[VPlan] Hoist predicated loads with complementary masks. (llvm#168373)

This patch adds a new VPlan transformation to hoist predicated loads, if
we can prove they execute unconditionally, i.e. there are 2 predicated
loads to the same address with complementary masks. Then we are
guaranteed to execute one of them on each iteration, allowing us to
remove the mask.

The transform groups masked replicating loads by their address SCEV,
then checks if there are 2 loads with complementary mask. If that is the
case, we check if there are any writes that may alias the load address
in the blocks between the first and last load with the same address.
The transforms operates after linearizing the CFG, but before
introducing replicate regions, which means this is just checking a chain
of consecutive blocks.

Currently this only uses noalias metadata to check for no-alias (using
the helpers added in llvm#166247).

Then we create an unpredicated VPReplicateRecipe at the position of the
first load, then replace all users of the grouped loads with it.

Small Alive2 proof for hoisting with complementary masks:
https://alive2.llvm.org/ce/z/kUx742

PR: llvm#168373

Author: fhahn

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -8336,6 +8336,7 @@ void LoopVectorizationPlanner::buildVPlansWithVPRecipes(ElementCount MinVF,
     if (auto Plan = tryToBuildVPlanWithVPRecipes(
             std::unique_ptr<VPlan>(VPlan0->duplicate()), SubRange, &LVer)) {
       // Now optimize the initial VPlan.
+      VPlanTransforms::hoistPredicatedLoads(*Plan, *PSE.getSE(), OrigLoop);
       VPlanTransforms::runPass(VPlanTransforms::truncateToMinimalBitwidths,
                                *Plan, CM.getMinimalBitwidths());
       VPlanTransforms::runPass(VPlanTransforms::optimize, *Plan);

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

@@ -139,6 +139,41 @@ bool VPlanTransforms::tryToConvertVPInstructionsToVPRecipes(
   return true;
 }
 
+// Check if a load can be hoisted by verifying it doesn't alias with any stores
+// in blocks between FirstBB and LastBB using scoped noalias metadata.
+static bool canHoistLoadWithNoAliasCheck(VPReplicateRecipe *Load,
+                                         VPBasicBlock *FirstBB,
+                                         VPBasicBlock *LastBB) {
+  // Get the load's memory location and check if it aliases with any stores
+  // using scoped noalias metadata.
+  auto LoadLoc = vputils::getMemoryLocation(*Load);
+  if (!LoadLoc || !LoadLoc->AATags.Scope)
+    return false;
+
+  const AAMDNodes &LoadAA = LoadLoc->AATags;
+  for (VPBlockBase *Block = FirstBB; Block;
+       Block = Block->getSingleSuccessor()) {
+    // This function assumes a simple linear chain of blocks. If there are
+    // multiple successors, we would need more complex analysis.
+    assert(Block->getNumSuccessors() <= 1 &&
+           "Expected at most one successor in block chain");
+    auto *VPBB = cast<VPBasicBlock>(Block);
+    for (VPRecipeBase &R : *VPBB) {
+      if (R.mayWriteToMemory()) {
+        auto Loc = vputils::getMemoryLocation(R);
+        // Bail out if we can't get the location or if the scoped noalias
+        // metadata indicates potential aliasing.
+        if (!Loc || ScopedNoAliasAAResult::mayAliasInScopes(
+                        LoadAA.Scope, Loc->AATags.NoAlias))
+          return false;
+      }
+    }
+    if (Block == LastBB)
+      break;
+  }
+  return true;
+}
+
 /// Return true if we do not know how to (mechanically) hoist or sink \p R out
 /// of a loop region.
 static bool cannotHoistOrSinkRecipe(const VPRecipeBase &R) {
@@ -4010,6 +4045,122 @@ void VPlanTransforms::hoistInvariantLoads(VPlan &Plan) {
   }
 }
 
