[PATCH] D76961: [BranchFolder] don't remove MBB's that have their address taken

[Eli Friedman via Phabricator via llvm-commits]

efriedma added a comment.

> Probably in InstrEmitter::EmitSpecialNode.

We normally construct the MachineFunction CFG much earlier; see SelectionDAGBuilder::visitCallBr.

It looks like the problem is that we split the "block" into two MBBs, and the successors end up attached to the second block, instead of the first.  Not sure where that happens, off the top of my head.

> Maybe blockaddress Constants could be generated as late as possible. Instead, we'd pass BasicBlock operands around in the Instruction level IR, then lower to MachineBasicBlock operands at the MachineInstr level, then only generate blockaddress operands very late when lowering to MCInst level?

I think the interaction between blockaddresses and indirectbr is fine.  The indirection of blockaddress constants is actually helpful when you're dealing with random instructions that aren't indirectbr/callbr.

For callbr, the indirection isn't so helpful, so yes, it would make sense to just use BasicBlock/MachineBasicBlock directly.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D76961/new/

https://reviews.llvm.org/D76961