mlir/lib/Transforms/Inliner.cpp

//===- Inliner.cpp - Pass to inline function calls ------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implements a basic inlining algorithm that operates bottom up over
// the Strongly Connect Components(SCCs) of the CallGraph. This enables a more
// incremental propagation of inlining decisions from the leafs to the roots of
// the callgraph.
//
//===----------------------------------------------------------------------===//

#include "PassDetail.h"
#include "mlir/Analysis/CallGraph.h"
#include "mlir/IR/PatternMatch.h"
#include "mlir/Interfaces/SideEffectInterfaces.h"
#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
#include "mlir/Transforms/InliningUtils.h"
#include "mlir/Transforms/Passes.h"
#include "llvm/ADT/SCCIterator.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/Parallel.h"

#define DEBUG_TYPE "inlining"

using namespace mlir;

//===----------------------------------------------------------------------===//
// Symbol Use Tracking
//===----------------------------------------------------------------------===//

/// Walk all of the used symbol callgraph nodes referenced with the given op.
static void walkReferencedSymbolNodes(
    Operation *op, CallGraph &cg, SymbolTableCollection &symbolTable,
    DenseMap<Attribute, CallGraphNode *> &resolvedRefs,
    function_ref<void(CallGraphNode *, Operation *)> callback) {
  auto symbolUses = SymbolTable::getSymbolUses(op);
  assert(symbolUses && "expected uses to be valid");

  Operation *symbolTableOp = op->getParentOp();
  for (const SymbolTable::SymbolUse &use : *symbolUses) {
    auto refIt = resolvedRefs.insert({use.getSymbolRef(), nullptr});
    CallGraphNode *&node = refIt.first->second;

    // If this is the first instance of this reference, try to resolve a
    // callgraph node for it.
    if (refIt.second) {
      auto *symbolOp = symbolTable.lookupNearestSymbolFrom(symbolTableOp,
                                                           use.getSymbolRef());
      auto callableOp = dyn_cast_or_null<CallableOpInterface>(symbolOp);
      if (!callableOp)
        continue;
      node = cg.lookupNode(callableOp.getCallableRegion());
    }
    if (node)
      callback(node, use.getUser());
  }
}

//===----------------------------------------------------------------------===//
// CGUseList

namespace {
/// This struct tracks the uses of callgraph nodes that can be dropped when
/// use_empty. It directly tracks and manages a use-list for all of the
/// call-graph nodes. This is necessary because many callgraph nodes are
/// referenced by SymbolRefAttr, which has no mechanism akin to the SSA `Use`
/// class.
struct CGUseList {
  /// This struct tracks the uses of callgraph nodes within a specific
  /// operation.
  struct CGUser {
    /// Any nodes referenced in the top-level attribute list of this user. We
    /// use a set here because the number of references does not matter.
    DenseSet<CallGraphNode *> topLevelUses;

    /// Uses of nodes referenced by nested operations.
    DenseMap<CallGraphNode *, int> innerUses;
  };

  CGUseList(Operation *op, CallGraph &cg, SymbolTableCollection &symbolTable);

  /// Drop uses of nodes referred to by the given call operation that resides
  /// within 'userNode'.
  void dropCallUses(CallGraphNode *userNode, Operation *callOp, CallGraph &cg);

  /// Remove the given node from the use list.
  void eraseNode(CallGraphNode *node);

  /// Returns true if the given callgraph node has no uses and can be pruned.
  bool isDead(CallGraphNode *node) const;

  /// Returns true if the given callgraph node has a single use and can be
  /// discarded.
  bool hasOneUseAndDiscardable(CallGraphNode *node) const;

  /// Recompute the uses held by the given callgraph node.
  void recomputeUses(CallGraphNode *node, CallGraph &cg);

  /// Merge the uses of 'lhs' with the uses of the 'rhs' after inlining a copy
  /// of 'lhs' into 'rhs'.
  void mergeUsesAfterInlining(CallGraphNode *lhs, CallGraphNode *rhs);

private:
  /// Decrement the uses of discardable nodes referenced by the given user.
  void decrementDiscardableUses(CGUser &uses);

  /// A mapping between a discardable callgraph node (that is a symbol) and the
  /// number of uses for this node.
  DenseMap<CallGraphNode *, int> discardableSymNodeUses;

