Re: Merging changes to volatile bitfield structs

["Phil Endecott" <spam_from_gcc_help_2 at chezphil dot org>]

John Love-Jensen wrote:
> Hi Phil,
>
> > Can anyone share their preferred approach to this sort of problem?
>
> Reg workingReg = reg;
> workingReg.a = 0;
> workingReg.b = 0;
> workingReg.c = 0;
> reg = workingReg;

Thanks, that's obviously a sensible approach.

One subtlety is that if my reg is 32 bits and all of my changes are 
within 8 bits of it, I would prefer to do a byte read-modify-write (or 
even just a byte write), while that will (I think) do a word read-modify-write.

Does anyone have any thoughts about the idea of not declaring it 
volatile and placing "memory barriers" around it?  Google tells me that 
I should use an asm statement that says it "clobbers memory" as a 
memory barrier.  So if I write

#define MEMORY_BARRIER asm("":"":"":"":"memory")  // haven't checked syntax

struct ctrl_regs* regs;
// regs gets initialised to point to some hardware
MEMORY_BARRIER
regs->a = 0;
regs->b = 0;
MEMORY_BARRIER

or something equivalent with a reference, will gcc do what I want?


Regards,

Phil.