[GlobalDCE] Add !llvm.used.metadata support for conditionally used global variables

As part of the optimization work to emit dead-strippable symbols for Swift and
enable GlobalDCE to remove those symbols when possible, we need to mark some
Swift symbols as conditionally used based on a set of other symbols. Background
info can be found here: https://bugs.swift.org/browse/SR-14509.

This diff adds a module-level name metadata !llvm.used.conditional, which
contains a list of metadata triplets, that describes the conditions under which
we allow removal of globals even if they are mentioned in @llvm.used. See the
included LangRef changes to details. The implementation is a strict addition,
it does not change any semantics and does not affect IR or passes in any way,
unless !llvm.used.conditional is present in the input IR.

Author: kubamracek

llvm/docs/LangRef.rst

@@ -7304,6 +7304,55 @@ For example, the following metadata section contains two library specifiers::
 Each library specifier will be handled independently by the consuming linker.
 The effect of the library specifiers are defined by the consuming linker.
 
+.. _llvm_used_conditional:
+
+Conditionally Used Globals
+==========================
+
+When a global variable is mentioned in the `@llvm.used` list, it's always going
+to be retained, but sometimes it's desirable to allow such a global variable to
+still be discardable by optimization passes. Using the `!llvm.used.conditional`
+named metadata allows expressing *when* it is legal to discard a global (from
+the `@llvm.used` list) in terms of liveness of other globals.
+
+If `!llvm.used.conditional` named metadata is present in IR of a module, it's
+expected to be a list of metadata triplets. Each triplet has the following form:
+
+1. The first element is the "target", the global variable from `@llvm.used` that
+   we are allowing removal of. If the global variable is not present in the
+   `@llvm.used` list, then there is no effect from using a entry in
+   `!llvm.used.conditional` for it.
+2. The second element is a "type", a boolean flag expressing what behavior do
+   we want if the list of "dependencies" (third element) contains more than one
+   element. `0` means if *any dependency* in the list is alive, the target symbol
+   must stay alive, otherwise removal is allowed (in other words: if all globals
+   from the dependency list are removed, the target can be removed too). `1`
+   means if *all dependencies* in the list are alive, the target must stay
+   alive, otherwise removal is allowed (in other words: if any global from the
+   dependency list is removed, the target can be removed too).
+3. The third element is a list of "dependencies", a list of globals.
+
+The following example expresses that the global `@record` (which is listed in
+`@llvm.used` otherwise it would be trivially discarded as nothing references it)
+is allowed to be optimized away, if *either of* `@a` or `@b` globals are
+themselves removable::
+
+    @record = internal global [...] {
+        @a, @b
+    }
+    @llvm.used = appending global [...] [ ..., @record ]
+
+    !1 = !{
+        @record,     ; target
+        1,           ; type
+        !{ @a, @b }  ; dependencies
+    }
+    !llvm.used.conditional = !{ !1 }
+
+The semantics of `!llvm.used.conditional` are only *allowing an optimization*,
+and are not requiring optimizations to follow them --- it is correct for any
+optimization/transformation to ignore `!llvm.used.conditional` or even drop it.
+
 .. _summary:
 
 ThinLTO Summary

llvm/include/llvm/Transforms/IPO/GlobalDCE.h

@@ -54,6 +54,7 @@ class GlobalDCEPass : public PassInfoMixin<GlobalDCEPass> {
   void UpdateGVDependencies(GlobalValue &GV);
   void MarkLive(GlobalValue &GV,
                 SmallVectorImpl<GlobalValue *> *Updates = nullptr);
+  void PropagateLivenessInGlobalValues();
   bool RemoveUnusedGlobalValue(GlobalValue &GV);
 
   // Dead virtual function elimination.
@@ -63,6 +64,9 @@ class GlobalDCEPass : public PassInfoMixin<GlobalDCEPass> {
   void ScanVTableLoad(Function *Caller, Metadata *TypeId, uint64_t CallOffset);
 
   void ComputeDependencies(Value *V, SmallPtrSetImpl<GlobalValue *> &U);
+
+  GlobalValue *TargetFromConditionalUsedIfLive(MDNode *M);
+  void PropagateLivenessToConditionallyUsed(Module &M);
 };
 
