src/passes/Inlining.cpp

/*
 * Copyright 2016 WebAssembly Community Group participants
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

//
// Inlining.
//
// This uses some simple heuristics to decide when to inline.
//
// Two versions are provided: inlining and inlining-optimizing. You
// probably want the optimizing version, which will optimize locations
// we inlined into, as inlining by itself creates a block to house the
// inlined code, some temp locals, etc., which can usually be removed
// by optimizations. Note that the two versions use the same heuristics,
// so we don't take into account the overhead if you don't optimize
// afterwards. The non-optimizing version is mainly useful for debugging,
// or if you intend to run a full set of optimizations anyhow on
// everything later.
//

#include <atomic>

#include "ir/debug.h"
#include "ir/literal-utils.h"
#include "ir/module-utils.h"
#include "ir/utils.h"
#include "parsing.h"
#include "pass.h"
#include "passes/opt-utils.h"
#include "wasm-builder.h"
#include "wasm.h"

namespace wasm {

// Useful into on a function, helping us decide if we can inline it
struct FunctionInfo {
  std::atomic<Index> refs;
  Index size;
  std::atomic<bool> lightweight;
  bool usedGlobally; // in a table or export

  FunctionInfo() {
    refs = 0;
    size = 0;
    lightweight = true;
    usedGlobally = false;
  }

  // See pass.h for how defaults for these options were chosen.
  bool worthInlining(PassOptions& options) {
    // if it's big, it's just not worth doing (TODO: investigate more)
    if (size > options.inlining.flexibleInlineMaxSize) {
      return false;
    }
    // if it's so small we have a guarantee that after we optimize the
    // size will not increase, inline it
    if (size <= options.inlining.alwaysInlineMaxSize) {
      return true;
    }
    // if it has one use, then inlining it would likely reduce code size
    // since we are just moving code around, + optimizing, so worth it
    // if small enough that we are pretty sure its ok
    // FIXME: move this check to be first in this function, since we should
    // return true if oneCallerInlineMaxSize is bigger than
    // flexibleInlineMaxSize (which it typically should be).
    if (refs == 1 && !usedGlobally &&
        size <= options.inlining.oneCallerInlineMaxSize) {
      return true;
    }
    // more than one use, so we can't eliminate it after inlining,
    // so only worth it if we really care about speed and don't care
    // about size, and if it's lightweight so a good candidate for
    // speeding us up.
    return options.optimizeLevel >= 3 && options.shrinkLevel == 0 &&
           lightweight;
  }
};

typedef std::unordered_map<Name, FunctionInfo> NameInfoMap;

struct FunctionInfoScanner
  : public WalkerPass<PostWalker<FunctionInfoScanner>> {
  bool isFunctionParallel() override { return true; }

  FunctionInfoScanner(NameInfoMap* infos) : infos(infos) {}

  FunctionInfoScanner* create() override {
    return new FunctionInfoScanner(infos);
  }

  void visitLoop(Loop* curr) {
    // having a loop is not lightweight
    (*infos)[getFunction()->name].lightweight = false;
  }

  void visitCall(Call* curr) {
    // can't add a new element in parallel
    assert(infos->count(curr->target) > 0);
    (*infos)[curr->target].refs++;
    // having a call is not lightweight
    (*infos)[getFunction()->name].lightweight = false;
  }

  void visitRefFunc(RefFunc* curr) {
    assert(infos->count(curr->func) > 0);
    (*infos)[curr->func].refs++;
  }

  void visitFunction(Function* curr) {
    (*infos)[curr->name].size = Measurer::measure(curr->body);
  }

private:
  NameInfoMap* infos;
};

struct InliningAction {
  Expression** callSite;
  Function* contents;

  InliningAction(Expression** callSite, Function* contents)
    : callSite(callSite), contents(contents) {}
};

struct InliningState {
  std::unordered_set<Name> worthInlining;
  // function name => actions that can be performed in it
  std::unordered_map<Name, std::vector<InliningAction>> actionsForFunction;
};

struct Planner : public WalkerPass<PostWalker<Planner>> {
  bool isFunctionParallel() override { return true; }

  Planner(InliningState* state) : state(state) {}

  Planner* create() override { return new Planner(state); }

  void visitCall(Call* curr) {
    // plan to inline if we know this is valid to inline, and if the call is
    // actually performed - if it is dead code, it's pointless to inline.
    // we also cannot inline ourselves.
    bool isUnreachable;
    if (curr->isReturn) {
      // Tail calls are only actually unreachable if an argument is
      isUnreachable = std::any_of(
        curr->operands.begin(), curr->operands.end(), [](Expression* op) {
          return op->type == Type::unreachable;
        });
    } else {
      isUnreachable = curr->type == Type::unreachable;
    }
    if (state->worthInlining.count(curr->target) && !isUnreachable &&
        curr->target != getFunction()->name) {
      // nest the call in a block. that way the location of the pointer to the
      // call will not change even if we inline multiple times into the same
      // function, otherwise call1(call2()) might be a problem
      auto* block = Builder(*getModule()).makeBlock(curr);
      replaceCurrent(block);
      // can't add a new element in parallel
      assert(state->actionsForFunction.count(getFunction()->name) > 0);
      state->actionsForFunction[getFunction()->name].emplace_back(
        &block->list[0], getModule()->getFunction(curr->target));
    }
  }

