[llvm-dev] An ambiguity in TBAA info format

[Ivan Kosarev via llvm-dev]

On 31/10/17 01:48, Hal Finkel wrote:
> On 10/30/2017 04:57 PM, Ivan Kosarev via llvm-dev wrote:
>> Hello,
>>
>> Consider these two TBAA access tags:
>>
>> !1 = !{!5, !5, i64 0}
>> !3 = !{!7, !7, i64 0}
>>
>> !5 = !{!"A", !9}
>> !7 = !{!"B", !9}
>
> I'd find this email less confusing if you'd write out all of your 
> examples using the current TBAA format (not using some forms that will 
> be auto-upgraded). I don't believe we generate any TBAA fields with 
> only two operands currently. Instead, we'd generate here:
>
> !5 = !{!"A", !9, i64 0}
>
> In the current scheme, scalar edges in the type-aliasing graph look 
> just like structure fields (just all at offset zero).

Correct, but the point is, we still support such two-field forms in 
input files so the new format shouldn't rely on them.

>
>>
>> The tag !1 describes an access to an object of type "A" and !3 
>> describes an access to object of type "B".
>>
>> Both the type descriptors, !5 and !7, refer to node !9 as their type 
>> group. A definition of that node could look like this:
>>
>> !9 = !{"omnipotent char", ...}
>>
>> We know that these two accesses should be considered no-alias as 
>> neither of them encloses the other; the least common type group for 
>> them is !9. TypeBasedAAResult::Aliases() and 
>> MDNode::getMostGenericTBAA() respond accordingly and all is good.
>>
>> Then, let's change the definition for the node !9:
>>
>> !9 = !{"int", ...}
>>
>> Now it doesn't look like a type group, but rather a structure member. 
>> And nodes !5 and !7 now look as descriptors for structure types, with 
>> their offset fields added during auto-upgrade:
>>
>> !5 = !{!"A", !9, i64 0}
>> !7 = !{!"B", !9, i64 0}
>>
>> We know that, being interpreted as structure accesses, they still 
>> should be considered no-alias. However, the least common type group 
>> for these types is likely to be the "omnipotent char" node, but 
>> certainly not the type of the field, which is "int".
>
> I'm not sure what "likely" means in this context. 
> MDNode::getMostGenericTBAA does not currently seem to have a special 
> case for the first member of structure types. Instead, it does not 
> deal with them at all. MDNode::getMostGenericTBAA looks as the access 
> type, which should be a scalar, and collects the paths from those 
> types up to the root. Then it returns a scalar access tag for the type 
> which is common, and most distant from the root, along those paths.
>
> Are you trying to extend this to do something else with the 
> struct/member information on the accesses?

Yes, sorry I didn't mention that MDNode::getMostGenericTBAA() currently 
only considers final access types.

This is supposed to get us closer to the support for aggregate accesses, 
but at this moment I'm trying to fix MDNode::getMostGenericTBAA() to 
handle struct-path accesses. We do this in TypeBasedAAResult::Aliases() 
and I see no reasons why we shouldn't do the same on merging of access 
tags. Another goal is to use the same function for matching couples of 
tags and finding most generic tags for them. Here's what it would look like:

https://reviews.llvm.org/differential/diff/120871/

But for this we need to able to distinct members from parent types.

Thanks,


>
>  -Hal
>
>>
>> The problem is that since the formats for the member-of-structure and 
>> member-of-type-group relationships match, 
>> MDNode::getMostGenericTBAA() cannot disambiguate between them and 
>> always treat first members of structure types as type groups.
>>
>> To resolve this issue I'm thinking of changing the format of type 
>> nodes so that all of them, except root ones, refer to their type 
>> groups with their first operand. The scalar types "A" and "B" 
>> mentioned above would then be rewritten as:
>>
>> !5 = !{!9, !"A"}
>> !7 = !{!9, !"B"}
>>
>> !9 = !{..., "omnipotent char"}
>>
>> and their structure versions would read:
>>
>> !5 = !{!9, !"A", !11, i64 0}
>> !7 = !{!9, !"B", !11, i64 0}
>>
>> !11 = !{!9, "int"}
>>
>> The new format can be easily recognized by considering the type of 
>> the first operand: a string would mean the old format and a metadata 
>> node would suggest the new convention.
>>
>> The question to the community is, are there any reasons that wouldn't 
>> work or not desirable? Or, are there better alternatives to the 
>> proposed solution?
>>
>> As usual, any comments are highly appreciated.
>>
>> Thanks,
>>
>