Fix mask bug when iteration doesn't start at 1, resolves #52.

Author: chriselrod

src/add_loads.jl

@@ -77,5 +77,5 @@ struct LoopValue end
 @inline SIMDPirates.vload(::LoopValue, i::Tuple{_MM{W}}, ::Unsigned) where {W} = SVec(SIMDPirates.vrangeincr(Val{W}(), @inbounds(i[1].i), Val{1}()))
 @inline VectorizationBase.load(::LoopValue, i::Integer) = i + one(i)
 @inline VectorizationBase.load(::LoopValue, i::Tuple{I}) where {I<:Integer} = @inbounds(i[1]) + one(I)
-@inline Base.eltype(::LoopValue) = Int8
+@inline Base.eltype(::LoopValue) = Int
 

src/condense_loopset.jl

@@ -380,7 +380,11 @@ function setup_call(ls::LoopSet, inline = Int8(2), U = zero(Int8), T = zero(Int8
     # Creating an anonymous function and calling it also achieves the outlining, while still
     # inlining the generated function into the loop preamble.
     if inline == Int8(2)
-        setup_call_inline(ls, U, T)
+        if num_loops(ls) == 1
+            iszero(U) ? lower(ls) : lower(ls, U, -one(U))
+        else
+            setup_call_inline(ls, U, T)
+        end
     else
         setup_call_noinline(ls, U, T)
     end

src/constructors.jl

@@ -134,7 +134,8 @@ macro avx(arg, q)
     @assert q.head === :for
     @assert arg.head === :(=)
     inline, U, T = check_macro_kwarg(arg)
-    esc(setup_call(LoopSet(q, __module__), inline, U, T))
+    ls = LoopSet(q, __module__)
+    esc(setup_call(ls, inline, U, T))
 end
 macro avx(arg1, arg2, q)
     @assert q.head === :for

src/lowering.jl

@@ -226,13 +226,21 @@ function determine_eltype(ls::LoopSet)
     end
     promote_q
 end
-function determine_width(ls::LoopSet, typeT::Symbol, unrolled::Symbol)
-    unrolledloop = getloop(ls, unrolled)
-    if isstaticloop(unrolledloop)
-        Expr(:call, lv(:pick_vector_width_val), Expr(:call, Expr(:curly, :Val, length(unrolledloop))), typeT)
-    else
-        Expr(:call, lv(:pick_vector_width_val), typeT)
+function determine_width(ls::LoopSet, vectorized::Symbol)
+    vloop = getloop(ls, vectorized)
+    vwidth_q = Expr(:call, lv(:pick_vector_width_val))
+    if isstaticloop(vloop)
+        push!(vwidth_q.args, Expr(:call, Expr(:curly, :Val, length(vloop))))
     end
+    push!(vwidth_q.args, ls.T)
+    # if length(ls.includedactualarrays) < 2
+    #     push!(vwidth_q.args, ls.T)
+    # else
+    #     for array ∈ ls.includedactualarrays
+    #         push!(vwidth_q.args, Expr(:call, :eltype, array))
+    #     end
+    # end
+    vwidth_q
 end
 function lower_unrolled!(
     q::Expr, ls::LoopSet, vectorized::Symbol, U::Int, T::Int, W::Symbol, typeT::Symbol, unrolledloop::Loop
@@ -329,7 +337,7 @@ end
 function definemask(loop::Loop, W::Symbol, allon::Bool)
     if isstaticloop(loop)
         maskexpr(W, length(loop), allon)
-    elseif loop.starthint == 0
+    elseif loop.startexact && loop.starthint == 0
         maskexpr(W, loop.stopsym, allon)
     else
         lexpr = if loop.startexact
@@ -349,7 +357,7 @@ using SIMDPirates: sizeequivalentfloat, sizeequivalentint
 function setup_preamble!(ls::LoopSet, W::Symbol, typeT::Symbol, vectorized::Symbol, unrolled::Symbol, tiled::Symbol, U::Int)
     # println("Setup preamble")
     length(ls.includedarrays) == 0 || push!(ls.preamble.args, Expr(:(=), typeT, determine_eltype(ls)))
-    push!(ls.preamble.args, Expr(:(=), W, determine_width(ls, typeT, unrolled)))
+    push!(ls.preamble.args, Expr(:(=), W, determine_width(ls, vectorized)))
     lower_licm_constants!(ls)
     pushpreamble!(ls, definemask(getloop(ls, vectorized), W, U > 1 && unrolled === vectorized))
     for op ∈ operations(ls)

test/miscellaneous.jl

@@ -415,7 +415,21 @@
         end
         accu
     end
-    
+    function test_for_with_different_index!(c, a, b, start_sample, num_samples)
+        @inbounds for i = start_sample:num_samples + start_sample - 1
+            c[i] = b[i] * a[i]
+        end
+    end
+    function test_for_with_different_indexavx!(c, a, b, start_sample, num_samples)
+        @avx for i = start_sample:num_samples + start_sample - 1
+            c[i] = b[i] * a[i]
+        end
+    end
+    function test_for_with_different_index_avx!(c, a, b, start_sample, num_samples)
+        @_avx for i = start_sample:num_samples + start_sample - 1
+            c[i] = b[i] * a[i]
+        end
+    end
 
     for T ∈ (Float32, Float64)
         @show T, @__LINE__
@@ -467,7 +481,15 @@
         fill!(y2, NaN); clenshawavx!(y2,x,c)
         @test y1 ≈ y2
 
-
+        C = randn(T, 199, 498);
+        start_sample = 29; num_samples = 800;
+        test_for_with_different_index!(B1, A, C, start_sample, num_samples)
+        test_for_with_different_indexavx!(B2, A, C, start_sample, num_samples)
+        r = start_sample:start_sample+num_samples - 1
+        @test view(vec(B1), r) == view(vec(B2), r)
+        fill!(B2, NaN); test_for_with_different_index_avx!(B2, A, C, start_sample, num_samples)
+        @test view(vec(B1), r) == view(vec(B2), r)
+        
         ni, nj, nk = (127, 113, 13)
         x = rand(T, ni, nj, nk);
         q1 = similar(x);