[llvm] [LoopVectorizer] Prune VFs based on plan register pressure (PR #132190)

[Sander de Smalen via llvm-commits]

================
@@ -7792,6 +7582,29 @@ VectorizationFactor LoopVectorizationPlanner::computeBestVF() {
   VectorizationFactor LegacyVF = selectVectorizationFactor();
   VPlan &BestPlan = getPlanFor(BestFactor.Width);
 
+  // VPlan calculates register pressure from the plan, so it can come to
+  // different conclusions than the legacy cost model.
+  bool RegUsageDeterminedVF = false;
+  if (BestFactor.Width != LegacyVF.Width) {
+    SmallVector<ElementCount, 1> LegacyVFs = {LegacyVF.Width};
+    SmallVector<ElementCount, 1> VFs = {BestFactor.Width};
+
+    auto LegacyRUs =
+        calculateRegisterUsage(getPlanFor(LegacyVF.Width), LegacyVFs, TTI);
+    auto RUs = calculateRegisterUsage(BestPlan, VFs, TTI);
+
+    auto GetMaxUsage = [](SmallMapVector<unsigned, unsigned, 4> MaxLocalUsers) {
+      unsigned Max = 0;
+      for (auto Pair : MaxLocalUsers)
+        if (Pair.second > Max)
+          Max = Pair.second;
+      return Max;
+    };
+    unsigned MaxLegacyRegUsage = GetMaxUsage(LegacyRUs[0].MaxLocalUsers);
+    unsigned MaxRegUsage = GetMaxUsage(RUs[0].MaxLocalUsers);
+    RegUsageDeterminedVF = MaxRegUsage <= MaxLegacyRegUsage;
----------------
sdesmalen-arm wrote:

I had a closer look into this, and ignore my suggestion above. Instead, I think it is sufficient to check that the LegacyVF that would otherwise have been chosen has a register usage that exceeds the maximum number of registers. Otherwise, the cost would have chosen the same VF.

That means, this can be simplified to:
```suggestion
    auto &MLU = LegacyRUs[0].MaxLocalUsers;
    RegUsageDeterminedVF = any_of(MLU, [&](decltype(MLU.front()) &LU) {
      return LU.second > TTI.getNumberOfRegisters(LU.first);
    });
```

https://github.com/llvm/llvm-project/pull/132190