Make reshape(::Bitarray,...) type stable (#21670)

This works around some of the performance regressions flagged in #20993. The instability itself is not new nor is it the cause of the regression. The trouble is just that there are a few more inlined functions now, each of which needs a dynamic dispatch.

It is worth noting that 0.6 (with this patch) is significantly faster than 0.5 (with this patch). I measure an improvement of ~30%.

base/bitarray.jl

@@ -472,10 +472,13 @@ function copy!(dest::BitArray, src::Array)
     return unsafe_copy!(dest, 1, src, 1, length(src))
 end
 
-function reshape(B::BitArray, dims::NTuple{N,Int}) where N
+function reshape(B::BitArray{N}, dims::NTuple{N,Int}) where N
+    return dims == size(B) ? B : _bitreshape(B, dims)
+end
+reshape(B::BitArray, dims::Tuple{Vararg{Int}}) = _bitreshape(B, dims)
+function _bitreshape(B::BitArray, dims::NTuple{N,Int}) where N
     prod(dims) == length(B) ||
         throw(DimensionMismatch("new dimensions $(dims) must be consistent with array size $(length(B))"))
-    dims == size(B) && return B
     Br = BitArray{N}(ntuple(i->0,Val{N})...)
     Br.chunks = B.chunks
     Br.len = prod(dims)

test/bitarray.jl

@@ -143,6 +143,10 @@ timesofar("conversions")
         @check_bit_operation reshape(b1, (n2,n1)) BitMatrix
         @test_throws DimensionMismatch reshape(b1, (1,n1))
 
+        @test @inferred(reshape(b1, n1*n2)) == @inferred(reshape(b1, (n1*n2,))) == @inferred(reshape(b1, Val{1})) == @inferred(reshape(b1, :))
+        @test @inferred(reshape(b1, n1, n2)) === @inferred(reshape(b1, Val{2})) === b1
+        @test @inferred(reshape(b1, n2, :)) == @inferred(reshape(b1, (n2, n1))) != @inferred(reshape(b1, Val{2}))
+
         b1 = bitrand(s1, s2, s3, s4)
         @check_bit_operation reshape(b1, (s3,s1,s2,s4)) BitArray{4}
         @test_throws DimensionMismatch reshape(b1, (1,n1))