[ARM] FP16: support vector INT_TO_FP and FP_TO_INT

[SjoerdMeijer]

Summary

This adds codegen support for the different vcvt_f16 variants.

[samparker]

LGTM

[SjoerdMeijer]

Also added HasFullFP16 check to LowerVectorFP_TO_INT.

lib/Target/ARM/ARMISelLowering.cpp

if (Subtarget->hasNEON()) {
     setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
     setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
     setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
     // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
     // a destination type that is wider than the source, and nor does
     // 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);
 
     // NEON does not have single instruction CTPOP for vectors with element
     // types wider than 8-bits.  However, custom lowering can leverage the
     // v8i8/v16i8 vcnt instruction.
     setOperationAction(ISD::CTPOP,      MVT::v2i32, Custom);
static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
   SDLoc dl(Op);
 
   if (Op.getValueType().getVectorElementType() == MVT::i32) {
     if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
       return Op;
     return DAG.UnrollVectorOp(Op.getNode());
   }
 
-  assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
-         "Invalid type for custom lowering!");
-  if (VT != MVT::v4i16)
+  EVT NewTy;
+  const EVT OpTy = Op.getOperand(0).getValueType();
+  if (OpTy == MVT::v4f32)
+    NewTy = MVT::v4i32;
+  else if (OpTy == MVT::v4f16)
+    NewTy = MVT::v4i16;
+  else if (OpTy == MVT::v8f16)
+    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);
 }
 
 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
   EVT VT = Op.getValueType();
   if (VT.isVector())
     return LowerVectorFP_TO_INT(Op, DAG);
   if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
   SDLoc dl(Op);
 
   if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
     if (VT.getVectorElementType() == MVT::f32)
       return Op;
     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;
   unsigned Opc;
   switch (Op.getOpcode()) {
   default: llvm_unreachable("Invalid opcode!");
   case ISD::SINT_TO_FP:
     CastOpc = ISD::SIGN_EXTEND;
     Opc = ISD::SINT_TO_FP;
     break;
   case ISD::UINT_TO_FP:
     CastOpc = ISD::ZERO_EXTEND;
     Opc = ISD::UINT_TO_FP;
     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);
 }
 
 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
   EVT VT = Op.getValueType();
   if (VT.isVector())
     return LowerVectorINT_TO_FP(Op, DAG);
   if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {

test/CodeGen/ARM/armv8.2a-fp16-vector-intrinsics.ll

 ; CHECK:         vclt.f16 q0, q0, #0
 ; CHECK-NEXT:    bx lr
 entry:
   %0 = fcmp olt <8 x half> %a, zeroinitializer
   %vcltz.i = sext <8 x i1> %0 to <8 x i16>
   ret <8 x i16> %vcltz.i
 }
 
-; FIXME (PR38404)
-;
-;define dso_local <4 x half> @test_vcvt_f16_s16(<4 x i16> %a) {
-;entry:
-;  %vcvt.i = sitofp <4 x i16> %a to <4 x half>
-;  ret <4 x half> %vcvt.i
-;}
-;
-;define dso_local <8 x half> @test_vcvtq_f16_s16(<8 x i16> %a) {
-;entry:
-;  %vcvt.i = sitofp <8 x i16> %a to <8 x half>
-;  ret <8 x half> %vcvt.i
-;}
+define dso_local <4 x half> @test_vcvt_f16_s16(<4 x i16> %a) {
+; CHECK-LABEL: test_vcvt_f16_s16:
+; CHECK:         vcvt.f16.s16 d0, d0
+; CHECK-NEXT:    bx lr
+entry:
+  %vcvt.i = sitofp <4 x i16> %a to <4 x half>
+  ret <4 x half> %vcvt.i
+}
 
-;define dso_local <4 x half> @test_vcvt_f16_u16(<4 x i16> %a) {
-;entry:
-;  %vcvt.i = uitofp <4 x i16> %a to <4 x half>
-;  ret <4 x half> %vcvt.i
-;}
+define dso_local <8 x half> @test_vcvtq_f16_s16(<8 x i16> %a) {
+; CHECK-LABEL: test_vcvtq_f16_s16:
+; CHECK:         vcvt.f16.s16 q0, q0
+; CHECK-NEXT:    bx lr
+entry:
+  %vcvt.i = sitofp <8 x i16> %a to <8 x half>
+  ret <8 x half> %vcvt.i
+}
 
