Improve M1 native support

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.12.11"
+version = "0.12.12"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"

src/codegen/lower_load.jl

@@ -1,4 +1,5 @@
 function prefetchisagoodidea(ls::LoopSet, op::Operation, td::UnrollArgs)
+    # ((Sys.ARCH === :x86_64) || (Sys.ARCH === :i686)) || return false
     # return false
     @unpack u₁, u₁loopsym, u₂loopsym, vloopsym, vstep, u₂max, suffix = td
     length(loopdependencies(op)) ≤ 1 && return 0
@@ -40,6 +41,8 @@ function prefetchisagoodidea(ls::LoopSet, op::Operation, td::UnrollArgs)
     0
 end
 function add_prefetches!(q::Expr, ls::LoopSet, op::Operation, td::UnrollArgs, prefetchind::Int)
+    # TODO: maybe prefetch for non-x86_64?
+    ((Sys.ARCH === :x86_64) || (Sys.ARCH === :i686)) || return nothing
     @unpack u₁, u₁loopsym, u₂loopsym, vloopsym, u₂max = td
     # we should only be here if `unitsride(vloop)`
     dontskip = (cache_lnsze(ls) ÷ reg_size(ls)) - 1

src/modeling/determinestrategy.jl

@@ -543,17 +543,16 @@ function solve_unroll(
     W::Int, vloopsym::Symbol, rounduᵢ::Int
 )
     (u₁step, u₂step) = if rounduᵢ == 1 # max is to safeguard against some weird arch I've never heard of.
-        (max(1, cache_lnsze(ls) ÷ reg_size(ls)), 1)
+        (clamp(cache_lnsze(ls) ÷ reg_size(ls), 1, 4), 1)
     elseif rounduᵢ == 2
-        (1, max(1,cache_lnsze(ls) ÷ reg_size(ls)))
+        (1, clamp(cache_lnsze(ls) ÷ reg_size(ls), 1, 4))
     elseif rounduᵢ == -1
         (8 ÷ ls.vector_width[], 1)
     elseif rounduᵢ == -2
         (1, 8 ÷ ls.vector_width[])
     else
         (1, 1)
     end
-    # @show u₁step, u₂step
     u₁loop = getloop(ls, u₁loopsym)
     u₂loop = getloop(ls, u₂loopsym)
     solve_unroll(