rvsdg2rhls.cpp

Go to the documentation of this file.

/*
 * Copyright 2021 David Metz <david.c.metz@ntnu.no>
 * See COPYING for terms of redistribution.
 */
 
#include <jlm/hls/backend/rvsdg2rhls/add-buffers.hpp>
#include <jlm/hls/backend/rvsdg2rhls/add-forks.hpp>
#include <jlm/hls/backend/rvsdg2rhls/add-prints.hpp>
#include <jlm/hls/backend/rvsdg2rhls/add-sinks.hpp>
#include <jlm/hls/backend/rvsdg2rhls/alloca-conv.hpp>
#include <jlm/hls/backend/rvsdg2rhls/check-rhls.hpp>
#include <jlm/hls/backend/rvsdg2rhls/decouple-mem-state.hpp>
#include <jlm/hls/backend/rvsdg2rhls/distribute-constants.hpp>
#include <jlm/hls/backend/rvsdg2rhls/GammaConversion.hpp>
#include <jlm/hls/backend/rvsdg2rhls/instrument-ref.hpp>
#include <jlm/hls/backend/rvsdg2rhls/mem-conv.hpp>
#include <jlm/hls/backend/rvsdg2rhls/mem-queue.hpp>
#include <jlm/hls/backend/rvsdg2rhls/mem-sep.hpp>
#include <jlm/hls/backend/rvsdg2rhls/memstate-conv.hpp>
#include <jlm/hls/backend/rvsdg2rhls/merge-gamma.hpp>
#include <jlm/hls/backend/rvsdg2rhls/remove-redundant-buf.hpp>
#include <jlm/hls/backend/rvsdg2rhls/rhls-dne.hpp>
#include <jlm/hls/backend/rvsdg2rhls/rvsdg2rhls.hpp>
#include <jlm/hls/backend/rvsdg2rhls/stream-conv.hpp>
#include <jlm/hls/backend/rvsdg2rhls/ThetaConversion.hpp>
#include <jlm/hls/backend/rvsdg2rhls/UnusedStateRemoval.hpp>
#include <jlm/hls/opt/cne.hpp>
#include <jlm/hls/opt/IOBarrierRemoval.hpp>
#include <jlm/hls/opt/IOStateElimination.hpp>
#include <jlm/hls/util/view.hpp>
#include <jlm/llvm/backend/IpGraphToLlvmConverter.hpp>
#include <jlm/llvm/backend/RvsdgToIpGraphConverter.hpp>
#include <jlm/llvm/DotWriter.hpp>
#include <jlm/llvm/ir/CallingConvention.hpp>
#include <jlm/llvm/ir/CallSummary.hpp>
#include <jlm/llvm/ir/operators/alloca.hpp>
#include <jlm/llvm/ir/operators/call.hpp>
#include <jlm/llvm/ir/operators/delta.hpp>
#include <jlm/llvm/ir/operators/IntegerOperations.hpp>
#include <jlm/llvm/ir/RvsdgModule.hpp>
#include <jlm/llvm/opt/DeadNodeElimination.hpp>
#include <jlm/llvm/opt/inlining.hpp>
#include <jlm/llvm/opt/InvariantValueRedirection.hpp>
#include <jlm/llvm/opt/LoopUnswitching.hpp>
#include <jlm/llvm/opt/NodeReduction.hpp>
#include <jlm/llvm/opt/PredicateCorrelation.hpp>
#include <jlm/rvsdg/Transformation.hpp>
#include <jlm/rvsdg/traverser.hpp>
#include <jlm/rvsdg/view.hpp>
#include <jlm/util/Statistics.hpp>
#include <llvm/IR/LLVMContext.h>
#include <llvm/IR/Module.h>
#include <llvm/Support/raw_ostream.h>
#include <llvm/Support/SourceMgr.h>
 
