[SelectionDAG] Widening the result of INSERT_SUBVECTOR

[CarolineConcatto]

Summary

Widens the result and first input vector because they have the same size.
The subvector to be inserted is widened in the operand widen function.

[sdesmalen]

llvm/test/CodeGen/AArch64/sve-insert-vector.ll

425

This should be v1?

[david-arm]

llvm/test/CodeGen/AArch64/sve-insert-vector.ll

407

Is it worth having an equivalent one for floats, i.e.

define <vscale x 3 x float> @insert_nxv3f32_nxv2f32(<vscale x 2 x float> %sv0)

?

407

Also, I wonder if it's worth having a test where we also pass in a <vscale x 3 x i32> vector, i.e.

define <vscale x 3 x i32> @insert_nxv3i32_nxv2i32_2(<vscale x 3 x i32> %sv0, <vscale x 2 x i32> %sv1) {
  %v0 = call <vscale x 3 x i32> @llvm.experimental.vector.insert.nxv3i32.nxv2i32(<vscale x 3 x i32> %sv0, <vscale x 2 x i32> %sv1, i64 0)
  ret <vscale x 3 x i32> %v0

}

I realise this may be covered by one of the tests below, but I just wanted to make sure that a DAG combine wasn't hiding a bug somewhere that's all.

[CarolineConcatto]

• Fix return value in the test insert_nxv6i32_nxv2i32

[CarolineConcatto]

• Add new tests insert_nxv3i32_nxv2i32_2 and insert_nxv3f32_nxv2f32

[david-arm]

llvm/test/CodeGen/AArch64/sve-insert-vector.ll

466

Hi @CarolineConcatto, did you mean to pass in %sv1 here instead?

[CarolineConcatto]

• Fix insert_nxv6i32_nxv2i32 to use sv0 and sv1 as input instead of only sv0

[CarolineConcatto]

llvm/test/CodeGen/AArch64/sve-insert-vector.ll

466

Thank you David!

[sdesmalen]

LGTM with nit addressed! Thanks @CarolineConcatto

llvm/test/CodeGen/AArch64/sve-insert-vector.ll

428

nit: please add nounwind attributes to the functions to avoid all the .cfi directives.

[david-arm]

LGTM! Thanks for making the changes @CarolineConcatto. :)

llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h

private:
   // Widen Vector Result Promotion.
   void WidenVectorResult(SDNode *N, unsigned ResNo);
   SDValue WidenVecRes_MERGE_VALUES(SDNode* N, unsigned ResNo);
   SDValue WidenVecRes_BITCAST(SDNode* N);
   SDValue WidenVecRes_BUILD_VECTOR(SDNode* N);
   SDValue WidenVecRes_CONCAT_VECTORS(SDNode* N);
   SDValue WidenVecRes_EXTEND_VECTOR_INREG(SDNode* N);
   SDValue WidenVecRes_EXTRACT_SUBVECTOR(SDNode* N);
+  SDValue WidenVecRes_INSERT_SUBVECTOR(SDNode *N);
   SDValue WidenVecRes_INSERT_VECTOR_ELT(SDNode* N);
   SDValue WidenVecRes_LOAD(SDNode* N);
   SDValue WidenVecRes_MLOAD(MaskedLoadSDNode* N);
   SDValue WidenVecRes_MGATHER(MaskedGatherSDNode* N);
   SDValue WidenVecRes_ScalarOp(SDNode* N);
   SDValue WidenVecRes_SELECT(SDNode* N);
   SDValue WidenVSELECTMask(SDNode *N);
   SDValue WidenVecRes_SELECT_CC(SDNode* N);

llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

#ifndef NDEBUG
     dbgs() << "\n";
 #endif
     llvm_unreachable("Do not know how to widen the result of this operator!");
 
   case ISD::MERGE_VALUES:      Res = WidenVecRes_MERGE_VALUES(N, ResNo); break;
   case ISD::BITCAST:           Res = WidenVecRes_BITCAST(N); break;
   case ISD::BUILD_VECTOR:      Res = WidenVecRes_BUILD_VECTOR(N); break;
   case ISD::CONCAT_VECTORS:    Res = WidenVecRes_CONCAT_VECTORS(N); break;
+  case ISD::INSERT_SUBVECTOR:
+    Res = WidenVecRes_INSERT_SUBVECTOR(N);
+    break;
   case ISD::EXTRACT_SUBVECTOR: Res = WidenVecRes_EXTRACT_SUBVECTOR(N); break;
   case ISD::INSERT_VECTOR_ELT: Res = WidenVecRes_INSERT_VECTOR_ELT(N); break;
   case ISD::LOAD:              Res = WidenVecRes_LOAD(N); break;
   case ISD::SPLAT_VECTOR:
   case ISD::SCALAR_TO_VECTOR:
     Res = WidenVecRes_ScalarOp(N);
     break;
   case ISD::SIGN_EXTEND_INREG: Res = WidenVecRes_InregOp(N); break;
for (unsigned j = 0; j < NumInElts; ++j)
                                DAG.getVectorIdxConstant(j, dl));
   }
   SDValue UndefVal = DAG.getUNDEF(EltVT);
   for (; Idx < WidenNumElts; ++Idx)
     Ops[Idx] = UndefVal;
   return DAG.getBuildVector(WidenVT, dl, Ops);
 }
 
