[llvm] [LSR] Optimize lsr-term-fold if the operand of condition is PHINode (PR #96048)

[Philip Reames via llvm-commits]

================
@@ -6830,15 +6830,18 @@ canFoldTermCondOfLoop(Loop *L, ScalarEvolution &SE, DominatorTree &DT,
     return std::nullopt;
   }
 
-  if (!SE.hasLoopInvariantBackedgeTakenCount(L)) {
-    LLVM_DEBUG(dbgs() << "Cannot fold on backedge that is loop variant\n");
-    return std::nullopt;
-  }
-
   BasicBlock *LoopLatch = L->getLoopLatch();
   BranchInst *BI = dyn_cast<BranchInst>(LoopLatch->getTerminator());
   if (!BI || BI->isUnconditional())
     return std::nullopt;
+
+  if (!SE.hasLoopInvariantBackedgeTakenCount(L)) {
+    if (isa<SCEVCouldNotCompute>(SE.getExitCount(L, BI->getParent()))) {
----------------
preames wrote:

This change has nothing to do with the patch description.  Extending term folding to multiple exit loops is quite subtle, and will absolutely require it's own review.

https://github.com/llvm/llvm-project/pull/96048