Remove broken register passing optimization.

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <elrodc@gmail.com>"]
-version = "0.12.168"
+version = "0.12.169"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"

src/codegen/lower_threads.jl

@@ -495,7 +495,7 @@ function thread_one_loops_expr(
     $AM,
     $LPSYM,
     Val(typeof(var"#avx#call#args#")),
-    flatten_to_tuple(var"#avx#call#args#")...
+    var"#avx#call#args#"...
   ))
   update_return_values = if length(ls.outer_reductions) > 0
     retv = loopset_return_value(ls, Val(false))
@@ -555,7 +555,7 @@ function thread_one_loops_expr(
               $AM,
               $LPSYM,
               StaticType{typeof(var"##lbvargs#to_launch##")}(),
-              flatten_to_tuple(var"##lbvargs#to_launch##"),
+              var"##lbvargs#to_launch##",
               var"#thread#id#"
             )
 
@@ -744,7 +744,7 @@ function thread_two_loops_expr(
     $AM,
     $LPSYM,
     Val(typeof(var"#avx#call#args#")),
-    flatten_to_tuple(var"#avx#call#args#")...
+    var"#avx#call#args#"...
   ))
   update_return_values = if length(ls.outer_reductions) > 0
     retv = loopset_return_value(ls, Val(false))
@@ -867,7 +867,7 @@ function thread_two_loops_expr(
               $AM,
               $LPSYM,
               StaticType{typeof(var"##lbvargs#to_launch##")}(),
-              flatten_to_tuple(var"##lbvargs#to_launch##"),
+              var"##lbvargs#to_launch##",
               var"#thread#id#"
             )
             var"#thread#mask#" >>>= var"#trailzing#zeros#"

src/condense_loopset.jl

@@ -4,71 +4,6 @@
 Base.:|(u::Unsigned, it::IndexType) = u | UInt8(it)
 Base.:(==)(u::Unsigned, it::IndexType) = (u % UInt8) == UInt8(it)
 
-function _append_fields!(t::Expr, body::Expr, sym::Symbol, ::Type{T}) where {T}
-  for f ∈ 1:fieldcount(T)
-    TF = fieldtype(T, f)
-    Base.issingletontype(TF) && continue
-    gfcall = Expr(:call, getfield, sym, f)
-    if fieldcount(TF) ≡ 0
-      push!(t.args, gfcall)
-    elseif TF <: DataType
-      push!(t.args, Expr(:call, Expr(:curly, lv(:StaticType), gfcall)))
-    else
-      newsym = gensym(sym)
-      push!(body.args, Expr(:(=), newsym, gfcall))
-      _append_fields!(t, body, newsym, TF)
-    end
-  end
-  return nothing
-end
-@generated function flatten_to_tuple(r::T) where {T}
-  body = Expr(:block, Expr(:meta, :inline))
-  t = Expr(:tuple)
-  if Base.issingletontype(T)
-    nothing
-  elseif fieldcount(T) ≡ 0
-    push!(t.args, :r)
-  elseif T <: DataType
-    push!(t.args, Expr(:call, Expr(:curly, lv(:StaticType), :r)))
-  else
-    _append_fields!(t, body, :r, T)
-  end
-  push!(body.args, t)
-  body
-end
-function rebuild_fields(offset::Int, ::Type{T}) where {T}
-  call = (T <: Tuple) ? Expr(:tuple) : Expr(:new, T)
-  for f ∈ 1:fieldcount(T)
-    TF = fieldtype(T, f)
-    if Base.issingletontype(TF)
-      push!(call.args, TF.instance)
-    elseif fieldcount(TF) ≡ 0
-      push!(call.args, Expr(:call, getfield, :t, (offset += 1)))
-    elseif TF <: DataType
-      push!(
-        call.args,
-        Expr(:call, lv(:gettype), Expr(:call, getfield, :t, (offset += 1)))
-      )
-    else
-      arg, offset = rebuild_fields(offset, TF)
-      push!(call.args, arg)
-    end
-  end
-  return call, offset
-end
-@generated function reassemble_tuple(::Type{T}, t::Tuple) where {T}
-  if Base.issingletontype(T)
-    return T.instance
-  elseif fieldcount(T) ≡ 0
-    call = Expr(:call, getfield, :t, 1)
-  elseif T <: DataType
-    call = Expr(:call, lv(:gettype), Expr(:call, getfield, :t, 1))
-  else
-    call, _ = rebuild_fields(0, T)
-  end
-  Expr(:block, Expr(:meta, :inline), call)
-end
-
 """
     ArrayRefStruct
 
@@ -893,9 +828,9 @@ function generate_call_types(
       )
     )
     if manyarg
-      push!(q.args, Expr(:call, lv(:flatten_to_tuple), vargsym))
+      push!(q.args, vargsym)
     else
-      push!(q.args, Expr(:(...), Expr(:call, lv(:flatten_to_tuple), vargsym)))
+      push!(q.args, Expr(:(...), vargsym))
     end
     Expr(:block, Expr(:(=), vargsym, Expr(:tuple, lbarg, extra_args)))
   end

src/reconstruct_loopset.jl

@@ -1111,7 +1111,7 @@ Execute an `@turbo` block. The block's code is represented via the arguments:
   ::Val{var"#AM#"},
   ::Val{var"#LPSYM#"},
   ::Val{Tuple{var"#LB#",var"#V#"}},
-  var"#flattened#var#arguments#"::Vararg{Any,var"#num#vargs#"}
+  var"#lv#tuple#args#"::Vararg{Any,var"#num#vargs#"}
 ) where {
   var"#UNROLL#",
   var"#OPS#",
@@ -1132,15 +1132,6 @@ Execute an `@turbo` block. The block's code is represented via the arguments:
     var"#V#".parameters,
     var"#UNROLL#"
   )
-  pushfirst!(
-    ls.preamble.args,
-    :(
-      var"#lv#tuple#args#" = reassemble_tuple(
-        Tuple{var"#LB#",var"#V#"},
-        var"#flattened#var#arguments#"
-      )
-    )
-  )
   post = hoist_constant_memory_accesses!(ls)
   # q = @show(avx_body(ls, var"#UNROLL#")); post === ls.preamble ? q : Expr(:block, q, post)
   q = if (var"#UNROLL#"[10] > 1) && length(var"#LPSYM#") == length(ls.loops)
@@ -1171,7 +1162,7 @@ end
   ::Val{var"#AM#"},
   ::Val{var"#LPSYM#"},
   ::Val{Tuple{var"#LB#",var"#V#"}},
-  var"#flattened#var#arguments#"::Tuple{Vararg{Any,var"#num#vargs#"}}
+  var"#lv#tuple#args#"::Tuple{Vararg{Any,var"#num#vargs#"}}
 ) where {
   var"#UNROLL#",
   var"#OPS#",
@@ -1192,15 +1183,6 @@ end
     var"#V#".parameters,
     var"#UNROLL#"
   )
-  pushfirst!(
-    ls.preamble.args,
-    :(
-      var"#lv#tuple#args#" = reassemble_tuple(
-        Tuple{var"#LB#",var"#V#"},
-        var"#flattened#var#arguments#"
-      )
-    )
-  )
   post = hoist_constant_memory_accesses!(ls)
   # q = @show(avx_body(ls, var"#UNROLL#")); post === ls.preamble ? q : Expr(:block, q, post)
   q = if (var"#UNROLL#"[10] > 1) && length(var"#LPSYM#") == length(ls.loops)