  /// A mapping between a callgraph node and the symbol callgraph nodes that it
  /// uses.
  DenseMap<CallGraphNode *, CGUser> nodeUses;

  /// A symbol table to use when resolving call lookups.
  SymbolTableCollection &symbolTable;
};
} // end anonymous namespace

CGUseList::CGUseList(Operation *op, CallGraph &cg,
                     SymbolTableCollection &symbolTable)
    : symbolTable(symbolTable) {
  /// A set of callgraph nodes that are always known to be live during inlining.
  DenseMap<Attribute, CallGraphNode *> alwaysLiveNodes;

  // Walk each of the symbol tables looking for discardable callgraph nodes.
  auto walkFn = [&](Operation *symbolTableOp, bool allUsesVisible) {
    for (Operation &op : symbolTableOp->getRegion(0).getOps()) {
      // If this is a callgraph operation, check to see if it is discardable.
      if (auto callable = dyn_cast<CallableOpInterface>(&op)) {
        if (auto *node = cg.lookupNode(callable.getCallableRegion())) {
          SymbolOpInterface symbol = dyn_cast<SymbolOpInterface>(&op);
          if (symbol && (allUsesVisible || symbol.isPrivate()) &&
              symbol.canDiscardOnUseEmpty()) {
            discardableSymNodeUses.try_emplace(node, 0);
          }
          continue;
        }
      }
      // Otherwise, check for any referenced nodes. These will be always-live.
      walkReferencedSymbolNodes(&op, cg, symbolTable, alwaysLiveNodes,
                                [](CallGraphNode *, Operation *) {});
    }
  };
  SymbolTable::walkSymbolTables(op, /*allSymUsesVisible=*/!op->getBlock(),
                                walkFn);

  // Drop the use information for any discardable nodes that are always live.
  for (auto &it : alwaysLiveNodes)
    discardableSymNodeUses.erase(it.second);

  // Compute the uses for each of the callable nodes in the graph.
  for (CallGraphNode *node : cg)
    recomputeUses(node, cg);
}

void CGUseList::dropCallUses(CallGraphNode *userNode, Operation *callOp,
                             CallGraph &cg) {
  auto &userRefs = nodeUses[userNode].innerUses;
  auto walkFn = [&](CallGraphNode *node, Operation *user) {
    auto parentIt = userRefs.find(node);
    if (parentIt == userRefs.end())
      return;
    --parentIt->second;
    --discardableSymNodeUses[node];
  };
  DenseMap<Attribute, CallGraphNode *> resolvedRefs;
  walkReferencedSymbolNodes(callOp, cg, symbolTable, resolvedRefs, walkFn);
}

void CGUseList::eraseNode(CallGraphNode *node) {
  // Drop all child nodes.
  for (auto &edge : *node)
    if (edge.isChild())
      eraseNode(edge.getTarget());

  // Drop the uses held by this node and erase it.
  auto useIt = nodeUses.find(node);
  assert(useIt != nodeUses.end() && "expected node to be valid");
  decrementDiscardableUses(useIt->getSecond());
  nodeUses.erase(useIt);
  discardableSymNodeUses.erase(node);
}

bool CGUseList::isDead(CallGraphNode *node) const {
  // If the parent operation isn't a symbol, simply check normal SSA deadness.
  Operation *nodeOp = node->getCallableRegion()->getParentOp();
  if (!isa<SymbolOpInterface>(nodeOp))
    return MemoryEffectOpInterface::hasNoEffect(nodeOp) && nodeOp->use_empty();

  // Otherwise, check the number of symbol uses.
  auto symbolIt = discardableSymNodeUses.find(node);
  return symbolIt != discardableSymNodeUses.end() && symbolIt->second == 0;
}

bool CGUseList::hasOneUseAndDiscardable(CallGraphNode *node) const {
  // If this isn't a symbol node, check for side-effects and SSA use count.
  Operation *nodeOp = node->getCallableRegion()->getParentOp();
  if (!isa<SymbolOpInterface>(nodeOp))
    return MemoryEffectOpInterface::hasNoEffect(nodeOp) && nodeOp->hasOneUse();

