fix: reproduce and fix motoko-base/issues/653 (#4712)

Cf. fixes dfinity/motoko-base#653

It appears that `a <> Big_int.zero_big_int` blows up when
`a` is zero. Weird. Eliminating the comparison doesn't hurt
correctness, and is probably more efficient too, as negating
zero should be cheap!

@timohanke feel free to review too!

src/mo_values/numerics.ml

@@ -41,7 +41,7 @@ let bigint_of_double (f : Wasm.F64.t) : Big_int.big_int =
 
   (* Fraction part of IEEE 754 double, 52 bits from the float, with an implicit
      1 at the 53rd bit *)
-  let frac = Int64.(logor (shift_right_logical (shift_left bits 12) 12) (shift_left (of_int 1) 52)) in
+  let frac = Int64.(logor (shift_right_logical (shift_left bits 12) 12) (shift_left one 52)) in
 
   if Int64.(equal exp bits_11) then
     (* Exponent is fully set: NaN or inf *)
@@ -55,7 +55,7 @@ let bigint_of_double (f : Wasm.F64.t) : Big_int.big_int =
 
   let a = Big_int.big_int_of_int64 frac in
 
-  let a = if Int64.(compare exp (of_int 0)) < 0 then
+  let a = if Int64.(compare exp zero) < 0 then
     (* Exponent < 0, shift right *)
     Big_int.(shift_right_big_int a (- (Int64.to_int exp)))
   else
@@ -64,7 +64,7 @@ let bigint_of_double (f : Wasm.F64.t) : Big_int.big_int =
   in
 
   (* Negate the number if sign bit is set (double is negative) *)
-  if Int64.shift_right_logical bits 63 = Int64.of_int 1 && a <> Big_int.zero_big_int then
+  if Int64.(shift_right_logical bits 63 = one) then
     Big_int.minus_big_int a
   else
     a

test/run/issue-base653.mo

@@ -0,0 +1,4 @@
+import { floatCopySign; floatToInt64 } = "mo:⛔";
+
+assert 0 == floatToInt64(-0.5);
+assert 0 == floatToInt64(floatCopySign(0.0, -1.0));