Wrong floating-point operations

[cardillan]

Mindcode currently performs optimizations which may change the outcome of floating-point operations performed by the program.

Example:

const c = 1e20;
param p = 1e20;
param x = 1;

println(p + x - p);
println(c + x - c);

compiles to

set p 1E20
set x 1
op add *tmp0 p x
op sub *tmp1 *tmp0 p
print "Unoptimized: "
print *tmp1
print "\nOptimized: "
print x
print "\n"

and outputs

Unoptimized: 0
Optimized: 1

Due to limited precision of mlog variables, the value of 1e20 + 1 is 1e20, and after subtracting 1e20, the result is 0. However, when the value is constant, Mindcode removes both addition and subtraction, as they're opposite operations in an ideal (algebraic) sense. This leaves the original value of 1 intact.

The problem is caused by the Data Flow Optimization (in constant folding), the Expression Optimization and perhaps some other.

The issue affects programs which perform numerical computations, computations with non-integer values or with integer values values exceeding 2⁵³. This should be quite unusual for a typical mlog program, so most programs probably won't be affected.