#include <regex>
 
namespace jlm::hls
{
 
void
split_opt(llvm::LlvmRvsdgModule & rm)
{
  // TODO: figure out which optimizations to use here
  jlm::llvm::DeadNodeElimination dne;
  CommonNodeElimination cne;
  constexpr llvm::InvariantValueRedirection::Configuration ivrConfiguration;
  llvm::InvariantValueRedirection invariantValueRedirection(std::move(ivrConfiguration));
  llvm::LoopUnswitching loopUnswitching(llvm::LoopUnswitchingDefaultHeuristic::create());
  jlm::llvm::NodeReduction red;
  jlm::util::StatisticsCollector statisticsCollector;
  loopUnswitching.Run(rm, statisticsCollector);
  dne.Run(rm, statisticsCollector);
  cne.Run(rm, statisticsCollector);
  invariantValueRedirection.Run(rm, statisticsCollector);
  red.Run(rm, statisticsCollector);
  dne.Run(rm, statisticsCollector);
}
 
void
pre_opt(jlm::llvm::LlvmRvsdgModule & rm)
{
  // TODO: figure out which optimizations to use here
  jlm::llvm::DeadNodeElimination dne;
  CommonNodeElimination cne;
  constexpr llvm::InvariantValueRedirection::Configuration ivrConfiguration;
  llvm::InvariantValueRedirection invariantValueRedirection(std::move(ivrConfiguration));
  llvm::LoopUnswitching loopUnswitching(llvm::LoopUnswitchingDefaultHeuristic::create());
  jlm::util::StatisticsCollector statisticsCollector;
  loopUnswitching.Run(rm, statisticsCollector);
  dne.Run(rm, statisticsCollector);
  cne.Run(rm, statisticsCollector);
  invariantValueRedirection.Run(rm, statisticsCollector);
  dne.Run(rm, statisticsCollector);
  cne.Run(rm, statisticsCollector);
  dne.Run(rm, statisticsCollector);
}
 
bool
function_match(rvsdg::LambdaNode * ln, const std::string & function_name)
{
  const std::regex fn_regex(function_name);
  if (std::regex_match(
          dynamic_cast<llvm::LlvmLambdaOperation &>(ln->GetOperation()).name(),
          fn_regex))
  { // TODO: handle C++ name mangling
    return true;
  }
  return false;
}
 
const jlm::rvsdg::Output *
trace_call(jlm::rvsdg::Input * input)
{
  auto graph = input->region()->graph();
 
  auto argument = dynamic_cast<const rvsdg::RegionArgument *>(input->origin());
  const rvsdg::Output * result = nullptr;
  if (auto theta = rvsdg::TryGetOwnerNode<rvsdg::ThetaNode>(*input->origin()))
  {
    result = trace_call(theta->MapOutputLoopVar(*input->origin()).input);
  }
  else if (argument == nullptr)
  {
    result = input->origin();
  }
  else if (argument->region() == &graph->GetRootRegion())
  {
    result = argument;
  }
  else
  {
    JLM_ASSERT(argument->input() != nullptr);
    result = trace_call(argument->input());
  }
  return result;
}
 
void
inline_calls(rvsdg::Region * region)
{
  for (auto & node : rvsdg::TopDownTraverser(region))
  {
    if (auto structnode = dynamic_cast<rvsdg::StructuralNode *>(node))
    {
      for (size_t n = 0; n < structnode->nsubregions(); n++)
      {
        inline_calls(structnode->subregion(n));
      }
    }
    else if (dynamic_cast<const llvm::CallOperation *>(&(node->GetOperation())))
    {
      auto traced = jlm::hls::trace_call(node->input(0));
      auto so = dynamic_cast<const rvsdg::StructuralOutput *>(traced);
      if (!so)
      {
        if (auto graphImport = dynamic_cast<const llvm::LlvmGraphImport *>(traced))
        {
          if (graphImport->Name().rfind("decouple_", 0) == 0)
          {
            // can't inline pseudo functions used for decoupling
            continue;
          }
          if (graphImport->Name().rfind("hls_", 0) == 0)
          {
            // can't inline pseudo functions used for streaming
            continue;
          }
          throw util::Error("can not inline external function " + graphImport->Name());
        }
      }
      JLM_ASSERT(rvsdg::is<rvsdg::LambdaOperation>(so->node()));
      auto ln = dynamic_cast<const rvsdg::StructuralOutput *>(traced)->node();
      llvm::FunctionInlining::inlineCall(
          *dynamic_cast<rvsdg::SimpleNode *>(node),
          *dynamic_cast<const rvsdg::LambdaNode *>(ln));
      // restart for this region
      inline_calls(region);
      return;
    }
  }
}
 
