Why is folding @llvm.assumes into operand bundle assumes behind --enable-knowledge-retention?

[philnik777]

Currently you have to pass -mllvm --enable-knowledge-retention for LLVM to fold

  %cmp = icmp ne ptr %ptr, null
  call void @llvm.assume(i1 %cmp)

into

  call void @llvm.assume(i1 true) [ "nonnull"(ptr %ptr) ]

AFAICT the operand bundles seem to be strictly superior for optimization compared to an equivalent instruction based version. Is there a reason this fold isn't enabled by default/unconditionally?

[llvmbot]

@llvm/issue-subscribers-llvm-mc

Author: Nikolas Klauser (philnik777)

Currently you have to pass `-mllvm --enable-knowledge-retention` for LLVM to fold ```llvm %cmp = icmp ne ptr %ptr, null call void @llvm.assume(i1 %cmp) ``` into ```llvm call void @llvm.assume(i1 true) [ "nonnull"(ptr %ptr) ] ```

AFAICT the operand bundles seem to be strictly superior for optimization compared to an equivalent instruction based version. Is there a reason this fold isn't enabled by default/unconditionally?