jitlayers: replace sharedbytes intern pool with one that respects ali…

…gnment

The llvm optimizations may increase alignment beyond the initial
MAX_ALIGN. This pool's alignment was previously only
`sizeof(struct { atomic<int> RefCount; size_t Length; char Data[]; })`
however, potentially resulting in segfaults at runtime.

Fixes #52118

src/jitlayers.cpp

@@ -1410,10 +1410,12 @@ namespace {
 
     struct JITPointersT {
 
-        JITPointersT(orc::ExecutionSession &ES) JL_NOTSAFEPOINT : ES(ES) {}
+        JITPointersT(SharedBytesT &SharedBytes, std::mutex &Lock) JL_NOTSAFEPOINT
+            : SharedBytes(SharedBytes), Lock(Lock) {}
 
         Expected<orc::ThreadSafeModule> operator()(orc::ThreadSafeModule TSM, orc::MaterializationResponsibility &R) JL_NOTSAFEPOINT {
             TSM.withModuleDo([&](Module &M) JL_NOTSAFEPOINT {
+                std::lock_guard<std::mutex> locked(Lock);
                 for (auto &GV : make_early_inc_range(M.globals())) {
                     if (auto *Shared = getSharedBytes(GV)) {
                         ++InternedGlobals;
@@ -1429,10 +1431,11 @@ namespace {
             return std::move(TSM);
         }
 
+    private:
         // optimize memory by turning long strings into memoized copies, instead of
         // making a copy per object file of output.
-        // we memoize them using the ExecutionSession's string pool;
-        // this makes it unsafe to call clearDeadEntries() on the pool.
+        // we memoize them using a StringSet with a custom-alignment allocator
+        // to ensure they are properly aligned
         Constant *getSharedBytes(GlobalVariable &GV) JL_NOTSAFEPOINT {
             // We could probably technically get away with
             // interning even external linkage globals,
@@ -1458,11 +1461,17 @@ namespace {
                 // Cutoff, since we don't want to intern small strings
                 return nullptr;
             }
-            auto Interned = *ES.intern(Data);
+            Align Required = GV.getAlign().valueOrOne();
+            Align Preferred = MaxAlignedAlloc::alignment(Data.size());
+            if (Required > Preferred)
+                return nullptr;
+            StringRef Interned = SharedBytes.insert(Data).first->getKey();
+            assert(llvm::isAddrAligned(Preferred, Interned.data()));
             return literal_static_pointer_val(Interned.data(), GV.getType());
         }
 
-        orc::ExecutionSession &ES;
+        SharedBytesT &SharedBytes;
+        std::mutex &Lock;
     };
 }
 
@@ -1696,7 +1705,7 @@ JuliaOJIT::JuliaOJIT()
 #endif
     LockLayer(ObjectLayer),
     CompileLayer(ES, LockLayer, std::make_unique<CompilerT<N_optlevels>>(orc::irManglingOptionsFromTargetOptions(TM->Options), *TM)),
-    JITPointersLayer(ES, CompileLayer, orc::IRTransformLayer::TransformFunction(JITPointersT(ES))),
+    JITPointersLayer(ES, CompileLayer, orc::IRTransformLayer::TransformFunction(JITPointersT(SharedBytes, RLST_mutex))),
     OptimizeLayer(ES, JITPointersLayer, orc::IRTransformLayer::TransformFunction(OptimizerT<N_optlevels>(*TM, PrintLLVMTimers))),
     OptSelLayer(ES, OptimizeLayer, orc::IRTransformLayer::TransformFunction(selectOptLevel)),
     DepsVerifyLayer(ES, OptSelLayer, orc::IRTransformLayer::TransformFunction(validateExternRelocations)),

src/jitlayers.h

@@ -1,6 +1,8 @@
 // This file is a part of Julia. License is MIT: https://julialang.org/license
 
 #include <llvm/ADT/MapVector.h>
+#include <llvm/ADT/StringSet.h>
+#include <llvm/Support/AllocatorBase.h>
 
