[Mlir-commits] [mlir] [mlir][CallGraph] Fix abstract edge connected to external node (PR #116177)

[Markus Böck]

https://github.com/zero9178 commented:

Conceptually makes a lot of sense to me, although the CI failure seems to be real.

I am also wondering whether the logic currently takes "address taken" into account. Consider e.g.:
```mlir
func.func @foo() -> (() -> ()) {
   %0 = func.constant @private_func
   return %0
}
```
does this properly get a call edge from an unknown or external caller to the private function?

LLVM e.g. actually uses the external caller node both for public functions but also ones whose address is taken. If we don't have any edges to functions whose addresses are taken then we are now at a risk of having a too optimistic rather than the conservatice callgraph we previously had until now

https://github.com/llvm/llvm-project/pull/116177