[webkit-dev] PSA: Bit fields won't be packed on Windows if you mix types

[Eric Seidel]

For core classes like these we should COMPILE_ASSERT their size.  We
do this for many, but not all of these classes.

On Thu, Mar 29, 2012 at 1:21 AM, Ryosuke Niwa <rniwa at webkit.org> wrote:
> Hi,
>
> Unlike gcc and clang, MSVC pads each consecutive member variables of the
> same type in bitfields. e.g. if you have:
> struct AB {
>     unsigned m_1 : 31;
>     bool m_2 : 1;
> }
> then MSVC pads m_1 and allocates sizeof(unsigned) * 2 for AB whereas gcc and
> clang only allocate sizeof(unsigned) * 1 for AB.
>
> This is not a compiler bug. It's a spec. compliant behavior, and may in fact
> have a better run-time performance in some situations. However, for core
> objects like RenderObject and InlineBox, allocating extra 4-8 bytes per each
> object is prohibitory expensive.
>
> In such cases, please use the same POD type for all bitfield member
> variables. (Storage class classifiers and variable qualifiers seem to have
> no effect on how variables are packed; e.g. mutable, const, etc...). For
> example, MSVC will allocate sizeof(unsigned) * 1 for AB if we rewrite the
> above code as:
> struct AB {
>     unsigned m_1 : 31;
>     unsigned m_2 : 1;
> }
>
> When you're making this change, be sure to audit all code that assigns a
> non-boolean value to m_2 because implicit type coercion into boolean is no
> longer in effect. For example,
>
> AB ab;
> ab.m_2 = 2;
> puts(ab.m_2 ? "true" : "false");
>
> will print "true" before the change and will print "false" after the change.
> An easy way to ensure you've audited all such code is to add getters and
> setters for all bitfield member variables or wrap them in a special
> structure or class as done
> in http://trac.webkit.org/changeset/103353 and http://trac.webkit.org/changeset/104254.
>
> Best regards,
> Ryosuke Niwa
> Software Engineer
> Google Inc.
>
>
>
> _______________________________________________
> webkit-dev mailing list
> webkit-dev at lists.webkit.org
> http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
>