 #include <llvm/IR/LLVMContext.h>
 #include <llvm/IR/Constants.h>
@@ -292,6 +294,44 @@ static const inline char *name_from_method_instance(jl_method_instance_t *li) JL
     return jl_is_method(li->def.method) ? jl_symbol_name(li->def.method->name) : "top-level scope";
 }
 
+template <size_t offset = 0>
+class MaxAlignedAllocImpl
+    : public AllocatorBase<MaxAlignedAllocImpl<offset>> {
+
+public:
+    MaxAlignedAllocImpl() JL_NOTSAFEPOINT = default;
+
+    static Align alignment(size_t Size) JL_NOTSAFEPOINT {
+        // Define the maximum alignment we expect to require, from offset bytes off
+        // the returned pointer, this is >= alignof(std::max_align_t), which is too
+        // small often to actually use.
+        const size_t MaxAlignment = JL_CACHE_BYTE_ALIGNMENT;
+        return Align(std::min((size_t)llvm::PowerOf2Ceil(Size), MaxAlignment));
+    }
+
+    LLVM_ATTRIBUTE_RETURNS_NONNULL void *Allocate(size_t Size, Align Alignment) {
+        Align MaxAlign = alignment(Size);
+        assert(Alignment < MaxAlign); (void)Alignment;
+        return jl_gc_perm_alloc(Size, 0, MaxAlign.value(), offset);
+    }
+
+    inline LLVM_ATTRIBUTE_RETURNS_NONNULL
+    void * Allocate(size_t Size, size_t Alignment) {
+        return Allocate(Size, Align(Alignment));
+    }
+
+    // Pull in base class overloads.
+    using AllocatorBase<MaxAlignedAllocImpl>::Allocate;
+
+    void Deallocate(const void *Ptr, size_t Size, size_t /*Alignment*/) { abort(); }
+
+    // Pull in base class overloads.
+    using AllocatorBase<MaxAlignedAllocImpl>::Deallocate;
+
+private:
+};
+using MaxAlignedAlloc = MaxAlignedAllocImpl<>;
+
 typedef JITSymbol JL_JITSymbol;
 // The type that is similar to SymbolInfo on LLVM 4.0 is actually
 // `JITEvaluatedSymbol`. However, we only use this type when a JITSymbol
@@ -300,6 +340,7 @@ typedef JITSymbol JL_SymbolInfo;
 
 using CompilerResultT = Expected<std::unique_ptr<llvm::MemoryBuffer>>;
 using OptimizerResultT = Expected<orc::ThreadSafeModule>;
+using SharedBytesT = StringSet<MaxAlignedAllocImpl<sizeof(StringSet<>::MapEntryTy)>>;
 
 class JuliaOJIT {
 public:
@@ -516,6 +557,7 @@ class JuliaOJIT {
 
     // Note that this is a safepoint due to jl_get_library_ and jl_dlsym calls
     void optimizeDLSyms(Module &M);
+
 private:
 
     const std::unique_ptr<TargetMachine> TM;
@@ -529,6 +571,7 @@ class JuliaOJIT {
     std::mutex RLST_mutex{};
     int RLST_inc = 0;
     DenseMap<void*, std::string> ReverseLocalSymbolTable;
+    SharedBytesT SharedBytes;
 
     std::unique_ptr<DLSymOptimizer> DLSymOpt;
 

src/julia_internal.h

@@ -349,7 +349,7 @@ JL_DLLEXPORT int jl_gc_classify_pools(size_t sz, int *osize) JL_NOTSAFEPOINT;
 extern uv_mutex_t gc_perm_lock;
 void *jl_gc_perm_alloc_nolock(size_t sz, int zero,
     unsigned align, unsigned offset) JL_NOTSAFEPOINT;
-void *jl_gc_perm_alloc(size_t sz, int zero,
+JL_DLLEXPORT void *jl_gc_perm_alloc(size_t sz, int zero,
     unsigned align, unsigned offset) JL_NOTSAFEPOINT;
 void gc_sweep_sysimg(void);