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

[Fabian Ritter via llvm-branch-commits]

ritter-x2a wrote:

> This seems semantically ambiguous.
> 
> In GlobalISel, you have G_PTR_ADD, and inbounds on that has an obvious meaning; G_PTR_ADD has basically the same semantics as getelementptr. But in SelectionDAG, we don't have that; we just have a plain ISD::ADD. How do you tell which operand is the pointer? The obvious heuristic is outright wrong: you can't assume a value is being used as a pointer just because its value is equal to the address of some variable.

I see your point. We could define the semantics of an inbounds ISD::ADD to be "The result is an address into an allocated object and (exactly) one of the operands is an address into the same allocated object."
I think this might be strong enough for the concrete goal that I have in mind (PR #131863), but I agree that it would be better to have a flag that is easier to make use of.

What do you think about a "ConstInBoundsOffset" flag instead (suggestions for better names would be much appreciated), meaning "One operand is a constant offset and the other is an address pointing into the same allocated object as the result"?
While a lot narrower in scope, it would be easier to make use of that.

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