CodeGen: emit IR-level f16 conversion intrinsics as fptrunc/fpext

This makes the first stage DAG for @llvm.convert.to.fp16 an fptrunc,
and correspondingly @llvm.convert.from.fp16 an fpext. The legalisation
path is now uniform, regardless of the input IR:

  fptrunc -> FP_TO_FP16 (if f16 illegal) -> libcall
  fpext -> FP16_TO_FP (if f16 illegal) -> libcall

Each target should be able to select the version that best matches its
operations and not be required to duplicate patterns for both fptrunc
and FP_TO_FP16 (for example).

As a result we can remove some redundant AArch64 patterns.

llvm-svn: 213507

3 files changed, 21 insertions, 5 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp
index e3c034c8a15..649dd7a349f 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp
@@ -373,6 +373,13 @@ SDValue DAGTypeLegalizer::SoftenFloatRes_FNEG(SDNode *N) {
 SDValue DAGTypeLegalizer::SoftenFloatRes_FP_EXTEND(SDNode *N) {
   EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
   SDValue Op = N->getOperand(0);
+
+  // There's only a libcall for f16 -> f32, so proceed in two stages. Also, it's
+  // entirely possible for both f16 and f32 to be legal, so use the fully
+  // hard-float FP_EXTEND rather than FP16_TO_FP.
+  if (Op.getValueType() == MVT::f16 && N->getValueType(0) != MVT::f32)
+    Op = DAG.getNode(ISD::FP_EXTEND, SDLoc(N), MVT::f32, Op);
+
   RTLIB::Libcall LC = RTLIB::getFPEXT(Op.getValueType(), N->getValueType(0));
   assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported FP_EXTEND!");
   return TLI.makeLibCall(DAG, LC, NVT, &Op, 1, false, SDLoc(N)).first;
@@ -511,6 +518,9 @@ SDValue DAGTypeLegalizer::SoftenFloatRes_FSUB(SDNode *N) {
 
 SDValue DAGTypeLegalizer::SoftenFloatRes_FTRUNC(SDNode *N) {
   EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+  if (N->getValueType(0) == MVT::f16)
+    return DAG.getNode(ISD::FP_TO_FP16, SDLoc(N), NVT, N->getOperand(0));
+
   SDValue Op = GetSoftenedFloat(N->getOperand(0));
   return TLI.makeLibCall(DAG, GetFPLibCall(N->getValueType(0),
                                            RTLIB::TRUNC_F32,
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index 189b56d5a93..c07b5e6a736 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -5155,13 +5155,16 @@ SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
     return nullptr;
   }
   case Intrinsic::convert_to_fp16:
-    setValue(&I, DAG.getNode(ISD::FP_TO_FP16, sdl,
-                             MVT::i16, getValue(I.getArgOperand(0))));
+    setValue(&I, DAG.getNode(ISD::BITCAST, sdl, MVT::i16,
+                             DAG.getNode(ISD::FP_ROUND, sdl, MVT::f16,
+                                         getValue(I.getArgOperand(0)),
+                                         DAG.getTargetConstant(0, MVT::i32))));
     return nullptr;
   case Intrinsic::convert_from_fp16:
     setValue(&I,
-             DAG.getNode(ISD::FP16_TO_FP, sdl, TLI->getValueType(I.getType()),
-                         getValue(I.getArgOperand(0))));
+             DAG.getNode(ISD::FP_EXTEND, sdl, TLI->getValueType(I.getType()),
+                         DAG.getNode(ISD::BITCAST, sdl, MVT::f16,
+                                     getValue(I.getArgOperand(0)))));
     return nullptr;
   case Intrinsic::pcmarker: {
     SDValue Tmp = getValue(I.getArgOperand(0));
diff --git a/llvm/lib/CodeGen/TargetLoweringBase.cpp b/llvm/lib/CodeGen/TargetLoweringBase.cpp
index 42cd4bf1fe1..e80ef7176c2 100644
--- a/llvm/lib/CodeGen/TargetLoweringBase.cpp
+++ b/llvm/lib/CodeGen/TargetLoweringBase.cpp
@@ -422,7 +422,10 @@ static void InitLibcallCallingConvs(CallingConv::ID *CCs) {
 /// getFPEXT - Return the FPEXT_*_* value for the given types, or
 /// UNKNOWN_LIBCALL if there is none.
 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) {
-  if (OpVT == MVT::f32) {
+  if (OpVT == MVT::f16) {
+    if (RetVT == MVT::f32)
+      return FPEXT_F16_F32;
+  } else if (OpVT == MVT::f32) {
     if (RetVT == MVT::f64)
       return FPEXT_F32_F64;
     if (RetVT == MVT::f128)