  // Otherwise, check the number of symbol uses.
  auto symbolIt = discardableSymNodeUses.find(node);
  return symbolIt != discardableSymNodeUses.end() && symbolIt->second == 1;
}

void CGUseList::recomputeUses(CallGraphNode *node, CallGraph &cg) {
  Operation *parentOp = node->getCallableRegion()->getParentOp();
  CGUser &uses = nodeUses[node];
  decrementDiscardableUses(uses);

  // Collect the new discardable uses within this node.
  uses = CGUser();
  DenseMap<Attribute, CallGraphNode *> resolvedRefs;
  auto walkFn = [&](CallGraphNode *refNode, Operation *user) {
    auto discardSymIt = discardableSymNodeUses.find(refNode);
    if (discardSymIt == discardableSymNodeUses.end())
      return;

    if (user != parentOp)
      ++uses.innerUses[refNode];
    else if (!uses.topLevelUses.insert(refNode).second)
      return;
    ++discardSymIt->second;
  };
  walkReferencedSymbolNodes(parentOp, cg, symbolTable, resolvedRefs, walkFn);
}

void CGUseList::mergeUsesAfterInlining(CallGraphNode *lhs, CallGraphNode *rhs) {
  auto &lhsUses = nodeUses[lhs], &rhsUses = nodeUses[rhs];
  for (auto &useIt : lhsUses.innerUses) {
    rhsUses.innerUses[useIt.first] += useIt.second;
    discardableSymNodeUses[useIt.first] += useIt.second;
  }
}

void CGUseList::decrementDiscardableUses(CGUser &uses) {
  for (CallGraphNode *node : uses.topLevelUses)
    --discardableSymNodeUses[node];
  for (auto &it : uses.innerUses)
    discardableSymNodeUses[it.first] -= it.second;
}

//===----------------------------------------------------------------------===//
// CallGraph traversal
//===----------------------------------------------------------------------===//

namespace {
/// This class represents a specific callgraph SCC.
class CallGraphSCC {
public:
  CallGraphSCC(llvm::scc_iterator<const CallGraph *> &parentIterator)
      : parentIterator(parentIterator) {}
  /// Return a range over the nodes within this SCC.
  std::vector<CallGraphNode *>::iterator begin() { return nodes.begin(); }
  std::vector<CallGraphNode *>::iterator end() { return nodes.end(); }

  /// Reset the nodes of this SCC with those provided.
  void reset(const std::vector<CallGraphNode *> &newNodes) { nodes = newNodes; }

  /// Remove the given node from this SCC.
  void remove(CallGraphNode *node) {
    auto it = llvm::find(nodes, node);
    if (it != nodes.end()) {
      nodes.erase(it);
      parentIterator.ReplaceNode(node, nullptr);
    }
  }

private:
  std::vector<CallGraphNode *> nodes;
  llvm::scc_iterator<const CallGraph *> &parentIterator;
};
} // end anonymous namespace

/// Run a given transformation over the SCCs of the callgraph in a bottom up
/// traversal.
static void
runTransformOnCGSCCs(const CallGraph &cg,
                     function_ref<void(CallGraphSCC &)> sccTransformer) {
  llvm::scc_iterator<const CallGraph *> cgi = llvm::scc_begin(&cg);
  CallGraphSCC currentSCC(cgi);
  while (!cgi.isAtEnd()) {
    // Copy the current SCC and increment so that the transformer can modify the
    // SCC without invalidating our iterator.
    currentSCC.reset(*cgi);
    ++cgi;
    sccTransformer(currentSCC);
  }
}

namespace {
/// This struct represents a resolved call to a given callgraph node. Given that
/// the call does not actually contain a direct reference to the
/// Region(CallGraphNode) that it is dispatching to, we need to resolve them
/// explicitly.
struct ResolvedCall {
  ResolvedCall(CallOpInterface call, CallGraphNode *sourceNode,
               CallGraphNode *targetNode)
      : call(call), sourceNode(sourceNode), targetNode(targetNode) {}
  CallOpInterface call;
  CallGraphNode *sourceNode, *targetNode;
};
} // end anonymous namespace

