[llvm] [NFC][VPlan] Split `makeMemOpWideningDecisions` into subpasses (PR #182593)

[Luke Lau via llvm-commits]

================
@@ -6924,38 +6924,79 @@ void VPlanTransforms::makeMemOpWideningDecisions(
     }
   }
 
-  VPBasicBlock *MiddleVPBB = Plan.getMiddleBlock();
+  // Few helpers to process different kinds of memory operations.
+
+  // To be used as argument to `VPlanTransforms::runPass` which explicitly
+  // specified pass name, hence `VPlan &` parameter.
+  auto ProcessSubset = [&](VPlan &, auto ProcessVPInst) {
+    SmallVector<VPInstruction *> RemainingMemOps;
+    for (VPInstruction *VPI : MemOps) {
+      if (!ProcessVPInst(VPI))
+        RemainingMemOps.push_back(VPI);
+    }
+
+    MemOps.clear();
+    std::swap(MemOps, RemainingMemOps);
+  };
+
+  auto ReplaceWith = [&](VPInstruction *VPI, VPRecipeBase *New) {
+    New->insertBefore(VPI);
+    if (VPI->getOpcode() == Instruction::Load)
+      VPI->replaceAllUsesWith(New->getVPSingleValue());
+    VPI->eraseFromParent();
+
+    // VPI has been processed.
+    return true;
+  };
+
+  auto Scalarize = [&](VPInstruction *VPI) {
+    return ReplaceWith(VPI, RecipeBuilder.handleReplication(VPI, Range));
+  };
+
+  auto *MiddleVPBB = Plan.getMiddleBlock();
   VPBuilder FinalRedStoresBuilder(MiddleVPBB, MiddleVPBB->getFirstNonPhi());
+  VPlanTransforms::runPass(
+      "lowerMemoryIdioms", ProcessSubset, Plan, [&](VPInstruction *VPI) {
+        if (RecipeBuilder.replaceWithFinalIfReductionStore(
+                VPI, FinalRedStoresBuilder))
+          return true;
+
+        // Filter out scalar VPlan for the remaining idioms.
+        if (LoopVectorizationPlanner::getDecisionAndClampRange(
+                [](ElementCount VF) { return VF.isScalar(); }, Range))
+          return false;
----------------
lukel97 wrote:

How come we filter out the scalar VPlan twice, once here and once in line 6976? Can we split the `replaceWithFinalIfReductionStore` into its own subpass, then filter out the scalar VPlans, and then continue with the rest of the histogram + `scalarizeMemOpsWithIrregularTypes` lowering?

https://github.com/llvm/llvm-project/pull/182593