- densifies zeros

[willow-ahrens]

A - A results in all the zero entries getting converted to -0.0 and being stored explicitly 😬

julia> A = fsprand(5, 5, 0.5)
5×5 Tensor{SparseCOOLevel{2, Tuple{Int64, Int64}, Vector{Int64}, Tuple{Vector{Int64}, Vector{Int64}}, ElementLevel{0.0, Float64, Int64, Vector{Float64}}}}:
 0.0       0.836373  0.73536  0.0         0.83067
 0.0       0.0       0.0      0.0         0.390398
 0.677004  0.0       0.0      0.0         0.605428
 0.0       0.0       0.0      0.00907705  0.728524
 0.0       0.0       0.0      0.878594    0.0

julia> A - A
5×5 Tensor{SparseDictLevel{Int64, Vector{Int64}, Vector{Int64}, Vector{Int64}, Dict{Tuple{Int64, Int64}, Int64}, Vector{Int64}, SparseDictLevel{Int64, Vector{Int64}, Vector{Int64}, Vector{Int64}, Dict{Tuple{Int64, Int64}, Int64}, Vector{Int64}, ElementLevel{0.0, Float64, Int64, Vector{Float64}}}}}:
 -0.0   0.0   0.0  -0.0   0.0
 -0.0  -0.0  -0.0  -0.0   0.0
  0.0  -0.0  -0.0  -0.0   0.0
 -0.0  -0.0  -0.0   0.0   0.0
 -0.0  -0.0  -0.0   0.0  -0.0
 
 julia> countstored(A - A)
25

[willow-ahrens]

After #383, the handling of -0.0 has proven to be a continuing issue. The problem is that there's two ways to evaluate 0.0 + (-0.0) in Finch. We could follow floating point semantics, which gives 0.0. We could also use the fact that 0.0 is an annihilator, which gives -0.0. I'm beginning to think the right choice may be to defer to actual evaluation semantics whenever possible.

[willow-ahrens]

After some thought, we can summarize this issue as:

• -0.0 + 0.0 = 0.0 (from floating point semantics)
• -0.0 + 0.0 = -0.0 (from x + 0 = x rule)
  We need one of these to win out over the other when we do constant propagation in Finch.
  I think that we can resolve this by overloading floating point semantics within Finch such that 0.0 + x = 0.0 and 0.0 * x = 0.0, whenever we evaluate floating point at compile time.

@wraith1995 I feel like you would appreciate this.