[VPlan] Use BlockFrequencyInfo in getPredBlockCostDivisor

llvm/include/llvm/Transforms/Vectorize/LoopVectorizationLegality.h

@@ -355,7 +355,7 @@ class LoopVectorizationLegality {
 
   /// Return true if the block BB needs to be predicated in order for the loop
   /// to be vectorized.
-  bool blockNeedsPredication(BasicBlock *BB) const;
+  bool blockNeedsPredication(const BasicBlock *BB) const;
 
   /// Check if this pointer is consecutive when vectorizing. This happens
   /// when the last index of the GEP is the induction variable, or that the

llvm/include/llvm/Transforms/Vectorize/LoopVectorize.h

@@ -145,7 +145,7 @@ struct LoopVectorizePass : public PassInfoMixin<LoopVectorizePass> {
   LoopInfo *LI;
   TargetTransformInfo *TTI;
   DominatorTree *DT;
-  BlockFrequencyInfo *BFI;
+  std::function<BlockFrequencyInfo *()> GetBFI;
   TargetLibraryInfo *TLI;
   DemandedBits *DB;
   AssumptionCache *AC;

llvm/lib/Transforms/Vectorize/LoopVectorizationLegality.cpp

@@ -1423,7 +1423,8 @@ bool LoopVectorizationLegality::isFixedOrderRecurrence(
   return FixedOrderRecurrences.count(Phi);
 }
 
-bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) const {
+bool LoopVectorizationLegality::blockNeedsPredication(
+    const BasicBlock *BB) const {
   // When vectorizing early exits, create predicates for the latch block only.
   // The early exiting block must be a direct predecessor of the latch at the
   // moment.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -873,12 +873,14 @@ class LoopVectorizationCostModel {
                              const TargetTransformInfo &TTI,
                              const TargetLibraryInfo *TLI, DemandedBits *DB,
                              AssumptionCache *AC,
-                             OptimizationRemarkEmitter *ORE, const Function *F,
-                             const LoopVectorizeHints *Hints,
+                             OptimizationRemarkEmitter *ORE,
+                             std::function<BlockFrequencyInfo *()> GetBFI,
+                             const Function *F, const LoopVectorizeHints *Hints,
                              InterleavedAccessInfo &IAI, bool OptForSize)
       : ScalarEpilogueStatus(SEL), TheLoop(L), PSE(PSE), LI(LI), Legal(Legal),
-        TTI(TTI), TLI(TLI), DB(DB), AC(AC), ORE(ORE), TheFunction(F),
-        Hints(Hints), InterleaveInfo(IAI), OptForSize(OptForSize) {
+        TTI(TTI), TLI(TLI), DB(DB), AC(AC), ORE(ORE), GetBFI(GetBFI),
+        TheFunction(F), Hints(Hints), InterleaveInfo(IAI),
+        OptForSize(OptForSize) {
     if (TTI.supportsScalableVectors() || ForceTargetSupportsScalableVectors)
       initializeVScaleForTuning();
     CostKind = F->hasMinSize() ? TTI::TCK_CodeSize : TTI::TCK_RecipThroughput;
@@ -1234,21 +1236,12 @@ class LoopVectorizationCostModel {
   /// optimizing for code size it will just be 1 as code size costs don't depend
   /// on execution probabilities.
   ///
-  /// TODO: We should use actual block probability here, if available.
-  /// Currently, we always assume predicated blocks have a 50% chance of
-  /// executing, apart from blocks that are only predicated due to tail folding.
+  /// Note that if a block wasn't originally predicated but was predicated due
+  /// to tail folding, the divisor will still be 1 because it will execute for
+  /// every iteration of the loop header.
   inline unsigned
   getPredBlockCostDivisor(TargetTransformInfo::TargetCostKind CostKind,
-                          BasicBlock *BB) const {
-    // If a block wasn't originally predicated but was predicated due to
-    // e.g. tail folding, don't divide the cost. Tail folded loops may still be
-    // predicated in the final vector loop iteration, but for most loops that
-    // don't have low trip counts we can expect their probability to be close to
-    // zero.
-    if (!Legal->blockNeedsPredication(BB))
-      return 1;
-    return CostKind == TTI::TCK_CodeSize ? 1 : 2;
-  }
+                          const BasicBlock *BB) const;
 
