On Mon, 24 Oct 2022, Andre Vieira (lists) wrote:

> 
> On 24/10/2022 13:46, Richard Biener wrote:
> > On Mon, 24 Oct 2022, Andre Vieira (lists) wrote:
> >
> >> On 24/10/2022 08:17, Richard Biener wrote:
> >>> Can you check why vect_find_stmt_data_reference doesn't trip on the
> >>>
> >>>     if (TREE_CODE (DR_REF (dr)) == COMPONENT_REF
> >>>         && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (dr), 1)))
> >>>       {
> >>>         free_data_ref (dr);
> >>>         return opt_result::failure_at (stmt,
> >>>                                        "not vectorized:"
> >>>                                        " statement is an unsupported"
> >>>                                        " bitfield access %G", stmt);
> >>>       }
> >> It used to, which is why this test didn't trigger the error before my
> >> patch,
> >> but we lower it to BIT_FIELD_REFs in ifcvt now so it is no longer a
> >> DECL_BIT_FIELD.
> >>
> >> But that is a red-herring, if you change the test structure's 'type Int24
> >> is
> >> mod 2**24;' to 'type Int24 is mod 2**32;', thus making the field we access
> >> a
> >> normal 32-bit integer, the field no longer is a DECL_BIT_FIELD and thus my
> >> lowering does nothing. However, you will still get the failure because the
> >> field before it is a packed 4-bit field, making the offset to the field we
> >> are
> >> accessing less than BITS_PER_UNIT.
> > Hmm, so the _intent_ of DECL_BIT_FIELD_REPRESENTATIVE is to definitely
> > _not_ be a DECL_BIT_FIELD (well, that's the whole point!).   So this
> > shows an issue with setting up DECL_BIT_FIELD_REPRESENTATIVE?  Of course
> > for a type with an alignment less than BITS_PER_UNIT (is StructB actually
> > such a type?) there cannot be a representative that isn't, so maybe
> > we should then set DECL_BIT_FIELD on it with a condition like Eric
> > mentions?
> I could do this, but it would not resolve the latent issue as I could still
> reproduce it without using any of the bitfield lowering code, see below.
> >
> >>> ?  I think we should amend this check and I guess that
> >>> checking multiple_p on DECL_FIELD_BIT_OFFSET should be enough?
> >> That won't work either, unless we do the same walk-through the full access
> >> as
> >> we do in get_inner_reference.
> > I suppose we should not "if-convert" bit field accesses with a
> > DECL_BIT_FIELD representative.  There isn't any benefit doing that
> > (not for general bitfield lowering either).
> Changing if-convert would merely change this testcase but we could still
> trigger using a different structure type, changing the size of Int24 to 32
> bits rather than 24:
> package Loop_Optimization23_Pkg is
>   type Nibble is mod 2**4;
>   type Int24  is mod 2**32;  -- Changed this from 24->32
>   type StructA is record
>     a : Nibble;
>     b : Int24;
>   end record;
>   pragma Pack(StructA);
>   type StructB is record
>     a : Nibble;
>     b : StructA;
>   end record;
>   pragma Pack(StructB);
>   type ArrayOfStructB is array(0..100) of StructB;
>   procedure Foo (X : in out ArrayOfStructB);
> end Loop_Optimization23_Pkg;
> 
> This would yield a DR_REF (dr): (*x_7(D))[_1].b.b  where the last 'b' isn't a
> DECL_BIT_FIELD anymore, but the first one still is and still has the
> non-multiple of BITS_PER_UNIT offset. Thus passing the
> vect_find_stmt_data_reference check and triggering the
> vect_check_gather_scatter failure. So unless we go and make sure we always set
> the DECL_BIT_FIELD on all subsequent accesses of a DECL_BIT_FIELD 'struct'
> (which is odd enough on its own) then we are better off catching the issue in
> vect_check_gather_scatter ?

But it's not only an issue with scatter-gather, other load/store handling
assumes it can create a pointer to the start of the access and thus
requires BITS_PER_UNIT alignment for each of them.  So we need to fail
at data-ref analysis somehow.

Richard.