Switch aarch64-darwin codegen to JITLink (ObjectLinkingLayer) and small code model

This fixes #41440, #43285 and similar issues, which stem from
CodeModel::Large not being correctly implemented on MachO/ARM64.

Requires LLVM 13.x or Git main (tested: 1dd5e6fed5db with patches
from the JuliaLang/llvm-project julia-release/13.x branch, available
at https://github.com/dnadlinger/llvm-project/commits/julia-main).

Requires an LLVM patch to pass through __eh_frame unwind information,
without which backtraces silently won't work:
llvm/llvm-project#52921

```
diff --git a/llvm/lib/ExecutionEngine/JITLink/MachO_arm64.cpp b/llvm/lib/ExecutionEngine/JITLink/MachO_arm64.cpp
index f2a029d35cd5..4d958b302ff9 100644
--- a/llvm/lib/ExecutionEngine/JITLink/MachO_arm64.cpp
+++ b/llvm/lib/ExecutionEngine/JITLink/MachO_arm64.cpp
@@ -705,6 +705,10 @@ void link_MachO_arm64(std::unique_ptr<LinkGraph> G,
     Config.PrePrunePasses.push_back(
         CompactUnwindSplitter("__LD,__compact_unwind"));

+    Config.PrePrunePasses.push_back(EHFrameSplitter("__TEXT,__eh_frame"));
+    Config.PrePrunePasses.push_back(EHFrameEdgeFixer("__TEXT,__eh_frame",
+        8, Delta64, Delta32, NegDelta32));
+
     // Add an in-place GOT/Stubs pass.
     Config.PostPrunePasses.push_back(
         PerGraphGOTAndPLTStubsBuilder_MachO_arm64::asPass);
```

Author: dnadlinger

src/codegen.cpp

@@ -8103,7 +8103,12 @@ extern "C" void jl_init_llvm(void)
     }
     // Allocate a target...
     Optional<CodeModel::Model> codemodel =
-#ifdef _P64
+#if defined(JL_USE_JITLINK)
+        // JITLink can patch up relocations between far objects so we can use the
+        // small code model – which is good, as the arge code model is unmaintained
+        // on MachO/AArch64.
+        CodeModel::Small;
+#elif defined(_P64)
         // Make sure we are using the large code model on 64bit
         // Let LLVM pick a default suitable for jitting on 32bit
         CodeModel::Large;
@@ -8145,13 +8150,15 @@ extern "C" void jl_init_llvm(void)
     }
 #endif
     if (jl_using_gdb_jitevents)
-        jl_ExecutionEngine->RegisterJITEventListener(JITEventListener::createGDBRegistrationListener());
+        jl_ExecutionEngine->enableJITDebuggingSupport();
 
 #if defined(JL_USE_INTEL_JITEVENTS) || \
     defined(JL_USE_OPROFILE_JITEVENTS) || \
     defined(JL_USE_PERF_JITEVENTS)
+#ifdef JL_USE_JITLINK
+#error "JIT profiling support (JL_USE_*_JITEVENTS) not yet available on platforms that use JITLink"
+#else
     const char *jit_profiling = getenv("ENABLE_JITPROFILING");
-#endif
 
 #if defined(JL_USE_INTEL_JITEVENTS)
     if (jit_profiling && atoi(jit_profiling)) {
@@ -8184,6 +8191,8 @@ extern "C" void jl_init_llvm(void)
 #ifdef JL_USE_PERF_JITEVENTS
     if (jl_using_perf_jitevents)
         jl_ExecutionEngine->RegisterJITEventListener(JITEventListener::createPerfJITEventListener());
+#endif
+#endif
 #endif
 
     cl::PrintOptionValues();

src/debuginfo.cpp

@@ -113,12 +113,6 @@ void jl_add_code_in_flight(StringRef name, jl_code_instance_t *codeinst, const D
 }
 
 
-#ifdef _OS_WINDOWS_
-#if defined(_CPU_X86_64_)
-void *lookupWriteAddressFor(RTDyldMemoryManager *memmgr, void *rt_addr);
-#endif
-#endif
-
 #if defined(_OS_WINDOWS_)
 static void create_PRUNTIME_FUNCTION(uint8_t *Code, size_t Size, StringRef fnname,
                                      uint8_t *Section, size_t Allocated, uint8_t *UnwindData)
@@ -177,7 +171,14 @@ struct revcomp {
 };
 
