This is similar to how SCCP (sparse conditional constant propagation) works in that it only considers that edges that are known possible to execute. In that case the edges are marked executable explicitly by the algorithm once you know you could take a certain path, as opposed to marking edges to see what would happen if only certain edges were considered executable.

To be blunt though, I don't think the idea is very interesting for a few reasons:

- It doesn't address how you decide which edges to speculatively consider executable.

- Optimizers intentionally do not play the "what if" game and try optimizing things multiple different ways to see what wins or what gains might be had. It simply gets too expensive (in compile time) too quickly.

- There are already other ways to achieve similar effect, e.g. superblock formation or simple tail duplication (which can be done for specific effect, e.g. see branch threading in LLVM).