hashes/hash casts 64-bit floats to 32-bit ints on 32-bit architecture

[pwste-dawid-k]

Nim Version

2.2.2

Description

The hash proc for floats casts float to int even if these types are not of the same size.

Current Output

Expected Output

Valid hash or error

Known Workarounds

Custom hash proc:

proc hash*(x: float): Hash {.inline.} =
  let y = x + 0.0
  when nimvm:
    # workaround a JS VM bug: bug #16547
    result = hashWangYi1(cast[int64](float64(y)))
  else:
    when not defined(js):
      when sizeof(system.int) == 4:
        result = hashWangYi1(cast[int64](float64(y)))
      else:
        result = hashWangYi1(cast[Hash](y))
    else:
      result = hashWangYiJS(toBits(y))

Additional Information

No response

[Araq]

So? What exactly is the problem with this?

[pwste-dawid-k]

assert hash(0.0) != hash(1.0) fails because both floats seem to get truncated to 0 before before hashWangYi1 is called.

[Araq]

It's a hash value, it cannot guarantee uniqueness. But what you say might indicate an avoidable weakness.