Set series of dependent operations to have the same name. This solution is going to be broken in some cases, but at least fixes #259.

Author: chriselrod

src/codegen/loopstartstopmanager.jl

@@ -425,7 +425,14 @@ function pushgespind!(
       if index_by_index
         if gespsymbol === Symbol("")
           if constoffset == 0
-            push!(gespinds.args, Expr(:call, GlobalRef(VectorizationBase, :NullStep)))
+            ns = Expr(:call, GlobalRef(VectorizationBase, :NullStep))
+            if fromgsp
+              loop = getloop(ls, ind)
+              if loop.rangesym ≢ Symbol("")
+                ns = Expr(:call, lv(:similardims), loop.rangesym, ns)
+              end
+            end
+            push!(gespinds.args, ns)
           else
             push!(gespinds.args, staticexpr(constoffset))
           end

src/condense_loopset.jl

@@ -279,7 +279,7 @@ val(x) = Expr(:call, Expr(:curly, :Val, x))
 @inline gespf1(x::StridedPointer{T,1}, i::Tuple{Zero}) where {T} = x
 @inline gespf1(x::StridedBitPointer{T,1}, i::Tuple{Zero}) where {T} = x
 @generated function gespf1(x::StridedPointer{T,N,C,B,R}, i::Tuple{I}) where {T,N,I<:Integer,C,B,R}
-  I === Zero && return :x
+  # I === Zero && return :x
   ri = 0; rm = typemax(Int)
   for (i, r) ∈ enumerate(R)
     if r < rm
@@ -295,6 +295,34 @@ val(x) = Expr(:call, Expr(:curly, :Val, x))
     StridedPointer{$T,1,$(C===1 ? 1 : 0),$(B===1 ? 1 : 0),$(R[ri],)}(ptr, (getfield(getfield(x,:strd), $ri, 1),), (Zero(),))
   end
 end
+@generated function gespf1(x::StridedPointer{T,N,C,B,R}, ::Tuple{VectorizationBase.NullStep}) where {T,N,C,B,R}
+  ri = 0; rm = typemax(Int)
+  for (i, r) ∈ enumerate(R)
+    if r < rm
+      rm = r
+      ri = i
+    end
+  end
+  ri = max(1, ri)
+  quote
+    $(Expr(:meta,:inline))
+    StridedPointer{$T,1,$(C===1 ? 1 : 0),$(B===1 ? 1 : 0),$(R[ri],)}(pointer(x), (getfield(getfield(x,:strd), $ri, 1),), (getfield(getfield(x,:offsets), $ri, 1),))
+  end
+end
+@generated function gespf1(x::StridedBitPointer{N,C,B,R}, ::Tuple{VectorizationBase.NullStep}) where {N,C,B,R}
+  ri = 0; rm = typemax(Int)
+  for (i, r) ∈ enumerate(R)
+    if r < rm
+      rm = r
+      ri = i
+    end
+  end
+  ri = max(1, ri)
+  quote
+    $(Expr(:meta,:inline))
+    StridedBitPointer{1,$(C===1 ? 1 : 0),$(B===1 ? 1 : 0),$(R[ri],)}(pointer(x), (getfield(getfield(x,:strd), $ri, 1),), (getfield(getfield(x,:offsets), $ri, 1),))
+  end
+end
 @generated function gespf1(x::StridedBitPointer{T,N,C,B,R}, i::Tuple{I}) where {T,N,I<:Integer,C,B,R}
   I === Zero && return :x
   quote

src/parse/add_compute.jl

@@ -154,21 +154,27 @@ function add_reduced_deps!(op::Operation, reduceddeps::Vector{Symbol})
 end
 
 function substitute_op_in_parents!(
-    vparents::Vector{Operation}, replacer::Operation, replacee::Operation, reduceddeps::Vector{Symbol}
+    vparents::Vector{Operation}, replacer::Operation, replacee::Operation, reduceddeps::Vector{Symbol}, reductsym::Symbol
 )
-    found = false
-    for i ∈ eachindex(vparents)
-        opp = vparents[i]
-        if opp === replacee
-            vparents[i] = replacer
-            found = true
-        else
-            fopp = substitute_op_in_parents!(parents(opp), replacer, replacee, reduceddeps)
-            fopp && add_reduced_deps!(opp, reduceddeps)
-            found |= fopp
-        end
+  found = false
+  for i ∈ eachindex(vparents)
+    opp = vparents[i]
+    if opp === replacee
+      vparents[i] = replacer
+      found = true
+    else
+      fopp = substitute_op_in_parents!(parents(opp), replacer, replacee, reduceddeps, reductsym)
+      if fopp
+        add_reduced_deps!(opp, reduceddeps)
+        # FIXME: https://github.com/JuliaSIMD/LoopVectorization.jl/issues/259
+        # 
+        opp.variable = reductsym
+        opp.mangledvariable = Symbol("##", reductsym, :_)
+      end
+      found |= fopp
     end
-    found
+  end
+  found
 end
 
 
@@ -216,7 +222,7 @@ function add_reduction_update_parent!(
             update_deps!(deps, reduceddeps, reductinit)#parent) # deps and reduced deps will not be disjoint
         end
     elseif !isouterreduction && reductinit !== parent
-        substitute_op_in_parents!(vparents, reductinit, parent, reduceddeps)
+        substitute_op_in_parents!(vparents, reductinit, parent, reduceddeps, reductsym)
     end
     update_reduction_status!(vparents, reduceddeps, name(reductinit))
     # this is the op added by add_compute

test/gemv.jl

@@ -227,6 +227,13 @@ using Test
         end
     end
 
+  function depchain_with_different_deps!(c1,c2,A,b)
+    @avx for j in axes(A,1), k in axes(A,2)
+      c1[j] += A[j,k] * b[k]
+      c2[j] += A[j,k] * b[k] - 0 * b[j]
+    end
+  end
+
     M, K, N = 51, 49, 61
     for T ∈ (Float32, Float64, Int32, Int64)
         @show T, @__LINE__
@@ -324,5 +331,9 @@ using Test
         @test Y0 ≈ Y1
         @test dY0 ≈ dY1
 
+        Y2 = zeros(TC, N, 2); Y3 = zeros(TC,N,2);
+        depchain_with_different_deps!(Y2, Y3, A, b)
+        @test view(Y2,:,1) == view(Y3,:,1)
+        @test view(Y2,:,1) ≈ A*b
     end
 end