LinearMapInfo.cpp

//===--- LinearMapInfo.cpp ------------------------------------*- C++ -*---===//
//
// This source file is part of the Swift.org open source project
//
// Copyright (c) 2019 - 2020 Apple Inc. and the Swift project authors
// Licensed under Apache License v2.0 with Runtime Library Exception
//
// See https://swift.org/LICENSE.txt for license information
// See https://swift.org/CONTRIBUTORS.txt for the list of Swift project authors
//
//===----------------------------------------------------------------------===//
//
// Linear map tuple and branching trace enum information for differentiation.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "differentiation"

#include "swift/SIL/ApplySite.h"
#include "swift/SILOptimizer/Differentiation/LinearMapInfo.h"
#include "swift/SILOptimizer/Differentiation/ADContext.h"

#include "swift/AST/DeclContext.h"
#include "swift/AST/ParameterList.h"
#include "swift/AST/SourceFile.h"
#include "swift/Basic/Assertions.h"
#include "swift/SIL/LoopInfo.h"

namespace swift {
namespace autodiff {

//===----------------------------------------------------------------------===//
// Local helpers
//===----------------------------------------------------------------------===//

/// Clone the generic parameters of the given generic signature and return a new
/// `GenericParamList`.
static GenericParamList *cloneGenericParameters(ASTContext &ctx,
                                                DeclContext *dc,
                                                CanGenericSignature sig) {
  SmallVector<GenericTypeParamDecl *, 2> clonedParams;
  for (auto paramType : sig.getGenericParams()) {
    auto *clonedParam = GenericTypeParamDecl::createImplicit(
        dc, paramType->getName(), paramType->getDepth(), paramType->getIndex(),
        paramType->getParamKind());
    clonedParam->setDeclContext(dc);
    clonedParams.push_back(clonedParam);
  }
  return GenericParamList::create(ctx, SourceLoc(), clonedParams, SourceLoc());
}

//===----------------------------------------------------------------------===//
// LinearMapInfo methods
//===----------------------------------------------------------------------===//

LinearMapInfo::LinearMapInfo(ADContext &context, AutoDiffLinearMapKind kind,
                             SILFunction *original, SILFunction *derivative,
                             const AutoDiffConfig &config,
                             const DifferentiableActivityInfo &activityInfo,
                             SILLoopInfo *loopInfo)
    : kind(kind), original(original), derivative(derivative),
      activityInfo(activityInfo), loopInfo(loopInfo), config(config),
      synthesizedFile(context.getOrCreateSynthesizedFile(original)),
      typeConverter(context.getTypeConverter()) {
  generateDifferentiationDataStructures(context, derivative);
}

SILType LinearMapInfo::remapTypeInDerivative(SILType ty) {
  if (ty.hasArchetype())
    return derivative->mapTypeIntoContext(ty.mapTypeOutOfContext());
  return derivative->mapTypeIntoContext(ty);
}

