signbit

[timotheecour]

/cc @xflywind

implementation for js:

when true:
  proc signbit(x: float): bool =
    asm """
    const buffer = new ArrayBuffer(8);
    const floatBuffer = new Float64Array(buffer);
    const uintBuffer = new Uint32Array(buffer);
    floatBuffer[0] = `x`;
    `result` = (uintBuffer[1] >> 31) != 0;
    """
  doAssert not signbit(0.0)
  doAssert signbit(-0.0)
  doAssert signbit(-0.1)
  doAssert not signbit(0.1)

for c: signbit importc
for vm: vmops

[ringabout]

With nim-lang#16403

proc toBitsImpl(x: float): array[2, uint32] =
  asm """
  const buffer = new ArrayBuffer(8);
  const floatBuffer = new Float64Array(buffer);
  const uintBuffer = new Uint32Array(buffer);
  floatBuffer[0] = `x`;
  `result` = uintBuffer
  """


proc signbit(x: float): bool =
  let uintBuffer = toBitsImpl(x)
  result = (uintBuffer[1] shr 31) != 0;

doAssert not signbit(0.0)
doAssert signbit(-0.0)
doAssert signbit(-0.1)
doAssert not signbit(0.1)

[ringabout]

We could use signbit to refactor copySign if needed.

[timotheecour]

We could use signbit to refactor copySign if needed.

ya i was just about to suggest that; it will also make copySign correct for NaN; let's do that after nim-lang#16592

EDIT:

• and also this:

    when nimvm:
      discard
    else:
      when not defined(js):
        doAssert copySign(-1.0, -NaN) == 1.0
        doAssert copySign(10.0, -NaN) == 10.0
        doAssert copySign(1.0, copySign(NaN, -1.0)) == -1.0 # fails in VM

=> should become:

    doAssert copySign(-1.0, -NaN) == 1.0
    # doAssert copySign(10.0, -NaN) == -10.0 # pending https://github.com/timotheecour/Nim/issues/499
    doAssert copySign(1.0, copySign(NaN, -1.0)) == -1.0 # fails in VM

(ie one test is wrong, and the 2 other should work in js,vm,c)

• and this: in runnableExamples:

    # fails in VM and JS backend
    doAssert copySign(1.0, copySign(NaN, -1.0)) == -1.0

=>

    doAssert copySign(1.0, copySign(NaN, -1.0)) == -1.0
    doAssert copySign(NaN, -1.0).signbit

[timotheecour]

@xflywind I think nim-lang#16592 closes this (but your PR mentioned Ref, not close; not sure why?)