+SDValue DAGTypeLegalizer::WidenVecRes_INSERT_SUBVECTOR(SDNode *N) {
+  EVT VT = N->getValueType(0);
+  EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
+  SDValue InOp1 = GetWidenedVector(N->getOperand(0));
+  SDValue InOp2 = N->getOperand(1);
+  SDValue Idx = N->getOperand(2);
+  SDLoc dl(N);
+  return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WidenVT, InOp1, InOp2, Idx);
+}
+
 SDValue DAGTypeLegalizer::WidenVecRes_EXTRACT_SUBVECTOR(SDNode *N) {
   EVT      VT = N->getValueType(0);
   EVT      EltVT = VT.getVectorElementType();
   EVT      WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
   SDValue  InOp = N->getOperand(0);
   SDValue  Idx  = N->getOperand(1);
   SDLoc dl(N);
 

llvm/test/CodeGen/AArch64/sve-insert-vector.ll

 declare <vscale x 16 x i64> @llvm.experimental.vector.insert.nxv8i64.nxv16i64(<vscale x 16 x i64>, <vscale x 8 x i64>, i64)
 declare <vscale x 16 x i64> @llvm.experimental.vector.insert.v2i64.nxv16i64(<vscale x 16 x i64>, <2 x i64>, i64)
 declare <vscale x 4 x i32> @llvm.experimental.vector.insert.nxv4i32.nxv1i32(<vscale x 4 x i32>, <vscale x 1 x i32>, i64)
 declare <vscale x 6 x i16> @llvm.experimental.vector.insert.nxv6i16.nxv1i16(<vscale x 6 x i16>, <vscale x 1 x i16>, i64)
 
 declare <vscale x 8 x i16> @llvm.experimental.vector.insert.nxv8i16.nxv2i16(<vscale x 8 x i16>, <vscale x 2 x i16>, i64)
 
 declare <vscale x 8 x half> @llvm.experimental.vector.insert.nxv8f16.nxv2f16(<vscale x 8 x half>, <vscale x 2 x half>, i64)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;;  Upacked types that need results widening
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+define <vscale x 3 x i32> @insert_nxv3i32_nxv2i32(<vscale x 2 x i32> %sv0) {
+; CHECK-LABEL: insert_nxv3i32_nxv2i32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    uunpkhi z1.d, z0.s
+; CHECK-NEXT:    uzp1 z0.s, z0.s, z1.s
+; CHECK-NEXT:    ret
+  %v0 = call <vscale x 3 x i32> @llvm.experimental.vector.insert.nxv3i32.nxv2i32(<vscale x 3 x i32> undef, <vscale x 2 x i32> %sv0, i64 0)
+ ret <vscale x 3 x i32> %v0
+}
+
+define <vscale x 6 x i32>  @insert_nxv6i32_nxv2i32(<vscale x 2 x i32> %sv0, <vscale x 2 x i32> %sv1) {
+; CHECK-LABEL: insert_nxv6i32_nxv2i32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    uunpkhi z1.d, z0.s
+; CHECK-NEXT:    uzp1 z0.s, z0.s, z1.s
+; CHECK-NEXT:    ret
+  %v0 = call <vscale x 6 x i32> @llvm.experimental.vector.insert.nxv6i32.nxv2i32(<vscale x 6 x i32> undef, <vscale x 2 x i32> %sv0, i64 0)
+  %v1 = call <vscale x 6 x i32> @llvm.experimental.vector.insert.nxv6i32.nxv2i32(<vscale x 6 x i32> %v0, <vscale x 2 x i32> %sv0, i64 2)
+  ret <vscale x 6 x i32> %v0
+}
+
+;; This only works because the input vector is undef and index is zero
+define  <vscale x 6 x i32> @insert_nxv6i32_nxv3i32(<vscale x 3 x i32> %sv0) {
+; CHECK-LABEL: insert_nxv6i32_nxv3i32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ret
+  %v0 = call <vscale x 6 x i32> @llvm.experimental.vector.insert.nxv6i32.nxv3i32(<vscale x 6 x i32> undef, <vscale x 3 x i32> %sv0, i64 0)
+  ret <vscale x 6 x i32> %v0
+}
+
+define <vscale x 12 x i32> @insert_nxv12i32_nxv4i32(<vscale x 4 x i32> %sv0, <vscale x 4 x i32> %sv1, <vscale x 4 x i32> %sv2) {
+; CHECK-LABEL: insert_nxv12i32_nxv4i32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ret
+  %v0 = call <vscale x 12 x i32> @llvm.experimental.vector.insert.nxv4i32.nxv12i32(<vscale x 12 x i32> undef, <vscale x 4 x i32> %sv0, i64 0)
+  %v1 = call <vscale x 12 x i32> @llvm.experimental.vector.insert.nxv4i32.nxv12i32(<vscale x 12 x i32> %v0, <vscale x 4 x i32> %sv1, i64 4)
+  %v2 = call <vscale x 12 x i32> @llvm.experimental.vector.insert.nxv4i32.nxv12i32(<vscale x 12 x i32> %v1, <vscale x 4 x i32> %sv2, i64 8)
+  ret <vscale x 12 x i32> %v2
+}
+
+declare <vscale x 3 x i32> @llvm.experimental.vector.insert.nxv3i32.nxv2i32(<vscale x 3 x i32>, <vscale x 2 x i32>, i64)
+declare <vscale x 6 x i32> @llvm.experimental.vector.insert.nxv6i32.nxv2i32(<vscale x 6 x i32>, <vscale x 2 x i32>, i64)
+declare <vscale x 6 x i32> @llvm.experimental.vector.insert.nxv6i32.nxv3i32(<vscale x 6 x i32>, <vscale x 3 x i32>, i64)
+declare <vscale x 12 x i32> @llvm.experimental.vector.insert.nxv4i32.nxv12i32(<vscale x 12 x i32>, <vscale x 4 x i32>, i64)