 }

llvm/include/llvm/Transforms/Utils/ModuleUtils.h

@@ -13,6 +13,7 @@
 #ifndef LLVM_TRANSFORMS_UTILS_MODULEUTILS_H
 #define LLVM_TRANSFORMS_UTILS_MODULEUTILS_H
 
+#include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
 #include <utility> // for std::pair
@@ -24,6 +25,7 @@ class Module;
 class Function;
 class FunctionCallee;
 class GlobalValue;
+class GlobalVariable;
 class Constant;
 class Value;
 class Type;
@@ -78,6 +80,10 @@ void appendToUsed(Module &M, ArrayRef<GlobalValue *> Values);
 /// Adds global values to the llvm.compiler.used list.
 void appendToCompilerUsed(Module &M, ArrayRef<GlobalValue *> Values);
 
+/// Replaces llvm.used or llvm.compiler.used list with a new set of values.
+GlobalVariable *setUsedInitializer(GlobalVariable &V,
+                                   const SmallPtrSetImpl<GlobalValue *> &Init);
+
 /// Filter out potentially dead comdat functions where other entries keep the
 /// entire comdat group alive.
 ///

llvm/lib/IR/Verifier.cpp

@@ -841,6 +841,32 @@ void Verifier::visitNamedMDNode(const NamedMDNode &NMD) {
 
     visitMDNode(*MD, AreDebugLocsAllowed::Yes);
   }
+
+  if (NMD.getName() == "llvm.used.conditional") {
+    for (const MDNode *MD : NMD.operands()) {
+      Assert(MD->getNumOperands() == 3, "invalid llvm.used.conditional member");
+      auto *TargetMD = MD->getOperand(0).get();
+      if (TargetMD != nullptr) {
+        Assert(mdconst::dyn_extract<GlobalValue>(TargetMD),
+               "invalid llvm.used.conditional member");
+      }
+      auto *TypeMD = mdconst::extract_or_null<ConstantInt>(MD->getOperand(1));
+      int64_t Type = TypeMD->getValue().getSExtValue();
+      Assert(Type == 0 || Type == 1, "invalid llvm.used.conditional member");
+      auto *DependenciesMD = dyn_cast<MDNode>(MD->getOperand(2).get());
+      Assert(DependenciesMD, "invalid llvm.used.conditional member");
+      Assert(DependenciesMD->getNumOperands() > 0,
+             "invalid llvm.used.conditional member");
+      for (auto &DependencyMD : DependenciesMD->operands()) {
+        auto *Dependency = DependencyMD.get();
+        if (Dependency != nullptr) {
+          auto *C =
+              mdconst::dyn_extract<Constant>(Dependency)->stripPointerCasts();
+          Assert(isa<GlobalValue>(C), "invalid llvm.used.conditional member");
+        }
+      }
+    }
+  }
 }
 
 void Verifier::visitMDNode(const MDNode &MD, AreDebugLocsAllowed AllowLocs) {

llvm/lib/Transforms/IPO/GlobalDCE.cpp

@@ -28,6 +28,7 @@
 #include "llvm/Transforms/IPO.h"
 #include "llvm/Transforms/Utils/CtorUtils.h"
 #include "llvm/Transforms/Utils/GlobalStatus.h"
+#include "llvm/Transforms/Utils/ModuleUtils.h"
 
 using namespace llvm;
 
@@ -157,6 +158,18 @@ void GlobalDCEPass::MarkLive(GlobalValue &GV,
   }
 }
 
+void GlobalDCEPass::PropagateLivenessInGlobalValues() {
+  // Propagate liveness from collected Global Values through the computed
+  // dependencies.
+  SmallVector<GlobalValue *, 8> NewLiveGVs{AliveGlobals.begin(),
+                                           AliveGlobals.end()};
+  while (!NewLiveGVs.empty()) {
+    GlobalValue *LGV = NewLiveGVs.pop_back_val();
+    for (auto *GVD : GVDependencies[LGV])
+      MarkLive(*GVD, &NewLiveGVs);
+  }
+}
+
 void GlobalDCEPass::ScanVTables(Module &M) {
   SmallVector<MDNode *, 2> Types;
   LLVM_DEBUG(dbgs() << "Building type info -> vtable map\n");
@@ -286,6 +299,129 @@ void GlobalDCEPass::AddVirtualFunctionDependencies(Module &M) {
   );
 }
 
