[VPlan] Use BlockFrequencyInfo in getPredBlockCostDivisor (#158690)

In 531.deepsjeng_r from SPEC CPU 2017 there's a loop that we
unprofitably loop vectorize on RISC-V.

The loop looks something like:

```c
  for (int i = 0; i < n; i++) {
    if (x0[i] == a)
      if (x1[i] == b)
        if (x2[i] == c)
          // do stuff...
  }
```

Because it's so deeply nested the actual inner level of the loop rarely
gets executed. However we still deem it profitable to vectorize, which
due to the if-conversion means we now always execute the body.

This stems from the fact that `getPredBlockCostDivisor` currently
assumes that blocks have 50% chance of being executed as a heuristic.

We can fix this by using BlockFrequencyInfo, which gives a more accurate
estimate of the innermost block being executed 12.5% of the time. We can
then calculate the probability as `HeaderFrequency / BlockFrequency`.

Fixing the cost here gives a 7% speedup for 531.deepsjeng_r on RISC-V.

Whilst there's a lot of changes in the in-tree tests, this doesn't
affect llvm-test-suite or SPEC CPU 2017 that much:

- On armv9-a -flto -O3 there's 0.0%/0.2% more geomean loops vectorized
on llvm-test-suite/SPEC CPU 2017.
- On x86-64 -flto -O3 **with PGO** there's 0.9%/0% less geomean loops
vectorized on llvm-test-suite/SPEC CPU 2017.

Overall geomean compile time impact is 0.03% on stage1-ReleaseLTO:
https://llvm-compile-time-tracker.com/compare.php?from=9eee396c58d2e24beb93c460141170def328776d&to=32fbff48f965d03b51549fdf9bbc4ca06473b623&stat=instructions%3Au

Author: lukel97

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

@@ -364,7 +364,7 @@ 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

@@ -1443,7 +1443,8 @@ 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

@@ -146,6 +146,7 @@
 #include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h"
 #include <algorithm>
 #include <cassert>
+#include <cmath>
 #include <cstdint>
 #include <functional>
 #include <iterator>
@@ -873,12 +874,14 @@ 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;
@@ -1219,7 +1222,7 @@ class LoopVectorizationCostModel {
   /// for which our chosen predication strategy is scalarization (i.e. we
   /// don't have an alternate strategy such as masking available).
   /// \p VF is the vectorization factor that will be used to vectorize \p I.
-  bool isScalarWithPredication(Instruction *I, ElementCount VF) const;
+  bool isScalarWithPredication(Instruction *I, ElementCount VF);
 
   /// Returns true if \p I is an instruction that needs to be predicated
   /// at runtime.  The result is independent of the predication mechanism.
@@ -1234,29 +1237,19 @@ 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);
 
   /// Return the costs for our two available strategies for lowering a
   /// div/rem operation which requires speculating at least one lane.
   /// First result is for scalarization (will be invalid for scalable
   /// vectors); second is for the safe-divisor strategy.
   std::pair<InstructionCost, InstructionCost>
-  getDivRemSpeculationCost(Instruction *I,
-                           ElementCount VF) const;
+  getDivRemSpeculationCost(Instruction *I, ElementCount VF);
 
   /// Returns true if \p I is a memory instruction with consecutive memory
   /// access that can be widened.
@@ -1729,6 +1722,20 @@ 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;
+  /// The BlockFrequencyInfo returned from GetBFI.
+  BlockFrequencyInfo *BFI = nullptr;
+  /// Returns the BlockFrequencyInfo for the function if cached, otherwise
+  /// fetches it via GetBFI. Avoids an indirect call to the std::function.
+  BlockFrequencyInfo &getBFI() {
+    if (!BFI)
+      BFI = &GetBFI();
+    return *BFI;
+  }
+
   const Function *TheFunction;
 
   /// Loop Vectorize Hint.
@@ -2792,8 +2799,8 @@ void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) {
   Scalars[VF].insert_range(Worklist);
 }
 
-bool LoopVectorizationCostModel::isScalarWithPredication(
-    Instruction *I, ElementCount VF) const {
+bool LoopVectorizationCostModel::isScalarWithPredication(Instruction *I,
+                                                         ElementCount VF) {
   if (!isPredicatedInst(I))
     return false;
 
@@ -2886,9 +2893,26 @@ bool LoopVectorizationCostModel::isPredicatedInst(Instruction *I) const {
   }
 }
 
