make llvm backend do better dce and dae (#1144)

Previously, the llvm optimization passes weren't terribly effective at
removing dead code and dead function arguments. We add some custom llvm
passes that make this work a bit more effectively.

There are two passes added:

* RemoveDeadKFunctions deletes function call instructions which call
side-effecting functions that are "effectively" pure, i.e., they exist
only to construct pure terms and return them.
* MustTailDeadArgElimination is a minorly modified version of the
deadargelim LLVM pass which fixes a bug preventing it from applying dead
argument elimination to musttail tail calls.

bin/llvm-kompile-clang

@@ -152,9 +152,9 @@ if ! $save_temps; then
 fi
 
 if [[ "$OSTYPE" == "darwin"* ]]; then
-  set_visibility_hidden="$LIBDIR/libSetVisibilityHidden.dylib"
+  passes="$LIBDIR/libKLLVMPass.dylib"
 else
-  set_visibility_hidden="$LIBDIR/libSetVisibilityHidden.so"
+  passes="$LIBDIR/libKLLVMPass.so"
 fi
 
 # On macOS, we get libunwind supplied as part of the developer tools in the OS,
@@ -176,7 +176,6 @@ fi
 if [ "$main" != "python_ast" ]; then
   if [ "$lto" = "lto" ]; then
     flags+=("-flto")
-    files=("$LIBDIR"/llvm/*.ll)
   else
     files=()
     if $compile; then
@@ -186,7 +185,7 @@ if [ "$main" != "python_ast" ]; then
       fi
       if $visibility_hidden; then
         modhidden="$tmpdir/hidden.bc"
-        run @OPT@ "$modopt" -load-pass-plugin "$set_visibility_hidden" -set-visibility-hidden -o "$modhidden"
+        run @OPT@ "$modopt" -load-pass-plugin "$passes" -set-visibility-hidden -o "$modhidden"
         modopt="$modhidden"
       fi
       run @LLC@ \

include/kllvm/ast/attribute_set.h

@@ -51,6 +51,7 @@ class attribute_set {
     Functional,
     Hook,
     Idem,
+    Impure,
     Label,
     Left,
     Location,

include/kllvm/codegen/CreateTerm.h

@@ -82,7 +82,8 @@ class create_term {
     */
   llvm::Value *create_function_call(
       std::string const &name, value_type return_cat,
-      std::vector<llvm::Value *> const &args, bool sret, bool tailcc);
+      std::vector<llvm::Value *> const &args, bool sret, bool tailcc,
+      bool impure);
 
   [[nodiscard]] llvm::BasicBlock *get_current_block() const {
     return current_block_;

