[llvm-branch-commits] [llvm] [SDAG] Introduce inbounds flag for pointer arithmetic (PR #131862)

[Eli Friedman via llvm-branch-commits]

efriedma-quic wrote:

Saying "one side is inbounds of the other side" is basically useless, as far as I can tell, for almost any transform.

The other possibility you mentioned is that we say one side is a constant, the other is not, and the non-constant side must be a pointer?  That seems fragile.  The semantics of the add depend on the "syntactic" form of the operands, which could be rewritten.  Not really great.

Maybe we could consider adding "ISD::PTRADD"?  Lowers to ISD::ADD by default, but targets that want to do weird things with pointer arithmetic could do them.

One other concern, which applies to basically any formulation of this: Since SelectionDAG doesn't have a distinct pointer type, you can't tell whether the pointer operand was produced by an inttoptr.  So in some cases, you have an operation marked "inbounds", but it's ambiguous which object it's actually inbounds to.  This isn't really a problem at the moment because we do IR-level transforms that remove inttoptr anyway, but if we ever do resolve the IR-level issues, we should have some idea for how we propagate the fix to SelectionDAG.

https://github.com/llvm/llvm-project/pull/131862