 
-class JuliaJITEventListener
+// Central registry for resolving function addresses to `jl_method_instance_t`s and
+// originating `ObjectFile`s (for the DWARF debug info).
+//
+// A global singleton instance is notified by the JIT whenever a new object is emitted,
+// and later queried by the various function info APIs. We also use the chance to handle
+// some platform-specific unwind info registration (which is unrelated to the query
+// functionality).
+class JITObjectRegistry
 {
     std::map<size_t, ObjectInfo, revcomp> objectmap;
     std::map<size_t, std::pair<size_t, jl_method_instance_t *>, revcomp> linfomap;
@@ -194,44 +195,16 @@ class JuliaJITEventListener
         return linfo;
     }
 
-    void NotifyObjectEmitted(const object::ObjectFile &Object,
-                             const RuntimeDyld::LoadedObjectInfo &L,
-                             RTDyldMemoryManager *memmgr)
+    void registerJITObject(const object::ObjectFile &Object,
+                           std::function<uint64_t(const StringRef &)> getLoadAddress,
+                           std::function<void*(void*)> lookupWriteAddress)
     {
         jl_ptls_t ptls = jl_current_task->ptls;
         // This function modify codeinst->fptr in GC safe region.
         // This should be fine since the GC won't scan this field.
         int8_t gc_state = jl_gc_safe_enter(ptls);
 
-        auto SavedObject = L.getObjectForDebug(Object).takeBinary();
-        // If the debug object is unavailable, save (a copy of) the original object
-        // for our backtraces.
-        // This copy seems unfortunate, but there doesn't seem to be a way to take
-        // ownership of the original buffer.
-        if (!SavedObject.first) {
-            auto NewBuffer = MemoryBuffer::getMemBufferCopy(
-                    Object.getData(), Object.getFileName());
-            auto NewObj = cantFail(object::ObjectFile::createObjectFile(NewBuffer->getMemBufferRef()));
-            SavedObject = std::make_pair(std::move(NewObj), std::move(NewBuffer));
-        }
-        const object::ObjectFile &debugObj = *SavedObject.first.release();
-        SavedObject.second.release();
-
-        object::section_iterator EndSection = debugObj.section_end();
-        StringMap<object::SectionRef> loadedSections;
-        for (const object::SectionRef &lSection: Object.sections()) {
-            auto sName = lSection.getName();
-            if (sName) {
-                bool inserted = loadedSections.insert(std::make_pair(*sName, lSection)).second;
-                assert(inserted); (void)inserted;
-            }
-        }
-        auto getLoadAddress = [&] (const StringRef &sName) -> uint64_t {
-            auto search = loadedSections.find(sName);
-            if (search == loadedSections.end())
-                return 0;
-            return L.getSectionLoadAddress(search->second);
-        };
+        object::section_iterator EndSection = Object.section_end();
 
 #ifdef _CPU_ARM_
         // ARM does not have/use .eh_frame
@@ -290,7 +263,7 @@ class JuliaJITEventListener
         uint8_t *UnwindData = NULL;
 #if defined(_CPU_X86_64_)
         uint8_t *catchjmp = NULL;
-        for (const object::SymbolRef &sym_iter : debugObj.symbols()) {
+        for (const object::SymbolRef &sym_iter : Object.symbols()) {
             StringRef sName = cantFail(sym_iter.getName());
             uint8_t **pAddr = NULL;
             if (sName.equals("__UnwindData")) {
@@ -313,9 +286,8 @@ class JuliaJITEventListener
                 SectionAddrCheck = SectionAddr;
                 SectionLoadCheck = SectionLoadAddr;
                 SectionWriteCheck = SectionLoadAddr;
-                if (memmgr)
-                    SectionWriteCheck = (uintptr_t)lookupWriteAddressFor(memmgr,
-                            (void*)SectionLoadAddr);
+                if (lookupWriteAddress)
+                    SectionWriteCheck = (uintptr_t)lookupWriteAddress((void*)SectionLoadAddr);
                 Addr += SectionWriteCheck - SectionLoadAddr;
                 *pAddr = (uint8_t*)Addr;
             }
@@ -343,7 +315,7 @@ class JuliaJITEventListener
 #endif // defined(_OS_X86_64_)
 #endif // defined(_OS_WINDOWS_)
 