EnumDecl *
LinearMapInfo::createBranchingTraceDecl(SILBasicBlock *originalBB,
                                        CanGenericSignature genericSig) {
  assert(originalBB->getParent() == original);
  auto &astCtx = original->getASTContext();
  auto &file = getSynthesizedFile();
  // Create a branching trace enum.
  Mangle::ASTMangler mangler(astCtx);
  auto config = this->config.withGenericSignature(genericSig);
  auto enumName = mangler.mangleAutoDiffGeneratedDeclaration(
      AutoDiffGeneratedDeclarationKind::BranchingTraceEnum,
      original->getName().str(), originalBB->getDebugID(), kind, config);
  auto enumId = astCtx.getIdentifier(enumName);
  auto loc = original->getLocation().getSourceLoc();
  GenericParamList *genericParams = nullptr;
  if (genericSig)
    genericParams = cloneGenericParameters(astCtx, &file, genericSig);
  auto *branchingTraceDecl = new (astCtx) EnumDecl(
      /*EnumLoc*/ SourceLoc(), /*Name*/ enumId, /*NameLoc*/ loc,
      /*Inherited*/ {}, /*GenericParams*/ genericParams, /*DC*/ &file);
  // Note: must mark enum as implicit to satisfy assertion in
  // `Parser::parseDeclListDelayed`.
  branchingTraceDecl->setImplicit();
  if (genericSig)
    branchingTraceDecl->setGenericSignature(genericSig);
  switch (original->getEffectiveSymbolLinkage()) {
  case swift::SILLinkage::Public:
  case swift::SILLinkage::PublicNonABI:
    // Branching trace enums shall not be resilient.
    branchingTraceDecl->getAttrs().add(new (astCtx) FrozenAttr(/*implicit*/ true));
    branchingTraceDecl->getAttrs().add(new (astCtx) UsableFromInlineAttr(/*Implicit*/ true));
    LLVM_FALLTHROUGH;
  case swift::SILLinkage::Hidden:
  case swift::SILLinkage::Shared:
    branchingTraceDecl->setAccess(AccessLevel::Internal);
    break;
  case swift::SILLinkage::Private:
    branchingTraceDecl->setAccess(AccessLevel::FilePrivate);
    break;
  default:
    // When the original function has external linkage, we create an internal
    // struct for use by our own module. This is necessary for cross-cell
    // differentiation in Jupyter.
    // TODO: Add a test in the compiler that exercises a similar situation as
    // cross-cell differentiation in Jupyter.
    branchingTraceDecl->setAccess(AccessLevel::Internal);
  }
  file.addTopLevelDecl(branchingTraceDecl);
  file.getParentModule()->clearLookupCache();

  return branchingTraceDecl;
}

void LinearMapInfo::populateBranchingTraceDecl(SILBasicBlock *originalBB,
                                               SILLoopInfo *loopInfo) {
  auto &astCtx = original->getASTContext();
  auto *moduleDecl = original->getModule().getSwiftModule();
  auto loc = original->getLocation().getSourceLoc();
  auto *branchingTraceDecl = getBranchingTraceDecl(originalBB);

  // Add basic block enum cases.
  for (auto *predBB : originalBB->getPredecessorBlocks()) {
    // Create dummy declaration representing enum case parameter.
    auto *decl = new (astCtx)
        ParamDecl(loc, loc, Identifier(), loc, Identifier(), moduleDecl);
    decl->setSpecifier(ParamDecl::Specifier::Default);
    // If predecessor block is in a loop, its linear map tuple will be
    // indirectly referenced in memory owned by the context object. The payload
    // is just a raw pointer.
    if (loopInfo->getLoopFor(predBB)) {
      heapAllocatedContext = true;
      decl->setInterfaceType(astCtx.TheRawPointerType);
    } else { // Otherwise the payload is the linear map tuple.
      auto *linearMapStructTy = getLinearMapTupleType(predBB);
      // Do not create entries for unreachable predecessors
      if (!linearMapStructTy)
        continue;
      auto canLinearMapStructTy = linearMapStructTy->getCanonicalType();
      decl->setInterfaceType(
          canLinearMapStructTy->hasArchetype()
              ? canLinearMapStructTy->mapTypeOutOfContext() : canLinearMapStructTy);
    }
    // Create enum element and enum case declarations.
    auto *paramList = ParameterList::create(astCtx, {decl});
    auto bbId = "bb" + std::to_string(predBB->getDebugID());
    auto *enumEltDecl = new (astCtx) EnumElementDecl(
        /*IdentifierLoc*/ loc, DeclName(astCtx.getIdentifier(bbId)), paramList,
        loc, /*RawValueExpr*/ nullptr, branchingTraceDecl);
    enumEltDecl->setImplicit();
    auto *enumCaseDecl = EnumCaseDecl::create(
        /*CaseLoc*/ loc, {enumEltDecl}, branchingTraceDecl);
    enumCaseDecl->setImplicit();
    branchingTraceDecl->addMember(enumEltDecl);
    branchingTraceDecl->addMember(enumCaseDecl);
    // Record enum element declaration.
    branchingTraceEnumCases.insert({{predBB, originalBB}, enumEltDecl});
  }
}


Type LinearMapInfo::getLinearMapType(ADContext &context, FullApplySite fai) {
  SmallVector<SILValue, 4> allResults;
  SmallVector<unsigned, 8> activeParamIndices;
  SmallVector<unsigned, 8> activeResultIndices;
  collectMinimalIndicesForFunctionCall(fai, config, activityInfo, allResults,
                                       activeParamIndices, activeResultIndices);

