[patch] Fix virtual operands creation in vectorizer

[Dorit Nuzman]

This patch fixes the over-conservativeness in the aliasing information that
the vectorizer sets for the vector pointers it creates. Instead of blindly
adding all the subvars as may-aliases of the new tag, we call
get_ref_base_and_extent and overlap_subvar to add only the potentially
overlapping subvars as may-aliases of the new tag. Looking at the resulting
vops created for the vector store in the attached testcase, you can see
that with this patch only the one relevant subvar appears as a virtual def:

  #   SFT.33_52 = V_MAY_DEF <SFT.33_85>;
  MEM[base: D.2686_175] = vect_cst_.79_103;

instead of all the subvars before:

  #   SFT.29_132 = V_MAY_DEF <SFT.29_110>;
  #   SFT.30_133 = V_MAY_DEF <SFT.30_114>;
  #   SFT.31_134 = V_MAY_DEF <SFT.31_118>;
  #   SFT.32_135 = V_MAY_DEF <SFT.32_122>;
  #   SFT.33_136 = V_MAY_DEF <SFT.33_85>;
  #   SFT.34_137 = V_MAY_DEF <SFT.34_128>;
  MEM[base: D.2686_4] = vect_cst_.79_103;

Bootsrapped on powerpc-linux and tested on the vectorizer testcases.
I couldn't test bootstrap with vectorization enabled because of the failure
described here:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26197#c11

Comments?

thanks,
dorit

        * tree-ssa-alias.c (new_type_alias): Takes additional argument.
Calls
        get_ref_base_and_extent and overlap_subvar to add only relevant
        subvars as may-aliases.
        * tree-vect-transform.c (vect_create_data_ref_ptr): Call
new_type_alias
        with additional argument.
        * tree-flow.h (new_type_alias): Takes additional argument.

(See attached file: vect_improved_aliasing.patch.txt) (See attached file:
vect-test.c)


Richard Guenther <rguenther@suse.de> wrote on 12/01/2006 16:11:32:

> On Thu, 12 Jan 2006, Dorit Nuzman wrote:
>
> > Does this patch make sense?
> >
> > The patch introduces one change to the function that creates new
type-tags
> > for the newly created vector pointers: instead of adding all the
subvars of
> > 'var' as may-aliases of the new tag, we add only the subvar whose
offset is
> > equal to 'field_bit_offset' (a new argument to the function).
>
> You should use get_ref_base_and_extent () to find the range of possible
> uses and add all subvars that offset_overlaps_with_access ().  You can
> look at get_constraint_for_component_ref () to get an idea how this
> should work (though that fn only adds the first subvar found).
>
> Richard.
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: vect_improved_aliasing.patch.txt
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20060228/c74109ed/attachment.txt>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: vect-test.c
Type: application/octet-stream
Size: 1616 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20060228/c74109ed/attachment.obj>