/// Collect all of the callable operations within the given range of blocks. If
/// `traverseNestedCGNodes` is true, this will also collect call operations
/// inside of nested callgraph nodes.
static void collectCallOps(iterator_range<Region::iterator> blocks,
                           CallGraphNode *sourceNode, CallGraph &cg,
                           SymbolTableCollection &symbolTable,
                           SmallVectorImpl<ResolvedCall> &calls,
                           bool traverseNestedCGNodes) {
  SmallVector<std::pair<Block *, CallGraphNode *>, 8> worklist;
  auto addToWorklist = [&](CallGraphNode *node,
                           iterator_range<Region::iterator> blocks) {
    for (Block &block : blocks)
      worklist.emplace_back(&block, node);
  };

  addToWorklist(sourceNode, blocks);
  while (!worklist.empty()) {
    Block *block;
    std::tie(block, sourceNode) = worklist.pop_back_val();

    for (Operation &op : *block) {
      if (auto call = dyn_cast<CallOpInterface>(op)) {
        // TODO: Support inlining nested call references.
        CallInterfaceCallable callable = call.getCallableForCallee();
        if (SymbolRefAttr symRef = callable.dyn_cast<SymbolRefAttr>()) {
          if (!symRef.isa<FlatSymbolRefAttr>())
            continue;
        }

        CallGraphNode *targetNode = cg.resolveCallable(call, symbolTable);
        if (!targetNode->isExternal())
          calls.emplace_back(call, sourceNode, targetNode);
        continue;
      }

      // If this is not a call, traverse the nested regions. If
      // `traverseNestedCGNodes` is false, then don't traverse nested call graph
      // regions.
      for (auto &nestedRegion : op.getRegions()) {
        CallGraphNode *nestedNode = cg.lookupNode(&nestedRegion);
        if (traverseNestedCGNodes || !nestedNode)
          addToWorklist(nestedNode ? nestedNode : sourceNode, nestedRegion);
      }
    }
  }
}

//===----------------------------------------------------------------------===//
// Inliner
//===----------------------------------------------------------------------===//
namespace {
/// This class provides a specialization of the main inlining interface.
struct Inliner : public InlinerInterface {
  Inliner(MLIRContext *context, CallGraph &cg,
          SymbolTableCollection &symbolTable)
      : InlinerInterface(context), cg(cg), symbolTable(symbolTable) {}

  /// Process a set of blocks that have been inlined. This callback is invoked
  /// *before* inlined terminator operations have been processed.
  void
  processInlinedBlocks(iterator_range<Region::iterator> inlinedBlocks) final {
    // Find the closest callgraph node from the first block.
    CallGraphNode *node;
    Region *region = inlinedBlocks.begin()->getParent();
    while (!(node = cg.lookupNode(region))) {
      region = region->getParentRegion();
      assert(region && "expected valid parent node");
    }

    collectCallOps(inlinedBlocks, node, cg, symbolTable, calls,
                   /*traverseNestedCGNodes=*/true);
  }

  /// Mark the given callgraph node for deletion.
  void markForDeletion(CallGraphNode *node) { deadNodes.insert(node); }

  /// This method properly disposes of callables that became dead during
  /// inlining. This should not be called while iterating over the SCCs.
  void eraseDeadCallables() {
    for (CallGraphNode *node : deadNodes)
      node->getCallableRegion()->getParentOp()->erase();
  }

  /// The set of callables known to be dead.
  SmallPtrSet<CallGraphNode *, 8> deadNodes;

  /// The current set of call instructions to consider for inlining.
  SmallVector<ResolvedCall, 8> calls;

  /// The callgraph being operated on.
  CallGraph &cg;

  /// A symbol table to use when resolving call lookups.
  SymbolTableCollection &symbolTable;
};
} // namespace

/// Returns true if the given call should be inlined.
static bool shouldInline(ResolvedCall &resolvedCall) {
  // Don't allow inlining terminator calls. We currently don't support this
  // case.
  if (resolvedCall.call.getOperation()->isKnownTerminator())
    return false;

  // Don't allow inlining if the target is an ancestor of the call. This
  // prevents inlining recursively.
  if (resolvedCall.targetNode->getCallableRegion()->isAncestor(
          resolvedCall.call.getParentRegion()))
    return false;

