[cxx-mem-model] bitfield tests

[Richard Guenther]

On Tue, Mar 29, 2011 at 7:00 PM, Aldy Hernandez <aldyh@redhat.com> wrote:
> [Language lawyers, please correct me if I have mis-interpreted the upcoming
> standard in any way.]
>
> In the C++ memory model, contiguous bitfields comprise a single memory
> location, so it's fair game to bit twiddle them when setting them.  For
> example:
>
>        struct {
>                unsigned int a : 4;
>                unsigned int b : 4;
>                unsigned int c : 4;
>        };
>
> In the above example, you can touch <b> and <c> while setting <a>.  No race
> there.
>
> However, non contiguous bitfields are a different story:
>
>        struct {
>                unsigned int a : 4;
>                char b;
>                unsigned int c : 6;
>        };
>
> Here we have 3 distinct memory locations, so you can't touch <b> or <c>
> while setting <a>.  No bit twiddling allowed.
>
> Similarly for bitfields separated by a zero-length bitfield:
>
>        struct {
>                unsigned int a : 4;
>                int : 0;
>                unsigned int c : 6;
>        };
>
> In the above example, <a> and <c> are distinct memory locations.
>
> Also, a structure/union boundary will also separate previously contiguous
> bit sequences:
>
>        struct {
>                unsigned int a : 4;
>                struct { unsigned int b : 4 } BBB;
>                unsigned int c : 4;
>        };
>
> Here we have 3 distinct memory locations, so again, we can't clobber <b> or
> <c> while setting <a>.
>
> The patch below adds a non-contiguous bit test (bitfields-2.C) which passes
> on x86, but upon assembly inspection, fails on PPC64, s390, and Alpha.
>  These 3 architectures bit-twiddle their way out of the problem.

The memory model is not implementable on strict-alignment targets
that do not have a byte store operation.  But we previously said that ;)

Also consider global vars

char a;
char b;

accessing them on strict-align targets may access adjacent globals
(that's a problem anyway, also with alias analysis).

Richard.

> There is also a similar test already in the testsuite (bitfields.C) which is
> similar except one field is a volatile.  This test fails on x86-64 as well
> and is the subject of PR48128 which Jakub is currently tackling.
>
> As soon as Jakub finishes with PR48128, I will be working on getting these
> bitfield tests working in a C++ memory model fashion.
>
> Committing to branch.
>