For non loopedindex, check that parent ops match as symbols matching may not be enough

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.12.51"
+version = "0.12.52"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"

src/codegen/lower_memory_common.jl

@@ -1,8 +1,8 @@
 function parentind(ind::Symbol, op::Operation)
-    for (id,opp) ∈ enumerate(parents(op))
-        name(opp) === ind && return id
-    end
-    -1
+  for (id,opp) ∈ enumerate(parents(op))
+    name(opp) === ind && return id
+  end
+  -1
 end
 function symbolind(ind::Symbol, op::Operation, td::UnrollArgs, ls::LoopSet)
     id = parentind(ind, op)

src/modeling/operations.jl

@@ -227,23 +227,23 @@ isu₂unrolled(op::Operation) = op.u₂unrolled
 isvectorized(op::Operation) = op.vectorized
 
 function matches(op1::Operation, op2::Operation)
-    op1 === op2 && return true
-    op1.instruction === op2.instruction || return false
-    op1.node_type == op2.node_type || return false
-    if isconstant(op1)
-        return iszero(length(loopdependencies(op1))) && iszero(length(loopdependencies(op2))) && (mangledvar(op1) === mangledvar(op2))
-    end
-    op1.dependencies == op2.dependencies || return false
-    op2.reduced_deps == op2.reduced_deps || return false
-    if accesses_memory(op1)
-        op1.ref == op2.ref || return false
-    end
-    nparents = length(parents(op1))
-    nparents == length(parents(op2)) || return false
-    for p ∈ 1:nparents
-        matches(op1.parents[p], op2.parents[p]) || return false
-    end
-    true
+  op1 === op2 && return true
+  op1.instruction === op2.instruction || return false
+  op1.node_type == op2.node_type || return false
+  if isconstant(op1)
+    return iszero(length(loopdependencies(op1))) && iszero(length(loopdependencies(op2))) && (mangledvar(op1) === mangledvar(op2))
+  end
+  op1.dependencies == op2.dependencies || return false
+  op2.reduced_deps == op2.reduced_deps || return false
+  if accesses_memory(op1)
+    op1.ref == op2.ref || return false
+  end
+  nparents = length(parents(op1))
+  nparents == length(parents(op2)) || return false
+  for p ∈ 1:nparents
+    matches(op1.parents[p], op2.parents[p]) || return false
+  end
+  true
 end
 
  # negligible save on allocations for operations that don't need these (eg, constants).

src/parse/add_loads.jl

@@ -1,24 +1,43 @@
 function maybeaddref!(ls::LoopSet, op::Operation)
-    ref = op.ref
-    id = findfirst(==(ref), ls.refs_aliasing_syms)
-    # try to CSE
-    if id === nothing
-        push!(ls.syms_aliasing_refs, name(op))
-        push!(ls.refs_aliasing_syms, ref)
-        0
-    else
-        id
-    end
+  ref = op.ref
+  id = findfirst(==(ref), ls.refs_aliasing_syms)
+  # try to CSE
+  if id === nothing
+    push!(ls.syms_aliasing_refs, name(op))
+    push!(ls.refs_aliasing_syms, ref)
+    0
+  else
+    id
+  end
 end
 
 function add_load!(ls::LoopSet, op::Operation, actualarray::Bool = true)
-    @assert isload(op)
-    if (id = maybeaddref!(ls, op)) > 0 # try to CSE
-        opp = ls.opdict[ls.syms_aliasing_refs[id]] # throw an error if not found.
-        return isstore(opp) ? getop(ls, first(parents(opp))) : opp
+  @assert isload(op)
+  if (id = maybeaddref!(ls, op)) > 0 # try to CSE
+    opp = ls.opdict[ls.syms_aliasing_refs[id]] # throw an error if not found.
+    op_inds = getindicesonly(op)
+    li = op.ref.loopedindex
+    allmatch = true
+    parents_op = parents(op)
+    parents_opp = parents(opp)
+    for i ∈ eachindex(op_inds)
+      li[i] && continue
+      ind = op_inds[i]
+      if (id_op = parentind(ind, op)) > 0
+        if (id_opp = parentind(ind, opp)) > 0
+          if !matches(parents_op[id_op], parents_opp[id_opp])
+            allmatch = false
+            break
+          end
+        end
+      end
     end
-    add_vptr!(ls, op.ref.ref.array, vptr(op), actualarray)
-    pushop!(ls, op, name(op))
+    if allmatch
+      return isstore(opp) ? getop(ls, first(parents(opp))) : opp
+    end
+  end
+  add_vptr!(ls, op.ref.ref.array, vptr(op), actualarray)
+  pushop!(ls, op, name(op))
 end
 
 function add_load!(