+// Returns the intersection of metadata from a group of loads.
+static VPIRMetadata getCommonLoadMetadata(ArrayRef<VPReplicateRecipe *> Loads) {
+  VPIRMetadata CommonMetadata = *Loads.front();
+  for (VPReplicateRecipe *Load : drop_begin(Loads))
+    CommonMetadata.intersect(*Load);
+  return CommonMetadata;
+}
+
+void VPlanTransforms::hoistPredicatedLoads(VPlan &Plan, ScalarEvolution &SE,
+                                           const Loop *L) {
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  VPTypeAnalysis TypeInfo(Plan);
+  VPDominatorTree VPDT(Plan);
+
+  // Group predicated loads by their address SCEV.
+  DenseMap<const SCEV *, SmallVector<VPReplicateRecipe *>> LoadsByAddress;
+  for (VPBlockBase *Block : vp_depth_first_shallow(LoopRegion->getEntry())) {
+    auto *VPBB = cast<VPBasicBlock>(Block);
+    for (VPRecipeBase &R : *VPBB) {
+      auto *RepR = dyn_cast<VPReplicateRecipe>(&R);
+      if (!RepR || RepR->getOpcode() != Instruction::Load ||
+          !RepR->isPredicated())
+        continue;
+
+      VPValue *Addr = RepR->getOperand(0);
+      const SCEV *AddrSCEV = vputils::getSCEVExprForVPValue(Addr, SE, L);
+      if (!isa<SCEVCouldNotCompute>(AddrSCEV))
+        LoadsByAddress[AddrSCEV].push_back(RepR);
+    }
+  }
+
+  // For each address, collect loads with complementary masks, sort by
+  // dominance, and use the earliest load.
+  for (auto &[Addr, Loads] : LoadsByAddress) {
+    if (Loads.size() < 2)
+      continue;
+
+    // Collect groups of loads with complementary masks.
+    SmallVector<SmallVector<VPReplicateRecipe *, 4>> LoadGroups;
+    for (VPReplicateRecipe *&LoadI : Loads) {
+      if (!LoadI)
+        continue;
+
+      VPValue *MaskI = LoadI->getMask();
+      Type *TypeI = TypeInfo.inferScalarType(LoadI);
+      SmallVector<VPReplicateRecipe *, 4> Group;
+      Group.push_back(LoadI);
+      LoadI = nullptr;
+
+      // Find all loads with the same type.
+      for (VPReplicateRecipe *&LoadJ : Loads) {
+        if (!LoadJ)
+          continue;
+
+        Type *TypeJ = TypeInfo.inferScalarType(LoadJ);
+        if (TypeI == TypeJ) {
+          Group.push_back(LoadJ);
+          LoadJ = nullptr;
+        }
+      }
+
+      // Check if any load in the group has a complementary mask with another,
+      // that is M1 == NOT(M2) or M2 == NOT(M1).
+      bool HasComplementaryMask =
+          any_of(drop_begin(Group), [MaskI](VPReplicateRecipe *Load) {
+            VPValue *MaskJ = Load->getMask();
+            return match(MaskI, m_Not(m_Specific(MaskJ))) ||
+                   match(MaskJ, m_Not(m_Specific(MaskI)));
+          });
+
+      if (HasComplementaryMask)
+        LoadGroups.push_back(std::move(Group));
+    }
+
+    // For each group, check memory dependencies and hoist the earliest load.
+    for (auto &Group : LoadGroups) {
+      // Sort loads by dominance order, with earliest (most dominating) first.
+      sort(Group, [&VPDT](VPReplicateRecipe *A, VPReplicateRecipe *B) {
+        return VPDT.properlyDominates(A, B);
+      });
+
+      VPReplicateRecipe *EarliestLoad = Group.front();
+      VPBasicBlock *FirstBB = EarliestLoad->getParent();
+      VPBasicBlock *LastBB = Group.back()->getParent();
+
+      // Check that the load doesn't alias with stores between first and last.
+      if (!canHoistLoadWithNoAliasCheck(EarliestLoad, FirstBB, LastBB))
+        continue;
+
+      // Find the load with minimum alignment to use.
+      auto *LoadWithMinAlign =
+          *min_element(Group, [](VPReplicateRecipe *A, VPReplicateRecipe *B) {
+            return cast<LoadInst>(A->getUnderlyingInstr())->getAlign() <
+                   cast<LoadInst>(B->getUnderlyingInstr())->getAlign();
+          });
+
+      // Collect common metadata from all loads in the group.
+      VPIRMetadata CommonMetadata = getCommonLoadMetadata(Group);
+
+      // Create an unpredicated load with minimum alignment using the earliest
+      // dominating address and common metadata.
+      auto *UnpredicatedLoad = new VPReplicateRecipe(
+          LoadWithMinAlign->getUnderlyingInstr(), EarliestLoad->getOperand(0),
+          /*IsSingleScalar=*/false, /*Mask=*/nullptr, /*Flags=*/{},
+          CommonMetadata);
+      UnpredicatedLoad->insertBefore(EarliestLoad);
+
+      // Replace all loads in the group with the unpredicated load.
+      for (VPReplicateRecipe *Load : Group) {
+        Load->replaceAllUsesWith(UnpredicatedLoad);
+        Load->eraseFromParent();
+      }
+    }
+  }
+}
+
 void VPlanTransforms::materializeConstantVectorTripCount(
     VPlan &Plan, ElementCount BestVF, unsigned BestUF,
     PredicatedScalarEvolution &PSE) {

llvm/lib/Transforms/Vectorize/VPlanTransforms.h

@@ -314,6 +314,12 @@ struct VPlanTransforms {
   /// plan using noalias metadata.
   static void hoistInvariantLoads(VPlan &Plan);
 
+  /// Hoist predicated loads from the same address to the loop entry block, if
+  /// they are guaranteed to execute on both paths (i.e., in replicate regions
+  /// with complementary masks P and NOT P).
+  static void hoistPredicatedLoads(VPlan &Plan, ScalarEvolution &SE,
+                                   const Loop *L);
+
   // Materialize vector trip counts for constants early if it can simply be
   // computed as (Original TC / VF * UF) * VF * UF.
   static void