[PATCH] D102110: [Inliner] Fix noalias metadata handling for instructions simplified during cloning (PR50270)

[Nikita Popov via Phabricator via llvm-commits]

nikic added inline comments.


================
Comment at: llvm/lib/Transforms/Utils/InlineFunction.cpp:997
   // instructions, add the alias scope metadata.
   for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
        VMI != VMIE; ++VMI) {
----------------
nikic wrote:
> jeroen.dobbelaere wrote:
> > Just wondering: don't we have a similar issue here ?
> Yes, unfortunately we do. It took me a while, but I found a suitable `llvm.masked.load` fold to exploit:
> 
> ```
> define <2 x i8> @callee(<2 x i8>* %ptr1, <2 x i8>* noalias %ptr2, <2 x i1> %mask, <2 x i8> %passthru) {
>   %ret = call <2 x i8> @llvm.masked.load.v2i8(<2 x i8>* %ptr1, i32 1, <2 x i1> %mask, <2 x i8> %passthru)
>   store <2 x i8> zeroinitializer, <2 x i8>* %ptr2
>   ret <2 x i8> %ret
> }
> 
> define void @caller(<2 x i8>* %ptr1, <2 x i8>* %ptr2) {
>   %passthru = load <2 x i8>, <2 x i8>* %ptr2
>   call <2 x i8> @callee(<2 x i8>* %ptr1, <2 x i8>* %ptr2, <2 x i1> zeroinitializer, <2 x i8> %passthru)
>   ret void
> }
> 
> declare <2 x i8> @llvm.masked.load.v2i8(<2 x i8>*, i32, <2 x i1>, <2 x i8>)
> ```
> Results in:
> ```
> define <2 x i8> @callee(<2 x i8>* %ptr1, <2 x i8>* noalias %ptr2, <2 x i1> %mask, <2 x i8> %passthru) {
>   %ret = call <2 x i8> @llvm.masked.load.v2i8.p0v2i8(<2 x i8>* %ptr1, i32 1, <2 x i1> %mask, <2 x i8> %passthru)
>   store <2 x i8> zeroinitializer, <2 x i8>* %ptr2, align 2
>   ret <2 x i8> %ret
> }
> 
> define void @caller(<2 x i8>* %ptr1, <2 x i8>* %ptr2) {
>   %passthru = load <2 x i8>, <2 x i8>* %ptr2, align 2, !noalias !0
>   call void @llvm.experimental.noalias.scope.decl(metadata !0)
>   store <2 x i8> zeroinitializer, <2 x i8>* %ptr2, align 2, !alias.scope !0
>   ret void
> }
> ```
> 
> In this case, we don't even need the incorrect annotation to go on an instruction from the caller, it can also happen with a callee instruction:
> 
> ```
> define <2 x i8> @callee(<2 x i8>* %ptr1, <2 x i8>* noalias %ptr2, <2 x i1> %mask) {
>   %passthru = load <2 x i8>, <2 x i8>* %ptr2
>   %ret = call <2 x i8> @llvm.masked.load.v2i8(<2 x i8>* %ptr1, i32 1, <2 x i1> %mask, <2 x i8> %passthru)
>   store <2 x i8> zeroinitializer, <2 x i8>* %ptr2
>   ret <2 x i8> %ret
> }
> 
> define void @caller(<2 x i8>* %ptr1, <2 x i8>* %ptr2) {
>   call <2 x i8> @callee(<2 x i8>* %ptr1, <2 x i8>* %ptr2, <2 x i1> zeroinitializer)
>   ret void
> }
> 
> declare <2 x i8> @llvm.masked.load.v2i8(<2 x i8>*, i32, <2 x i1>, <2 x i8>)
> ```
> Results in:
> ```
> define <2 x i8> @callee(<2 x i8>* %ptr1, <2 x i8>* noalias %ptr2, <2 x i1> %mask) {
>   %passthru = load <2 x i8>, <2 x i8>* %ptr2, align 2
>   %ret = call <2 x i8> @llvm.masked.load.v2i8.p0v2i8(<2 x i8>* %ptr1, i32 1, <2 x i1> %mask, <2 x i8> %passthru)
>   store <2 x i8> zeroinitializer, <2 x i8>* %ptr2, align 2
>   ret <2 x i8> %ret
> }
> 
> define void @caller(<2 x i8>* %ptr1, <2 x i8>* %ptr2) {
>   call void @llvm.experimental.noalias.scope.decl(metadata !0)
>   %passthru.i = load <2 x i8>, <2 x i8>* %ptr2, align 2, !alias.scope !0, !noalias !0
>   store <2 x i8> zeroinitializer, <2 x i8>* %ptr2, align 2, !alias.scope !0
>   ret void
> }
> ```
> 
> Note how the metadata is claiming that the load does not alias ... with itself.
> 
> So, in this case the problem is a bit different from the other two: We have to not just prevent annotating instructions from the caller, we should prevent annotating //any// instructions for which some kind of folding occurred.
Anyone know how to fix this nicely? I don't think we have any information on whether an instruction was simplified or not at this point. The only thing that comes to mind is to go through all instructions and check which ones occur in the map multiple times.

Do we need to change the cloning implementation to track simplification info in a side table?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D102110/new/

https://reviews.llvm.org/D102110