  // Otherwise, inline.
  return true;
}

/// Attempt to inline calls within the given scc. This function returns
/// success if any calls were inlined, failure otherwise.
static LogicalResult inlineCallsInSCC(Inliner &inliner, CGUseList &useList,
                                      CallGraphSCC &currentSCC) {
  CallGraph &cg = inliner.cg;
  auto &calls = inliner.calls;

  // A set of dead nodes to remove after inlining.
  SmallVector<CallGraphNode *, 1> deadNodes;

  // Collect all of the direct calls within the nodes of the current SCC. We
  // don't traverse nested callgraph nodes, because they are handled separately
  // likely within a different SCC.
  for (CallGraphNode *node : currentSCC) {
    if (node->isExternal())
      continue;

    // Don't collect calls if the node is already dead.
    if (useList.isDead(node)) {
      deadNodes.push_back(node);
    } else {
      collectCallOps(*node->getCallableRegion(), node, cg, inliner.symbolTable,
                     calls, /*traverseNestedCGNodes=*/false);
    }
  }

  // Try to inline each of the call operations. Don't cache the end iterator
  // here as more calls may be added during inlining.
  bool inlinedAnyCalls = false;
  for (unsigned i = 0; i != calls.size(); ++i) {
    ResolvedCall it = calls[i];
    bool doInline = shouldInline(it);
    CallOpInterface call = it.call;
    LLVM_DEBUG({
      if (doInline)
        llvm::dbgs() << "* Inlining call: " << call << "\n";
      else
        llvm::dbgs() << "* Not inlining call: " << call << "\n";
    });
    if (!doInline)
      continue;
    Region *targetRegion = it.targetNode->getCallableRegion();

    // If this is the last call to the target node and the node is discardable,
    // then inline it in-place and delete the node if successful.
    bool inlineInPlace = useList.hasOneUseAndDiscardable(it.targetNode);

    LogicalResult inlineResult = inlineCall(
        inliner, call, cast<CallableOpInterface>(targetRegion->getParentOp()),
        targetRegion, /*shouldCloneInlinedRegion=*/!inlineInPlace);
    if (failed(inlineResult)) {
      LLVM_DEBUG(llvm::dbgs() << "** Failed to inline\n");
      continue;
    }
    inlinedAnyCalls = true;

    // If the inlining was successful, Merge the new uses into the source node.
    useList.dropCallUses(it.sourceNode, call.getOperation(), cg);
    useList.mergeUsesAfterInlining(it.targetNode, it.sourceNode);

    // then erase the call.
    call.erase();

    // If we inlined in place, mark the node for deletion.
    if (inlineInPlace) {
      useList.eraseNode(it.targetNode);
      deadNodes.push_back(it.targetNode);
    }
  }

  for (CallGraphNode *node : deadNodes) {
    currentSCC.remove(node);
    inliner.markForDeletion(node);
  }
  calls.clear();
  return success(inlinedAnyCalls);
}

/// Canonicalize the nodes within the given SCC with the given set of
/// canonicalization patterns.
static void canonicalizeSCC(CallGraph &cg, CGUseList &useList,
                            CallGraphSCC &currentSCC, MLIRContext *context,
                            const FrozenRewritePatternList &canonPatterns) {
  // Collect the sets of nodes to canonicalize.
  SmallVector<CallGraphNode *, 4> nodesToCanonicalize;
  for (auto *node : currentSCC) {
    // Don't canonicalize the external node, it has no valid callable region.
    if (node->isExternal())
      continue;

    // Don't canonicalize nodes with children. Nodes with children
    // require special handling as we may remove the node during
    // canonicalization. In the future, we should be able to handle this
    // case with proper node deletion tracking.
    if (node->hasChildren())
      continue;

    // We also won't apply canonicalizations for nodes that are not
    // isolated. This avoids potentially mutating the regions of nodes defined
    // above, this is also a stipulation of the 'applyPatternsAndFoldGreedily'
    // driver.
    auto *region = node->getCallableRegion();
    if (!region->getParentOp()->isKnownIsolatedFromAbove())
      continue;
    nodesToCanonicalize.push_back(node);
  }
  if (nodesToCanonicalize.empty())
    return;