  // Check if there are any active results or arguments. If not, skip
  // this instruction.
  auto hasActiveResults = llvm::any_of(allResults, [&](SILValue res) {
    return activityInfo.isActive(res, config);
  });
  bool hasActiveSemanticResultArgument = false;
  bool hasActiveArguments = false;
  auto numIndirectResults = fai.getNumIndirectSILResults();
  for (auto argIdx : range(fai.getSubstCalleeConv().getNumParameters())) {
    auto arg = fai.getArgumentsWithoutIndirectResults()[argIdx];
    if (activityInfo.isActive(arg, config)) {
      hasActiveArguments = true;
      auto paramInfo = fai.getSubstCalleeConv().getParamInfoForSILArg(
          numIndirectResults + argIdx);
      if (paramInfo.isAutoDiffSemanticResult())
        hasActiveSemanticResultArgument = true;
    }
  }
  if (!hasActiveArguments)
    return {};
  if (!hasActiveResults && !hasActiveSemanticResultArgument)
    return {};

  // Compute differentiability parameters.
  // - If the callee has `@differentiable` function type, use differentiation
  //   parameters from the function type.
  // - Otherwise, use the active parameters.
  IndexSubset *parameters;
  auto origFnSubstTy = fai.getSubstCalleeType();
  auto remappedOrigFnSubstTy =
      remapTypeInDerivative(SILType::getPrimitiveObjectType(origFnSubstTy))
          .castTo<SILFunctionType>()
          ->getUnsubstitutedType(original->getModule());
  if (remappedOrigFnSubstTy->isDifferentiable()) {
    parameters = remappedOrigFnSubstTy->getDifferentiabilityParameterIndices();
  } else {
    parameters = IndexSubset::get(
        original->getASTContext(),
        fai.getArgumentsWithoutIndirectResults().size(), activeParamIndices);
  }
  // Compute differentiability results.
  auto *results = IndexSubset::get(original->getASTContext(),
                                   remappedOrigFnSubstTy->getNumAutoDiffSemanticResults(),
                                   activeResultIndices);
  // Create autodiff indices for the `apply` instruction.
  AutoDiffConfig applyConfig(parameters, results);

  // Check for non-differentiable original function type.
  auto checkNondifferentiableOriginalFunctionType =
    [&](CanSILFunctionType origFnTy) {
    // Check non-differentiable arguments.
    for (auto paramIndex : applyConfig.parameterIndices->getIndices()) {
      auto remappedParamType =
          origFnTy->getParameters()[paramIndex].getSILStorageInterfaceType();
      if (!remappedParamType.isDifferentiable(derivative->getModule()))
        return true;
    }
    // Check non-differentiable results.
    unsigned firstSemanticParamResultIdx = origFnTy->getNumResults();
    unsigned firstYieldResultIndex = origFnTy->getNumResults() +
      origFnTy->getNumAutoDiffSemanticResultsParameters();
    for (auto resultIndex : applyConfig.resultIndices->getIndices()) {
      SILType remappedResultType;
      if (resultIndex >= firstYieldResultIndex) {
        auto yieldResultIdx = resultIndex - firstYieldResultIndex;
        const auto& yield = origFnTy->getYields()[yieldResultIdx];
        // We do not have a good way to differentiate direct yields
        if (!yield.isAutoDiffSemanticResult())
          return true;
        remappedResultType = yield.getSILStorageInterfaceType();
      } else if (resultIndex >= firstSemanticParamResultIdx) {
        auto semanticResultArgIdx = resultIndex - firstSemanticParamResultIdx;
        auto semanticResultArg =
            *std::next(fai.getAutoDiffSemanticResultArguments().begin(),
                       semanticResultArgIdx);
        remappedResultType = semanticResultArg->getType();
      } else {
        remappedResultType =
            origFnTy->getResults()[resultIndex].getSILStorageInterfaceType();
      }
      if (!remappedResultType.isDifferentiable(derivative->getModule()))
        return true;
    }
    return false;
  };
  if (checkNondifferentiableOriginalFunctionType(remappedOrigFnSubstTy))
    return nullptr;

