On 17 July 2012 03:24, Catalin Patulea <catal...@google.com> wrote:
> I'm just sending this in as a draft for now. There are two outstanding issues:
>
> 1) What is the recommended type for bitfields? A quick grep shows inconsistent
> usage of unsigned int/signed int/uintNN_t across the codebase. Note that
> mantissa must be 63 bits so int won't do on either 32 or 64-bit builds.
I think our current position is "don't use bitfields for anything that
has to match an externally defined layout". This is because on Windows
we have to compile with -mms-bitfields [for compatibility with Windows
APIs], which can lead to different layouts for structs with bitfields.
For instance in your struct:
> + struct {
> + unsigned long long mantissa:63;
> + unsigned int one:1;
> + unsigned int exponent:15;
> + unsigned int negative:1;
> + unsigned int empty:16;
> + } ieee;
the Windows build would not put the fields 'mantissa' and
'one' in the same 64 bit word, it would insert a padding bit.
-- PMM