private:
  InliningState* state;
};

struct Updater : public PostWalker<Updater> {
  Module* module;
  std::map<Index, Index> localMapping;
  Name returnName;
  Builder* builder;
  void visitReturn(Return* curr) {
    replaceCurrent(builder->makeBreak(returnName, curr->value));
  }
  // Return calls in inlined functions should only break out of the scope of
  // the inlined code, not the entire function they are being inlined into. To
  // achieve this, make the call a non-return call and add a break. This does
  // not cause unbounded stack growth because inlining and return calling both
  // avoid creating a new stack frame.
  template<typename T> void handleReturnCall(T* curr, Type targetType) {
    curr->isReturn = false;
    curr->type = targetType;
    if (targetType.isConcrete()) {
      replaceCurrent(builder->makeBreak(returnName, curr));
    } else {
      replaceCurrent(builder->blockify(curr, builder->makeBreak(returnName)));
    }
  }
  void visitCall(Call* curr) {
    if (curr->isReturn) {
      handleReturnCall(curr, module->getFunction(curr->target)->sig.results);
    }
  }
  void visitCallIndirect(CallIndirect* curr) {
    if (curr->isReturn) {
      handleReturnCall(curr, curr->sig.results);
    }
  }
  void visitLocalGet(LocalGet* curr) {
    curr->index = localMapping[curr->index];
  }
  void visitLocalSet(LocalSet* curr) {
    curr->index = localMapping[curr->index];
  }
};

// Core inlining logic. Modifies the outside function (adding locals as
// needed), and returns the inlined code.
static Expression*
doInlining(Module* module, Function* into, const InliningAction& action) {
  Function* from = action.contents;
  auto* call = (*action.callSite)->cast<Call>();
  // Works for return_call, too
  Type retType = module->getFunction(call->target)->sig.results;
  Builder builder(*module);
  auto* block = builder.makeBlock();
  block->name = Name(std::string("__inlined_func$") + from->name.str);
  if (call->isReturn) {
    if (retType.isConcrete()) {
      *action.callSite = builder.makeReturn(block);
    } else {
      *action.callSite = builder.makeSequence(block, builder.makeReturn());
    }
  } else {
    *action.callSite = block;
  }
  // Prepare to update the inlined code's locals and other things.
  Updater updater;
  updater.module = module;
  updater.returnName = block->name;
  updater.builder = &builder;
  // Set up a locals mapping
  for (Index i = 0; i < from->getNumLocals(); i++) {
    updater.localMapping[i] = builder.addVar(into, from->getLocalType(i));
  }
  // Assign the operands into the params
  for (Index i = 0; i < from->sig.params.size(); i++) {
    block->list.push_back(
      builder.makeLocalSet(updater.localMapping[i], call->operands[i]));
  }
  // Zero out the vars (as we may be in a loop, and may depend on their
  // zero-init value
  for (Index i = 0; i < from->vars.size(); i++) {
    block->list.push_back(
      builder.makeLocalSet(updater.localMapping[from->getVarIndexBase() + i],
                           LiteralUtils::makeZero(from->vars[i], *module)));
  }
  // Generate and update the inlined contents
  auto* contents = ExpressionManipulator::copy(from->body, *module);
  if (!from->debugLocations.empty()) {
    debug::copyDebugInfo(from->body, contents, from, into);
  }
  updater.walk(contents);
  block->list.push_back(contents);
  block->type = retType;
  // If the function returned a value, we just set the block containing the
  // inlined code to have that type. or, if the function was void and
  // contained void, that is fine too. a bad case is a void function in which
  // we have unreachable code, so we would be replacing a void call with an
  // unreachable.
  if (contents->type == Type::unreachable && block->type == Type::none) {
    // Make the block reachable by adding a break to it
    block->list.push_back(builder.makeBreak(block->name));
  }
  return block;
}

struct Inlining : public Pass {
  // whether to optimize where we inline
  bool optimize = false;

  // the information for each function. recomputed in each iteraction
  NameInfoMap infos;