  AutoDiffDerivativeFunctionKind derivativeFnKind(kind);
  auto derivativeFnType =
      remappedOrigFnSubstTy
          ->getAutoDiffDerivativeFunctionType(
              parameters, results, derivativeFnKind, context.getTypeConverter(),
              LookUpConformanceInModule())
          ->getUnsubstitutedType(original->getModule());

  auto linearMapSILType = derivativeFnType->getAllResultsInterfaceType();
  if (auto tupleType = linearMapSILType.getAs<TupleType>()) {
    linearMapSILType = SILType::getPrimitiveObjectType(
        tupleType.getElementType(tupleType->getElements().size() - 1));
  }
  if (auto fnTy = linearMapSILType.getAs<SILFunctionType>()) {
    linearMapSILType = SILType::getPrimitiveObjectType(
        fnTy->getUnsubstitutedType(original->getModule()));
  }

  // IRGen requires decls to have AST types (not `SILFunctionType`), so we
  // convert the `SILFunctionType` of the linear map to a `FunctionType` with
  // the same parameters and results.
  auto silFnTy = linearMapSILType.castTo<SILFunctionType>();
  SmallVector<AnyFunctionType::Param, 8> params;
  for (auto &param : silFnTy->getParameters()) {
    ParameterTypeFlags flags;
    if (param.isAutoDiffSemanticResult())
      flags = flags.withInOut(true);

    params.push_back(
        AnyFunctionType::Param(param.getInterfaceType(), Identifier(), flags));
  }

  AnyFunctionType *astFnTy;
  if (auto genSig = silFnTy->getSubstGenericSignature()) {
    // FIXME: Verify ExtInfo state is correct, not working by accident.
    GenericFunctionType::ExtInfo info;
    astFnTy = GenericFunctionType::get(
        genSig, params, silFnTy->getAllResultsInterfaceType().getASTType(),
        info);
  } else {
    FunctionType::ExtInfo info;
    astFnTy = FunctionType::get(
        params, silFnTy->getAllResultsInterfaceType().getASTType(), info);
  }

  if (astFnTy->hasArchetype())
    return astFnTy->mapTypeOutOfContext();

  return astFnTy;
}

void LinearMapInfo::generateDifferentiationDataStructures(
    ADContext &context, SILFunction *derivativeFn) {
  auto &astCtx = original->getASTContext();
  // Get the derivative function generic signature.
  CanGenericSignature derivativeFnGenSig = nullptr;
  if (auto *derivativeFnGenEnv = derivativeFn->getGenericEnvironment())
    derivativeFnGenSig =
        derivativeFnGenEnv->getGenericSignature().getCanonicalSignature();

  // Create branching trace enum for each original block and add it as a field
  // in the corresponding struct.
  StringRef traceEnumFieldName;
  switch (kind) {
  case AutoDiffLinearMapKind::Differential:
    traceEnumFieldName = "successor";
    break;
  case AutoDiffLinearMapKind::Pullback:
    traceEnumFieldName = "predecessor";
    break;
  }

  for (auto &origBB : *original) {
    auto *traceEnum =
        createBranchingTraceDecl(&origBB, derivativeFnGenSig);
    branchingTraceDecls.insert({&origBB, traceEnum});
  }

  // Add linear map fields to the linear map tuples.
  //
  // Now we need to be very careful as we're having a very subtle
  // chicken-and-egg problem. We need lowered branch trace enum type for the
  // linear map typle type. However branch trace enum type lowering depends on
  // the lowering of its elements (at very least, the type classification of
  // being trivial / non-trivial). As the lowering is cached we need to ensure
  // we compute lowered type for the branch trace enum when the corresponding
  // EnumDecl is fully complete: we cannot add more entries without causing some
  // very subtle issues later on. However, the elements of the enum are linear
  // map tuples of predecessors, that correspondingly may contain branch trace
  // enums of corresponding predecessor BBs.
  //
  // Traverse all BBs in reverse post-order traversal order to ensure we process
  // each BB before its predecessors.
  llvm::ReversePostOrderTraversal<SILFunction *> RPOT(original);
  for (auto Iter = RPOT.begin(), E = RPOT.end(); Iter != E; ++Iter) {
    auto *origBB = *Iter;
    SmallVector<TupleTypeElt, 4> linearTupleTypes;
    if (!origBB->isEntry()) {
      populateBranchingTraceDecl(origBB, loopInfo);

