Make register spills double cost rather than return infinite, to allow specifying tile sizes even when the optimizer thinks it is inadvisable.

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.6.21"
+version = "0.6.22"
 
 [deps]
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"

src/determinestrategy.jl

@@ -499,11 +499,11 @@ function evaluate_cost_tile(
             reg_pressure[4] += rp
         end
     end
-    sum(reg_pressure) > VectorizationBase.REGISTER_COUNT && return 0, 0, Inf
+    costpenalty = (sum(reg_pressure) > VectorizationBase.REGISTER_COUNT) ? 2 : 1
     # @show order, vectorized cost_vec reg_pressure
     # @show solve_tilesize(ls, unrolled, tiled, cost_vec, reg_pressure)
     U, T, tcost = solve_tilesize(ls, unrolled, tiled, cost_vec, reg_pressure, W, vectorized)
-    U, T, tcost + stride_penalty(ls, order)
+    U, T, costpenalty * tcost + stride_penalty(ls, order)
 end
 
 

src/lowering.jl

@@ -353,7 +353,12 @@ function lower(ls::LoopSet)#, prependinlineORorUnroll = 0)
 end
 function lower(ls::LoopSet, U, T)#, prependinlineORorUnroll = 0)
     num_loops(ls) == 1 && @assert T == -1
-    order, unrolled, tiled, vectorized, _U, _T = choose_order(ls)
+    if T == -1
+        order, vectorized, c = choose_unroll_order(ls, Inf)
+        unrolled = first(order); tiled = Symbol("##undefined##")
+    else
+        order, unrolled, tiled, vectorized, _U, _T, c = choose_tile(ls)
+    end
     lower(ls, order, unrolled, tiled, vectorized, U, T)
 end
 

test/offsetarrays.jl

@@ -60,7 +60,6 @@ using Test
     function avx2d!(out::AbstractMatrix, A::AbstractMatrix, kern)
         rng1k, rng2k = axes(kern)
         rng1,  rng2  = axes(out)
-        # Manually unpack the OffsetArray
         for j in rng2, i in rng1
             tmp = zero(eltype(out))
             @avx for jk in rng2k, ik in rng1k
@@ -73,7 +72,6 @@ using Test
     function avx2douter!(out::AbstractMatrix, A::AbstractMatrix, kern)
         rng1k, rng2k = axes(kern)
         rng1,  rng2  = axes(out)
-        # Manually unpack the OffsetArray
         @avx for j in rng2, i in rng1
             tmp = zero(eltype(out))
             for jk in rng2k, ik in rng1k
@@ -84,6 +82,7 @@ using Test
         out
     end
 
+
 
     struct SizedOffsetMatrix{T,LR,UR,LC,RC} <: AbstractMatrix{T}
         data::Matrix{T}