rvsdg::DeltaNode *
rename_delta(rvsdg::DeltaNode * odn)
{
  auto op = util::assertedCast<const llvm::LlvmDeltaOperation>(&odn->GetOperation());
  auto name = op->name();
  std::replace_if(
      name.begin(),
      name.end(),
      [](char c)
      {
        return c == '.';
      },
      '_');
  std::cout << "renaming delta node " << op->name() << " to " << name << "\n";
  auto db = rvsdg::DeltaNode::Create(
      odn->region(),
      llvm::LlvmDeltaOperation::Create(
          odn->Type(),
          name,
          llvm::Linkage::externalLinkage,
          "",
          op->constant(),
          op->getAlignment()));
  /* add dependencies */
  rvsdg::SubstitutionMap rmap;
  for (auto ctxVar : odn->GetContextVars())
  {
    auto input = ctxVar.input;
    auto nd = db->AddContextVar(*input->origin()).inner;
    rmap.insert(ctxVar.inner, nd);
  }
 
  /* copy subregion */
  odn->subregion()->copy(db->subregion(), rmap);
 
  auto result = &rmap.lookup(*odn->subregion()->result(0)->origin());
  auto data = &db->finalize(result);
 
  odn->output().divert_users(data);
  jlm::rvsdg::remove(odn);
  return rvsdg::TryGetOwnerNode<rvsdg::DeltaNode>(*data);
}
 
rvsdg::LambdaNode *
change_linkage(rvsdg::LambdaNode * ln, llvm::Linkage link)
{
  const auto & op = dynamic_cast<llvm::LlvmLambdaOperation &>(ln->GetOperation());
  auto lambda = rvsdg::LambdaNode::Create(
      *ln->region(),
      llvm::LlvmLambdaOperation::Create(
          op.Type(),
          op.name(),
          link,
          op.callingConvention(),
          op.attributes()));
 
  /* add context variables */
  rvsdg::SubstitutionMap subregionmap;
  for (const auto & cv : ln->GetContextVars())
  {
    auto origin = cv.input->origin();
    auto newcv = lambda->AddContextVar(*origin);
    subregionmap.insert(cv.inner, newcv.inner);
  }
  /* collect function arguments */
  auto args = ln->GetFunctionArguments();
  auto newArgs = lambda->GetFunctionArguments();
  JLM_ASSERT(args.size() == newArgs.size());
  for (std::size_t n = 0; n < args.size(); ++n)
  {
    subregionmap.insert(args[n], newArgs[n]);
  }
 
  /* copy subregion */
  ln->subregion()->copy(lambda->subregion(), subregionmap);
 
  /* collect function results */
  std::vector<jlm::rvsdg::Output *> results;
  for (auto result : ln->GetFunctionResults())
    results.push_back(&subregionmap.lookup(*result->origin()));
 
  /* finalize lambda */
  lambda->finalize(results);
 
  divert_users(ln, outputs(lambda));
  jlm::rvsdg::remove(ln);
 
  return lambda;
}
 
std::unique_ptr<jlm::llvm::LlvmRvsdgModule>
split_hls_function(llvm::LlvmRvsdgModule & rm, const std::string & function_name)
{
  // TODO: use a different datastructure for rhls?
  // create a copy of rm
  auto rhls =
      llvm::LlvmRvsdgModule::Create(rm.SourceFileName(), rm.TargetTriple(), rm.DataLayout());
  std::cout << "processing " << rm.SourceFileName().name() << "\n";
  auto root = &rm.Rvsdg().GetRootRegion();
  for (auto node : rvsdg::TopDownTraverser(root))
  {
    if (auto ln = dynamic_cast<rvsdg::LambdaNode *>(node))
    {
      if (!function_match(ln, function_name))
      {
        continue;
      }
      inline_calls(ln->subregion());
      split_opt(rm);
 