  // Canonicalize each of the nodes within the SCC in parallel.
  // NOTE: This is simple now, because we don't enable canonicalizing nodes
  // within children. When we remove this restriction, this logic will need to
  // be reworked.
  if (context->isMultithreadingEnabled()) {
    ParallelDiagnosticHandler canonicalizationHandler(context);
    llvm::parallelForEachN(
        /*Begin=*/0, /*End=*/nodesToCanonicalize.size(), [&](size_t index) {
          // Set the order for this thread so that diagnostics will be properly
          // ordered.
          canonicalizationHandler.setOrderIDForThread(index);

          // Apply the canonicalization patterns to this region.
          auto *node = nodesToCanonicalize[index];
          applyPatternsAndFoldGreedily(*node->getCallableRegion(),
                                       canonPatterns);

          // Make sure to reset the order ID for the diagnostic handler, as this
          // thread may be used in a different context.
          canonicalizationHandler.eraseOrderIDForThread();
        });
  } else {
    for (CallGraphNode *node : nodesToCanonicalize)
      applyPatternsAndFoldGreedily(*node->getCallableRegion(), canonPatterns);
  }

  // Recompute the uses held by each of the nodes.
  for (CallGraphNode *node : nodesToCanonicalize)
    useList.recomputeUses(node, cg);
}

//===----------------------------------------------------------------------===//
// InlinerPass
//===----------------------------------------------------------------------===//

namespace {
struct InlinerPass : public InlinerBase<InlinerPass> {
  void runOnOperation() override;

  /// Attempt to inline calls within the given scc, and run canonicalizations
  /// with the given patterns, until a fixed point is reached. This allows for
  /// the inlining of newly devirtualized calls.
  void inlineSCC(Inliner &inliner, CGUseList &useList, CallGraphSCC &currentSCC,
                 MLIRContext *context,
                 const FrozenRewritePatternList &canonPatterns);
};
} // end anonymous namespace

void InlinerPass::runOnOperation() {
  CallGraph &cg = getAnalysis<CallGraph>();
  auto *context = &getContext();

  // The inliner should only be run on operations that define a symbol table,
  // as the callgraph will need to resolve references.
  Operation *op = getOperation();
  if (!op->hasTrait<OpTrait::SymbolTable>()) {
    op->emitOpError() << " was scheduled to run under the inliner, but does "
                         "not define a symbol table";
    return signalPassFailure();
  }

  // Collect a set of canonicalization patterns to use when simplifying
  // callable regions within an SCC.
  OwningRewritePatternList canonPatterns;
  for (auto *op : context->getRegisteredOperations())
    op->getCanonicalizationPatterns(canonPatterns, context);
  FrozenRewritePatternList frozenCanonPatterns(std::move(canonPatterns));

  // Run the inline transform in post-order over the SCCs in the callgraph.
  SymbolTableCollection symbolTable;
  Inliner inliner(context, cg, symbolTable);
  CGUseList useList(getOperation(), cg, symbolTable);
  runTransformOnCGSCCs(cg, [&](CallGraphSCC &scc) {
    inlineSCC(inliner, useList, scc, context, frozenCanonPatterns);
  });

  // After inlining, make sure to erase any callables proven to be dead.
  inliner.eraseDeadCallables();
}

void InlinerPass::inlineSCC(Inliner &inliner, CGUseList &useList,
                            CallGraphSCC &currentSCC, MLIRContext *context,
                            const FrozenRewritePatternList &canonPatterns) {
  // If we successfully inlined any calls, run some simplifications on the
  // nodes of the scc. Continue attempting to inline until we reach a fixed
  // point, or a maximum iteration count. We canonicalize here as it may
  // devirtualize new calls, as well as give us a better cost model.
  unsigned iterationCount = 0;
  while (succeeded(inlineCallsInSCC(inliner, useList, currentSCC))) {
    // If we aren't allowing simplifications or the max iteration count was
    // reached, then bail out early.
    if (disableCanonicalization || ++iterationCount >= maxInliningIterations)
      break;
    canonicalizeSCC(inliner.cg, useList, currentSCC, context, canonPatterns);
  }
}

std::unique_ptr<Pass> mlir::createInlinerPass() {
  return std::make_unique<InlinerPass>();
}