Nonlinear: change how constraints are parsed

[odow]

To minimize differences when refactoring, we kept the same parsing behavior as JuMP. That means, constraints like x^2 <= 1 get parsed into x^2 - 1 <= 0.0:

MathOptInterface.jl/src/Nonlinear/parse.jl

Lines 242 to 259 in 1af7064

	function _expr_to_constraint(expr::Expr)
	if isexpr(expr, :comparison)
	@assert expr.args[2] == expr.args[4]
	@assert expr.args[2] in (:<=, :>=)
	lhs, body, rhs =
	_normalize_constraint_expr(expr.args[1], expr.args[3], expr.args[5])
	return body, MOI.Interval(lhs, rhs)
	end
	lhs, rhs = _normalize_constraint_expr(expr.args[2], expr.args[3])
	if expr.args[1] == :<=
	return :($lhs - $rhs), MOI.LessThan(0.0)
	elseif expr.args[1] == :>=
	return :($lhs - $rhs), MOI.GreaterThan(0.0)
	else
	@assert expr.args[1] == :(==)
	return :($lhs - $rhs), MOI.EqualTo(0.0)
	end
	end

Changing this behavior would lead to more informative RHSs being passed to the solvers, and remove an additional operation. However, doing so would break JuMP's tests and printing, which have things like this:

MathOptInterface.jl/test/Nonlinear/Nonlinear.jl

Lines 296 to 303 in 1af7064

	model = Nonlinear.Model()
	x = MOI.VariableIndex(1)
	input = :($x^2 + 1 <= 1.0)
	c = Nonlinear.add_constraint(model, input)
	@test model[c].set == MOI.LessThan(0.0)
	evaluator = Nonlinear.Evaluator(model)
	MOI.initialize(evaluator, [:ExprGraph])
	@test MOI.constraint_expr(evaluator, 1) == :((x[$x]^2 + 1) - 1.0 <= 0.0)

As suggested by @blegat #1804 (comment), one option is to remove the ability to specify constraints as a <= b, and instead force add_constraint(model, f, set), and then make JuMP keep the existing behavior.

[blegat]

constraints like x^2 <= 1 get parsed into x^2 - 1 <= 0.0

That's the opposite right?

Tha's a good argument, we don't want MOI to have to do these choices. Everything is explicit in MOI

[odow]

That's the opposite right?

No, at the moment that's what JuMP does.