[LoopInterchange] Ignore the cost-model, force interchange if legal

llvm/lib/Transforms/Scalar/LoopInterchange.cpp

@@ -78,6 +78,7 @@ enum class RuleTy {
   PerLoopCacheAnalysis,
   PerInstrOrderCost,
   ForVectorization,
+  Ignore
 };
 
 } // end anonymous namespace
@@ -106,14 +107,20 @@ static cl::list<RuleTy> Profitabilities(
                clEnumValN(RuleTy::PerInstrOrderCost, "instorder",
                           "Prioritize the IVs order of each instruction"),
                clEnumValN(RuleTy::ForVectorization, "vectorize",
-                          "Prioritize vectorization")));
+                          "Prioritize vectorization"),
+               clEnumValN(RuleTy::Ignore, "ignore",
+                          "Ignore profitability, force interchange (does not "
+                          "work with other options)")));
 
 #ifndef NDEBUG
-static bool noDuplicateRules(ArrayRef<RuleTy> Rules) {
+static bool noDuplicateRulesAndIgnore(ArrayRef<RuleTy> Rules) {
   SmallSet<RuleTy, 4> Set;
-  for (RuleTy Rule : Rules)
+  for (RuleTy Rule : Rules) {
     if (!Set.insert(Rule).second)
       return false;
+    if (Rule == RuleTy::Ignore)
+      return false;
+  }
   return true;
 }
 
@@ -1286,6 +1293,13 @@ std::optional<bool> LoopInterchangeProfitability::isProfitableForVectorization(
 bool LoopInterchangeProfitability::isProfitable(
     const Loop *InnerLoop, const Loop *OuterLoop, unsigned InnerLoopId,
     unsigned OuterLoopId, CharMatrix &DepMatrix, CacheCostManager &CCM) {
+
+  // Return true if interchange is forced and the cost-model ignored.
+  if (Profitabilities.size() == 1 && Profitabilities[0] == RuleTy::Ignore)
+    return true;
+  assert(noDuplicateRulesAndIgnore(Profitabilities) &&
+         "Duplicate rules and option 'ignore' are not allowed");
+
   // 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
@@ -1294,7 +1308,6 @@ bool LoopInterchangeProfitability::isProfitable(
   // 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.
-  assert(noDuplicateRules(Profitabilities) && "Detect duplicate rules");
   std::optional<bool> shouldInterchange;
   for (RuleTy RT : Profitabilities) {
     switch (RT) {
@@ -1311,6 +1324,9 @@ bool LoopInterchangeProfitability::isProfitable(
       shouldInterchange =
           isProfitableForVectorization(InnerLoopId, OuterLoopId, DepMatrix);
       break;
+    case RuleTy::Ignore:
+      llvm_unreachable("Option 'ignore' is not supported with other options");
+      break;
     }
 
     // If this rule could determine the profitability, don't call subsequent

llvm/test/Transforms/LoopInterchange/force-interchange.ll

@@ -0,0 +1,43 @@
+; RUN: opt < %s -passes=loop-interchange -pass-remarks-output=%t -disable-output -loop-interchange-profitabilities=ignore -S
+; RUN: FileCheck --input-file=%t %s
+
+; There should be no reason to interchange this, unless it is forced.
+;
+;     for (int i = 0; i<1024; i++)
+;       for (int j = 0; j<1024; j++)
+;         A[i][j] = 42;
+;
+; CHECK:      --- !Passed
+; CHECK-NEXT: Pass:            loop-interchange
+; CHECK-NEXT: Name:            Interchanged
+; CHECK-NEXT: Function:        f
+; CHECK-NEXT: Args:
+; CHECK-NEXT:   - String:          Loop interchanged with enclosing loop.
+; CHECK-NEXT: ...
+
+@A = dso_local local_unnamed_addr global [1024 x [1024 x i32]] zeroinitializer, align 4
+
+define dso_local void @f() local_unnamed_addr #0 {
+entry:
+  br label %outer.header
+
+outer.header:
+  %i = phi i64 [ 0, %entry ], [ %i.next, %inner.header ]
+  br label %inner.body
+
+inner.header:
+  %i.next = add nuw nsw i64 %i, 1
+  %exitcond20.not = icmp eq i64 %i.next, 1024
+  br i1 %exitcond20.not, label %exit, label %outer.header
+
+inner.body:
+  %j = phi i64 [ 0, %outer.header ], [ %j.next, %inner.body ]
+  %arrayidx6 = getelementptr inbounds nuw [1024 x [1024 x i32]], ptr @A, i64 0, i64 %i, i64 %j
+  store i32 42, ptr %arrayidx6, align 4
+  %j.next = add nuw nsw i64 %j, 1
+  %exitcond.not = icmp eq i64 %j.next, 1024
+  br i1 %exitcond.not, label %inner.header, label %inner.body
+
+exit:
+  ret void
+}