Don't reject reorders based on compute without children, fixes #338.

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.12.75"
+version = "0.12.76"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"

src/codegen/lowering.jl

@@ -917,7 +917,6 @@ function lower(ls::LoopSet, inline::Int = -1)
 end
 function lower(ls::LoopSet, u₁::Int, u₂::Int, v::Int, inline::Int)
   fill_offset_memop_collection!(ls)
-  fill_children!(ls)
   if u₂ > 1
     @assert num_loops(ls) > 1 "There is only $(num_loops(ls)) loop, but specified blocking parameter u₂ is $u₂."
     order, u₁loop, u₂loop, vectorized, _u₁, _u₂, c, shouldinline = choose_tile(ls, store_load_deps(operations(ls)), v)

src/condense_loopset.jl

@@ -316,12 +316,12 @@ function skip_constant(instr::Instruction)
 end
 
 function add_reassigned_syms!(q::Expr, ls::LoopSet)
-    for op ∈ operations(ls)
-        if isconstant(op)
-            instr = instruction(op)
-            skip_constant(instr) || push!(q.args, instr.instr)
-        end
+  for op ∈ operations(ls)
+    if isconstant(op)
+      instr = instruction(op)
+      skip_constant(instr) || push!(q.args, instr.instr)
     end
+  end
 end
 function add_external_functions!(q::Expr, ls::LoopSet)
     for op ∈ operations(ls)
@@ -625,7 +625,7 @@ function generate_call_types(
   ops = operations(ls)
   for op ∈ ops
     instr::Instruction = instruction(op)
-    if (isconstant(op) && (instr == LOOPCONSTANT)) && (!roots[identifier(op)])
+    if ((isconstant(op) && (instr == LOOPCONSTANT)) && (!roots[identifier(op)]))
       instr = op.instruction = DROPPEDCONSTANT
     end
     push!(operation_descriptions.args, QuoteNode(instr.mod))

src/modeling/costs.jl

@@ -267,6 +267,7 @@ const COST = Dict{Symbol,InstructionCost}(
     :prefetch1 => InstructionCost(0,0.0,0.0,0),
     :prefetch2 => InstructionCost(0,0.0,0.0,0),
   :convert => InstructionCost(4,0.5),
+  :oftype => InstructionCost(4,0.5),
   :vpermilps177 => InstructionCost(1, 1.0),
   :vmovsldup => InstructionCost(1, 1.0),
   :vmovshdup => InstructionCost(1, 1.0),

src/modeling/graphs.jl

@@ -1306,7 +1306,7 @@ end
 
 function accept_reorder_according_to_tracked_reductions(ls::LoopSet, reordered::Symbol)
   for op ∈ operations(ls)
-    if reordered ∈ loopdependencies(op)
+    if (reordered ∈ loopdependencies(op)) && !(iscompute(op) & iszero(length(children(op))))
       for opp ∈ parents(op)
         (iscompute(opp) && isanouterreduction(ls, opp)) && return 0x00
       end
@@ -1367,6 +1367,7 @@ end
 
 offsetloadcollection(ls::LoopSet) = ls.omop
 function fill_offset_memop_collection!(ls::LoopSet)
+  fill_children!(ls)
   omop = offsetloadcollection(ls)
   ops = operations(ls)
   num_ops = length(ops)