remove aggressive const prop (was added as an attempt to fix non-specialization on functions)

Author: chriselrod

src/LoopVectorization.jl

@@ -112,6 +112,8 @@ LoopVectorization
 include("precompile.jl")
 _precompile_()
 
+_vreduce(+, Float64[1.0])
+
 # import ChainRulesCore, ForwardDiff
 # include("vmap_grad.jl")
 function __init__()

src/reconstruct_loopset.jl

@@ -683,12 +683,6 @@ function _turbo_loopset(
     avx_loopset!(ls, instr, ops, arsv, tovector(AMsv), tovector(LPSYMsv), LBsv, vargs)
 end
 
-@static if VERSION ≥ v"1.7.0-DEV.421"
-  using Base: @aggressive_constprop
-else
-  macro aggressive_constprop(ex); esc(ex); end
-end
-
 """
     _turbo_!(unroll, ops, arf, am, lpsym, lb, vargs...)
 
@@ -708,7 +702,7 @@ Execute an `@turbo` block. The block's code is represented via the arguments:
   `StaticLowerUnitRange(1)` because the lower bound of the iterator can be determined to be 1.
 - `vargs...` holds the encoded pointers of all the arrays (see `VectorizationBase`'s various pointer types).
 """
-@aggressive_constprop @generated function _turbo_!(
+@generated function _turbo_!(
     ::Val{var"#UNROLL#"}, ::Val{var"#OPS#"}, ::Val{var"#ARF#"}, ::Val{var"#AM#"}, ::Val{var"#LPSYM#"}, ::Val{Tuple{var"#LB#",var"#V#"}}, var"#flattened#var#arguments#"::Vararg{Any,var"#num#vargs#"}
 ) where {var"#UNROLL#", var"#OPS#", var"#ARF#", var"#AM#", var"#LPSYM#", var"#LB#", var"#V#", var"#num#vargs#"}
   1 + 1 # Irrelevant line you can comment out/in to force recompilation...

src/simdfunctionals/mapreduce.jl

@@ -103,7 +103,7 @@ for (op, init) in zip((:+, :max, :min), (:zero, :typemin, :typemax))
         end
     end
 
-    @eval function _vreduce_dims!(out, ::typeof($op), Rpre, is, Rpost, arg)
+    @eval @inline function _vreduce_dims!(out, ::typeof($op), Rpre, is, Rpost, arg)
         s = $init(first(arg))
         @turbo for Ipost in Rpost, Ipre in Rpre
             accum = s
@@ -115,7 +115,7 @@ for (op, init) in zip((:+, :max, :min), (:zero, :typemin, :typemax))
         return out
     end
 
-    @eval function _vreduce(::typeof($op), arg)
+    @eval @inline function _vreduce(::typeof($op), arg)
         s = $init(first(arg))
         @turbo for i in eachindex(arg)
             s = $op(s, arg[i])

src/user_api_conveniences.jl

@@ -23,3 +23,12 @@ end
 end
 matmul_params() = matmul_params(register_size(), register_count(), cache_linesize())
 
+# function dotturbo(x,y)
+#   s = zero(promote_type(eltype(x),eltype(y)))
+#   @turbo for i ∈ eachindex(x,y)
+#     s += x[i]*y[i]
+#   end
+#   s
+# end
+# dotturbo(Float64[1.0], Float64[2.0])
+