DAGCombiner: Fold away vector extract of insert with the same index

[arsenm]

Summary

This only really matters when the index is non-constant since the constant case already gets taken care of by other combines.

[arsenm]

r274569

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
   // extract_vector_elt (v2i32 (bitcast i64:x)), 0 -> i32 (trunc i64:x)
   if (ConstEltNo && InVec.getOpcode() == ISD::BITCAST && InVec.hasOneUse() &&
       ConstEltNo->isNullValue() && VT.isInteger()) {
     SDValue BCSrc = InVec.getOperand(0);
     if (BCSrc.getValueType().isScalarInteger())
       return DAG.getNode(ISD::TRUNCATE, SDLoc(N), NVT, BCSrc);
   }
 
+  // extract_vector_elt (insert_vector_elt vec, val, idx), idx) -> val
+  //
+  // This only really matters if the index is non-constant since other combines
+  // on the constant elements already work.
+  if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT &&
+      EltNo == InVec.getOperand(2))
+    return InVec.getOperand(1);
+
   // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
   // We only perform this optimization before the op legalization phase because
   // we may introduce new vector instructions which are not backed by TD
   // patterns. For example on AVX, extracting elements from a wide vector
   // without using extract_subvector. However, if we can find an underlying
   // scalar value, then we can always use that.
   if (ConstEltNo && InVec.getOpcode() == ISD::VECTOR_SHUFFLE) {
     int NumElem = VT.getVectorNumElements();

test/CodeGen/AMDGPU/vector-extract-insert.ll

+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+
+; Test that when extracting the same unknown vector index from an
+; insertelement the dynamic indexing is folded away.
+
+declare i32 @llvm.amdgcn.workitem.id.x() #0
+
+; No dynamic indexing required
+; GCN-LABEL: {{^}}extract_insert_same_dynelt_v4i32:
+; GCN: s_load_dword [[VAL:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xd{{$}}
+; GCN-NOT buffer_load_dword
+; GCN-NOT: [[VAL]]
+; GCN: v_mov_b32_e32 [[VVAL:v[0-9]+]], [[VAL]]
+; GCN-NOT: [[VVAL]]
+; GCN: buffer_store_dword [[VVAL]]
+define void @extract_insert_same_dynelt_v4i32(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %val, i32 %idx) #1 {
+  %id = call i32 @llvm.amdgcn.workitem.id.x()
+  %id.ext = sext i32 %id to i64
+  %gep.in = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %in, i64 %id.ext
+  %gep.out = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %id.ext
+  %vec = load <4 x i32>, <4 x i32> addrspace(1)* %gep.in
+  %insert = insertelement <4 x i32> %vec, i32 %val, i32 %idx
+  %extract = extractelement <4 x i32> %insert, i32 %idx
+  store i32 %extract, i32 addrspace(1)* %gep.out
+  ret void
+}
+
+; GCN-LABEL: {{^}}extract_insert_different_dynelt_v4i32:
+; GCN: buffer_load_dwordx4
+; GCN: v_movreld_b32
+; GCN: v_movrels_b32
+; GCN: buffer_store_dword v
+define void @extract_insert_different_dynelt_v4i32(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %val, i32 %idx0, i32 %idx1) #1 {
+  %id = call i32 @llvm.amdgcn.workitem.id.x()
+  %id.ext = sext i32 %id to i64
+  %gep.in = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %in, i64 %id.ext
+  %gep.out = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %id.ext
+  %vec = load <4 x i32>, <4 x i32> addrspace(1)* %gep.in
+  %insert = insertelement <4 x i32> %vec, i32 %val, i32 %idx0
+  %extract = extractelement <4 x i32> %insert, i32 %idx1
+  store i32 %extract, i32 addrspace(1)* %gep.out
+  ret void
+}
+
+; GCN-LABEL: {{^}}extract_insert_same_elt2_v4i32:
+; GCN: s_load_dword [[VAL:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xd{{$}}
+; GCN-NOT buffer_load_dword
+; GCN-NOT: [[VAL]]
+; GCN: v_mov_b32_e32 [[VVAL:v[0-9]+]], [[VAL]]
+; GCN-NOT: [[VVAL]]
+; GCN: buffer_store_dword [[VVAL]]
+define void @extract_insert_same_elt2_v4i32(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %val, i32 %idx) #1 {
+  %id = call i32 @llvm.amdgcn.workitem.id.x()
+  %id.ext = sext i32 %id to i64
+  %gep.in = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %in, i64 %id.ext
+  %gep.out = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %id.ext
+  %vec = load <4 x i32>, <4 x i32> addrspace(1)* %gep.in
+  %insert = insertelement <4 x i32> %vec, i32 %val, i32 %idx
+  %extract = extractelement <4 x i32> %insert, i32 %idx
+  store i32 %extract, i32 addrspace(1)* %gep.out
+  ret void
+}
+
+
+attributes #0 = { nounwind readnone }
+attributes #1 = { nounwind }
 No newline at end of file