-        auto symbols = object::computeSymbolSizes(debugObj);
+        auto symbols = object::computeSymbolSizes(Object);
         bool first = true;
         for (const auto &sym_size : symbols) {
             const object::SymbolRef &sym_iter = sym_size.first;
@@ -380,7 +352,7 @@ class JuliaJITEventListener
                 if (codeinst)
                     linfomap[Addr] = std::make_pair(Size, codeinst->def);
                 if (first) {
-                    ObjectInfo tmp = {&debugObj,
+                    ObjectInfo tmp = {&Object,
                         (size_t)SectionSize,
                         (ptrdiff_t)(SectionAddr - SectionLoadAddr),
                         *Section,
@@ -394,22 +366,18 @@ class JuliaJITEventListener
         jl_gc_safe_leave(ptls, gc_state);
     }
 
-    // must implement if we ever start freeing code
-    // virtual void NotifyFreeingObject(const ObjectImage &Object) {}
-    // virtual void NotifyFreeingObject(const object::ObjectFile &Obj) {}
-
     std::map<size_t, ObjectInfo, revcomp>& getObjectMap() JL_NOTSAFEPOINT
     {
         return objectmap;
     }
 };
 
-static JuliaJITEventListener jl_jit_events;
-JL_DLLEXPORT void ORCNotifyObjectEmitted(const object::ObjectFile &Object,
-                                         const RuntimeDyld::LoadedObjectInfo &L,
-                                         RTDyldMemoryManager *memmgr)
+static JITObjectRegistry jl_jit_object_registry;
+void jl_register_jit_object(const object::ObjectFile &Object,
+                            std::function<uint64_t(const StringRef &)> getLoadAddress,
+                            std::function<void *(void *)> lookupWriteAddress)
 {
-    jl_jit_events.NotifyObjectEmitted(Object, L, memmgr);
+    jl_jit_object_registry.registerJITObject(Object, getLoadAddress, lookupWriteAddress);
 }
 
 // TODO: convert the safe names from aotcomile.cpp:makeSafeName back into symbols
@@ -1178,7 +1146,7 @@ int jl_DI_for_fptr(uint64_t fptr, uint64_t *symsize, int64_t *slide,
 {
     int found = 0;
     uv_rwlock_wrlock(&threadsafe);
-    std::map<size_t, ObjectInfo, revcomp> &objmap = jl_jit_events.getObjectMap();
+    std::map<size_t, ObjectInfo, revcomp> &objmap = jl_jit_object_registry.getObjectMap();
     std::map<size_t, ObjectInfo, revcomp>::iterator fit = objmap.lower_bound(fptr);
 
     if (symsize)
@@ -1212,7 +1180,7 @@ extern "C" JL_DLLEXPORT int jl_getFunctionInfo_impl(jl_frame_t **frames_out, siz
     int64_t slide;
     uint64_t symsize;
     if (jl_DI_for_fptr(pointer, &symsize, &slide, &Section, &context)) {
-        frames[0].linfo = jl_jit_events.lookupLinfo(pointer);
+        frames[0].linfo = jl_jit_object_registry.lookupLinfo(pointer);
         int nf = lookup_pointer(Section, context, frames_out, pointer, slide, true, noInline);
         return nf;
     }
@@ -1221,7 +1189,7 @@ extern "C" JL_DLLEXPORT int jl_getFunctionInfo_impl(jl_frame_t **frames_out, siz
 
 extern "C" jl_method_instance_t *jl_gdblookuplinfo(void *p) JL_NOTSAFEPOINT
 {
-    return jl_jit_events.lookupLinfo((size_t)p);
+    return jl_jit_object_registry.lookupLinfo((size_t)p);
 }
 
 #if (defined(_OS_LINUX_) || defined(_OS_FREEBSD_) || (defined(_OS_DARWIN_) && defined(LLVM_SHLIB)))
@@ -1643,7 +1611,7 @@ uint64_t jl_getUnwindInfo_impl(uint64_t dwAddr)
 {
     // Might be called from unmanaged thread
     uv_rwlock_rdlock(&threadsafe);
-    std::map<size_t, ObjectInfo, revcomp> &objmap = jl_jit_events.getObjectMap();
+    std::map<size_t, ObjectInfo, revcomp> &objmap = jl_jit_object_registry.getObjectMap();
     std::map<size_t, ObjectInfo, revcomp>::iterator it = objmap.lower_bound(dwAddr);
     uint64_t ipstart = 0; // ip of the start of the section (if found)
     if (it != objmap.end() && dwAddr < it->first + it->second.SectionSize) {