[DAG] foldCONCAT_VECTORS - fold concat_vectors(v1xX insertelt(v,e,0), ...) -> build_vector(e,...) (#163420)

Extend the foldCONCAT_VECTORS BUILD_VECTOR construction to handle cases where the scalars have come from <1 x X> vector insertions

Fixes #163023

Author: RKSimon

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

@@ -6405,8 +6405,9 @@ static SDValue foldCONCAT_VECTORS(const SDLoc &DL, EVT VT,
   if (VT.isScalableVector())
     return SDValue();
 
-  // A CONCAT_VECTOR with all UNDEF/BUILD_VECTOR operands can be
-  // simplified to one big BUILD_VECTOR.
+  // A CONCAT_VECTOR of scalar sources, such as UNDEF, BUILD_VECTOR and
+  // single-element INSERT_VECTOR_ELT operands can be simplified to one big
+  // BUILD_VECTOR.
   // FIXME: Add support for SCALAR_TO_VECTOR as well.
   EVT SVT = VT.getScalarType();
   SmallVector<SDValue, 16> Elts;
@@ -6416,6 +6417,10 @@ static SDValue foldCONCAT_VECTORS(const SDLoc &DL, EVT VT,
       Elts.append(OpVT.getVectorNumElements(), DAG.getUNDEF(SVT));
     else if (Op.getOpcode() == ISD::BUILD_VECTOR)
       Elts.append(Op->op_begin(), Op->op_end());
+    else if (Op.getOpcode() == ISD::INSERT_VECTOR_ELT &&
+             OpVT.getVectorNumElements() == 1 &&
+             isNullConstant(Op.getOperand(2)))
+      Elts.push_back(Op.getOperand(1));
     else
       return SDValue();
   }

llvm/test/CodeGen/X86/masked_gather_scatter.ll

@@ -4768,16 +4768,10 @@ declare void @llvm.masked.scatter.v8f32.v8p0(<8 x float>, <8 x ptr>, i32 immarg,
 define <16 x i32> @pr163023(ptr %a0, <16 x i32> %a1) {
 ; X64-LABEL: pr163023:
 ; X64:       # %bb.0:
-; X64-NEXT:    vpmovsxdq %ymm0, %zmm1
-; X64-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
-; X64-NEXT:    vpmovsxdq %ymm0, %zmm0
 ; X64-NEXT:    kxnorw %k0, %k0, %k1
-; X64-NEXT:    vpxor %xmm2, %xmm2, %xmm2
-; X64-NEXT:    vpxor %xmm3, %xmm3, %xmm3
-; X64-NEXT:    kxnorw %k0, %k0, %k2
-; X64-NEXT:    vpgatherqd (%rdi,%zmm0), %ymm3 {%k2}
-; X64-NEXT:    vpgatherqd (%rdi,%zmm1), %ymm2 {%k1}
-; X64-NEXT:    vinserti64x4 $1, %ymm3, %zmm2, %zmm0
+; X64-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; X64-NEXT:    vpgatherdd (%rdi,%zmm0), %zmm1 {%k1}
+; X64-NEXT:    vmovdqa64 %zmm1, %zmm0
 ; X64-NEXT:    retq
 ;
 ; X86-LABEL: pr163023: