From 8a87d1f8d6b594f1a2d027c1d1eab22dc441cdb3 Mon Sep 17 00:00:00 2001
From: Axel Naumann <Axel.Naumann@cern.ch>
Date: Thu, 12 Jan 2023 16:20:40 +0100
Subject: [PATCH] JIT: find existing weak symbol without materializer:

m_JIT.getSymbolAddress() invokes the symbol materializers, which compile (which
is sort of okay) but also try autoloading (which totally is not okay).

Instead, implement a function to search existing JIT symbols.

This can be accelerated by looking up the whole set of symbols, instead of doing it
symbol by symbol. I leave that refactoring for later...
---
 lib/Interpreter/BackendPasses.cpp  | 15 ++++++++++++---
 lib/Interpreter/IncrementalJIT.cpp | 15 ++++++++++++++-
 lib/Interpreter/IncrementalJIT.h   |  4 ++++
 3 files changed, 30 insertions(+), 4 deletions(-)

diff --git a/lib/Interpreter/BackendPasses.cpp b/lib/Interpreter/BackendPasses.cpp
index 514180a4..f1a04b65 100644
--- a/lib/Interpreter/BackendPasses.cpp
+++ b/lib/Interpreter/BackendPasses.cpp
@@ -267,9 +267,17 @@ namespace {
       if (!GV.isDiscardableIfUnused(LT) || !GV.isWeakForLinker(LT))
         return false;
 
-      // Find the symbol as existing, previously compiled symbol in the JIT,
-      // or in shared libraries (without auto-loading).
-      return (m_JIT.getSymbolAddress(GV.getName(), /*IncludeHostSyms*/ true));
+      // Find the symbol as existing, previously compiled symbol in the JIT...
+      if (m_JIT.doesSymbolAlreadyExist(GV.getName()))
+        return true;
+
+      // ...or in shared libraries (without auto-loading).
+      std::string Name = GV.getName().str();
+#if !defined(_WIN32)
+        return llvm::sys::DynamicLibrary::SearchForAddressOfSymbol(Name);
+#else
+        return platform::DLSym(Name);
+#endif
     }
 
     bool runOnVar(GlobalVariable& GV) {
@@ -310,6 +318,7 @@ namespace {
 
     bool runOnModule(Module &M) override {
       bool ret = false;
+      // FIXME: use SymbolLookupSet, rather than looking up symbol by symbol.
       for (auto &&F: M)
         ret |= runOnFunc(F);
       for (auto &&G: M.globals())
diff --git a/lib/Interpreter/IncrementalJIT.cpp b/lib/Interpreter/IncrementalJIT.cpp
index f6aa0265..e736c091 100644
--- a/lib/Interpreter/IncrementalJIT.cpp
+++ b/lib/Interpreter/IncrementalJIT.cpp
@@ -504,6 +504,10 @@ llvm::Error IncrementalJIT::removeModule(const Transaction& T) {
   m_ResourceTrackers.erase(&T);
   if (Error Err = RT->remove())
     return Err;
+  auto iMod = m_CompiledModules.find(T.m_CompiledModule);
+  if (iMod != m_CompiledModules.end())
+    m_CompiledModules.erase(iMod);
+
   return llvm::Error::success();
 }
 
@@ -550,7 +554,7 @@ IncrementalJIT::addOrReplaceDefinition(StringRef Name,
   return KnownAddr;
 }
 
-void* IncrementalJIT::getSymbolAddress(StringRef Name, bool IncludeHostSymbols) {
+void* IncrementalJIT::getSymbolAddress(StringRef Name, bool IncludeHostSymbols){
   std::unique_lock<SharedAtomicFlag> G(SkipHostProcessLookup, std::defer_lock);
   if (!IncludeHostSymbols)
     G.lock();
@@ -577,4 +581,13 @@ void* IncrementalJIT::getSymbolAddress(StringRef Name, bool IncludeHostSymbols)
   return jitTargetAddressToPointer<void*>(Symbol->getAddress());
 }
 
+bool IncrementalJIT::doesSymbolAlreadyExist(StringRef UnmangledName) {
+  auto Name = Jit->mangle(UnmangledName);
+  for (auto &&M: m_CompiledModules) {
+    if (M.first->getNamedValue(Name))
+      return true;
+  }
+  return false;
+}
+
 } // namespace cling
diff --git a/lib/Interpreter/IncrementalJIT.h b/lib/Interpreter/IncrementalJIT.h
index 6d067a6d..c05458c2 100644
--- a/lib/Interpreter/IncrementalJIT.h
+++ b/lib/Interpreter/IncrementalJIT.h
@@ -78,6 +78,10 @@ public:
   /// should include symbols from the host process (via dlsym) or not.
   void* getSymbolAddress(llvm::StringRef Name, bool IncludeHostSymbols);
 
+  /// @brief Check whether the JIT already has emitted or knows how to emit
+  /// a symbol based on its IR name (as coming from clang's mangler).
+  bool doesSymbolAlreadyExist(llvm::StringRef UnmangledName);
+
   /// Inject a symbol with a known address. Name is not linker mangled, i.e.
   /// as known by the IR.
   llvm::JITTargetAddress addOrReplaceDefinition(llvm::StringRef Name,