[PATCH] D80249: CodeGen: Don't lazily construct MachineFunctionInfo

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added a comment.

In D80249#2158886 <https://reviews.llvm.org/D80249#2158886>, @nhaehnle wrote:

> MFI is mutable, and so referring back to the MachineFunction as part of the mutation is potentially something reasonable to do. So the point isn't to use data from the MachineFunction during construction of the MFI, but rather to allow the "child" object to have a backlink to its "parent".

MFI is always queried from the function itself, so there's no real context where you need the backlink. Having the MachineFunction in the constructor provides a huge footgun which I've had to fix the consequences of at least 5 different times. If you allow people to write their own quick and dirty machine analyses in the constructor, they will. Doing that at a random point causes confusion and hard to predict differences in target behavior.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D80249/new/

https://reviews.llvm.org/D80249