Merge branch 'master' into patch-1

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.12.10"
+version = "0.12.11"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"
@@ -24,11 +24,11 @@ DocStringExtensions = "0.8"
 IfElse = "0.1"
 OffsetArrays = "1.4.1, 1.5"
 Requires = "1"
-SLEEFPirates = "0.6.12"
+SLEEFPirates = "0.6.14"
 Static = "0.2"
 ThreadingUtilities = "0.4.1"
 UnPack = "1"
-VectorizationBase = "0.19.19"
+VectorizationBase = "0.19.22"
 julia = "1.5"
 
 [extras]

src/codegen/lower_load.jl

@@ -81,6 +81,7 @@ function add_prefetches!(q::Expr, ls::LoopSet, op::Operation, td::UnrollArgs, pr
         (ind == u₁loopsym) && (i = j)
     end
     push!(q.args, Expr(:call, lv(:prefetch0), gptr, copy(inds)))
+    i == 0 && return
     for u ∈ 1:u₁-1
         # for u ∈ umin:min(umin,U-1)
         # (u₁loopsym === vloopsym && !iszero(u & dontskip)) && continue

src/condense_loopset.jl

@@ -394,20 +394,10 @@ end
 Returns true if the element type is supported.
 """
 @inline check_type(::Type{T}) where {T <: NativeTypes} = true
-function check_type(::Type{T}) where {T}
-    # @info """`LoopVectorization.check_type` returned `false`, because `LoopVectorization.check_type(::$(T)) == false`.
-    #     `LoopVectorization` currently only supports `T <: $(NativeTypes)`.
-    #     Therefore compiling a probably slow `@inbounds @fastmath` fallback loop.""" maxlog=1
-    false
-end
+@inline check_type(::Type{T}) where {T} = false
 @inline check_device(::ArrayInterface.CPUPointer) = true
 @inline check_device(::ArrayInterface.CPUTuple) = true
-function check_device(x)
-    # @info """`LoopVectorization.check_args` returned `false`, because `ArrayInterface.device(::$(typeof(x))) == $x`
-    #     `LoopVectorization` normally requires `ArrayInterface.CPUPointer` (exceptions include ranges, `BitVector`s, and
-    #     `BitArray`s whose number of rows is a multiple of 8). Therefore compiling a probably slow `@inbounds @fastmath` fallback loop.""" maxlog=1
-    false
-end
+@inline check_device(x) = false
 
 function check_args_call(ls::LoopSet)
     q = Expr(:call, lv(:check_args))

src/reconstruct_loopset.jl

@@ -623,7 +623,7 @@ Execute an `@avx` block. The block's code is represented via the arguments:
 @generated function _avx_!(
     ::Val{UNROLL}, ::Val{OPS}, ::Val{ARF}, ::Val{AM}, ::Val{LPSYM}, var"#lv#tuple#args#"::Tuple{LB,V}
 ) where {UNROLL, OPS, ARF, AM, LPSYM, LB, V}
-    # 1 + 1 # Irrelevant line you can comment out/in to force recompilation...
+    1 + 1 # Irrelevant line you can comment out/in to force recompilation...
     ls = _avx_loopset(OPS, ARF, AM, LPSYM, LB.parameters, V.parameters, UNROLL)
     # return @show avx_body(ls, UNROLL)
     if last(UNROLL) > 1