This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Make nonoverlapping_component_refs work with duplicated main variants

> For consistency yes I guess but IIRC they cannot really appear in 

OK, i tought that if I put SVE into structures, we may end up with
> > +      /* Different fields of the same record type cannot overlap.
> > +	 ??? Bitfields can overlap at RTL level so punt on them.  */
> > +      if (DECL_BIT_FIELD (field1) && DECL_BIT_FIELD (field2))
> > +	return 0;
> > +
> don't you need the DECL_BIT_FIELD_REPRESENTATIVE check here as well?
> I'd do
>         if (DECL_BIT_FIELD_REPRESENTATIVE (field1))
>           field1 = DECL_BIT_FIELD_REPRESENTATIVE (field1);
>         if (DECL_BIT_FIELD_REPRESENTATIVE (field2))
>           field2 = DECL_BIT_FIELD_REPRESENTATIVE (field2);
> thus use the representative for the overlap check.  It might
> be the case that we can improve here and if we do this
> can do the DECL_BIT_FIELD check after this (hoping the
> representative doesn't have it set).

> > +      if (tree_int_cst_equal (DECL_FIELD_OFFSET (field1),
> > +			      DECL_FIELD_OFFSET (field2))
> > +	  && tree_int_cst_equal (DECL_FIELD_BIT_OFFSET (field1),
> > +				 DECL_FIELD_BIT_OFFSET (field2)))
> > +	return 0;
> In gimple_compare_field_offset this was fast-pathed for
> DECL_OFFSET_ALIGN (f1) == DECL_OFFSET_ALIGN (f2) so I suggest to
> do that here as well.  Note that DECL_FIELD_OFFSET can be
> a non-constant which means you cannot use tree_int_cst_equal
> unconditionally here but you have to use operand_equal_p.

tree_int_cst_equal will return false if offsets are not INTEGER_CST.
I was not sure if I can safely use operand_equal_p.  What happens for
fields with variable offsets when I inline two copies of same function
which takes size as parameter and make the size different? Will I get
here proper SSA name so operand_equal_p will work?

If so, I still see no point for fast-path for DECL_OFFSET_ALIGN. In many
cases BIT_OFFSET will be just 0, so even if offset alignments are
different we are likely going to hit this fast path avoiding parsing
trees later.
> > +      /* Note that it may be possible to use component_ref_field_offset
> > +	 which would provide offsets as trees. However constructing and folding
> > +	 trees is expensive and does not seem to be worth the compile time
> > +	 cost.  */
> > +
> > +      poly_uint64 offset1, offset2;
> > +      poly_uint64 bit_offset1, bit_offset2;
> > +      poly_uint64 size1, size2;
> I think you need poly_offset_int here since you convert to bits below.
> The gimple_compare_field_offset checking way looks cheaper btw, so
> I wonder why you don't simply call it but replicate things here?
> When do we expect to have partially overlapping field decls?  Even
> when considering canonical type merging?

Because the types I am comparing may not have same canonical types.

nonoverlapping_component_refs_since_match_p is called when we prove that
base pointers are the same (even with -fno-strict-aliasing).  In such
cases the access paths may be based on completely different types. The
point of nonoverlapping_component_refs_since_match_p is to match them as
far as possible when they are semantically equivalent in hope to get
non-overlapping refs in the last step.

This is stronger than the get_base_ref_and_extend based check in
presence of non-constant ARRAY_REFs.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]