The previous version was good, just not optimal. Because the input was
an unsigned 64-bit number, compilers needed to generate extra code to
deal with HW instructions that only convert 64-bit signed input. And
that was useless because a double uniformly distributed from 0 to 1 can
only have 53 bits of randomness.
The previous implementation did exactly what the Microsoft libstdc++ and
libc++ implementations do. In my opinion, those implementations have an
imperfect distribution, which is corrected in this commit. In those, all
random input bigger than 0x20000000000000 has a different frequency
compared to input below that mark. For example, both 0x20000000000000
and 0x20000000000001 produce the same result (4.8828125e-4).
What's more, for the libc++ and MSVC implementations, input between
0xfffffffffffff001 and 0xffffffffffffffff results in 1.0 (probability 1
in 2⁵³), even though the Standard is very clear that the result should
be strictly less than 1. GCC 7's libstdc++ doesn't have this issue,
whereas the versions before would enter an infinite loop.
Change-Id: Ib17dde1a1dbb49a7bba8fffd14eced3c375dd2ec
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
c90d9f95d2