+unsigned LoopVectorizationCostModel::getPredBlockCostDivisor(
+    TargetTransformInfo::TargetCostKind CostKind, const BasicBlock *BB) {
+  if (CostKind == TTI::TCK_CodeSize)
+    return 1;
+  // If the block wasn't originally predicated then return early to avoid
+  // computing BlockFrequencyInfo unnecessarily.
+  if (!Legal->blockNeedsPredication(BB))
+    return 1;
+
+  uint64_t HeaderFreq =
+      getBFI().getBlockFreq(TheLoop->getHeader()).getFrequency();
+  uint64_t BBFreq = getBFI().getBlockFreq(BB).getFrequency();
+  assert(HeaderFreq >= BBFreq &&
+         "Header has smaller block freq than dominated BB?");
+  return std::round((double)HeaderFreq / BBFreq);
+}
+
 std::pair<InstructionCost, InstructionCost>
 LoopVectorizationCostModel::getDivRemSpeculationCost(Instruction *I,
-                                                    ElementCount VF) const {
+                                                     ElementCount VF) {
   assert(I->getOpcode() == Instruction::UDiv ||
          I->getOpcode() == Instruction::SDiv ||
          I->getOpcode() == Instruction::SRem ||
@@ -9182,8 +9206,9 @@ 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");
@@ -9196,8 +9221,8 @@ 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.
@@ -9897,8 +9922,10 @@ 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;
@@ -9932,7 +9959,8 @@ 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.");
 
@@ -10035,7 +10063,7 @@ 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);
@@ -10353,9 +10381,9 @@ 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]() -> BlockFrequencyInfo & {
+    return AM.getResult<BlockFrequencyAnalysis>(F);
+  };
   LoopVectorizeResult Result = runImpl(F);
   if (!Result.MadeAnyChange)
     return PreservedAnalyses::all();

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/predicated-costs.ll

@@ -385,6 +385,59 @@ attributes #1 = { "target-cpu"="neoverse-v2" }
 !1 = !{!"llvm.loop.mustprogress"}
 !2 = !{!"llvm.loop.vectorize.predicate.enable", i1 true}
 !3 = !{!"llvm.loop.vectorize.enable", i1 true}
+
+; BFI computes if is taken 20 times, and loop 32 times. Make sure we round the
+; divisor up to 2 so that we don't vectorize the loop unprofitably.
+define void @round_scalar_pred_divisor(ptr %dst, double %x) {
+; CHECK-LABEL: define void @round_scalar_pred_divisor(
+; CHECK-SAME: ptr [[DST:%.*]], double [[X:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*]]:
+; CHECK-NEXT:    br label %[[LOOP:.*]]
+; CHECK:       [[LOOP]]:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LATCH:.*]] ]
+; CHECK-NEXT:    [[C:%.*]] = fcmp une double [[X]], 0.000000e+00
+; CHECK-NEXT:    br i1 [[C]], label %[[IF:.*]], label %[[LATCH]]
+; CHECK:       [[IF]]:
+; CHECK-NEXT:    [[TRUNC:%.*]] = trunc i64 [[IV]] to i32
+; CHECK-NEXT:    [[UITOFP:%.*]] = uitofp i32 [[TRUNC]] to double
+; CHECK-NEXT:    [[SIN:%.*]] = tail call double @llvm.sin.f64(double [[UITOFP]])
+; CHECK-NEXT:    [[FPTRUNC:%.*]] = fptrunc double [[SIN]] to float
+; CHECK-NEXT:    br label %[[LATCH]]
+; CHECK:       [[LATCH]]:
+; CHECK-NEXT:    [[PHI:%.*]] = phi float [ [[FPTRUNC]], %[[IF]] ], [ 0.000000e+00, %[[LOOP]] ]
+; CHECK-NEXT:    store float [[PHI]], ptr [[DST]], align 4
+; CHECK-NEXT:    [[IV_NEXT]] = add i64 [[IV]], 1
+; CHECK-NEXT:    [[EC:%.*]] = icmp eq i64 [[IV]], 1024
+; CHECK-NEXT:    br i1 [[EC]], label %[[EXIT:.*]], label %[[LOOP]]
+; CHECK:       [[EXIT]]:
+; CHECK-NEXT:    ret void
+;
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %latch ]
+  %c = fcmp une double %x, 0.0
+  br i1 %c, label %if, label %latch
+
+if:
+  %trunc = trunc i64 %iv to i32
+  %uitofp = uitofp i32 %trunc to double
+  %sin = tail call double @llvm.sin(double %uitofp)
+  %fptrunc = fptrunc double %sin to float
+  br label %latch
+
+latch:
+  %phi = phi float [ %fptrunc, %if ], [ 0.0, %loop ]
+  store float %phi, ptr %dst
+  %iv.next = add i64 %iv, 1
+  %ec = icmp eq i64 %iv, 1024
+  br i1 %ec, label %exit, label %loop
+
+exit:
+  ret void
+}
+
 ;.
 ; CHECK: [[META0]] = !{[[META1:![0-9]+]]}
 ; CHECK: [[META1]] = distinct !{[[META1]], [[META2:![0-9]+]]}

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)