[llvm] [CFG] Avoid introducing complex phi when removing empty blocks (PR #104887)

Nikita Popov via llvm-commits llvm-commits at lists.llvm.org
Tue Aug 20 00:59:42 PDT 2024


================
@@ -1040,6 +1045,42 @@ CanRedirectPredsOfEmptyBBToSucc(BasicBlock *BB, BasicBlock *Succ,
   return true;
 }
 
+// Check whether removing BB will make the phis in its Succ will have too
+// many incoming entries. This function does not check whether BB is foldable
+// or not.
+static bool introduceTooComplexPhi(BasicBlock *BB) {
+  // Check BB only has phi and an unconditional branch
+  BranchInst *Branch = dyn_cast<BranchInst>(BB->getFirstNonPHIOrDbg(true));
+  assert(Branch && Branch->isUnconditional() && "BB is not an empty block");
----------------
nikic wrote:

Branch is unused outside the assertion, and getFirstNonPHIOrDbg() may be expensive.

I think it would be better to pass in BasicBlock *Succ to this function, like many other parts of this transform. The fact that it's an empty block with single successor is verified elsewhere.

https://github.com/llvm/llvm-project/pull/104887


More information about the llvm-commits mailing list