNode(Op.getOpcode(), dl, NewTy, Op.getOperand(0));
   return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
 }
 
@@ -4882,9 +4898,21 @@ static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
     return DAG.UnrollVectorOp(Op.getNode());
   }
 
-  assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
+  assert((Op.getOperand(0).getValueType() == MVT::v4i16 ||
+          Op.getOperand(0).getValueType() == MVT::v8i16) &&
          "Invalid type for custom lowering!");
-  if (VT != MVT::v4f32)
+
+  const bool HasFullFP16 =
+    static_cast<const ARMSubtarget&>(DAG.getSubtarget()).hasFullFP16();
+
+  EVT DestVecType;
+  if (VT == MVT::v4f32)
+    DestVecType = MVT::v4i32;
+  else if (VT == MVT::v4f16 && HasFullFP16)
+    DestVecType = MVT::v4i16;
+  else if (VT == MVT::v8f16 && HasFullFP16)
+    DestVecType = MVT::v8i16;
+  else
     return DAG.UnrollVectorOp(Op.getNode());
 
   unsigned CastOpc;
@@ -4901,7 +4929,7 @@ static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
     break;
   }
 
-  Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
+  Op = DAG.getNode(CastOpc, dl, DestVecType, Op.getOperand(0));
   return DAG.getNode(Opc, dl, VT, Op);
 }