      rvsdg::SubstitutionMap smap;
      for (size_t i = 0; i < ln->ninputs(); ++i)
      {
        auto orig_node_output = dynamic_cast<rvsdg::NodeOutput *>(ln->input(i)->origin());
        if (!orig_node_output)
        {
          // handle decouple stuff
          auto oldGraphImport = dynamic_cast<llvm::LlvmGraphImport *>(ln->input(i)->origin());
          auto & newGraphImport = oldGraphImport->Copy(rhls->Rvsdg().GetRootRegion(), nullptr);
          smap.insert(ln->input(i)->origin(), &newGraphImport);
          continue;
        }
        auto orig_node = orig_node_output->node();
        if (auto oln = dynamic_cast<rvsdg::LambdaNode *>(orig_node))
        {
          throw util::Error(
              "Inlining of function "
              + dynamic_cast<llvm::LlvmLambdaOperation &>(oln->GetOperation()).name()
              + " not supported");
        }
        else if (auto odn = dynamic_cast<rvsdg::DeltaNode *>(orig_node))
        {
          auto op = util::assertedCast<const llvm::LlvmDeltaOperation>(&odn->GetOperation());
          // modify name to not contain .
          if (op->name().find('.') != std::string::npos)
          {
            odn = rename_delta(odn);
            op = util::assertedCast<const llvm::LlvmDeltaOperation>(&odn->GetOperation());
          }
          std::cout << "delta node " << op->name() << ": " << op->Type()->debug_string() << "\n";
          // add import for delta to rhls
          auto & graphImport = llvm::LlvmGraphImport::createGlobalImport(
              rhls->Rvsdg(),
              op->Type(),
              llvm::PointerType::Create(),
              op->name(),
              llvm::Linkage::externalLinkage,
              op->constant(),
              op->getAlignment());
          smap.insert(ln->input(i)->origin(), &graphImport);
          // add export for delta to rm
          // TODO: check if not already exported and maybe adjust linkage?
          rvsdg::GraphExport::Create(odn->output(), op->name());
        }
        else
        {
          throw util::Error("Unsupported node type: " + orig_node->DebugString());
        }
      }
      // copy function into rhls
      auto new_ln = ln->copy(&rhls->Rvsdg().GetRootRegion(), smap);
      new_ln = change_linkage(new_ln, llvm::Linkage::externalLinkage);
      auto oldExport = jlm::llvm::ComputeCallSummary(*ln).GetRvsdgExport();
      rvsdg::GraphExport::Create(*new_ln->output(), oldExport ? oldExport->Name() : "");
      // add function as input to rm and remove it
      const auto & op = dynamic_cast<llvm::LlvmLambdaOperation &>(ln->GetOperation());
      auto & graphImport = llvm::LlvmGraphImport::createFunctionImport(
          rm.Rvsdg(),
          op.Type(),
          op.name(),
          llvm::Linkage::externalLinkage, // TODO: change linkage?
          llvm::CallingConvention::Default);
      ln->output()->divert_users(&graphImport);
      remove(ln);
      std::cout << "function "
                << dynamic_cast<llvm::LlvmLambdaOperation &>(new_ln->GetOperation()).name()
                << " extracted for HLS\n";
      return rhls;
    }
  }
  throw util::Error("HLS function " + function_name + " not found");
}
 
void
rvsdg2ref(llvm::LlvmRvsdgModule & rhls, const util::FilePath & path)
{
  dump_ref(rhls, path);
}
 