include/kllvm/codegen/MustTailDeadArgElimination.h

@@ -0,0 +1,160 @@
+//===- DeadArgumentElimination.h - Eliminate Dead Args ----------*- C++ -*-===//
+//
+// 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
+//
+// Minor modification made by Pi Squared Inc to support tailcc musttail calls.
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass deletes dead arguments from internal functions.  Dead argument
+// elimination removes arguments which are directly dead, as well as arguments
+// only passed into function calls as dead arguments of other functions.  This
+// pass also deletes dead return values in a similar way.
+//
+// This pass is often useful as a cleanup pass to run after aggressive
+// interprocedural passes, which add possibly-dead arguments or return values.
+//
+//===----------------------------------------------------------------------===//
+
+// NOLINTBEGIN
+
+#ifndef LLVM_TRANSFORMS_IPO_DEADARGUMENTELIMINATION_H
+#define LLVM_TRANSFORMS_IPO_DEADARGUMENTELIMINATION_H
+
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/Twine.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/PassManager.h"
+#include "llvm/Pass.h"
+#include <map>
+#include <set>
+#include <string>
+#include <tuple>
+
+using namespace llvm;
+
+namespace llvm {
+
+class Module;
+class Use;
+class Value;
+
+} // namespace llvm
+
+namespace kllvm {
+
+/// Eliminate dead arguments (and return values) from functions.
+class DeadArgumentEliminationPass
+    : public PassInfoMixin<DeadArgumentEliminationPass> {
+public:
+#if LLVM_VERSION_MAJOR == 16
+  /// Struct that represents (part of) either a return value or a function
+  /// argument.  Used so that arguments and return values can be used
+  /// interchangeably.
+  struct RetOrArg {
+    Function const *F;
+    unsigned Idx;
+    bool IsArg;
+
+    RetOrArg(Function const *F, unsigned Idx, bool IsArg)
+        : F(F)
+        , Idx(Idx)
+        , IsArg(IsArg) { }
+
+    /// Make RetOrArg comparable, so we can put it into a map.
+    bool operator<(RetOrArg const &O) const {
+      return std::tie(F, Idx, IsArg) < std::tie(O.F, O.Idx, O.IsArg);
+    }
+
+    /// Make RetOrArg comparable, so we can easily iterate the multimap.
+    bool operator==(RetOrArg const &O) const {
+      return F == O.F && Idx == O.Idx && IsArg == O.IsArg;
+    }
+
+    std::string getDescription() const {
+      return (Twine(IsArg ? "Argument #" : "Return value #") + Twine(Idx)
+              + " of function " + F->getName())
+          .str();
+    }
+  };
+
+  /// During our initial pass over the program, we determine that things are
+  /// either alive or maybe alive. We don't mark anything explicitly dead (even
+  /// if we know they are), since anything not alive with no registered uses
+  /// (in Uses) will never be marked alive and will thus become dead in the end.
+  enum Liveness { Live, MaybeLive };
+
+  DeadArgumentEliminationPass(bool ShouldHackArguments = false)
+      : ShouldHackArguments(ShouldHackArguments) { }
+#endif
+  PreservedAnalyses run(Module &M, ModuleAnalysisManager &);
+#if LLVM_VERSION_MAJOR == 16
+  /// Convenience wrapper
+  RetOrArg createRet(Function const *F, unsigned Idx) {
+    return RetOrArg(F, Idx, false);
+  }
+
+  /// Convenience wrapper
+  RetOrArg createArg(Function const *F, unsigned Idx) {
+    return RetOrArg(F, Idx, true);
+  }
+
+  using UseMap = std::multimap<RetOrArg, RetOrArg>;
+
+  /// This maps a return value or argument to any MaybeLive return values or
+  /// arguments it uses. This allows the MaybeLive values to be marked live
+  /// when any of its users is marked live.
+  /// For example (indices are left out for clarity):
+  ///  - Uses[ret F] = ret G
+  ///    This means that F calls G, and F returns the value returned by G.
+  ///  - Uses[arg F] = ret G
+  ///    This means that some function calls G and passes its result as an
+  ///    argument to F.
+  ///  - Uses[ret F] = arg F
+  ///    This means that F returns one of its own arguments.
+  ///  - Uses[arg F] = arg G
+  ///    This means that G calls F and passes one of its own (G's) arguments
+  ///    directly to F.
+  UseMap Uses;
+
+  using LiveSet = std::set<RetOrArg>;
+  using LiveFuncSet = std::set<Function const *>;
+
+  /// This set contains all values that have been determined to be live.
+  LiveSet LiveValues;
+
+  /// This set contains all values that are cannot be changed in any way.
+  LiveFuncSet LiveFunctions;
+
+  using UseVector = SmallVector<RetOrArg, 5>;
+
+  /// This allows this pass to do double-duty as the dead arg hacking pass
+  /// (used only by bugpoint).
+  bool ShouldHackArguments = false;
+
+private:
+  Liveness markIfNotLive(RetOrArg Use, UseVector &MaybeLiveUses);
+  Liveness
+  surveyUse(Use const *U, UseVector &MaybeLiveUses, unsigned RetValNum = -1U);
+  Liveness surveyUses(Value const *V, UseVector &MaybeLiveUses);
+
+  void surveyFunction(Function const &F);
+  bool isLive(RetOrArg const &RA);
+  void
+  markValue(RetOrArg const &RA, Liveness L, UseVector const &MaybeLiveUses);
+  void markLive(RetOrArg const &RA);
+  void markLive(Function const &F);
+  void propagateLiveness(RetOrArg const &RA);
+  bool removeDeadStuffFromFunction(Function *F);
+  bool deleteDeadVarargs(Function &F);
+  bool removeDeadArgumentsFromCallers(Function &F);
+#endif
+};
+
+} // namespace kllvm
+
+#endif // LLVM_TRANSFORMS_IPO_DEADARGUMENTELIMINATION_H
+
+// NOLINTEND

