Unrolling tweaks

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.12.12"
+version = "0.12.13"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"

docs/src/assets/bench_filter2d_dynamic_v2.png

@@ -184,7 +184,7 @@ function lower_no_unroll(ls::LoopSet, us::UnrollSpecification, n::Int, inclmask:
     # q = if align_loop
     #     Expr(:block, align_inner_loop_expr(ls, us, loop), Expr(:while, tc, body))
     # elseif nisvectorized
-    if loopisstatic && (isone(length(loop) ÷ W) || (n ≤ 3 && length(loop) ≤ 8W && allinteriorunrolled(ls, us, n)))
+    if loopisstatic && (!ls.loadelimination) && (isone(length(loop) ÷ W) || (n ≤ 3 && length(loop) ≤ 8W && allinteriorunrolled(ls, us, n)))
         q = Expr(:block)
         for _ ∈ 1:(length(loop) ÷ W)
             push!(q.args, body)

docs/src/assets/bench_selfdot_v2.png

@@ -580,7 +580,7 @@ function solve_unroll(
     if isstaticloop(u₂loop)
         if u₂loopsym !== vloopsym && u₂L ≤ 4
             u₁ = max(1, solve_unroll_constT(reg_pressure, u₂L))
-            u₁ = isstaticloop(u₁loop) ? maybedemotesize(u₁, u₁L) : u₁
+            u₁ = isstaticloop(u₁loop) ? maybedemotesize(u₁, u₁loopsym === vloopsym ? cld(u₁L,W) : u₁L) : u₁
             return u₁, u₂L, unroll_cost(cost_vec, u₁, u₂L, u₁L, u₂L)
         end
         u₂Ltemp = u₂loopsym === vloopsym ? cld(u₂L, W) : u₂L
@@ -589,7 +589,7 @@ function solve_unroll(
     if isstaticloop(u₁loop)
         if u₁loopsym !== vloopsym && u₁L ≤ 4
             u₂ = max(1, solve_unroll_constU(reg_pressure, u₁L))
-            u₂ = isstaticloop(u₂loop) ? maybedemotesize(u₂, u₂L) : u₂
+            u₂ = isstaticloop(u₂loop) ? maybedemotesize(u₂, u₂loopsym === vloopsym ? cld(u₂L,W) : u₂L) : u₂
             return u₁L, u₂, unroll_cost(cost_vec, u₁L, u₂, u₁L, u₂L)
         end
         u₁Ltemp = u₁loopsym === vloopsym ? cld(u₁L, W) : u₁L