init sym for CartesianIndices{0} loops correctly more often

Author: chriselrod

src/transforms.jl

@@ -40,6 +40,55 @@ function return_empty_reductinit(op::Operation, var::Symbol)
 end
 
 
+
+function constant_symbol!(ls::LoopSet, op::Operation)
+  # hack
+  # relowers, but should make it work
+  # TODO: DRY with `lower_licm_constants!` from `src/codegen/lower_constants.jl`
+  skip_constant(instruction(op)) || return instruction(op).instr
+  idcheck = identifier(op)
+  symname = constantopname(op)
+  for (id, sym) ∈ ls.preamble_symsym
+    (idcheck ≢ nothing) && ((idcheck == id) && continue)
+    pushpreamble!(ls, Expr(:(=), symname, sym))
+    return symname
+    # setconstantop!(ls, op,  sym)
+    # setconstantop!(ls, op, Expr(:call, lv(:maybeconvert), ls.T, sym))
+  end
+  for (id,(intval,intsz,signed)) ∈ ls.preamble_symint
+    (idcheck ≢ nothing) && ((idcheck == id) && continue)
+    if intsz == 1
+      pushpreamble!(ls, Expr(:(=), symname, intval % Bool))
+    else
+      pushpreamble!(ls, Expr(:(=), symname, sizeequivalent_symint_expr(intval, signed)))
+    end
+    return symname
+  end
+  for (id,floatval) ∈ ls.preamble_symfloat
+    (idcheck ≢ nothing) && ((idcheck == id) && continue)
+    pushpreamble!(ls, Expr(:(=), symname, Expr(:call, lv(:sizeequivalentfloat), ELTYPESYMBOL, floatval)))
+    return symname
+  end
+  for (id,typ) ∈ ls.preamble_zeros
+    (idcheck ≢ nothing) && ((idcheck == id) && continue)
+    instruction(op) === LOOPCONSTANT || continue
+    if typ == IntOrFloat
+      pushpreamble!(ls, Expr(:(=), symname, Expr(:call, :zero, ELTYPESYMBOL)))
+    elseif typ == HardInt
+      pushpreamble!(ls, Expr(:(=), symname, Expr(:call, lv(:zerointeger), ELTYPESYMBOL)))
+    else#if typ == HardFloat
+      pushpreamble!(ls, Expr(:(=), symname, Expr(:call, lv(:zerofloat), ELTYPESYMBOL)))
+    end
+    return symname
+  end
+  for (id,f) ∈ ls.preamble_funcofeltypes
+    (idcheck ≢ nothing) && ((idcheck == id) && continue)
+    pushpreamble!(ls, Expr(:(=), symname, Expr(:call, reduction_zero(f), ELTYPESYMBOL)))
+    return symname
+  end
+  throw("Constant operation symbol not found.")
+end
+
 function hoist_constant_store!(q::Expr, ls::LoopSet, op::Operation)
   op.instruction = DROPPEDCONSTANT
   op.node_type = constant
@@ -52,7 +101,10 @@ function hoist_constant_store!(q::Expr, ls::LoopSet, op::Operation)
   end
   push!(ls.outer_reductions, identifier(opr))
 
-  init = return_empty_reductinit(opr, name(opr)).instruction.instr
+  initop = return_empty_reductinit(opr, name(opr))
+  # @show last(ls.preamble.args)
+  init = constant_symbol!(ls, initop)
+  # @show last(ls.preamble.args)
   pushpreamble!(ls, Expr(:(=), outer_reduct_init_typename(opr), Expr(:call, lv(:typeof), init)))
   qpre = Expr(:block)
   push!(q.args, Expr(:call, lv(:unsafe_store!), Expr(:call, lv(:pointer), op.ref.ptr), outer_reduction_to_scalar_reduceq!(qpre, opr, init)))

test/miscellaneous.jl

@@ -1182,6 +1182,22 @@ end
         end
         out
     end
+    function smoothdim_kernel_tile_avx_2!(out, z, src::AbstractArray, kernel::AbstractVector, Rpre::CartesianIndices, axout_tile, Rpost::CartesianIndices)
+        axkernel = axes(kernel, 1)
+        for Ipost in Rpost
+            for i in axout_tile
+                @turbo for Ipre in Rpre
+                    tmp = zero(eltype(out))
+                    # tmp = convert(eltype(out), z)    # failing to hoist this leads to an "UndefVarError: tmp not defined"
+                    for j in axkernel
+                        tmp += oftype(z, src[Ipre,i+j,Ipost])*kernel[j]
+                    end
+                    out[Ipre,i,Ipost] = tmp
+                end
+            end
+        end
+        out
+    end
 
 
     for T ∈ (UInt8,Float32, Float64)
@@ -1229,6 +1245,8 @@ end
                 smoothdim_kernel_tile!(    dest1, float(zero(T)), src, kernel, Rpre, axes(dest1, d), Rpost);
                 smoothdim_kernel_tile_avx!(dest2, float(zero(T)), src, kernel, Rpre, axes(dest2, d), Rpost);
                 @test dest1 ≈ dest2
+                fill!(dest2,NaN); smoothdim_kernel_tile_avx_2!(dest2, float(zero(T)), src, kernel, Rpre, axes(dest2, d), Rpost);
+                @test dest1 ≈ dest2
             end
         end
     end