[llvm-dev] TBAA for subset of a loaded value

Sanjoy Das via llvm-dev llvm-dev at lists.llvm.org
Mon Apr 10 18:44:07 PDT 2017


Hi Keno,

On April 10, 2017 at 6:02:38 PM, Keno Fischer via llvm-dev
(llvm-dev at lists.llvm.org) wrote:
> I'm interested in what we can do about TBAA for loads that the
> compiler inserts when optimizing loads into smaller loads (e.g. what
> SROA does). I'm gonna set the stage by using a small C snippet,
> because I think C has the best-understood implementation of TBAA among
> the folks on the list. However, I was unable to actually come up with
> an example where this inhibits optimizations coming from C code. I
> myself have this problem as a real problem in a non-C frontend, which
> allows for more stringent AA. With that in mind, consider
>
> void foo(unsigned int *data, int128_t *value) {
> for (int i = 1; i < 100; ++i) {
> data[i] += (int)*value;
> }
> }
>
> This will get lowered to something like (pseudo llvm IR here)
>
> for.body:
> %l = load i128, i128* %value, !tbaa !1
> %t = trunc %l to i32
> %gep = ...
> %old = load i32, i32* %gep, !tbaa !2
> %new = add i32 %t, %old
> store i32 %new, %gep, !tbaa !2
>
> Now, it seems like a perfectly reasonable optimization for the
> optimizer (though it currently doesn't do that), to fold %l and %t to:
>
> %b = bitcast i128* %value to i32*
> %t = load i32, i32 *%b
>
> However, what can we say about the tbaa of %t at this point? With the
> current semantics, I'm not sure we're allowed to say anything (in this
> case maybe we're still allowed to say !tbaa !1, but consider what
> happens for say *value>>32, which we could fold to gep+load). However,
> that would mean that we lose the information that %value does not
> alias %data, so e.g. we can't perform LICM anymore.
>
> Now, as I said, I couldn't actually get this to be a problem, because
> C only uses scalar TBAA for rather small values. However, in my
> frontend, for certain memory access I know that there are no interior
> pointer, so I emit scalar TBAA for a larger struct. Then SROA comes
> along and splits it, but it drops the TBAA.
>
> So the main points of discussion are:
> 1) Is there anything we can say about a smaller load created from a
> TBAA annotated larger load in the current semantics?
> 2) If not, what is the best way to retain this information. E.g.
> should we have a TBAA node that says "this access was n bytes into a
> scalar TBAA of type !1"?
>

I'm not sure if we can do something with the TBAA metadata we have
*today*, but I think there is hope if we're willing to change our TBAA
metadata representation.

In particular, if we allow access types[0] of TBAA tags to be struct
types then we can probably solve the problem.

In your case, we'd start with %l tagged with a scalar access type,
!i128ty:

  %l = load i128, i128* %value, !tbaa !{!i128ty, !i128ty, i64 0}
  %t = trunc %l to i32

which after narrowing would be

  %l = load i32, i32* %value, !tbaa !{!i128ty, !i32ty, i64 0}
  %t = trunc %l to i32

and we'd also change !i128ty from a scalar type to a struct type,
containing a !i32ty at the appropriate offset.  Of course, the loads
and stores which mention !i128ty as their access type now have a
struct type as their access type, which is why we have to make the
generalization I mentioned.

[0]: http://llvm.org/docs/LangRef.html#semantics

-- Sanjoy


> I would very much appreciate any guidance on the right approach to
> this (perfectly willing and planning to code this up myself - but
> would appreciate discussion on direction).
>
> Thanks,
> Keno
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>


More information about the llvm-dev mailing list