Maybedemotesizes in full static loop unroll conditions.

Author: chriselrod

src/determinestrategy.jl

@@ -472,7 +472,7 @@ function solve_unroll(
     if isstaticloop(u₂loop)
         if u₂loopsym !== vectorized && u₂L ≤ 4
             u₁ = max(1, solve_unroll_constT(reg_pressure, u₂L))
-            u₁ = isstaticloop(u₁loop) ? min(u₁, u₁L) : u₁
+            u₁ = isstaticloop(u₁loop) ? maybedemotesize(u₁, u₁L) : u₁
             return u₁, u₂L, unroll_cost(cost_vec, u₁, u₂L, u₁L, u₂L)
         end
         u₂L = u₂loopsym === vectorized ? cld(u₂L,W) : u₂L
@@ -481,7 +481,7 @@ function solve_unroll(
     if isstaticloop(u₁loop)
         if u₁loopsym !== vectorized && u₁L ≤ 4
             u₂ = max(1, solve_unroll_constU(reg_pressure, u₁L))
-            u₂ = isstaticloop(u₂loop) ? min(u₂, u₂L) : u₂
+            u₂ = isstaticloop(u₂loop) ? maybedemotesize(u₂, u₂L) : u₂
             return u₁L, u₂, unroll_cost(cost_vec, u₁L, u₂, u₁L, u₂L)
         end
         u₁L = u₁loopsym === vectorized ? cld(u₁L,W) : u₁L

src/memory_ops_common.jl

@@ -72,7 +72,9 @@ function subset_vptr!(ls::LoopSet, vptr::Symbol, indnum::Int, ind, previndices,
             if loopindex[i]
                 append_loop_valdims!(valcall, getloop(ls, previndices[i+offset]))
             else
-                for loopdep ∈ loopdependencies(ls.opdict[previndices[i+offset]])
+                # assumes all valdims will be of equal length once expanded...
+                # A[I + J, constindex], I and J may be CartesianIndices. This requires they all be of same number of dims
+                let loopdep = first(loopdependencies(ls.opdict[previndices[i+offset]]))
                     append_loop_valdims!(valcall, getloop(ls, loopdep))
                 end
             end

src/reconstruct_loopset.jl

@@ -492,7 +492,7 @@ Execute an `@avx` block. The block's code is represented via the arguments:
 - `vargs...` holds the encoded pointers of all the arrays (see `VectorizationBase`'s various pointer types).
 """
 @generated function _avx_!(::Val{UNROLL}, ::Type{OPS}, ::Type{ARF}, ::Type{AM}, ::Type{LPSYM}, lb::LB, vargs...) where {UNROLL, OPS, ARF, AM, LPSYM, LB}
-    1 + 1 # Irrelevant line you can comment out/in to force recompilation...
+    # 1 + 1 # Irrelevant line you can comment out/in to force recompilation...
     ls = _avx_loopset(OPS.parameters, ARF.parameters, AM.parameters, LPSYM.parameters, LB.parameters, vargs)
     # @show avx_body(ls, UNROLL)
     # @show UNROLL, OPS, ARF, AM, LPSYM, LB