CodeGen: extend f16 conversions to permit types > float.

This makes the two intrinsics @llvm.convert.from.f16 and
@llvm.convert.to.f16 accept types other than simple "float". This is
only strictly needed for the truncate operation, since otherwise
double rounding occurs and there's no way to represent the strict IEEE
conversion. However, for symmetry we allow larger types in the extend
too.

During legalization, we can expand an "fp16_to_double" operation into
two extends for convenience, but abort when the truncate isn't legal. A new
libcall is probably needed here.

Even after this commit, various target tweaks are needed to actually use the
extended intrinsics. I've put these into separate commits for clarity, so there
are no actual tests of f64 conversion here.

llvm-svn: 213248

1 files changed, 20 insertions, 3 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
index bff540ad174..58f290e1137 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -1186,6 +1186,7 @@ void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
     if (Action != TargetLowering::Promote)
       Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
     break;
+  case ISD::FP_TO_FP16:
   case ISD::SINT_TO_FP:
   case ISD::UINT_TO_FP:
   case ISD::EXTRACT_VECTOR_ELT:
@@ -3513,10 +3514,26 @@ void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
                                       RTLIB::FMA_F80, RTLIB::FMA_F128,
                                       RTLIB::FMA_PPCF128));
     break;
-  case ISD::FP16_TO_FP32:
-    Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
+  case ISD::FP16_TO_FP: {
+    if (Node->getValueType(0) == MVT::f32) {
+      Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
+      break;
+    }
+
+    // We can extend to types bigger than f32 in two steps without changing the
+    // result. Since "f16 -> f32" is much more commonly available, give CodeGen
+    // the option of emitting that before resorting to a libcall.
+    SDValue Res =
+        DAG.getNode(ISD::FP16_TO_FP, dl, MVT::f32, Node->getOperand(0));
+    Results.push_back(
+        DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Res));
     break;
-  case ISD::FP32_TO_FP16:
+  }
+  case ISD::FP_TO_FP16:
+    // Can't use two-step truncation here because the rounding may be
+    // significant.
+    assert(Node->getOperand(0).getValueType() == MVT::f32 &&
+           "Don't know libcall for FPROUND_F64_F16");
     Results.push_back(ExpandLibCall(RTLIB::FPROUND_F32_F16, Node, false));
     break;
   case ISD::ConstantFP: {