llvm-ir: TBAA and struct copies

Project Infrastructure LLVM Dev List Archives
1

Hi,

I have a question about the current definition of TBAA (See [1]).

In the LLVM-IR code that we produce, we generate load/stores of struct types. (See [2] and [3] for a godbolt example showing the issue)

For following c-alike code:

  struct S { int dummy; short e, f; } x,y;
  struct S* p = &x;
  int foobar() {
    x.f=42;
    *p=y; //**** struct copy
    return x.f;
  }

We produce:

  [...]
  %struct.S = type { i32, i16, i16 }
  ; Function Attrs: norecurse nounwind uwtable
  define dso_local i32 @foobar() local_unnamed_addr #0 {
    store i16 42, i16* getelementptr inbounds (%struct.S, %struct.S* @x, i64 0, i32 2), align 2, !tbaa !2
    %1 = load %struct.S*, %struct.S** @p, align 8, !tbaa !8
    %2 = load %struct.S, %struct.S* @y, align 8, !tbaa !10 ;*************
    store %struct.S %2, %struct.S* %1, align 4, !tbaa !10 ; *************
    %3 = load i16, i16* getelementptr inbounds (%struct.S, %struct.S* @x, i64 0, i32 2), align 2, !tbaa !2
    %4 = sext i16 %3 to i32
    ret i32 %4
  }
  [...]
  !0 = !{i32 1, !"wchar_size", i32 4}
  !1 = !{!"clang version 9.0.0 (trunk 362464)"}
  !2 = !{!3, !7, i64 6}
  !3 = !{!"_ZTS1S", !4, i64 0, !7, i64 4, !7, i64 6}
  !4 = !{!"int", !5, i64 0}
  !5 = !{!"omnipotent char", !6, i64 0}
  !6 = !{!"Simple C++ TBAA"}
  !7 = !{!"short", !5, i64 0}
  !8 = !{!9, !9, i64 0}
  !9 = !{!"any pointer", !5, i64 0}
  !10 = !{!3, !3, i64 0} ; ***********************

In order to allow other optimizations, we have also attached TBAA information to the load/store of the struct (!tbaa !10).
The logical solution would be to construct '!tbaa !10' as shown:
  - the base type is '!3 (_ZTS1S)'
  - the access type is also '!3 (_ZTS1S)'
  - the offset is 'i64 0'

I observe following issues:
  - issue 1: the tbaa semantics will not detect aliasing between '!10' and '!2' (See [2] and [3]; the load i16 in bar_wrong should not be optimized away)
  - issue 2: according to the pure definition of the 'access tags', the base type and the access type can not be the same struct type.
      As such, the provided example could be found to be 'invalid'. Still, by adding an extra indirection, a similar 'valid' example (with wrong behavior)
      can be constructed. See [3]

  - For 'issue 2', I think that the definition of 'access tag' should be relaxed, allowing the description of a full copy' of a struct.
  - For 'issue 1', the lookup algorithm should be enhanced so that aliasing can be detected for these cases.

Is this a correct interpretation ? Any input is welcome !

Thanks,

Jeroen Dobbelaere

2

Hello Jeroen,

AFAIR, with the current TBAA format the access type shall never be an aggregate, so the !10 node looks suspicious.

Tried to reproduce this on 362464 with that same input snippet, but got something completely different, see below.

Regards,

3

Hi Ivan,

The code that we have is indeed different from what the 'standard llvm' expects. Let me explain:

in our version we came into this situation in two steps:
1) I added support for 'special types' that map directly to types supported by hardware.
These types are represented by a struct containing a single iXXX member, providing the necessary bits
of the type, and at the same time avoiding any change in the bits (if it would have been represented
by a pure iXXX). They need to be handled similarly as a fundamental type, so we provided support for that,
resulting in clang generating load/store instructions that work on a 'struct'.

In order to get this working (for tbaa), I had to relax the check for access type, as it now could see an aggregate type
(that behaved as a fundamental type).

2) later on, we noticed that sroa (if I remember correctly) produced better code if it could work on load/store instructions of aggregates instead of
llvm.memcpy. I adapted clang to also emit load/store pairs for small structs. Just like with the special 'struct type', we the load/store's are decorated with
the struct and the size of the struct as access TBAAAccessInfo information.

At that time all looked fine, but recently, csmith testing has been showing cases where alias analysis is now too aggressive.

An easy way out is of course to fall back to llvm.memcpy, but that blocks a lot of possible optimizations, as it loses all the TBAA information.

(In the original example: because of the llvm.memcpy, it does no matter if the *p=y copy is for a 'S' or for a different 'struct T'. In both cases, it blocks the optimization of the 'x.f' )

With the current understanding that I have about the problem and about the TBAA implementation, I am not sure that it is easy to fix with the default struct path tbaa.
With the 'new struct path tbaa', as it also tracks the access size, I do see a possible solution:

in TypeBasedAliasAnalysis.cpp, in function bool mayBeAccessToSubobjectOf(...):

https://github.com/llvm/llvm-project/blob/master/llvm/lib/Analysis/TypeBasedAliasAnalysis.cpp#L619

  if (BaseType.getNode() == SubobjectTag.getBaseType()) {
        bool SameMemberAccess = OffsetInBase == SubobjectTag.getOffset(); // **** HERE
        if (GenericTag) {
          *GenericTag = SameMemberAccess ? SubobjectTag.getNode() :
                                           createAccessTag(CommonType);
        }
        MayAlias = SameMemberAccess;
        return true;
  }

For the 'NewFormat', The 'MayAlias' should be based on overlap of {OffsetInBase, BaseTag.getSize} and {SubobjecTag.getOffset(), SubobjectTag.getSize()}
I am not sure if the same holds for the 'GenericTag'.

Do you think that would cause other problems ?

Thanks,

Jeroen Dobbelaere

4

Hi Jeroen,

> With the current understanding that I have about the problem and about the TBAA implementation, I am not sure that it is easy to fix with the default struct path tbaa.

To handle aggregate accesses correctly, we need to know the size of the access in addition to the offset. Since the current format doesn't store sizes, it doesn't seem possible to do much more than we already do about aggregate accesses. This is further complicated by the fact that calls like memcpy() have their own format of metadata nodes.

> For the 'NewFormat', The 'MayAlias' should be based on overlap of {OffsetInBase, BaseTag.getSize} and {SubobjecTag.getOffset(), SubobjectTag.getSize()}
> I am not sure if the same holds for the 'GenericTag'.

Correct. As to the generic tag, this needs some very careful analysis and extensive testing, but generally I would expect it to be built so that:
- its base type is the matching type in the considered access chains,
- its access type is the generic type of the two access types being matched and
- its range (offset + size) is the union of the matched access ranges.

Regards,

5

Thanks !

I'l try to come up with a patch to achieve this.

Greetings,

Jeroen Dobbelaere