float.__mod__ and float.__divmod__ produce incorrect results for negative numbers

[pauvrepetit]

Description

The current __mod__ implementation for all float types uses LLVM's frem instruction, which computes the IEEE 754 remainder (truncated toward zero). However, __floordiv__ is implemented as floor(a / b), which follows Python's floored division semantics. This mismatch causes the fundamental identity (a // b) * b + (a % b) == a to break when negative numbers are involved.

Example

a = -1.0
b = 3.0
print(a % b)       # Codon outputs: -1.0, expected (Python): 2.0
print(a // b)      # Codon outputs: -1.0, which is correct

[arshajii]

Hi @pauvrepetit, this is known and documented here: https://docs.exaloop.io/start/usage/#numerical-semantics -- in short, you can get Python semantics by using the -numerics=py flag. Note that we'll likely be changing this in Codon to use Python semantics by default prior to v1.0 (with support for the current C-semantics versions via new methods like .cdiv(..) / .cmod(..)).

[pauvrepetit]

@arshajii I'm aware of the implementation differences regarding floordiv/mod between Python and C++ for integer numbers.
The issue here is for the calculation of floating-point numbers. Consider two float numbers a and b, codon produced a // b and a % b cannot fulfill (a // b) * b + (a % b) == a, which is unreasonable for a programming language.