inline broadcast

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.12.7"
+version = "0.12.8"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"

src/broadcast.jl

@@ -409,7 +409,7 @@ end
     doaddref!(ls, storeop)
     resize!(ls.loop_order, num_loops(ls)) # num_loops may be greater than N, eg Product
     # return ls
-    Expr(:block, setup_call(ls, :(Base.Broadcast.materialize!(dest, bc)), LineNumberNode(0), inline, false, u₁, u₂, threads%Int), :dest)
+    Expr(:block, Expr(:meta,:inline), setup_call(ls, :(Base.Broadcast.materialize!(dest, bc)), LineNumberNode(0), inline, false, u₁, u₂, threads%Int), :dest)
 end
 @generated function vmaterialize!(
     dest′::Union{Adjoint{T,A},Transpose{T,A}}, bc::BC, ::Val{Mod}, ::Val{UNROLL}
@@ -428,7 +428,7 @@ end
     storeop = add_simple_store!(ls, :dest, ArrayReference(:dest, reverse(loopsyms)), elementbytes)
     doaddref!(ls, storeop)
     resize!(ls.loop_order, num_loops(ls)) # num_loops may be greater than N, eg Product
-    Expr(:block, setup_call(ls, :(Base.Broadcast.materialize!(dest′, bc)), LineNumberNode(0), inline, false, u₁, u₂, threads%Int), :dest′)
+    Expr(:block, Expr(:meta,:inline), setup_call(ls, :(Base.Broadcast.materialize!(dest′, bc)), LineNumberNode(0), inline, false, u₁, u₂, threads%Int), :dest′)
 end
 # these are marked `@inline` so the `@avx` itself can choose whether or not to inline.
 @generated function vmaterialize!(

src/constructors.jl

@@ -105,16 +105,21 @@ function check_macro_kwarg(arg, inline::Bool, check_empty::Bool, u₁::Int8, u
     end
     inline, check_empty, u₁, u₂, threads
 end
-function avx_macro(mod, src, q, args...)
-    q = macroexpand(mod, q)
-    inline = false; check_empty = false; u₁ = zero(Int8); u₂ = zero(Int8); threads = 1;
+function process_args(args; inline = false, check_empty = false, u₁ = zero(Int8), u₂ = zero(Int8), threads = 1)
     for arg ∈ args
         inline, check_empty, u₁, u₂, threads = check_macro_kwarg(arg, inline, check_empty, u₁, u₂, threads)
     end
+    inline, check_empty, u₁, u₂, threads
+end
+function avx_macro(mod, src, q, args...)
+    q = macroexpand(mod, q)
+    
     if q.head === :for
         ls = LoopSet(q, mod)
+        inline, check_empty, u₁, u₂, threads = process_args(args)
         esc(setup_call(ls, q, src, inline, check_empty, u₁, u₂, threads))
     else
+        inline, check_empty, u₁, u₂, threads = process_args(args, inline=true)
         substitute_broadcast(q, Symbol(mod), inline, u₁, u₂, threads)
     end
 end