[Mlir-commits] [mlir] [mlir] Add support for staged dataflow analyses (PR #192998)

[llvmlistbot at llvm.org]

Max191 wrote:

> I've been doing this with two solvers and no need for API changes or extra complexity. Solver1 runs the first stage of analyses. We then construct the second stage of analyses by passing const references to the analysis instances completed by Solver1 into their constructors. The second stage analyses can then query the results of the first stage without any possible interference between the stages.

Doing it this way will avoid running any mutually dependent analyses multiple times unnecessarily. For example, most analyses run together with DeadCodeAnalysis, so we would need to run it once for each solver. With this PR, we will only have to run it one time in the first phase, and then later phases can all access the saved analysis results without needing to run it again.

https://github.com/llvm/llvm-project/pull/192998