+static bool RemoveConditionalTargetsFromUsedList(Module &M) {
+  auto *Used = M.getGlobalVariable("llvm.used");
+  if (!Used)
+    return false;
+
+  auto *UsedConditional = M.getNamedMetadata("llvm.used.conditional");
+  if (!UsedConditional)
+    return false;
+  if (UsedConditional->getNumOperands() == 0)
+    return false;
+
+  // Construct a set of conditionally used targets.
+  SmallPtrSet<GlobalValue *, 8> Targets;
+  for (auto *M : UsedConditional->operands()) {
+    assert(M->getNumOperands() == 3);
+    auto *V = mdconst::extract_or_null<GlobalValue>(M->getOperand(0));
+    if (!V)
+      continue;
+    Targets.insert(V);
+  }
+
+  if (Targets.empty())
+    return false;
+
+  // Now remove all targets from @llvm.used.
+  SmallPtrSet<GlobalValue *, 8> NewUsedArray;
+  const ConstantArray *UsedList = cast<ConstantArray>(Used->getInitializer());
+  for (Value *Op : UsedList->operands()) {
+    GlobalValue *G = cast<GlobalValue>(Op->stripPointerCasts());
+    if (Targets.contains(G))
+      continue;
+    NewUsedArray.insert(G);
+  }
+  Used = setUsedInitializer(*Used, NewUsedArray);
+  return true;
+}
+
+// Parse one entry from !llvm.used.conditional list as a triplet of
+// { target, type, dependencies } and evaluate the conditional dependency, i.e.
+// check liveness of all dependencies and based on type conclude whether the
+// target is supposed to be declared alive. If yes, return the target, otherwise
+// return nullptr.
+GlobalValue *GlobalDCEPass::TargetFromConditionalUsedIfLive(MDNode *M) {
+  assert(M->getNumOperands() == 3);
+  auto *Target = mdconst::extract_or_null<GlobalValue>(M->getOperand(0));
+  if (!Target)
+    return nullptr;
+
+  auto *DependenciesMD = dyn_cast_or_null<MDNode>(M->getOperand(2).get());
+  SmallPtrSet<GlobalValue *, 8> Dependencies;
+  if (DependenciesMD == nullptr) {
+    Dependencies.insert(nullptr);
+  } else {
+    for (auto &DependencyMD : DependenciesMD->operands()) {
+      auto *Dependency = DependencyMD.get();
+      if (!Dependency)
+        continue;
+      auto *C =
+          mdconst::extract_or_null<Constant>(Dependency)->stripPointerCasts();
+      Dependencies.insert(cast<GlobalValue>(C));
+    }
+  }
+
+  bool AllDependenciesAlive = true;
+  bool AnyDependencyAlive = false;
+  for (auto *Dep : Dependencies) {
+    bool Live = AliveGlobals.count(Dep) != 0;
+    if (Live)
+      AnyDependencyAlive = true;
+    else
+      AllDependenciesAlive = false;
+  }
+
+  auto *Type = mdconst::extract_or_null<ConstantInt>(M->getOperand(1));
+  switch (Type->getValue().getSExtValue()) {
+  case 0:
+    return AnyDependencyAlive ? Target : nullptr;
+  case 1:
+    return AllDependenciesAlive ? Target : nullptr;
+  default:
+    llvm_unreachable("bad !llvm.used.conditional type");
+  }
+}
+
+void GlobalDCEPass::PropagateLivenessToConditionallyUsed(Module &M) {
+  auto *Used = M.getGlobalVariable("llvm.used");
+  auto *UsedConditional = M.getNamedMetadata("llvm.used.conditional");
+
+  SmallPtrSet<GlobalValue *, 8> NewUsedArray;
+  const ConstantArray *UsedList = cast<ConstantArray>(Used->getInitializer());
+  for (Value *Op : UsedList->operands()) {
+    NewUsedArray.insert(cast<GlobalValue>(Op->stripPointerCasts()));
+  }
+
+  // Repeat the liveness propagation iteraticely, one iteration might force
+  // other conditionally used globals to become alive.
+  while (true) {
+    PropagateLivenessInGlobalValues();
+
+    unsigned OldSize = NewUsedArray.size();
+    for (auto *M : UsedConditional->operands()) {
+      auto *Target = TargetFromConditionalUsedIfLive(M);
+      if (!Target) continue;
+
+      NewUsedArray.insert(Target);
+      MarkLive(*Target);
+      LLVM_DEBUG(dbgs() << "Conditionally used target alive: "
+                        << Target->getName() << "\n");
+    }
+
+    unsigned NewSize = NewUsedArray.size();
+    LLVM_DEBUG(dbgs() << "Conditionally used iteration end, old size: "
+                      << OldSize << " new size: " << NewSize << "\n");
+
+    // Stop the iteration once we reach a steady state (no new additions to
+    // @llvm.used).
+    if (NewSize == OldSize) break;
+  }
+
+  Used = setUsedInitializer(*Used, NewUsedArray);
+  MarkLive(*Used);
+}
+
 PreservedAnalyses GlobalDCEPass::run(Module &M, ModuleAnalysisManager &MAM) {
   bool Changed = false;
 
