Merge pull request #2764 from apple/fix-fp-round-20210107

[SelectionDAG] Don't scalarize vector fpround sources that don't need it.

Author: ornata

llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

@@ -318,10 +318,21 @@ SDValue DAGTypeLegalizer::ScalarizeVecRes_EXTRACT_SUBVECTOR(SDNode *N) {
 }
 
 SDValue DAGTypeLegalizer::ScalarizeVecRes_FP_ROUND(SDNode *N) {
-  EVT NewVT = N->getValueType(0).getVectorElementType();
-  SDValue Op = GetScalarizedVector(N->getOperand(0));
-  return DAG.getNode(ISD::FP_ROUND, SDLoc(N),
-                     NewVT, Op, N->getOperand(1));
+  SDLoc DL(N);
+  SDValue Op = N->getOperand(0);
+  EVT OpVT = Op.getValueType();
+  // The result needs scalarizing, but it's not a given that the source does.
+  // See similar logic in ScalarizeVecRes_UnaryOp.
+  if (getTypeAction(OpVT) == TargetLowering::TypeScalarizeVector) {
+    Op = GetScalarizedVector(Op);
+  } else {
+    EVT VT = OpVT.getVectorElementType();
+    Op = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Op,
+                     DAG.getVectorIdxConstant(0, DL));
+  }
+  return DAG.getNode(ISD::FP_ROUND, DL,
+                     N->getValueType(0).getVectorElementType(), Op,
+                     N->getOperand(1));
 }
 
 SDValue DAGTypeLegalizer::ScalarizeVecRes_FPOWI(SDNode *N) {

llvm/test/CodeGen/AArch64/arm64-vcvt_f.ll

@@ -199,6 +199,33 @@ define <2 x float> @test_vcvt_f32_f64(<2 x double> %v) nounwind readnone ssp {
   ret <2 x float> %vcvt1.i
 }
 
+define half @test_vcvt_f16_f32(<1 x float> %x) {
+; GENERIC-LABEL: test_vcvt_f16_f32:
+; GENERIC:       // %bb.0:
+; GENERIC-NEXT:    // kill: def $d0 killed $d0 def $q0
+; GENERIC-NEXT:    fcvt h0, s0
+; GENERIC-NEXT:    ret
+;
+; FAST-LABEL: test_vcvt_f16_f32:
+; FAST:       // %bb.0:
+; FAST-NEXT:    mov.16b v1, v0
+; FAST-NEXT:    // implicit-def: $q0
+; FAST-NEXT:    mov.16b v0, v1
+; FAST-NEXT:    // kill: def $s0 killed $s0 killed $q0
+; FAST-NEXT:    fcvt h0, s0
+; FAST-NEXT:    ret
+;
+; GISEL-LABEL: test_vcvt_f16_f32:
+; GISEL:       // %bb.0:
+; GISEL-NEXT:    fmov x8, d0
+; GISEL-NEXT:    fmov s0, w8
+; GISEL-NEXT:    fcvt h0, s0
+; GISEL-NEXT:    ret
+  %tmp = fptrunc <1 x float> %x to <1 x half>
+  %elt = extractelement <1 x half> %tmp, i32 0
+  ret half %elt
+}
+
 ; FALLBACK-NOT: remark{{.*}}G_FPEXT{{.*}}(in function: test_vcvt_high_f32_f64)
 ; FALLBACK-NOT: remark{{.*}}fpext{{.*}}(in function: test_vcvt_high_f32_f64)
 define <4 x float> @test_vcvt_high_f32_f64(<2 x float> %x, <2 x double> %v) nounwind readnone ssp {