[PATCH] D47023: Limit the number of phis in intptr/ptrint folding

[Michael Zolotukhin via Phabricator via llvm-commits]

mzolotukhin added a comment.

I thought about this transformation more, and I no longer think that we even need to move it to aggressive-instcombine (or another FunctionPass). What we need is to just change it from top-down to bottom-up: i.e. to start looking not from phi-nodes, but rather from `inttoptr` instructions. That is, the algorithm would look like:

  visitIntToPtr(Instruction &I) {
    Value *Def = I.getOperand()
    if (!Def.hasSingleUse())
       return;
    if (isa<PtrToInt>(Def)) {   // Simple case without phi - it's probably already handled somewhere else, but I'm putting it here for completeness
       I.replaceAllUsesWith(Def.getOperand());
    }
    if (isa<PHINode>(Def)) {    // Interesting case where we have a phi-node
       if (all operands are PtrToInt with a single use) {
         NewPHI = RewritePHI();
         I.replaceAllUsesWith(NewPHI);
      }
    }
  }

What do you think? Would it work?

PS: Internally we worked around the slowdown, so it's not pressing on us anymore.

Michael


Repository:
  rL LLVM

https://reviews.llvm.org/D47023