[llvm] [LoopInterchange] Add an option to prioritize vectorization (PR #131988)

[Ryotaro Kasuga via llvm-commits]

================
@@ -1193,22 +1198,53 @@ bool LoopInterchangeProfitability::isProfitable(
     unsigned OuterLoopId, CharMatrix &DepMatrix,
     const DenseMap<const Loop *, unsigned> &CostMap,
     std::unique_ptr<CacheCost> &CC) {
-  // isProfitable() is structured to avoid endless loop interchange.
-  // If loop cache analysis could decide the profitability then,
-  // profitability check will stop and return the analysis result.
-  // If cache analysis failed to analyze the loopnest (e.g.,
-  // due to delinearization issues) then only check whether it is
-  // profitable for InstrOrderCost. Likewise, if InstrOrderCost failed to
-  // analysis the profitability then only, isProfitableForVectorization
-  // will decide.
-  std::optional<bool> shouldInterchange =
-      isProfitablePerLoopCacheAnalysis(CostMap, CC);
-  if (!shouldInterchange.has_value()) {
-    shouldInterchange = isProfitablePerInstrOrderCost();
-    if (!shouldInterchange.has_value())
+  // isProfitable() is structured to avoid endless loop interchange. If the
+  // highest priority rule (isProfitablePerLoopCacheAnalysis by default) could
+  // decide the profitability then, profitability check will stop and return the
+  // analysis result. If it failed to determine it (e.g., cache analysis failed
+  // to analyze the loopnest due to delinearization issues) then go ahead the
+  // second highest priority rule (isProfitablePerInstrOrderCost by default).
+  // Likewise, if it failed to analysis the profitability then only, the last
+  // rule (isProfitableForVectorization by default) will decide.
+  enum class RuleTy {
+    PerLoopCacheAnalysis,
+    PerInstrOrderCost,
+    ForVectorization,
+  };
+
+  // We prefer cache cost to vectorization by default.
+  RuleTy RuleOrder[3] = {RuleTy::PerLoopCacheAnalysis,
+                         RuleTy::PerInstrOrderCost, RuleTy::ForVectorization};
+
+  // If we prefer vectorization to cache cost, change the order of application
+  // of each rule.
+  if (PrioritizeVectorization) {
+    RuleOrder[0] = RuleTy::ForVectorization;
+    RuleOrder[1] = RuleTy::PerLoopCacheAnalysis;
+    RuleOrder[2] = RuleTy::PerInstrOrderCost;
+  }
+
+  std::optional<bool> shouldInterchange;
----------------
kasuga-fj wrote:

I see, I don't have strong opinion about variable names, but avoiding unnecessary diffs seems better to me, so I'll revert this change. Thanks for your input.

https://github.com/llvm/llvm-project/pull/131988