      CanType traceEnumType = getBranchingTraceEnumLoweredType(origBB).getASTType();
      linearTupleTypes.emplace_back(traceEnumType,
                                    astCtx.getIdentifier(traceEnumFieldName));
    }

    if (isSemanticMemberAccessor(original)) {
      // Do not add linear map fields for semantic member accessors, which have
      // special-case pullback generation. Linear map tuples should be empty.
    } else {
      for (auto &inst : *origBB) {
        if (auto *ai = dyn_cast<ApplyInst>(&inst))
          // Skip array literal intrinsic applications since array literal
          // initialization is linear and handled separately.
          if (ArraySemanticsCall(ai, semantics::ARRAY_UNINITIALIZED_INTRINSIC) ||
              ArraySemanticsCall(ai, semantics::ARRAY_FINALIZE_INTRINSIC))
            continue;

        if (!isa<FullApplySite>(&inst))
          continue;

        FullApplySite fai(&inst);
        // Add linear map field to struct for active apply sites instructions.
        if (!shouldDifferentiateApplySite(fai))
          continue;

        LLVM_DEBUG(getADDebugStream()
                   << "Adding linear map tuple field for " << inst);
        if (Type linearMapType = getLinearMapType(context, fai)) {
          LLVM_DEBUG(getADDebugStream() << "Computed type: " << linearMapType << '\n');
          linearMapIndexMap.insert({fai, linearTupleTypes.size()});
          linearTupleTypes.emplace_back(linearMapType);
        }
      }
    }

    linearMapTuples.insert({origBB, TupleType::get(linearTupleTypes, astCtx)});
  }

  // Print generated linear map structs and branching trace enums.
  // These declarations do not show up with `-emit-sil` because they are
  // implicit. Instead, use `-Xllvm -debug-only=differentiation` to test
  // declarations with FileCheck.
  LLVM_DEBUG({
    auto &s = getADDebugStream();
    PrintOptions printOptions;
    printOptions.TypeDefinitions = true;
    printOptions.ExplodePatternBindingDecls = true;
    printOptions.SkipImplicit = false;
    s << "Generated linear map tuples and branching trace enums for @"
      << original->getName() << ":\n";
    for (auto &origBB : *original) {
      auto *linearMapTuple = getLinearMapTupleType(&origBB);
      linearMapTuple->print(s, printOptions);
      s << '\n';
    }

    for (auto &origBB : *original) {
      auto *traceEnum = getBranchingTraceDecl(&origBB);
      traceEnum->print(s, printOptions);
      s << '\n';
    }
  });
}

/// Returns a flag that indicates whether the `apply` instruction should be
/// differentiated, given the differentiation indices of the instruction's
/// parent function. Whether the `apply` should be differentiated is determined
/// sequentially from the following conditions:
/// 1. The instruction has an active `inout` argument.
/// 2. The instruction is a call to the array literal initialization intrinsic
///    ("array.uninitialized_intrinsic"), where the result is active and where
///    there is a `store` of an active value into the array's buffer.
/// 3. The instruction has both an active result (direct or indirect) and an
///    active argument.
bool LinearMapInfo::shouldDifferentiateApplySite(FullApplySite applySite) {
  // Function applications with an active inout argument should be
  // differentiated.
  for (auto inoutArg : applySite.getInoutArguments())
    if (activityInfo.isActive(inoutArg, config))
      return true;

  bool hasActiveDirectResults = false;
  forEachApplyDirectResult(applySite, [&](SILValue directResult) {
    hasActiveDirectResults |= activityInfo.isActive(directResult, config);
  });
  bool hasActiveIndirectResults =
      llvm::any_of(applySite.getIndirectSILResults(), [&](SILValue result) {
        return activityInfo.isActive(result, config);
      });
  bool hasActiveResults = hasActiveDirectResults || hasActiveIndirectResults;

