[PATCH] D124697: Distinguish between different forms of "address-taken" MachineBasicBlocks

[Eli Friedman via Phabricator via llvm-commits]

efriedma added inline comments.


================
Comment at: llvm/include/llvm/CodeGen/MachineBasicBlock.h:152
+  /// constant, this contains a pointer to that block.
+  BasicBlock *AddressTakenIRBlock = nullptr;
 
----------------
arsenm wrote:
> I meant go from the underlying IR block that's already there, not introduce a second reference. Blocks split from the end of of the block would not have an IR block reference
Directly hacking at getBasicBlock() would be basically impossible to land in a single patch. There are a bunch of places that use getBasicBlock(); we'd need to ensure that every existing use of getBasicBlock() doesn't depend on the existing splitting behavior. We'd need to land a patch similar to this, then audit/modify various users of getBasicBlock(), then hack at the code that creates MBBs.

I agree that the concept of the "corresponding IR block" is a bit fuzzy for a random basic block once you've started to run MachineFunction optimizations. But I'm not sure changing the rule for splitting BBs is helpful; we do a bunch of work to change the rules, and the result is that it's still fuzzy. I mean, knowing the BB tied to the initial MBB is useful specifically if you want some IR element tied to the beginning of the block, like a blockaddress or landingpad. But not so useful if you want to query the terminator the IR block used, like some other places do.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D124697/new/

https://reviews.llvm.org/D124697