[RISCV] Add lowerINSERT_SUBVECTORAsBUILD_VECTOR.

[HanKuanChen]

No description provided.

[llvmbot]

@llvm/pr-subscribers-backend-risc-v

Author: Han-Kuan Chen (HanKuanChen)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/110069.diff

2 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVISelLowering.cpp (+32)
• (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-insert-subvector.ll (+23)

diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index bf822eb2c6eeb5..0fe38839b7759a 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -10140,6 +10140,35 @@ SDValue RISCVTargetLowering::lowerVPREDUCE(SDValue Op,
       DAG.getConstantFP(APFloat::getNaN(ResVT.getFltSemantics()), DL, ResVT));
 }
 
+// Merge BUILD_VECTOR and insert_subvector into BUILD_VECTOR.
+// e.g.,
+// t7: v8i32 = BUILD_VECTOR 0, 0, 0, 0, 0, 0, 0, 0
+// t9: v16i32 = insert_subvector undef:v16i32, t7, i64<8>
+// ->
+// BUILD_VECTOR undef, undef, undef, undef, undef, undef, undef, undef, 0, 0, 0,
+// 0, 0, 0, 0, 0
+static SDValue lowerINSERT_SUBVECTORAsBUILD_VECTOR(SDValue Op,
+                                                   SelectionDAG &DAG) {
+  SDValue SubVec = Op.getOperand(1);
+  if (!Op.getOperand(0).isUndef())
+    return SDValue();
+  if (SubVec.getOpcode() != ISD::BUILD_VECTOR)
+    return SDValue();
+  SDLoc DL(Op);
+  MVT VecVT = Op.getSimpleValueType();
+  uint64_t InsertIndex = Op.getConstantOperandVal(2);
+  MVT SubVecVT = SubVec.getSimpleValueType();
+  MVT SubVecElementVT = SubVec.getOperand(0).getSimpleValueType();
+  unsigned SubVecNumElements = SubVecVT.getVectorNumElements();
+  SmallVector<SDValue> NewOps(VecVT.getVectorNumElements());
+  for (unsigned I = 0, E = VecVT.getVectorNumElements(); I != E; ++I)
+    if (I < InsertIndex || InsertIndex + SubVecNumElements <= I)
+      NewOps[I] = DAG.getUNDEF(SubVecElementVT);
+    else
+      NewOps[I] = SubVec.getOperand(I - InsertIndex);
+  return DAG.getNode(ISD::BUILD_VECTOR, DL, Op.getSimpleValueType(), NewOps);
+}
+
 SDValue RISCVTargetLowering::lowerINSERT_SUBVECTOR(SDValue Op,
                                                    SelectionDAG &DAG) const {
   SDValue Vec = Op.getOperand(0);
@@ -10191,6 +10220,9 @@ SDValue RISCVTargetLowering::lowerINSERT_SUBVECTOR(SDValue Op,
     }
   }
 
+  if (SDValue V = lowerINSERT_SUBVECTORAsBUILD_VECTOR(Op, DAG))
+    return V;
+
   // If the subvector vector is a fixed-length type and we don't know VLEN
   // exactly, we cannot use subregister manipulation to simplify the codegen; we
   // don't know which register of a LMUL group contains the specific subvector
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-insert-subvector.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-insert-subvector.ll
index e81f686a283032..347e7eeb6e8906 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-insert-subvector.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-insert-subvector.ll
@@ -894,3 +894,26 @@ define <4 x i32> @insert_extract_v8i32_v2i32_0(<2 x i32> %v) {
   %2 = call <4 x i32> @llvm.vector.extract.v4i32.v8i32(<8 x i32> %1, i64 0)
   ret <4 x i32> %2
 }
+
+define <16 x i32> @build_vector_insert(<16 x i32> %0) {
+; VLA-LABEL: build_vector_insert:
+; VLA:       # %bb.0: # %entry
+; VLA-NEXT:    vsetivli zero, 16, e32, m4, ta, ma
+; VLA-NEXT:    vmv.v.i v12, 0
+; VLA-NEXT:    vsetivli zero, 8, e32, m4, tu, ma
+; VLA-NEXT:    vmv.v.v v12, v8
+; VLA-NEXT:    vmv4r.v v8, v12
+; VLA-NEXT:    ret
+;
+; VLS-LABEL: build_vector_insert:
+; VLS:       # %bb.0: # %entry
+; VLS-NEXT:    vsetivli zero, 16, e32, m4, ta, ma
+; VLS-NEXT:    vmv.v.i v12, 0
+; VLS-NEXT:    vmv1r.v v10, v14
+; VLS-NEXT:    vmv1r.v v11, v15
+; VLS-NEXT:    ret
+entry:
+  %1 = call <16 x i32> @llvm.vector.insert.v16f32.v8f32(<16 x i32> poison, <8 x i32> zeroinitializer, i64 8)
+  %2 = shufflevector <16 x i32> %0, <16 x i32> %1, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
+  ret <16 x i32> %2
+}

[topperc]

Should this be a DAG combine? Typically lowering doesn't look at multiple nodes.

[topperc]

InstCombine already flattens this. Do you have other examples?

define <16 x i32> @build_vector_insert(<16 x i32> %0) {
entry:
  %1 = shufflevector <16 x i32> %0, <16 x i32> <i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
  ret <16 x i32> %1
}

[HanKuanChen]

Close it since instcombine can solve this.