Set initial pointer value for correct comparisons when they may be used to determine loop bounds. Fixes #290

Author: chriselrod

src/codegen/loopstartstopmanager.jl

@@ -428,7 +428,12 @@ function pushgespind!(
       if index_by_index
         if gespsymbol === Symbol("")
           if constoffset == 0
-            ns = Expr(:call, GlobalRef(VectorizationBase, :NullStep))
+            nostep = if fromgsp | (!index_by_index)
+              GlobalRef(VectorizationBase, :NullStep)
+            else
+              GlobalRef(ArrayInterface.Static, :Zero)
+            end
+            ns = Expr(:call, nostep)
             if fromgsp
               loop = getloop(ls, ind)
               if loop.rangesym ≢ Symbol("")

src/precompile.jl

@@ -1,7 +1,10 @@
 function _precompile_()
     ccall(:jl_generating_output, Cint, ()) == 1 || return nothing
     # Base.precompile(Tuple{typeof(which(_turbo_!,(Val{UNROLL},Val{OPS},Val{ARF},Val{AM},Val{LPSYM},Tuple{LB, V},)).generator.gen),Any,Any,Any,Any,Any,Any,Any,Any,Type,Type,Type,Type,Any,Any})   # time: 1.0198073
-    # Base.precompile(Tuple{typeof(gespf1),Any,Tuple{Any, VectorizationBase.NullStep}})   # time: 0.1096832
+  # Base.precompile(Tuple{typeof(gespf1),Any,Tuple{Any, VectorizationBase.NullStep}})   # time: 0.1096832
+    Base.precompile(avx_body, (LoopSet, Tuple{Bool,Int8,Int8,Bool,Int,Int,Int,Int,Int,Int,Int,UInt}))
+    Base.precompile(lower_and_split_loops, (LoopSet, Int))
+    Base.precompile(lower, (LoopSet, Int, Int, Int))
     Base.precompile(Tuple{typeof(turbo_macro),Module,LineNumberNode,Expr})   # time: 0.09183489
     Base.precompile(Tuple{typeof(gespf1),StridedPointer{Float64, 1, 1, 0, (1,), Tuple{StaticInt{8}}, Tuple{StaticInt{1}}},Tuple{StaticInt{1}}})   # time: 0.05469272
     Base.precompile(Tuple{typeof(zerorangestart),UnitRange{Int}})   # time: 0.04291692

test/iteration_bound_tests.jl

@@ -1,3 +1,4 @@
+using OffsetArrays
 
 @testset "Iteration Bound Tests" begin
   function masktest_incr1_none1start!(y,x)
@@ -19,4 +20,56 @@
   @test y == x
   masktest_incr2_none1start!(y,x)
   @test y == x .+ ifelse.((axes(x,1) .≤ 20) .& iseven.(axes(x,1)), 2, 0)
+
+
+  # issue #290
+  function my_gemm_noturbo!(out, s::Matrix{UInt8}, V, srows, scols, Vcols, μ)
+    k = srows >> 2
+    rem = srows & 3
+    fill!(out, 0)
+
+    for c in 1:Vcols
+      for i in 1:scols
+        for l in 1:k
+          block = s[l, i]
+          for p in 1:4
+            Aij = (block >> (2 * (p - 1))) & 3
+            out[i, c] += (((Aij >= 2) + (Aij == 3) + (Aij == 1) * μ[i]) *
+                          V[4 * (l - 1) + p, c])
+          end
+        end
+      end
+    end
+    nothing
+  end
+  function my_gemm!(out, s::Matrix{UInt8}, V, srows, scols, Vcols, μ)
+    k = srows >> 2
+    rem = srows & 3
+    fill!(out, 0)
+
+    @avx for c in 1:Vcols
+      for i in 1:scols
+        for l in 1:k
+          block = s[l, i]
+          for p in 1:4
+            Aij = (block >> (2 * (p - 1))) & 3
+            out[i, c] += (((Aij >= 2) + (Aij == 3) + (Aij == 1) * μ[i]) *
+                          V[4 * (l - 1) + p, c])
+          end
+        end
+      end
+    end
+    nothing
+  end
+
+  out_true = Matrix{Float64}(undef, 100, 100);
+  out_test1 = similar(out_true);
+  # out_test2 = zeros(100, 100)
+  μ = rand(100);
+  s = rand(UInt8, 100, 100);
+  V = rand(400, 100);
+
+  my_gemm_noturbo!(out_true, s, V, 400, 100, 100, μ)
+  my_gemm!(out_test1, s, V, 400, 100, 100, μ);
+  @test out_true ≈ out_test1
 end