include/kllvm/codegen/RemoveDeadKFunctions.h

@@ -0,0 +1,45 @@
+#ifndef REMOVE_DEAD_K_FUNCTIONS_H
+#define REMOVE_DEAD_K_FUNCTIONS_H
+
+#include "llvm/IR/Function.h"
+#include "llvm/IR/LegacyPassManager.h"
+#include "llvm/Pass.h"
+#include "llvm/Passes/PassBuilder.h"
+#include "llvm/Passes/PassPlugin.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/raw_ostream.h"
+
+using namespace llvm;
+
+namespace kllvm {
+
+bool run_remove_dead_k_functions(llvm::Function &f, TargetLibraryInfo *tli);
+
+struct legacy_remove_dead_k_functions : llvm::FunctionPass {
+  // NOLINTNEXTLINE(*-identifier-naming)
+  static char ID;
+  legacy_remove_dead_k_functions()
+      : llvm::FunctionPass(ID) { }
+  bool runOnFunction(llvm::Function &f) override {
+    TargetLibraryInfo *tli
+        = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(f);
+    return run_remove_dead_k_functions(f, tli);
+  }
+};
+
+struct remove_dead_k_functions : llvm::PassInfoMixin<remove_dead_k_functions> {
+  static llvm::PreservedAnalyses
+  run(llvm::Function &f, llvm::FunctionAnalysisManager &am) {
+    if (!run_remove_dead_k_functions(
+            f, &am.getResult<TargetLibraryAnalysis>(f))) {
+      return llvm::PreservedAnalyses::all();
+    }
+    llvm::PreservedAnalyses pa;
+    pa.preserveSet<llvm::CFGAnalyses>();
+    return pa;
+  }
+};
+
+} // namespace kllvm
+
+#endif

lib/CMakeLists.txt

@@ -3,7 +3,7 @@ add_subdirectory(ast)
 add_subdirectory(binary)
 add_subdirectory(codegen)
 add_subdirectory(printer)
-add_subdirectory(set-visibility-hidden)
+add_subdirectory(passes)
 
 separate_arguments(LLVM_DEFINITIONS_LIST NATIVE_COMMAND ${LLVM_DEFINITIONS})
 add_definitions(${LLVM_DEFINITIONS_LIST})

lib/ast/attribute_set.cpp

@@ -27,6 +27,7 @@ std::unordered_map<attribute_set::key, std::string> const &attribute_table() {
       {attribute_set::key::Functional, "functional"},
       {attribute_set::key::Hook, "hook"},
       {attribute_set::key::Idem, "idem"},
+      {attribute_set::key::Impure, "impure"},
       {attribute_set::key::Label, "label"},
       {attribute_set::key::Left, "left"},
       {attribute_set::key::Location,

lib/codegen/ApplyPasses.cpp

@@ -1,5 +1,7 @@
 #include <kllvm/codegen/ApplyPasses.h>
+#include <kllvm/codegen/MustTailDeadArgElimination.h>
 #include <kllvm/codegen/Options.h>
+#include <kllvm/codegen/RemoveDeadKFunctions.h>
 #include <kllvm/codegen/SetVisibilityHidden.h>
 
 #include "runtime/alloc_cpp.h"
@@ -93,6 +95,16 @@ void apply_kllvm_opt_passes(llvm::Module &mod, bool hidden_visibility) {
           pm.addPass(set_visibility_hidden());
         }
       });
+  pb.registerScalarOptimizerLateEPCallback(
+      [](llvm::FunctionPassManager &pm, OptimizationLevel level) {
+        pm.addPass(remove_dead_k_functions());
+      });
+  pb.registerOptimizerEarlyEPCallback(
+      [](llvm::ModulePassManager &pm, OptimizationLevel level) {
+        pm.addPass(DeadArgumentEliminationPass());
+        pm.addPass(
+            llvm::createModuleToFunctionPassAdaptor(remove_dead_k_functions()));
+      });
 
