[llvm] [VPlan] Optimize resume values of IVs together with other exit values. (PR #174239)

[David Sherwood via llvm-commits]

================
@@ -742,6 +751,30 @@ void VPlanTransforms::createHeaderPhiRecipes(
     PhiR->replaceAllUsesWith(HeaderPhiR);
     PhiR->eraseFromParent();
   }
+
+  for (const auto &[HeaderPhiR, ScalarPhiR] : zip_equal(
+           drop_begin(HeaderVPBB->phis()), Plan.getScalarPreheader()->phis())) {
+    auto *ResumePhiR = cast<VPPhi>(&ScalarPhiR);
+    if (auto *FOR = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&HeaderPhiR)) {
+      ResumePhiR->setName("scalar.recur.init");
+      auto *ExtractLastLane = cast<VPInstruction>(ResumePhiR->getOperand(0));
+      ExtractLastLane->setName("vector.recur.extract");
+      if (ExtractLastLane->getOpcode() == VPInstruction::ExitingIVValue) {
+        // The FOR's backedge value is shared with an induction and the extract
+        // chain was replaced. Create a fresh one from the backedge value.
----------------
david-arm wrote:

Why do we need to create a fresh one? It would be good to explain the rationale here. I presume it's because multiple uses of a ExitingIVValue can prevent optimisations?

https://github.com/llvm/llvm-project/pull/174239