Re: [PATCH] Fix PRs 52080, 52097 and 48124, rewrite bitfield expansion, enable the C++ memory model wrt bitfields everywhere

[Richard Guenther <rguenther at suse dot de>]

On Thu, 22 Mar 2012, Eric Botcazou wrote:

> > Yes, I have noticed that from what you can see in the code.  So the
> > issue is that get_bit_range tells us that it's ok to touch bits
> > outside of the field - but that's obviously required.  We may not
> > change bits that are not covered by the FIELD_DECL which now
> > somehow happens?  That sounds like a latent bug in the bitfield
> > expander to me - it should never _change_ bits outside of bitpos/bitsize
> > as returned by get_inner_reference.
> 
> It's the new C++0x memory model stuff:
> 
> void
> store_bit_field (rtx str_rtx, unsigned HOST_WIDE_INT bitsize,
> 		 unsigned HOST_WIDE_INT bitnum,
> 		 unsigned HOST_WIDE_INT bitregion_start,
> 		 unsigned HOST_WIDE_INT bitregion_end,
> 		 enum machine_mode fieldmode,
> 		 rtx value)
> {
>   /* Under the C++0x memory model, we must not touch bits outside the
>      bit region.  Adjust the address to start at the beginning of the
>      bit region.  */
>   if (MEM_P (str_rtx)
>       && bitregion_start > 0)
>     {
>       enum machine_mode bestmode;
>       enum machine_mode op_mode;
>       unsigned HOST_WIDE_INT offset;
> 
>       op_mode = mode_for_extraction (EP_insv, 3);
>       if (op_mode == MAX_MACHINE_MODE)
> 	op_mode = VOIDmode;
> 
>       offset = bitregion_start / BITS_PER_UNIT;
>       bitnum -= bitregion_start;
>       bitregion_end -= bitregion_start;
>       bitregion_start = 0;
> 
> which assumes that a bitfield group starts on a byte boundary.  So this is 
> probably indeed latent in GCC 4.7 as well.

But the bitfield group _does_ start on a byte boundary.  At least
that's what the new code in stor-layout ensures.

It's ok to assume the group starts on a byte boundary.  But it's not
ok to modify bits outside of the access, so the RMW cycle has to mask
them properly.

Richard.