@@ -315,6 +451,11 @@ PreservedAnalyses GlobalDCEPass::run(Module &M, ModuleAnalysisManager &MAM) {
   // might call, if we have that information.
   AddVirtualFunctionDependencies(M);
 
+  // Process the !llvm.used.conditional list and (temporarily, see below)
+  // remove all "targets" from @llvm.used. No effect if `!llvm.used.conditional`
+  // is not present in the module.
+  bool UsedConditionalPresent = RemoveConditionalTargetsFromUsedList(M);
+
   // Loop over the module, adding globals which are obviously necessary.
   for (GlobalObject &GO : M.global_objects()) {
     Changed |= RemoveUnusedGlobalValue(GO);
@@ -348,16 +489,14 @@ PreservedAnalyses GlobalDCEPass::run(Module &M, ModuleAnalysisManager &MAM) {
     UpdateGVDependencies(GIF);
   }
 
-  // Propagate liveness from collected Global Values through the computed
-  // dependencies.
-  SmallVector<GlobalValue *, 8> NewLiveGVs{AliveGlobals.begin(),
-                                           AliveGlobals.end()};
-  while (!NewLiveGVs.empty()) {
-    GlobalValue *LGV = NewLiveGVs.pop_back_val();
-    for (auto *GVD : GVDependencies[LGV])
-      MarkLive(*GVD, &NewLiveGVs);
+  // Step 2 of !llvm.used.conditional processing: If any conditionally used
+  // "targets" are alive, put them back into @llvm.used.
+  if (UsedConditionalPresent) {
+    PropagateLivenessToConditionallyUsed(M);
   }
 
+  PropagateLivenessInGlobalValues();
+
   // Now that all globals which are needed are in the AliveGlobals set, we loop
   // through the program, deleting those which are not alive.
   //

llvm/lib/Transforms/Utils/ModuleUtils.cpp

@@ -111,6 +111,44 @@ void llvm::appendToCompilerUsed(Module &M, ArrayRef<GlobalValue *> Values) {
   appendToUsedList(M, "llvm.compiler.used", Values);
 }
 
+static int compareNames(Constant *const *A, Constant *const *B) {
+  Value *AStripped = (*A)->stripPointerCasts();
+  Value *BStripped = (*B)->stripPointerCasts();
+  return AStripped->getName().compare(BStripped->getName());
+}
+
+GlobalVariable *
+llvm::setUsedInitializer(GlobalVariable &V,
+                         const SmallPtrSetImpl<GlobalValue *> &Init) {
+  if (Init.empty()) {
+    V.eraseFromParent();
+    return nullptr;
+  }
+
+  // Type of pointer to the array of pointers.
+  PointerType *Int8PtrTy = Type::getInt8PtrTy(V.getContext(), 0);
+
+  SmallVector<Constant *, 8> UsedArray;
+  for (GlobalValue *GV : Init) {
+    Constant *Cast =
+        ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, Int8PtrTy);
+    UsedArray.push_back(Cast);
+  }
+  // Sort to get deterministic order.
+  array_pod_sort(UsedArray.begin(), UsedArray.end(), compareNames);
+  ArrayType *ATy = ArrayType::get(Int8PtrTy, UsedArray.size());
+
+  Module *M = V.getParent();
+  V.removeFromParent();
+  GlobalVariable *NV =
+      new GlobalVariable(*M, ATy, false, GlobalValue::AppendingLinkage,
+                         ConstantArray::get(ATy, UsedArray), "");
+  NV->takeName(&V);
+  NV->setSection("llvm.metadata");
+  delete &V;
+  return NV;
+}
+
 FunctionCallee
 llvm::declareSanitizerInitFunction(Module &M, StringRef InitName,
                                    ArrayRef<Type *> InitArgTypes) {