[LV] Vectorize Epilogues for loops with small VF but high IC

- Consider MainLoopVF * IC when determining whether Epilogue
  Vectorization is profitable
- Allow the same VF for the Epilogue as for the main loop
- Use an upper bound for the trip count of the Epilogue
  when choosing the Epilogue VF

Author: juliannagele

llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h

@@ -471,6 +471,10 @@ class LoopVectorizationPlanner {
   bool isMoreProfitable(const VectorizationFactor &A,
                         const VectorizationFactor &B) const;
 
+  bool isMoreProfitable(const VectorizationFactor &A,
+                        const VectorizationFactor &B,
+                        const unsigned MaxTripCount) const;
+
   /// Determines if we have the infrastructure to vectorize the loop and its
   /// epilogue, assuming the main loop is vectorized by \p VF.
   bool isCandidateForEpilogueVectorization(const ElementCount VF) const;

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -1537,7 +1537,8 @@ class LoopVectorizationCostModel {
   /// Returns true if epilogue vectorization is considered profitable, and
   /// false otherwise.
   /// \p VF is the vectorization factor chosen for the original loop.
-  bool isEpilogueVectorizationProfitable(const ElementCount VF) const;
+  bool isEpilogueVectorizationProfitable(const ElementCount VF,
+                                         const unsigned Multiplier) const;
 
   /// Returns the execution time cost of an instruction for a given vector
   /// width. Vector width of one means scalar.
@@ -4282,12 +4283,11 @@ getVScaleForTuning(const Loop *L, const TargetTransformInfo &TTI) {
 }
 
 bool LoopVectorizationPlanner::isMoreProfitable(
-    const VectorizationFactor &A, const VectorizationFactor &B) const {
+    const VectorizationFactor &A, const VectorizationFactor &B,
+    const unsigned MaxTripCount) const {
   InstructionCost CostA = A.Cost;
   InstructionCost CostB = B.Cost;
 
-  unsigned MaxTripCount = PSE.getSE()->getSmallConstantMaxTripCount(OrigLoop);
-
   // Improve estimate for the vector width if it is scalable.
   unsigned EstimatedWidthA = A.Width.getKnownMinValue();
   unsigned EstimatedWidthB = B.Width.getKnownMinValue();
@@ -4336,6 +4336,13 @@ bool LoopVectorizationPlanner::isMoreProfitable(
   return CmpFn(RTCostA, RTCostB);
 }
 
+bool LoopVectorizationPlanner::isMoreProfitable(
+    const VectorizationFactor &A, const VectorizationFactor &B) const {
+  const unsigned MaxTripCount =
+      PSE.getSE()->getSmallConstantMaxTripCount(OrigLoop);
+  return LoopVectorizationPlanner::isMoreProfitable(A, B, MaxTripCount);
+}
+
 void LoopVectorizationPlanner::emitInvalidCostRemarks(
     OptimizationRemarkEmitter *ORE) {
   using RecipeVFPair = std::pair<VPRecipeBase *, ElementCount>;
@@ -4648,7 +4655,7 @@ bool LoopVectorizationPlanner::isCandidateForEpilogueVectorization(
 }
 
 bool LoopVectorizationCostModel::isEpilogueVectorizationProfitable(
-    const ElementCount VF) const {
+    const ElementCount VF, const unsigned Multiplier) const {
   // FIXME: We need a much better cost-model to take different parameters such
   // as register pressure, code size increase and cost of extra branches into
   // account. For now we apply a very crude heuristic and only consider loops
@@ -4663,9 +4670,6 @@ bool LoopVectorizationCostModel::isEpilogueVectorizationProfitable(
   if (TTI.getMaxInterleaveFactor(VF) <= 1)
     return false;
 
-  unsigned Multiplier = 1;
-  if (VF.isScalable())
-    Multiplier = getVScaleForTuning(TheLoop, TTI).value_or(1);
   if ((Multiplier * VF.getKnownMinValue()) >= EpilogueVectorizationMinVF)
     return true;
   return false;
@@ -4711,7 +4715,11 @@ VectorizationFactor LoopVectorizationPlanner::selectEpilogueVectorizationFactor(
     return Result;
   }
 
-  if (!CM.isEpilogueVectorizationProfitable(MainLoopVF)) {
+  unsigned Multiplier = IC;
+  if (MainLoopVF.isScalable())
+    Multiplier = getVScaleForTuning(OrigLoop, TTI).value_or(1);
+
+  if (!CM.isEpilogueVectorizationProfitable(MainLoopVF, Multiplier)) {
     LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization is not profitable for "
                          "this loop\n");
     return Result;
@@ -4730,16 +4738,20 @@ VectorizationFactor LoopVectorizationPlanner::selectEpilogueVectorizationFactor(
   ScalarEvolution &SE = *PSE.getSE();
   Type *TCType = Legal->getWidestInductionType();
   const SCEV *RemainingIterations = nullptr;
+  unsigned MaxTripCount = 0;
   for (auto &NextVF : ProfitableVFs) {
     // Skip candidate VFs without a corresponding VPlan.
     if (!hasPlanWithVF(NextVF.Width))
       continue;
 