  Index iterationNumber;

  void run(PassRunner* runner, Module* module) override {
    Index numFunctions = module->functions.size();
    // keep going while we inline, to handle nesting. TODO: optimize
    iterationNumber = 0;
    // no point to do more iterations than the number of functions, as
    // it means we infinitely recursing (which should
    // be very rare in practice, but it is possible that a recursive call
    // can look like it is worth inlining)
    while (iterationNumber <= numFunctions) {
#ifdef INLINING_DEBUG
      std::cout << "inlining loop iter " << iterationNumber
                << " (numFunctions: " << numFunctions << ")\n";
#endif
      calculateInfos(module);
      if (!iteration(runner, module)) {
        return;
      }
      iterationNumber++;
    }
  }

  void calculateInfos(Module* module) {
    infos.clear();
    // fill in info, as we operate on it in parallel (each function to its own
    // entry)
    for (auto& func : module->functions) {
      infos[func->name];
    }
    PassRunner runner(module);
    FunctionInfoScanner(&infos).run(&runner, module);
    // fill in global uses
    // anything exported or used in a table should not be inlined
    for (auto& ex : module->exports) {
      if (ex->kind == ExternalKind::Function) {
        infos[ex->value].usedGlobally = true;
      }
    }
    for (auto& segment : module->table.segments) {
      for (auto name : segment.data) {
        infos[name].usedGlobally = true;
      }
    }
  }

  bool iteration(PassRunner* runner, Module* module) {
    // decide which to inline
    InliningState state;
    ModuleUtils::iterDefinedFunctions(*module, [&](Function* func) {
      if (infos[func->name].worthInlining(runner->options)) {
        state.worthInlining.insert(func->name);
      }
    });
    if (state.worthInlining.size() == 0) {
      return false;
    }
    // fill in actionsForFunction, as we operate on it in parallel (each
    // function to its own entry)
    for (auto& func : module->functions) {
      state.actionsForFunction[func->name];
    }
    // find and plan inlinings
    Planner(&state).run(runner, module);
    // perform inlinings TODO: parallelize
    std::unordered_map<Name, Index> inlinedUses; // how many uses we inlined
    // which functions were inlined into
    std::unordered_set<Function*> inlinedInto;
    for (auto& func : module->functions) {
      // if we've inlined a function, don't inline into it in this iteration,
      // avoid risk of races
      // note that we do not risk stalling progress, as each iteration() will
      // inline at least one call before hitting this
      if (inlinedUses.count(func->name)) {
        continue;
      }
      for (auto& action : state.actionsForFunction[func->name]) {
        auto* inlinedFunction = action.contents;
        // if we've inlined into a function, don't inline it in this iteration,
        // avoid risk of races
        // note that we do not risk stalling progress, as each iteration() will
        // inline at least one call before hitting this
        if (inlinedInto.count(inlinedFunction)) {
          continue;
        }
        Name inlinedName = inlinedFunction->name;
#ifdef INLINING_DEBUG
        std::cout << "inline " << inlinedName << " into " << func->name << '\n';
#endif
        doInlining(module, func.get(), action);
        inlinedUses[inlinedName]++;
        inlinedInto.insert(func.get());
        assert(inlinedUses[inlinedName] <= infos[inlinedName].refs);
      }
    }
    // anything we inlined into may now have non-unique label names, fix it up
    for (auto func : inlinedInto) {
      wasm::UniqueNameMapper::uniquify(func->body);
    }
    if (optimize && inlinedInto.size() > 0) {
      OptUtils::optimizeAfterInlining(inlinedInto, module, runner);
    }
    // remove functions that we no longer need after inlining
    module->removeFunctions([&](Function* func) {
      auto name = func->name;
      auto& info = infos[name];
      return inlinedUses.count(name) && inlinedUses[name] == info.refs &&
             !info.usedGlobally;
    });
    // return whether we did any work
    return inlinedUses.size() > 0;
  }
};

Pass* createInliningPass() { return new Inlining(); }

Pass* createInliningOptimizingPass() {
  auto* ret = new Inlining();
  ret->optimize = true;
  return ret;
}

static const char* MAIN = "main";
static const char* ORIGINAL_MAIN = "__original_main";

// Inline __original_main into main, if they exist. This works around the odd
// thing that clang/llvm currently do, where __original_main contains the user's
// actual main (this is done as a workaround for main having two different
// possible signatures).
struct InlineMainPass : public Pass {
  void run(PassRunner* runner, Module* module) override {
    auto* main = module->getFunctionOrNull(MAIN);
    auto* originalMain = module->getFunctionOrNull(ORIGINAL_MAIN);
    if (!main || main->imported() || !originalMain ||
        originalMain->imported()) {
      return;
    }
    FindAllPointers<Call> calls(main->body);
    Expression** callSite = nullptr;
    for (auto* call : calls.list) {
      if ((*call)->cast<Call>()->target == ORIGINAL_MAIN) {
        if (callSite) {
          // More than one call site.
          return;
        }
        callSite = call;
      }
    }
    if (!callSite) {
      // No call at all.
      return;
    }
    doInlining(module, main, InliningAction(callSite, originalMain));
  }
};

Pass* createInlineMainPass() { return new InlineMainPass(); }

} // namespace wasm