Julia 1.9.0 doesn't drop zeros when assigning to a sparse array only when the elements are Big

[LilithHafner]

MWE:

julia> A = spzeros(BigInt, 3, 3); A[1] = 0; A
3×3 SparseMatrixCSC{BigInt, Int64} with 1 stored entry:
 0  ⋅  ⋅
 ⋅  ⋅  ⋅
 ⋅  ⋅  ⋅

This change was introduced in #296.

Originally posted by @YingboMa in JuliaLang/julia#49766

[LilithHafner]

This is to preserve egality when possible, an admittedly weak justification in this case

julia> A = spzeros(BigInt, 3, 3); x = big(0); A[1] = x;

julia> A[1] === x
true

julia> A[2] === x
false

@YingboMa, Did you report this issue as a curiosity, because it is causing measurable performance problems, or for some other reason?

[YingboMa]

It broke MTK.

[LilithHafner]

Do you have a MWE of something that broke there?

[YingboMa]

Does the MWE alone not qualify as a bug? What's the use case for egality other than a theoretically nice sounding property?

[LilithHafner]

The egality property makes reinterpret safe which allows us to fix #289. reinterpret only comes up in the non-isbitstype case.

Other folks might view it differently, but I think the example in the OP is ambiguous. Taken alone, I would describe it as odd but harmless. Much like this example (related to #294) from 1.8 which, alone, I would also describe as odd but harmless

julia> x = spzeros(BigFloat, 3)
3-element SparseVector{BigFloat, Int64} with 0 stored entries

julia> x[1] = -0.0
-0.0

julia> isless(x[1], 0)
false

julia> isless(-0.0, 0)
true

An example showing how the behavior in the OP breaks something (e.g. MTK) could certainly shift my thinking.

[Keno]

I think #390 is the real problem here. I don't think MTK would have noticed was it not for that.

[rayegun]

Is the iszero check happening before promotion? 🤔

[LilithHafner]

I think #390 is the real problem here.

Fortunately, that's due to a silly and fixed mistake.

Is the iszero check happening before promotion? 🤔

No, it's big(0) !== big(0).

[rayegun]

Oh because they're objects of course... What's the reasoning for egality instead of equality?

[rayegun]

If it's for 0.0 == -0.0 it seems to me that we should have _iszero(x) = iszero(x); _iszero(x::AbstractFloat) = ...

[LilithHafner]

It's also for dual numbers from forward diff (JuliaDiff/ForwardDiff.jl#542) and possibly other types with nontrivial zeros. The idea is that a sparse array is an optimization which should have different performance characteristics than Array but identical semantics, so if we can't recreate the passed zero exactly, we should store it.

[rayegun]

We can have it be the more conservative version, but allow for the overload?

_iszero(x) = x === zero(typeof(x)) Then we can support BigInt and other types? Insertion is still pretty expensive and should be avoided if possible, since we don't have pending insertions here.

[LilithHafner]

Do you have an example use-case where making _iszero less conservative results in measurable performance gains? I have a hard time coming up with example use cases for sparse arrays that use setindex! a whole bunch of times and are not already woefully inefficient.

[YingboMa]

Isn't it obvious? For example:

julia> using SparseArrays, LinearAlgebra, BenchmarkTools

julia> N = 500; A = diagm(0=>rand(BigFloat, N), 1=>rand(BigFloat, N-1));

julia> S = spzeros(BigFloat, N, N); S .= A; S[end, 1] = S[1, end] = 10; b = rand(N);

julia> @btime $S*$b;
  15.579 ms (1000003 allocations: 49.60 MiB)

julia> @btime $(dropzeros(S))*$b;
  102.375 μs (4007 allocations: 207.49 KiB)

julia> S
500×500 SparseMatrixCSC{BigFloat, Int64} with 250000 stored entries:
⎡⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎤
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎢⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎥
⎣⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⎦

but if we use Float64, S is actually sparse

julia> S = spzeros(Float64, N, N); S .= A; S[end, 1] = S[1, end] = 10; b = rand(N);

julia> S
500×500 SparseMatrixCSC{Float64, Int64} with 1001 stored entries:
⎡⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⎤
⎢⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⎥
⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⎥
⎣⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⎦

_iszero(x) = x === zero(typeof(x)) makes generic programming extremely hard.