-    // Skip candidate VFs with widths >= the estimate runtime VF (scalable
-    // vectors) or the VF of the main loop (fixed vectors).
+    // Skip candidate VFs with widths >= the (estimated) runtime VF (scalable
+    // vectors) or > the VF of the main loop (fixed vectors).
     if ((!NextVF.Width.isScalable() && MainLoopVF.isScalable() &&
          ElementCount::isKnownGE(NextVF.Width, EstimatedRuntimeVF)) ||
-        ElementCount::isKnownGE(NextVF.Width, MainLoopVF))
+        (NextVF.Width.isScalable() &&
+         ElementCount::isKnownGE(NextVF.Width, MainLoopVF)) ||
+        (!NextVF.Width.isScalable() && !MainLoopVF.isScalable() &&
+         ElementCount::isKnownGT(NextVF.Width, MainLoopVF)))
       continue;
 
     // If NextVF is greater than the number of remaining iterations, the
@@ -4750,6 +4762,14 @@ VectorizationFactor LoopVectorizationPlanner::selectEpilogueVectorizationFactor(
         const SCEV *TC = createTripCountSCEV(TCType, PSE, OrigLoop);
         RemainingIterations = SE.getURemExpr(
             TC, SE.getConstant(TCType, MainLoopVF.getKnownMinValue() * IC));
+        const APInt MaxRemainingIterations =
+            SE.getUnsignedRangeMax(RemainingIterations);
+        // Guard against huge trip counts.
+        if (MaxRemainingIterations.getActiveBits() <= 32) {
+          MaxTripCount = MaxRemainingIterations.getZExtValue();
+          LLVM_DEBUG(dbgs() << "LEV: Maximum Trip Count for Epilogue: "
+                            << MaxTripCount << "\n");
+        }
       }
       if (SE.isKnownPredicate(
               CmpInst::ICMP_UGT,
@@ -4758,7 +4778,8 @@ VectorizationFactor LoopVectorizationPlanner::selectEpilogueVectorizationFactor(
         continue;
     }
 
-    if (Result.Width.isScalar() || isMoreProfitable(NextVF, Result))
+    if (Result.Width.isScalar() ||
+        isMoreProfitable(NextVF, Result, MaxTripCount))
       Result = NextVF;
   }
 