  // TODO: Pattern match to make sure there is at least one `store` to the
  // array's active buffer.
  if (ArraySemanticsCall(applySite.getInstruction(),
                         semantics::ARRAY_UNINITIALIZED_INTRINSIC) &&
      hasActiveResults)
    return true;

  auto arguments = applySite.getArgumentsWithoutIndirectResults();
  bool hasActiveArguments = llvm::any_of(arguments, [&](SILValue arg) {
    return activityInfo.isActive(arg, config);
  });
  return hasActiveResults && hasActiveArguments;
}

static bool shouldDifferentiateInjectEnumAddr(
    const InjectEnumAddrInst &inject,
    const DifferentiableActivityInfo &activityInfo,
    const AutoDiffConfig &config) {
  SILValue en = inject.getOperand();
  for (auto use : en->getUses()) {
    auto *init = dyn_cast<InitEnumDataAddrInst>(use->getUser());
    if (init && activityInfo.isActive(init, config))
      return true;
  }
  return false;
}

/// Returns a flag indicating whether the instruction should be differentiated,
/// given the differentiation indices of the instruction's parent function.
/// Whether the instruction should be differentiated is determined sequentially
/// from any of the following conditions:
/// 1. The instruction is a full apply site and `shouldDifferentiateApplyInst`
///    returns true.
/// 2. The instruction has a source operand and a destination operand, both
///    being active.
/// 3. The instruction is an allocation instruction and has an active result.
/// 4. The instruction performs reference counting, lifetime ending, access
///    ending, or destroying on an active operand.
/// 5. The instruction creates an SSA copy of an active operand.
bool LinearMapInfo::shouldDifferentiateInstruction(SILInstruction *inst) {
  // A full apply site with an active argument and an active result (direct or
  // indirect) should be differentiated.
  if (FullApplySite::isa(inst))
    return shouldDifferentiateApplySite(FullApplySite(inst));
  // Anything with an active result and an active operand should be
  // differentiated.
  auto hasActiveOperands =
      llvm::any_of(inst->getAllOperands(), [&](Operand &op) {
        return activityInfo.isActive(op.get(), config);
      });
  auto hasActiveResults = llvm::any_of(inst->getResults(), [&](SILValue val) {
    return activityInfo.isActive(val, config);
  });
  if (hasActiveOperands && hasActiveResults)
    return true;
    // `store`-like instructions do not have an SSA result, but have two
    // operands that represent the source and the destination. We treat them as
    // the input and the output, respectively.
    // For `store`-like instructions whose destination is an element address
    // from an `array.uninitialized_intrinsic` application, return true if the
    // intrinsic application (representing the semantic destination) is active.
#define CHECK_INST_TYPE_ACTIVE_DEST(INST)                                      \
  if (auto *castInst = dyn_cast<INST##Inst>(inst))                             \
    return activityInfo.isActive(castInst->getDest(), config);
  CHECK_INST_TYPE_ACTIVE_DEST(Store)
  CHECK_INST_TYPE_ACTIVE_DEST(StoreBorrow)
  CHECK_INST_TYPE_ACTIVE_DEST(CopyAddr)
  CHECK_INST_TYPE_ACTIVE_DEST(UnconditionalCheckedCastAddr)
#undef CHECK_INST_TYPE_ACTIVE_DEST
  // Should differentiate any allocation instruction that has an active result.
  if ((isa<AllocationInst>(inst) && hasActiveResults))
    return true;
  // Should differentiate end_apply if the corresponding begin_apply is
  // differentiable
  if (auto *eai = dyn_cast<EndApplyInst>(inst))
    return shouldDifferentiateApplySite(eai->getBeginApply());
  if (hasActiveOperands) {
    // Should differentiate any instruction that performs reference counting,
    // lifetime ending, access ending, or destroying on an active operand.
    if (isa<RefCountingInst>(inst) || isa<EndAccessInst>(inst) ||
        isa<EndBorrowInst>(inst) || isa<DeallocationInst>(inst) ||
        isa<DestroyValueInst>(inst) || isa<DestroyAddrInst>(inst) ||
        isa<AbortApplyInst>(inst) ||
        isa<YieldInst>(inst))
      return true;
  }

  // Should differentiate `inject_enum_addr` if the corresponding
  // `init_enum_addr` has an active operand.
  if (auto inject = dyn_cast<InjectEnumAddrInst>(inst))
    if (shouldDifferentiateInjectEnumAddr(*inject, activityInfo, config))
      return true;

  return false;
}

} // end namespace autodiff
} // end namespace swift