> On 19 Dec 2025, at 7:41 pm, Richard Biener <[email protected]> wrote:
>
> External email: Use caution opening links or attachments
>
>
> On Fri, Dec 19, 2025 at 12:52 AM Kugan Vivekanandarajah
> <[email protected]> wrote:
>>
>> Hi,
>>
>> Thanks for the review.
>>
>>> On 15 Dec 2025, at 11:56 pm, Richard Biener <[email protected]> 
>>> wrote:
>>>
>>> External email: Use caution opening links or attachments
>>>
>>>
>>> On Sun, Dec 14, 2025 at 6:17 AM Kugan Vivekanandarajah
>>> <[email protected]> wrote:
>>>>
>>>> Hi,
>>>> This a patch fixes Bug 123067] by checking for partial aliasing in self 
>>>> write test in LICM.
>>>>
>>>> Bootstrapped and regression tested with no new regressions.
>>>>
>>>> gcc/ChangeLog:
>>>>
>>>> 2025-12-09  Kugan Vivekanandarajah  <[email protected]>
>>>>
>>>>       PR middle-end/123067
>>>>       * tree-ssa-loop-im.cc (is_self_write):
>>>>
>>>> gcc/testsuite/ChangeLog:
>>>>
>>>> 2025-12-09  Kugan Vivekanandarajah  <[email protected]>
>>>>
>>>>       PR middle-end/123067
>>>>       * gcc.dg/licm-self-write-partial-alias.c: New test.
>>>> Is this OK?
>>>
>>> +  /* Verify there is no partial aliasing.  */
>>> +  if (!mem_refs_may_alias_p (load_ref, store_ref,
>>> +                            &memory_accesses.ttae_cache, true))
>>> +    return true;  /* Disjoint: safe to hoist.  */
>>>
>>> this is redundant?  If they are not aliasing then the caller would
>>> already say so?
>>>
>>> +  /* They may alias. Verify exact same location.  */
>>> +  return (operand_equal_p (load_ref->mem.base, store_ref->mem.base, 0)
>>> +         && known_eq (load_ref->mem.size, store_ref->mem.size)
>>> +         && known_eq (load_ref->mem.offset, store_ref->mem.offset));
>>>
>>> this looks incomplete.  See mem_ref_hasher::equal.
>>> That is, dependent on ->ref_decomposed the compare should look different,
>>> merging .offset with the MEM_REF offset in  base.  Maybe we can factor
>>> out a helper like
>>>
>>> bool im_compare_access_position_and_size (ao_ref *ref1, ao_ref *ref2)
>>
>> Tried factoring out but this is  making it more complicated (due to the 
>> divergence). Here is the version I tested. Please let me know
>> If you want me to post the version with im_compare_access_position_and_size.
>
> I think this will now regress the case where the reference we want to optimize
> is variably indexed (with invariant index, of course).  I'm not sure we need
> the
>
> +         && load_ref->mem.volatile_p == store_ref->mem.volatile_p
> +         && (load_ref->mem.ref_alias_set == store_ref->mem.ref_alias_set
> +             /* We are not canonicalizing alias-sets but for the
> +                special-case we didn't canonicalize yet and the
> +                incoming ref is a alias-set zero MEM we pick
> +                the correct one already.  */
> +             || (!load_ref->ref_canonical
> +                 && (TREE_CODE (store_ref->mem.ref) == MEM_REF
> +                     || TREE_CODE (store_ref->mem.ref) == TARGET_MEM_REF)
> +                 && store_ref->mem.ref_alias_set == 0)
> +             /* Likewise if there's a canonical ref with alias-set zero.  */
> +             || (load_ref->ref_canonical
> +                 && load_ref->mem.ref_alias_set == 0)));
>
> part here, since the argument is not about TBAA.  It's just the
>
>  if (obj2->max_size_known_p ())
>    return (mem1->ref_decomposed
>            && ((TREE_CODE (mem1->mem.base) == MEM_REF
>                 && TREE_CODE (obj2->base) == MEM_REF
>                 && operand_equal_p (TREE_OPERAND (mem1->mem.base, 0),
>                                     TREE_OPERAND (obj2->base, 0), 0)
>                 && known_eq (mem_ref_offset (mem1->mem.base) *
> BITS_PER_UNIT + mem1->mem.offset,
>                              mem_ref_offset (obj2->base) *
> BITS_PER_UNIT + obj2->offset))
>                || (operand_equal_p (mem1->mem.base, obj2->base, 0)
>                    && known_eq (mem1->mem.offset, obj2->offset)))
>            && known_eq (mem1->mem.size, obj2->size)
>            && known_eq (mem1->mem.max_size, obj2->max_size)
> ..
>  else
>   return operand_equal_p (mem1->mem.ref, obj2->ref, 0);
>
> parts that are relevant.  You'll have to ensure max_size_known_p agrees
> or resort to alignment considerations to rule out partial overlaps.  I mostly
> suggested the factoring to have one place with the "delicate" handling.

Added a TODO with the link to the message and some context.
>
> But I see this is now very complicated so I'd say go with your original
> version which should be conservatively correct, just not perfect in
> allowing all opportunities.  I'll note down a TODO to try to factor this
> in a way that suits me which I guess is more effective than trying
> back-and-forth via reviews ;)

Attached is the patch that has the TODO and old simpler implementation.

Is this OK?

Thanks,
Kugan


>
> Thanks,
> Richard.
>
>> Thanks,
>> Kugan
>>
>>
>>
>>
>>>
>>> for this?
>>>
>>> Thanks,
>>> Richard.
>>>
>>>> Thanks,
>>>> Kugan
>>>>
>>>>
>>>>
>>>>
>>

Attachment: 0001-Bug-123067-V3-LICM-wrong-code.patch
Description: 0001-Bug-123067-V3-LICM-wrong-code.patch

Reply via email to