   /// Return the costs for our two available strategies for lowering a
   /// div/rem operation which requires speculating at least one lane.
@@ -1729,6 +1722,11 @@ class LoopVectorizationCostModel {
   /// Interface to emit optimization remarks.
   OptimizationRemarkEmitter *ORE;
 
+  /// A function to lazily fetch BlockFrequencyInfo. This avoids computing it
+  /// unless necessary, e.g. when the loop isn't legal to vectorize or when
+  /// there is no predication.
+  std::function<BlockFrequencyInfo *()> GetBFI;
+
   const Function *TheFunction;
 
   /// Loop Vectorize Hint.
@@ -2886,6 +2884,23 @@ bool LoopVectorizationCostModel::isPredicatedInst(Instruction *I) const {
   }
 }
 
+unsigned LoopVectorizationCostModel::getPredBlockCostDivisor(
+    TargetTransformInfo::TargetCostKind CostKind, const BasicBlock *BB) const {
+  // If the block wasn't originally predicated then return early to avoid
+  // computing BlockFrequencyInfo unnecessarily.
+  if (!Legal->blockNeedsPredication(BB))
+    return 1;
+  if (CostKind == TTI::TCK_CodeSize)
+    return 1;
+
+  BlockFrequencyInfo *BFI = GetBFI();
+  uint64_t HeaderFreq = BFI->getBlockFreq(TheLoop->getHeader()).getFrequency();
+  uint64_t BBFreq = BFI->getBlockFreq(BB).getFrequency();
+  assert(HeaderFreq >= BBFreq &&
+         "Header has smaller block freq than dominated BB?");
+  return HeaderFreq / BBFreq;
+}
+
 std::pair<InstructionCost, InstructionCost>
 LoopVectorizationCostModel::getDivRemSpeculationCost(Instruction *I,
                                                     ElementCount VF) const {
@@ -9166,8 +9181,9 @@ static bool processLoopInVPlanNativePath(
     Loop *L, PredicatedScalarEvolution &PSE, LoopInfo *LI, DominatorTree *DT,
     LoopVectorizationLegality *LVL, TargetTransformInfo *TTI,
     TargetLibraryInfo *TLI, DemandedBits *DB, AssumptionCache *AC,
-    OptimizationRemarkEmitter *ORE, bool OptForSize, LoopVectorizeHints &Hints,
-    LoopVectorizationRequirements &Requirements) {
+    OptimizationRemarkEmitter *ORE,
+    std::function<BlockFrequencyInfo *()> GetBFI, bool OptForSize,
+    LoopVectorizeHints &Hints, LoopVectorizationRequirements &Requirements) {
 
   if (isa<SCEVCouldNotCompute>(PSE.getBackedgeTakenCount())) {
     LLVM_DEBUG(dbgs() << "LV: cannot compute the outer-loop trip count\n");
@@ -9180,8 +9196,8 @@ static bool processLoopInVPlanNativePath(
   ScalarEpilogueLowering SEL =
       getScalarEpilogueLowering(F, L, Hints, OptForSize, TTI, TLI, *LVL, &IAI);
 
-  LoopVectorizationCostModel CM(SEL, L, PSE, LI, LVL, *TTI, TLI, DB, AC, ORE, F,
-                                &Hints, IAI, OptForSize);
+  LoopVectorizationCostModel CM(SEL, L, PSE, LI, LVL, *TTI, TLI, DB, AC, ORE,
+                                GetBFI, F, &Hints, IAI, OptForSize);
   // Use the planner for outer loop vectorization.
   // TODO: CM is not used at this point inside the planner. Turn CM into an
   // optional argument if we don't need it in the future.
@@ -9881,8 +9897,9 @@ bool LoopVectorizePass::processLoop(Loop *L) {
 