   // Create the pass manager.
   ModulePassManager mpm

lib/codegen/CreateTerm.cpp

@@ -795,6 +795,13 @@ llvm::Value *create_term::create_function_call(
   auto *return_sort = dynamic_cast<kore_composite_sort *>(
       pattern->get_constructor()->get_sort().get());
   auto return_cat = return_sort->get_category(definition_);
+  auto const &att = definition_->get_symbol_declarations()
+                        .at(pattern->get_constructor()->get_name())
+                        ->attributes();
+
+  bool impure = att.contains(attribute_set::key::Impure)
+                || !att.contains(attribute_set::key::Total);
+
   int i = 0;
   for (auto const &sort : pattern->get_constructor()->get_arguments()) {
     auto *concrete_sort = dynamic_cast<kore_composite_sort *>(sort.get());
@@ -832,12 +839,13 @@ llvm::Value *create_term::create_function_call(
     current_block_ = e.function_event_post(current_block_);
   }
 
-  return create_function_call(name, return_cat, args, sret, tailcc);
+  return create_function_call(name, return_cat, args, sret, tailcc, impure);
 }
 
 llvm::Value *create_term::create_function_call(
     std::string const &name, value_type return_cat,
-    std::vector<llvm::Value *> const &args, bool sret, bool tailcc) {
+    std::vector<llvm::Value *> const &args, bool sret, bool tailcc,
+    bool impure) {
   llvm::Type *return_type = getvalue_type(return_cat, module_);
   std::vector<llvm::Type *> types;
   bool collection = false;
@@ -872,6 +880,9 @@ llvm::Value *create_term::create_function_call(
   llvm::FunctionType *func_type
       = llvm::FunctionType::get(return_type, types, false);
   llvm::Function *func = get_or_insert_function(module_, name, func_type);
+  if (!impure) {
+    func->addFnAttr("kllvm-pure");
+  }
 
   auto *call = llvm::CallInst::Create(func, real_args, "", current_block_);
   set_debug_loc(call);
@@ -1173,6 +1184,7 @@ bool make_function(
   llvm::FunctionType *func_type
       = llvm::FunctionType::get(return_type, param_types, false);
   llvm::Function *apply_rule = get_or_insert_function(module, name, func_type);
+  apply_rule->setLinkage(llvm::GlobalValue::InternalLinkage);
   init_debug_axiom(axiom->attributes());
   std::string debug_name = name;
   if (axiom->attributes().contains(attribute_set::key::Label)) {

lib/codegen/Decision.cpp

@@ -440,7 +440,7 @@ void function_node::codegen(decision *d) {
       final_subst, d->definition_, d->current_block_, d->module_, false);
   auto *call = creator.create_function_call(
       function_, cat_, args, function_.substr(0, 5) == "hook_",
-      is_side_condition);
+      is_side_condition, false);
   call->setName(name_.substr(0, max_name_length));
   d->store(std::make_pair(name_, type_), call);
 
@@ -806,6 +806,13 @@ void make_eval_or_anywhere_function(
   // have one correct version of the function body after code generation
   // finishes.
   match_func->deleteBody();
+  auto const &att = definition->get_symbol_declarations()
+                        .at(function->get_name())
+                        ->attributes();
+  if (!att.contains(attribute_set::key::Impure)
+      && att.contains(attribute_set::key::Total)) {
+    match_func->addFnAttr("kllvm-pure");
+  }
   [[maybe_unused]] kore_symbol_declaration *symbol_decl
       = definition->get_symbol_declarations().at(function->get_name());
   init_debug_axiom(symbol_decl->attributes());

lib/passes/CMakeLists.txt

@@ -0,0 +1,23 @@
+add_library(KLLVMPassInternal
+  SetVisibilityHidden.cpp
+  RemoveDeadKFunctions.cpp
+  MustTailDeadArgElimination.cpp
+  PluginInfo.cpp
+)
+
+add_library(KLLVMPass MODULE
+  SetVisibilityHidden.cpp
+  RemoveDeadKFunctions.cpp
+  MustTailDeadArgElimination.cpp
+  PluginInfo.cpp
+)
+
+install(
+  TARGETS KLLVMPassInternal KLLVMPass
+  LIBRARY DESTINATION lib/kllvm
+)
+
+llvm_config(KLLVMPass
+  USE_SHARED true
+  core irreader passes
+)