broadcast fixes

Author: chriselrod

src/broadcast.jl

@@ -189,6 +189,7 @@ for f ∈ [ # groupedstridedpointer support
   :(ArrayInterface.contiguous_axis),
   :(ArrayInterface.contiguous_batch_size),
   :(ArrayInterface.device),
+  :(ArrayInterface.dense_dims),
   :(ArrayInterface.stride_rank),
   :(VectorizationBase.val_dense_dims),
   :(ArrayInterface.offsets),
@@ -204,6 +205,8 @@ function is_column_major(x)
   true
 end
 is_row_major(x) = is_column_major(reverse(x))
+_find_arg_least_greater(r::Vector{Int}, i) =
+  findmin(x -> x > i ? x : typemax(Int), r)
 # @inline _bytestrides(s,paren) = VectorizationBase.bytestrides(paren)
 function _strides_expr(
   @nospecialize(s),
@@ -215,19 +218,18 @@ function _strides_expr(
   q = Expr(:block, Expr(:meta, :inline))
   strd_tup = Expr(:tuple)
   ifel = GlobalRef(Core, :ifelse)
-  Nrange = 1:1:N # type stability w/ respect to reverse
+  Nrange = 1:N # type stability w/ respect to reverse
+  # Nrange = 1:1:N # type stability w/ respect to reverse
   use_stride_acc = true
   stride_acc::Int = 1
-  if is_column_major(R)
-    # elseif is_row_major(R)
-    #   Nrange = reverse(Nrange)
-  else # not worth my time optimizing this case at the moment...
-    # will write something generic stride-rank agnostic eventually
+  next, n = _find_arg_least_greater(R, 0)
+  n = findfirst(==(1), R)
+  if !D[n]
     use_stride_acc = false
     stride_acc = 0
   end
   sₙ_value::Int = 0
-  for n ∈ Nrange
+  for _n ∈ Nrange
     xₙ_type = x[n]
     xₙ_static = xₙ_type <: StaticInt
     xₙ_value::Int = xₙ_static ? (xₙ_type.parameters[1])::Int : 0
@@ -254,20 +256,22 @@ function _strides_expr(
         )
       end
     end
-    if (n ≠ last(Nrange)) && use_stride_acc
-      nnext = n + step(Nrange)
-      if D[nnext]
-        if xₙ_static & sₙ_static
-          stride_acc = xₙ_value * sₙ_value
-        elseif sₙ_static
-          if stride_acc ≠ 0
-            stride_acc *= sₙ_value
+    if (n ≠ N)
+      next, n = _find_arg_least_greater(R, next)
+      if use_stride_acc
+        if D[n]
+          if xₙ_static & sₙ_static
+            stride_acc = xₙ_value * sₙ_value
+          elseif sₙ_static
+            if stride_acc ≠ 0
+              stride_acc *= sₙ_value
+            end
+          else
+            stride_acc = 0
           end
         else
           stride_acc = 0
         end
-      else
-        stride_acc = 0
       end
     end
   end
@@ -675,6 +679,7 @@ end
   ::Val{UNROLL},
   ::Val{dontbc}
 ) where {T<:NativeTypes,N,BC<:Union{Broadcasted,Product},Mod,UNROLL,dontbc}
+  @show (dest) (BC)
   vmaterialize_fun(sizeof(T), N, BC, Mod, UNROLL, dontbc, false)
 end
 @generated function vmaterialize!(