std::unique_ptr<rvsdg::TransformationSequence>
createTransformationSequence(rvsdg::DotWriter & dotWriter, const bool dumpRvsdgGraphs)
{
  auto predicateCorrelation = std::make_shared<llvm::PredicateCorrelation>();
  auto deadNodeElimination = std::make_shared<llvm::DeadNodeElimination>();
  auto commonNodeElimination = std::make_shared<CommonNodeElimination>();
  auto invariantValueRedirection = std::make_shared<llvm::InvariantValueRedirection>(
      llvm::InvariantValueRedirection::Configuration());
  auto loopUnswitching =
      std::make_shared<llvm::LoopUnswitching>(llvm::LoopUnswitchingDefaultHeuristic::create());
  auto ioBarrierRemoval = std::make_shared<IOBarrierRemoval>();
  auto ioStateElimination = std::make_shared<IOStateElimination>();
  auto memoryStateSeparation = std::make_shared<MemoryStateSeparation>();
  auto gammaMerge = std::make_shared<GammaMerge>();
  auto unusedStateRemoval = std::make_shared<UnusedStateRemoval>();
  auto constantDistribution = std::make_shared<ConstantDistribution>();
  auto gammaNodeConversion = std::make_shared<GammaNodeConversion>();
  auto thetaNodeConversion = std::make_shared<ThetaNodeConversion>();
  auto rhlsDeadNodeElimination = std::make_shared<RhlsDeadNodeElimination>();
  auto allocaNodeConversion = std::make_shared<AllocaNodeConversion>();
  auto streamConversion = std::make_shared<StreamConversion>();
  auto addressQueueInsertion = std::make_shared<AddressQueueInsertion>();
  auto memoryStateDecoupling = std::make_shared<MemoryStateDecoupling>();
  auto memoryConverter = std::make_shared<MemoryConverter>();
  auto nodeReduction = std::make_shared<llvm::NodeReduction>();
  auto memoryStateSplitConversion = std::make_shared<MemoryStateSplitConversion>();
  auto redundantBufferElimination = std::make_shared<RedundantBufferElimination>();
  auto sinkInsertion = std::make_shared<SinkInsertion>();
  auto forkInsertion = std::make_shared<ForkInsertion>();
  auto bufferInsertion = std::make_shared<BufferInsertion>();
  auto rhlsVerification = std::make_shared<RhlsVerification>();
 
  // Use this transformation to dump HLS dot graphs at specific points in the sequence
  [[maybe_unused]] auto dumpDot = std::make_shared<DumpDotTransformation>();
 
  std::vector<std::shared_ptr<rvsdg::Transformation>> sequence({
      loopUnswitching,
      deadNodeElimination,
      commonNodeElimination,
      invariantValueRedirection,
      predicateCorrelation,
      deadNodeElimination,
      commonNodeElimination,
      deadNodeElimination,
      ioBarrierRemoval,
      ioStateElimination,
      memoryStateSeparation,
      gammaMerge,
      unusedStateRemoval,
      deadNodeElimination,
      loopUnswitching,
      commonNodeElimination,
      deadNodeElimination,
      gammaMerge,
      deadNodeElimination,
      unusedStateRemoval,
      constantDistribution,
      deadNodeElimination,
      gammaNodeConversion,
      thetaNodeConversion,
      commonNodeElimination,
      rhlsDeadNodeElimination,
      allocaNodeConversion,
      streamConversion,
      addressQueueInsertion,
      memoryStateDecoupling,
      unusedStateRemoval,
      memoryConverter,
      nodeReduction,
      memoryStateSplitConversion,
      redundantBufferElimination,
      sinkInsertion,
      forkInsertion,
      bufferInsertion,
      rhlsVerification,
  });
 
  return std::make_unique<rvsdg::TransformationSequence>(
      std::move(sequence),
      dotWriter,
      dumpRvsdgGraphs);
}
 
void
dump_ref(llvm::LlvmRvsdgModule & rhls, const util::FilePath & path)
{
  const std::unique_ptr<llvm::LlvmRvsdgModule> reference(
      static_cast<llvm::LlvmRvsdgModule *>(rhls.copy().release()));
  pre_opt(*reference);
  instrument_ref(*reference);
  for (size_t i = 0; i < reference->Rvsdg().GetRootRegion().narguments(); ++i)
  {
    auto graphImport = util::assertedCast<const llvm::LlvmGraphImport>(
        reference->Rvsdg().GetRootRegion().argument(i));
    std::cout << "impport " << graphImport->Name() << ": " << graphImport->Type()->debug_string()
              << "\n";
  }
  ::llvm::LLVMContext ctx;
  jlm::util::StatisticsCollector statisticsCollector;
  auto jm2 = llvm::RvsdgToIpGraphConverter::CreateAndConvertModule(*reference, statisticsCollector);
  auto lm2 = llvm::IpGraphToLlvmConverter::CreateAndConvertModule(*jm2, ctx);
  std::error_code EC;
  ::llvm::raw_fd_ostream os(path.to_str(), EC);
  lm2->print(os, nullptr);
}
 
}