On Tue, May 10, 2016 at 10:18:59AM +0200, Richard Biener wrote:
> There are two options - not apply this folding if it doesn't preserve
> TBAA behavior (first patch) or preserve TBAA behavior by wrapping
> the base into a MEM_REF (which also requires fixing
> reference_alias_ptr_type, sth which we need to do anyway).
> See the two patches below.
>
> Incidentially the folding (which I still consider premature) prevents
> CSE on the GIMPLE level as well:
>
> <bb 2>:
> MEM[(struct &)this_2(D)] ={v} {CLOBBER};
> _4 = &this_2(D)->m_str;
> MEM[(struct &)this_2(D)] ={v} {CLOBBER};
> MEM[(struct short_t &)this_2(D)].h.is_short = 1;
> MEM[(struct short_t &)this_2(D)].h.length = 0;
> MEM[(struct short_t &)this_2(D)].data[0] = 0;
> _19 = BIT_FIELD_REF <MEM[(void *)this_2(D)], 8, 0>;
> _20 = _19 & 1;
> if (_20 != 0)
>
> Here we can't CSE the is_short access and optimize the compare.
> So on trunk I am thinking of at least removing the compare
> against constant case. [I've removed the whole code twice in
> GCC history and it always got installed back ...]
>
> Bootstrap and regtest of both patches running on x86_64-unknown-linux-gnu.
>
> Any preference for GCC 6?
I think I prefer the second patch for the branch.
Jakub