llvm/test/Transforms/LoopVectorize/AArch64/deterministic-type-shrinkage.ll

@@ -16,7 +16,7 @@ define void @test_pr25490(i32 %n, ptr noalias nocapture %a, ptr noalias nocaptur
 ; CHECK-NEXT:    br i1 [[CMP_28]], label [[FOR_COND_CLEANUP:%.*]], label [[ITER_CHECK:%.*]]
 ; CHECK:       iter.check:
 ; CHECK-NEXT:    [[TMP0:%.*]] = zext i32 [[N]] to i64
-; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N]], 8
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[N]], 4
 ; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH:%.*]], label [[VECTOR_MAIN_LOOP_ITER_CHECK:%.*]]
 ; CHECK:       vector.main.loop.iter.check:
 ; CHECK-NEXT:    [[MIN_ITERS_CHECK1:%.*]] = icmp ult i32 [[N]], 16
@@ -50,33 +50,33 @@ define void @test_pr25490(i32 %n, ptr noalias nocapture %a, ptr noalias nocaptur
 ; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[N_VEC]], [[TMP0]]
 ; CHECK-NEXT:    br i1 [[CMP_N]], label [[FOR_COND_CLEANUP_LOOPEXIT:%.*]], label [[VEC_EPILOG_ITER_CHECK:%.*]]
 ; CHECK:       vec.epilog.iter.check:
-; CHECK-NEXT:    [[N_VEC_REMAINING:%.*]] = and i64 [[TMP0]], 8
-; CHECK-NEXT:    [[MIN_EPILOG_ITERS_CHECK_NOT_NOT:%.*]] = icmp eq i64 [[N_VEC_REMAINING]], 0
-; CHECK-NEXT:    br i1 [[MIN_EPILOG_ITERS_CHECK_NOT_NOT]], label [[VEC_EPILOG_SCALAR_PH]], label [[VEC_EPILOG_PH]]
+; CHECK-NEXT:    [[N_VEC_REMAINING:%.*]] = and i64 [[TMP0]], 12
+; CHECK-NEXT:    [[MIN_EPILOG_ITERS_CHECK:%.*]] = icmp eq i64 [[N_VEC_REMAINING]], 0
+; CHECK-NEXT:    br i1 [[MIN_EPILOG_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH]], label [[VEC_EPILOG_PH]]
 ; CHECK:       vec.epilog.ph:
 ; CHECK-NEXT:    [[VEC_EPILOG_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[VECTOR_MAIN_LOOP_ITER_CHECK]] ]
-; CHECK-NEXT:    [[N_VEC5:%.*]] = and i64 [[TMP0]], 4294967288
+; CHECK-NEXT:    [[N_VEC5:%.*]] = and i64 [[TMP0]], 4294967292
 ; CHECK-NEXT:    br label [[VEC_EPILOG_VECTOR_BODY:%.*]]
 ; CHECK:       vec.epilog.vector.body:
 ; CHECK-NEXT:    [[INDEX7:%.*]] = phi i64 [ [[VEC_EPILOG_RESUME_VAL]], [[VEC_EPILOG_PH]] ], [ [[INDEX_NEXT11:%.*]], [[VEC_EPILOG_VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i8, ptr [[C]], i64 [[INDEX7]]
-; CHECK-NEXT:    [[WIDE_LOAD8:%.*]] = load <8 x i8>, ptr [[TMP14]], align 1
+; CHECK-NEXT:    [[WIDE_LOAD8:%.*]] = load <4 x i8>, ptr [[TMP14]], align 1
 ; CHECK-NEXT:    [[TMP15:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[INDEX7]]
-; CHECK-NEXT:    [[WIDE_LOAD9:%.*]] = load <8 x i8>, ptr [[TMP15]], align 1
-; CHECK-NEXT:    [[TMP16:%.*]] = zext <8 x i8> [[WIDE_LOAD9]] to <8 x i16>
-; CHECK-NEXT:    [[TMP17:%.*]] = zext <8 x i8> [[WIDE_LOAD8]] to <8 x i16>
-; CHECK-NEXT:    [[TMP18:%.*]] = mul nuw <8 x i16> [[TMP16]], [[TMP17]]
-; CHECK-NEXT:    [[TMP19:%.*]] = lshr <8 x i16> [[TMP18]], <i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8>
-; CHECK-NEXT:    [[TMP20:%.*]] = trunc nuw <8 x i16> [[TMP19]] to <8 x i8>
-; CHECK-NEXT:    store <8 x i8> [[TMP20]], ptr [[TMP15]], align 1
+; CHECK-NEXT:    [[WIDE_LOAD9:%.*]] = load <4 x i8>, ptr [[TMP15]], align 1
+; CHECK-NEXT:    [[TMP16:%.*]] = zext <4 x i8> [[WIDE_LOAD9]] to <4 x i16>
+; CHECK-NEXT:    [[TMP17:%.*]] = zext <4 x i8> [[WIDE_LOAD8]] to <4 x i16>
+; CHECK-NEXT:    [[TMP18:%.*]] = mul nuw <4 x i16> [[TMP16]], [[TMP17]]
+; CHECK-NEXT:    [[TMP19:%.*]] = lshr <4 x i16> [[TMP18]], <i16 8, i16 8, i16 8, i16 8>
+; CHECK-NEXT:    [[TMP20:%.*]] = trunc nuw <4 x i16> [[TMP19]] to <4 x i8>
+; CHECK-NEXT:    store <4 x i8> [[TMP20]], ptr [[TMP15]], align 1
 ; CHECK-NEXT:    [[TMP21:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[INDEX7]]
-; CHECK-NEXT:    [[WIDE_LOAD10:%.*]] = load <8 x i8>, ptr [[TMP21]], align 1
-; CHECK-NEXT:    [[TMP22:%.*]] = zext <8 x i8> [[WIDE_LOAD10]] to <8 x i16>
-; CHECK-NEXT:    [[TMP23:%.*]] = mul nuw <8 x i16> [[TMP22]], [[TMP17]]
-; CHECK-NEXT:    [[TMP24:%.*]] = lshr <8 x i16> [[TMP23]], <i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8>
-; CHECK-NEXT:    [[TMP25:%.*]] = trunc nuw <8 x i16> [[TMP24]] to <8 x i8>
-; CHECK-NEXT:    store <8 x i8> [[TMP25]], ptr [[TMP21]], align 1
-; CHECK-NEXT:    [[INDEX_NEXT11]] = add nuw i64 [[INDEX7]], 8
+; CHECK-NEXT:    [[WIDE_LOAD10:%.*]] = load <4 x i8>, ptr [[TMP21]], align 1
+; CHECK-NEXT:    [[TMP22:%.*]] = zext <4 x i8> [[WIDE_LOAD10]] to <4 x i16>
+; CHECK-NEXT:    [[TMP23:%.*]] = mul nuw <4 x i16> [[TMP22]], [[TMP17]]
+; CHECK-NEXT:    [[TMP24:%.*]] = lshr <4 x i16> [[TMP23]], <i16 8, i16 8, i16 8, i16 8>
+; CHECK-NEXT:    [[TMP25:%.*]] = trunc nuw <4 x i16> [[TMP24]] to <4 x i8>
+; CHECK-NEXT:    store <4 x i8> [[TMP25]], ptr [[TMP21]], align 1
+; CHECK-NEXT:    [[INDEX_NEXT11]] = add nuw i64 [[INDEX7]], 4
 ; CHECK-NEXT:    [[TMP26:%.*]] = icmp eq i64 [[INDEX_NEXT11]], [[N_VEC5]]
 ; CHECK-NEXT:    br i1 [[TMP26]], label [[VEC_EPILOG_MIDDLE_BLOCK:%.*]], label [[VEC_EPILOG_VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
 ; CHECK:       vec.epilog.middle.block: