[PATCH 1/3] Call recompute_tree_invariant_for_addr_expr when replacing a constant in an ADDR_EXPR.

[Richard Guenther]

On Wed, Nov 24, 2010 at 1:13 AM, Sebastian Pop <sebpop@gmail.com> wrote:
> On Tue, Nov 23, 2010 at 06:45, Richard Guenther
> <richard.guenther@gmail.com> wrote:
>> On Tue, Nov 23, 2010 at 12:18 AM, Sebastian Pop <sebpop@gmail.com> wrote:
>>> Committed to the graphite branch.
>>>
>>> 2010-11-22  Sebastian Pop  <sebastian.pop@amd.com>
>>>
>>>        * sese.c (rename_uses): Call recompute_tree_invariant_for_addr_expr
>>>        when replacing a constant in an ADDR_EXPR.
>>> ---
>>>  gcc/ChangeLog.graphite |    5 +++++
>>>  gcc/sese.c             |   13 +++++++++++++
>>>  2 files changed, 18 insertions(+), 0 deletions(-)
>>>
>>> diff --git a/gcc/ChangeLog.graphite b/gcc/ChangeLog.graphite
>>> index c2da86f..e5df72e 100644
>>> --- a/gcc/ChangeLog.graphite
>>> +++ b/gcc/ChangeLog.graphite
>>> @@ -1,3 +1,8 @@
>>> +2010-11-22  Sebastian Pop  <sebastian.pop@amd.com>
>>> +
>>> +       * sese.c (rename_uses): Call recompute_tree_invariant_for_addr_expr
>>> +       when replacing a constant in an ADDR_EXPR.
>>> +
>>>  2010-10-20  Sebastian Pop  <sebastian.pop@amd.com>
>>>
>>>        * graphite-blocking.c (lst_do_strip_mine_loop): Extra parameter
>>> diff --git a/gcc/sese.c b/gcc/sese.c
>>> index ffcb51c..dbb0cd7 100644
>>> --- a/gcc/sese.c
>>> +++ b/gcc/sese.c
>>> @@ -547,6 +547,19 @@ rename_uses (gimple copy, htab_t rename_map, gimple_stmt_iterator *gsi_tgt,
>>>                                       true, NULL_TREE);
>>>       gsi_insert_seq_before (gsi_tgt, stmts, GSI_SAME_STMT);
>>>       replace_exp (use_p, new_expr);
>>> +
>>> +
>>> +      if (TREE_CODE (new_expr) == INTEGER_CST)
>>> +       {
>>> +         tree lhs = gimple_assign_lhs (copy);
>>> +         tree rhs = gimple_assign_rhs1 (copy);
>>> +
>>> +         if (TREE_CODE (lhs) == ADDR_EXPR)
>>
>> The lhs of an assign can never be ADDR_EXPR.
>>
>
> I removed this.
>
>>> +           recompute_tree_invariant_for_addr_expr (lhs);
>>> +         if (TREE_CODE (rhs) == ADDR_EXPR)
>>> +           recompute_tree_invariant_for_addr_expr (rhs);
>>
>> And I'm confused anyway - of new_expr is an INTEGER_CST, how
>> come it is an ADDR_EXPR?
>>
>
> rhs is an ADDR_EXPR and we replace some SSA_NAME that indexes into
> that array with a constant:
> for instance, I had an ICE when replacing "i" with "9" in A[i], as in
> the new form, A[9] is invariant.

Ah, I see.

Richard.

> Sebastian
>