-;define dso_local <8 x half> @test_vcvtq_f16_u16(<8 x i16> %a) {
-;entry:
-;  %vcvt.i = uitofp <8 x i16> %a to <8 x half>
-;  ret <8 x half> %vcvt.i
-;}
+define dso_local <4 x half> @test_vcvt_f16_u16(<4 x i16> %a) {
+; CHECK-LABEL: test_vcvt_f16_u16:
+; CHECK:         vcvt.f16.u16 d0, d0
+; CHECK-NEXT:    bx lr
+entry:
+  %vcvt.i = uitofp <4 x i16> %a to <4 x half>
+  ret <4 x half> %vcvt.i
+}
 
-;define dso_local <4 x i16> @test_vcvt_s16_f16(<4 x half> %a) {
-;entry:
-;  %vcvt.i = fptosi <4 x half> %a to <4 x i16>
-;  ret <4 x i16> %vcvt.i
-;}
+define dso_local <8 x half> @test_vcvtq_f16_u16(<8 x i16> %a) {
+; CHECK-LABEL: test_vcvtq_f16_u16:
+; CHECK:         vcvt.f16.u16 q0, q0
+; CHECK-NEXT:    bx lr
+entry:
+  %vcvt.i = uitofp <8 x i16> %a to <8 x half>
+  ret <8 x half> %vcvt.i
+}
 
-;define dso_local <8 x i16> @test_vcvtq_s16_f16(<8 x half> %a) {
-;entry:
-;  %vcvt.i = fptosi <8 x half> %a to <8 x i16>
-;  ret <8 x i16> %vcvt.i
-;}
+define dso_local <4 x i16> @test_vcvt_s16_f16(<4 x half> %a) {
+; CHECK-LABEL: test_vcvt_s16_f16:
+; CHECK:         vcvt.s16.f16 d0, d0
+; CHECK-NEXT:    bx lr
+entry:
+  %vcvt.i = fptosi <4 x half> %a to <4 x i16>
+  ret <4 x i16> %vcvt.i
+}
 
-;define dso_local <4 x i16> @test_vcvt_u16_f16(<4 x half> %a) {
-;entry:
-;  %vcvt.i = fptoui <4 x half> %a to <4 x i16>
-;  ret <4 x i16> %vcvt.i
-;}
+define dso_local <8 x i16> @test_vcvtq_s16_f16(<8 x half> %a) {
+; CHECK-LABEL: test_vcvtq_s16_f16:
+; CHECK:         vcvt.s16.f16 q0, q0
+; CHECK-NEXT:    bx lr
+entry:
+  %vcvt.i = fptosi <8 x half> %a to <8 x i16>
+  ret <8 x i16> %vcvt.i
+}
 
-;define dso_local <8 x i16> @test_vcvtq_u16_f16(<8 x half> %a) {
-;entry:
-;  %vcvt.i = fptoui <8 x half> %a to <8 x i16>
-;  ret <8 x i16> %vcvt.i
-;}
+define dso_local <4 x i16> @test_vcvt_u16_f16(<4 x half> %a) {
+; CHECK-LABEL: test_vcvt_u16_f16:
+; CHECK:         vcvt.u16.f16 d0, d0
+; CHECK-NEXT:    bx lr
+entry:
+  %vcvt.i = fptoui <4 x half> %a to <4 x i16>
+  ret <4 x i16> %vcvt.i
+}
+
+define dso_local <8 x i16> @test_vcvtq_u16_f16(<8 x half> %a) {
+; CHECK-LABEL: test_vcvtq_u16_f16:
+; CHECK:         vcvt.u16.f16 q0, q0
+; CHECK-NEXT:    bx lr
+entry:
+  %vcvt.i = fptoui <8 x half> %a to <8 x i16>
+  ret <8 x i16> %vcvt.i
+}
 
 define dso_local <4 x i16> @test_vcvta_s16_f16(<4 x half> %a) {
 ; CHECK-LABEL: test_vcvta_s16_f16:
 ; CHECK:         vcvta.s16.f16 d0, d0
 ; CHECK-NEXT:    bx lr
 entry:
   %vcvta_s16_v1.i = tail call <4 x i16> @llvm.arm.neon.vcvtas.v4i16.v4f16(<4 x half> %a)
   ret <4 x i16> %vcvta_s16_v1.i