   // Query this against the original loop and save it here because the profile
   // of the original loop header may change as the transformation happens.
-  bool OptForSize = llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI,
-                                                PGSOQueryType::IRPass);
+  bool OptForSize = llvm::shouldOptimizeForSize(
+      L->getHeader(), PSI, PSI && PSI->hasProfileSummary() ? GetBFI() : nullptr,
+      PGSOQueryType::IRPass);
 
   // Check if it is legal to vectorize the loop.
   LoopVectorizationRequirements Requirements;
@@ -9916,7 +9933,8 @@ bool LoopVectorizePass::processLoop(Loop *L) {
   // pipeline.
   if (!L->isInnermost())
     return processLoopInVPlanNativePath(L, PSE, LI, DT, &LVL, TTI, TLI, DB, AC,
-                                        ORE, OptForSize, Hints, Requirements);
+                                        ORE, GetBFI, OptForSize, Hints,
+                                        Requirements);
 
   assert(L->isInnermost() && "Inner loop expected.");
 
@@ -10019,7 +10037,7 @@ bool LoopVectorizePass::processLoop(Loop *L) {
 
   // Use the cost model.
   LoopVectorizationCostModel CM(SEL, L, PSE, LI, &LVL, *TTI, TLI, DB, AC, ORE,
-                                F, &Hints, IAI, OptForSize);
+                                GetBFI, F, &Hints, IAI, OptForSize);
   // Use the planner for vectorization.
   LoopVectorizationPlanner LVP(L, LI, DT, TLI, *TTI, &LVL, CM, IAI, PSE, Hints,
                                ORE);
@@ -10337,9 +10355,7 @@ PreservedAnalyses LoopVectorizePass::run(Function &F,
 
   auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F);
   PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
-  BFI = nullptr;
-  if (PSI && PSI->hasProfileSummary())
-    BFI = &AM.getResult<BlockFrequencyAnalysis>(F);
+  GetBFI = [&AM, &F]() { return &AM.getResult<BlockFrequencyAnalysis>(F); };
   LoopVectorizeResult Result = runImpl(F);
   if (!Result.MadeAnyChange)
     return PreservedAnalyses::all();

llvm/test/Transforms/LoopVectorize/AArch64/aarch64-predication.ll

@@ -25,7 +25,7 @@ define i64 @predicated_udiv_scalarized_operand(ptr %a, i64 %x) {
 ; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <2 x i64> [ zeroinitializer, [[ENTRY]] ], [ [[TMP17:%.*]], [[PRED_UDIV_CONTINUE2]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i64, ptr [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <2 x i64>, ptr [[TMP0]], align 4
-; CHECK-NEXT:    [[TMP2:%.*]] = icmp sgt <2 x i64> [[WIDE_LOAD]], zeroinitializer
+; CHECK-NEXT:    [[TMP2:%.*]] = icmp sgt <2 x i64> [[WIDE_LOAD]], splat (i64 1)
 ; CHECK-NEXT:    [[TMP3:%.*]] = extractelement <2 x i1> [[TMP2]], i64 0
 ; CHECK-NEXT:    br i1 [[TMP3]], label [[PRED_UDIV_IF:%.*]], label [[PRED_UDIV_CONTINUE:%.*]]
 ; CHECK:       pred.udiv.if:
@@ -65,7 +65,7 @@ for.body:
   %r = phi i64 [ 0, %entry ], [ %var6, %for.inc ]
   %var0 = getelementptr inbounds i64, ptr %a, i64 %i
   %var2 = load i64, ptr %var0, align 4
-  %cond0 = icmp sgt i64 %var2, 0
+  %cond0 = icmp sgt i64 %var2, 1
   br i1 %cond0, label %if.then, label %for.inc
 
 if.then:

llvm/test/Transforms/LoopVectorize/AArch64/conditional-branches-cost.ll

@@ -1208,8 +1208,7 @@ define void @pred_udiv_select_cost(ptr %A, ptr %B, ptr %C, i64 %n, i8 %y) #1 {
 ; DEFAULT-NEXT:    [[TMP0:%.*]] = add i64 [[N]], 1
 ; DEFAULT-NEXT:    [[TMP1:%.*]] = call i64 @llvm.vscale.i64()
 ; DEFAULT-NEXT:    [[TMP2:%.*]] = shl nuw i64 [[TMP1]], 2
-; DEFAULT-NEXT:    [[TMP3:%.*]] = call i64 @llvm.umax.i64(i64 [[TMP2]], i64 8)
-; DEFAULT-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], [[TMP3]]
+; DEFAULT-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], [[TMP2]]
 ; DEFAULT-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_MEMCHECK:.*]]
 ; DEFAULT:       [[VECTOR_MEMCHECK]]:
 ; DEFAULT-NEXT:    [[TMP4:%.*]] = call i64 @llvm.vscale.i64()

llvm/test/Transforms/LoopVectorize/AArch64/early_exit_costs.ll

@@ -57,8 +57,8 @@ define i64 @same_exit_block_pre_inc_use1_nosve() {
 ; CHECK-NEXT: Cost of 48 for VF 16: EMIT vp<{{.*}}> = first-active-lane ir<%cmp3>
 ; CHECK-NEXT: Cost of 0 for VF 16: EMIT vp<{{.*}}> = add
 ; CHECK-NEXT: Cost of 0 for VF 16: vp<{{.*}}> = DERIVED-IV
-; CHECK: LV: Minimum required TC for runtime checks to be profitable:160
-; CHECK-NEXT: LV: Vectorization is not beneficial: expected trip count < minimum profitable VF (64 < 160)
+; CHECK: LV: Minimum required TC for runtime checks to be profitable:128
+; CHECK-NEXT: LV: Vectorization is not beneficial: expected trip count < minimum profitable VF (64 < 128)
 ; CHECK-NEXT: LV: Too many memory checks needed.
 entry:
   %p1 = alloca [1024 x i8]
@@ -105,7 +105,7 @@ loop.header:
   %gep.src = getelementptr inbounds i64, ptr %src, i64 %iv
   %l = load i64, ptr %gep.src, align 1
   %t = trunc i64 %l to i1
-  br i1 %t, label %exit.0, label %loop.latch
+  br i1 %t, label %exit.0, label %loop.latch, !prof !0
 
 loop.latch:
   %iv.next = add i64 %iv, 1
@@ -120,4 +120,6 @@ exit.1:
   ret i64 0
 }
 
+!0 = !{!"branch_weights", i32 1, i32 1}
+
 attributes #1 = { "target-features"="+sve" vscale_range(1,16) }

llvm/test/Transforms/LoopVectorize/AArch64/masked-op-cost.ll

@@ -30,7 +30,7 @@ for.body:                                         ; preds = %for.body.preheader,
   %i.07 = phi i64 [ %inc, %for.inc ], [ 0, %for.body.preheader ]
   %arrayidx = getelementptr inbounds i32, ptr %b, i64 %i.07
   %0 = load i32, ptr %arrayidx, align 4
-  %tobool.not = icmp eq i32 %0, 0
+  %tobool.not = icmp eq i32 %0, 1
   br i1 %tobool.not, label %for.inc, label %if.then
 
 if.then:                                          ; preds = %for.body

llvm/test/Transforms/LoopVectorize/AArch64/simple_early_exit.ll

@@ -386,7 +386,7 @@ define i32 @diff_exit_block_needs_scev_check(i32 %end) {
 ; CHECK-NEXT:    [[TMP0:%.*]] = trunc i32 [[END]] to i10
 ; CHECK-NEXT:    [[TMP1:%.*]] = zext i10 [[TMP0]] to i64
 ; CHECK-NEXT:    [[UMAX1:%.*]] = call i64 @llvm.umax.i64(i64 [[TMP1]], i64 1)
-; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[UMAX1]], 12
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[UMAX1]], 8
 ; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]]
 ; CHECK:       vector.scevcheck:
 ; CHECK-NEXT:    [[UMAX:%.*]] = call i32 @llvm.umax.i32(i32 [[END_CLAMPED]], i32 1)