A potential issue in fhe.inputset or fhe.constant

[yuriname]

Description

• versions affected: concrete-python 2.11.0
• python version: 3.11.0

I met some problems in my project. After reducing it, I obtained this program below:

from concrete import fhe
import numpy as np

@fhe.compiler({'x1': 'encrypted', 'x2': 'encrypted'})
def debug_foo(x1, x2):
    x2_pow = x2 * x2
    con_secret = np.invert(-fhe.constant(4)).astype(np.int64)
    cond_val = (x2_pow > 0)
    res = con_secret >> cond_val
    return res

if __name__ == '__main__':
    x1, x2 = -12, 13
    
    expect_result = debug_foo(x1, x2)
    print(f"expect result: {expect_result}")

    inputset = fhe.inputset(fhe.int8, fhe.int8)
    circuit = debug_foo.compile(inputset)
    circuit.keygen()
    encrypted_x1, encrypted_x2 = circuit.encrypt(x1, x2)
    encrypted_result = circuit.run(encrypted_x1, encrypted_x2)
    result = circuit.decrypt(encrypted_result)
    print(f"concrete result: {result}")
    print(f"{circuit}")

Its final output does not match the expected plaintext output:

expect result: 1
concrete result: 3

I printed the circuit of this program, and the bit range looks correct:

 %0 = x2                            # EncryptedScalar<int8>          ∈ [-125, 124]
 %1 = multiply(%0, %0)              # EncryptedScalar<uint14>        ∈ [1, 15625]
 %2 = zeros()                       # EncryptedScalar<uint1>         ∈ [0, 0]
 %3 = 4                             # ClearScalar<uint3>             ∈ [4, 4]
 %4 = add(%2, %3)                   # EncryptedScalar<uint3>         ∈ [4, 4]
 %5 = negative(%4)                  # EncryptedScalar<int3>          ∈ [-4, -4]
 %6 = invert(%5)                    # EncryptedScalar<uint2>         ∈ [3, 3]
 %7 = astype(%6, dtype=int_)        # EncryptedScalar<uint2>         ∈ [3, 3]
 %8 = 0                             # ClearScalar<uint1>             ∈ [0, 0]
 %9 = greater(%1, %8)               # EncryptedScalar<uint1>         ∈ [1, 1]
%10 = right_shift(%7, %9)           # EncryptedScalar<uint1>         ∈ [1, 1]
return %10

Since my input x2=13 falls within the valid range of -125 to 124, I would expect the program to output a correct value.

I also find that if I use -4 rather than -fhe.constant(4), the output becomes consistent.
OR, if I use fhe.inputset(fhe.int8, fhe.int8, size=1000) to increase the number of samples, the output can also become consistent.

But this is still strange, because the circuit shows that the valid input range for x2 is [-125, 124], so I would expect inputs within this range to always yield a correct result. I wonder whether this is an issue with fhe.inputset or fhe.constant, or just an issue with how the circuit is printed.