[SobhanMP]

Maybe it sense to make an is_reconstructible_zero? The default implementation could be x === zero(typeof(x)), and then it could be overwritten for different types on an as-needed basis? I think for compatibility, this should go to Julia directly.

[LilithHafner]

I don't find that example particularly convincing because both the slow version and the proposed fix have a significant performance problem (using diagm instead of spdiagm).

julia> function slow(N)
           A = diagm(0=>rand(BigFloat, N), 1=>rand(BigFloat, N-1));
           S = spzeros(BigFloat, N, N); S .= A; S[end, 1] = S[1, end] = 10
           b = rand(N);
           S*b
       end
slow (generic function with 1 method)

julia> function assign_only_nonzeros(N)
           A = diagm(0=>rand(BigFloat, N), 1=>rand(BigFloat, N-1));
           S = spzeros(BigFloat, N, N); S[A .!= 0] .= A[A .!= 0]; S[end, 1] = S[1, end] = 10
           b = rand(N);
           S*b
       end
assign_only_nonzeros (generic function with 1 method)

julia> function better(N)
           A = spdiagm(0=>rand(BigFloat, N), 1=>rand(BigFloat, N-1));
           S = spzeros(BigFloat, N, N); S .= A; S[end, 1] = S[1, end] = 10
           b = rand(N);
           S*b
       end
better (generic function with 1 method)

julia> @btime slow(500);
  61.185 ms (1506044 allocations: 82.46 MiB)

julia> @btime assign_only_nonzeros(500);
  3.918 ms (12048 allocations: 2.59 MiB)

julia> @btime better(500);
  286.542 μs (6051 allocations: 443.64 KiB)

[YingboMa]

I don't find the proposed fix particularly convincing, because should people always write S[A .!= 0] .= A[A .!= 0] for generic programming? The point of the above example is not about diagm. This happens if one mixes dense and sparse computation.

[LilithHafner]

Apologies for the confusion. The name proposed was ambiguous. I renamed that function from proposed to assign_only_nonzeros to clarify.

The assign_only_nonzeros function models the runtime of the whole process if we implement the proposed modifications to iszero. I recommend using the better approach, which is 10x faster than that proposal.

[YingboMa]

This doesn't just happen with diagm though. This happens whenever one assigns dense matrices into a sparse matrix, if the eltype is big, and the underlaying issue is precisely due to the un-generic choice of _iszero(x) = x === zero(typeof(x)).

[LilithHafner]

In the event that an object iszero but is not egal to it's canonical zero, we have a choice: either assign for better API alignment with dense arrays, or don't assign for better performance. The tradeoff is API alignment with dense arrays vs. performance, not generic vs un-generic. I stand by the decision to choose to support the dense array API over performance in this case.

It is plausible that there exist examples of efficient code that stops being efficient after this change. If that is the case, then I would support @SobhanMP's proposal of defining is_reconstructible_zero(x) = x === zero(typeof(x)) and then special casing non-egal reconstructable zeros (such as big). If we do add such a function, though, I think it should go in SparseArrays.jl, not Base. In the uncommon event that packages choose to override this definition, they can do so in package extensions.

[YingboMa]

Yeah, this solution sounds good.

[YingboMa]

In MTK, we had to add an explicit dropzero! because of this change https://github.com/SciML/ModelingToolkit.jl/blob/05525b540820569c01d2249096dbbc5880d943d0/src/structural_transformation/bareiss.jl#L146-L148

julia> using ModelingToolkit, BenchmarkTools, SparseArrays

julia> A = BigInt.(sprand(Int, 100, 100, 0.01)) + I;

julia> @btime ModelingToolkit.bareiss!(copy($A)) # without `dropzero!`
  129.664 ms (5869620 allocations: 115.39 MiB)
(100, -469434422448344658861925147159840003015949584463799478228775210152597681412950869211899407566099105579794178943, false)

julia> @btime ModelingToolkit.bareiss!(copy($A))
  78.866 ms (3386748 allocations: 67.56 MiB)
(100, -469434422448344658861925147159840003015949584463799478228775210152597681412950869211899407566099105579794178943, false)