[SelectionDAG] Add support for FP_ROUND in WidenVectorOperand.

Summary:
This is used on AMDGPU for rounding from v3f64 (which is illegal) to
v3f32 (which is legal).

Subscribers: jvesely, nhaehnle, tpr, hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D69339

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
index 3763e88..6c47c5b 100644 (file)
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
@@ -4161,6 +4161,8 @@ bool DAGTypeLegalizer::WidenVectorOperand(SDNode *N, unsigned OpNo) {
 
   case ISD::FP_EXTEND:
   case ISD::STRICT_FP_EXTEND:
+  case ISD::FP_ROUND:
+  case ISD::STRICT_FP_ROUND:
   case ISD::FP_TO_SINT:
   case ISD::STRICT_FP_TO_SINT:
   case ISD::FP_TO_UINT:
@@ -4297,13 +4299,21 @@ SDValue DAGTypeLegalizer::WidenVecOp_Convert(SDNode *N) {
   if (TLI.isTypeLegal(WideVT) && !N->isStrictFPOpcode()) {
     SDValue Res;
     if (N->isStrictFPOpcode()) {
-      Res = DAG.getNode(Opcode, dl, { WideVT, MVT::Other }, 
-                        { N->getOperand(0), InOp });
+      if (Opcode == ISD::STRICT_FP_ROUND)
+        Res = DAG.getNode(Opcode, dl, { WideVT, MVT::Other },
+                          { N->getOperand(0), InOp, N->getOperand(2) });
+      else
+        Res = DAG.getNode(Opcode, dl, { WideVT, MVT::Other },
+                          { N->getOperand(0), InOp });
       // Legalize the chain result - switch anything that used the old chain to
       // use the new one.
       ReplaceValueWith(SDValue(N, 1), Res.getValue(1));
-    } else
-      Res = DAG.getNode(Opcode, dl, WideVT, InOp);
+    } else {
+      if (Opcode == ISD::FP_ROUND)
+        Res = DAG.getNode(Opcode, dl, WideVT, InOp, N->getOperand(1));
+      else
+        Res = DAG.getNode(Opcode, dl, WideVT, InOp);
+    }
     return DAG.getNode(
         ISD::EXTRACT_SUBVECTOR, dl, VT, Res,
         DAG.getConstant(0, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));

diff --git a/llvm/test/CodeGen/AMDGPU/fptrunc.ll b/llvm/test/CodeGen/AMDGPU/fptrunc.ll
index d9c5b7e..80f0a37 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/fptrunc.ll
+++ b/llvm/test/CodeGen/AMDGPU/fptrunc.ll
@@ -30,6 +30,16 @@ define amdgpu_kernel void @fptrunc_v2f64_to_v2f32(<2 x float> addrspace(1)* %out
   ret void
 }
 
+; FUNC-LABEL: {{^}}fptrunc_v3f64_to_v3f32:
+; GCN: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
+define amdgpu_kernel void @fptrunc_v3f64_to_v3f32(<3 x float> addrspace(1)* %out, <3 x double> %in) {
+  %result = fptrunc <3 x double> %in to <3 x float>
+  store <3 x float> %result, <3 x float> addrspace(1)* %out
+  ret void
+}
+
 ; FUNC-LABEL: {{^}}fptrunc_v4f64_to_v4f32:
 ; GCN: v_cvt_f32_f64_e32
 ; GCN: v_cvt_f32_f64_e32