Re: [PATCH 1/6]tree-ssa-dom.c: Normalize exprs, starting with ARRAY_REF to MEM_REF

[Richard Biener <richard dot guenther at gmail dot com>]

On Fri, Oct 30, 2015 at 6:35 AM, Jeff Law <law@redhat.com> wrote:
> On 10/29/2015 01:18 PM, Alan Lawrence wrote:
>>
>> This patch just teaches DOM that ARRAY_REFs can be equivalent to MEM_REFs
>> (with
>> pointer type to the array element type).
>>
>> gcc/ChangeLog:
>>
>>         * tree-ssa-dom.c (dom_normalize_single_rhs): New.
>>         (dom_normalize_gimple_stmt): New.
>>         (lookup_avail_expr): Call dom_normalize_gimple_stmt.
>
> Has this been tested?  Do you have tests where it can be shown to make a
> difference independent of the changes to tree-sra.c?
>
> The implementation looks fine, I just want to have some basic tests in the
> testsuite that show the benefit of this normalization.

Err, I think the implementation is extremely wasteful ...

> Similarly for patch #2. Interestingly enough we had code that made that kind
> of transformation during gimplification eons ago.  Presumably it was ripped
> out at some point because of the differences in aliasing properties.

Please have a look at how SCCVN handles variants of memory references.
You might even want to re-use it's copy_reference_ops_from_ref implementation
(and reference hashing).  Note that SCCVN needs to be able to reconstruct a
tree expression from its representation (for PRE) which DOM needs not, so
DOM might use a simpler form.  The basic idea is to accumulate constant
offset handled_components into a "offset component".  Thus

  a[2].i                     ->  (&a)->offset(8 + offsetof(i))
  MEM_REF[&a, 8].i ->  (&a)->offset(8 + offsetof(i))

for DOM you can probably do the copy_reference_ops_from_ref work
on-the-fly for hashing and comparing.  The main point will be to forgo
with the DOM way of hashing/comparing for memory references.

Richard.

>
>
> Jeff