[llvm] [VPlan] Skip uses-scalars restriction if one of ops needs broadcast. (PR #168246)

[Ramkumar Ramachandra via llvm-commits]

================
@@ -1476,32 +1476,36 @@ static void narrowToSingleScalarRecipes(VPlan &Plan) {
         continue;
       }
 
-      // Skip recipes that aren't single scalars or don't have only their
-      // scalar results used. In the latter case, we would introduce extra
-      // broadcasts.
-      if (!vputils::isSingleScalar(RepOrWidenR) ||
-          !all_of(RepOrWidenR->users(), [RepOrWidenR](const VPUser *U) {
-            if (auto *Store = dyn_cast<VPWidenStoreRecipe>(U)) {
-              // VPWidenStore doesn't have users, and stores are always
-              // profitable to widen: hence, permitting address and mask
-              // operands, and single-scalar stored values is an important leaf
-              // condition. The assert must hold as we checked the RepOrWidenR
-              // operand against vputils::isSingleScalar.
-              assert(RepOrWidenR != Store->getStoredValue() ||
-                     vputils::isSingleScalar(Store->getStoredValue()));
-              (void)Store;
-              return true;
-            }
-
-            if (auto *VPI = dyn_cast<VPInstruction>(U)) {
-              unsigned Opcode = VPI->getOpcode();
-              if (Opcode == VPInstruction::ExtractLastElement ||
-                  Opcode == VPInstruction::ExtractLastLanePerPart ||
-                  Opcode == VPInstruction::ExtractPenultimateElement)
-                return true;
-            }
-
-            return U->usesScalars(RepOrWidenR);
+      // Skip recipes that aren't single scalars.
+      if (!vputils::isSingleScalar(RepOrWidenR))
+        continue;
+
+      // Skip recipes for which conversion to single-scalar does introduce
+      // additional broadcasts. No extra broadcasts are needed, if either only
+      // the scalars of the recipe are used, or at least one of the operands
+      // would require a broadcast. In the latter case, the single-scalar may
+      // need to be broadcasted, but another broadcast is removed.
+      if (!all_of(RepOrWidenR->users(),
+                  [RepOrWidenR](const VPUser *U) {
+                    if (auto *VPI = dyn_cast<VPInstruction>(U)) {
+                      unsigned Opcode = VPI->getOpcode();
+                      if (Opcode == VPInstruction::ExtractLastElement ||
+                          Opcode == VPInstruction::ExtractLastLanePerPart ||
+                          Opcode == VPInstruction::ExtractPenultimateElement)
+                        return true;
+                    }
+
----------------
artagnon wrote:

nit: strip newline?

https://github.com/llvm/llvm-project/pull/168246