src/jitlayers.cpp - master - githubmirror/julia

src/jitlayers.cpp
2 727 строк108 KB
Keno Fischer
jit: bypass the object cache while a jl_dump_llvm_opt hook is installed (#62619)
05 авг 2026, 22:54
05 авг 2026, 22:541c98883
// This file is a part of Julia. License is MIT: https://julialang.org/license

#include "llvm-version.h"
#include "platform.h"
#include <pthread.h>
#include <stdint.h>
#include <string>

#include "llvm/IR/Mangler.h"
#include <llvm/ADT/BitmaskEnum.h>
#include <llvm/ADT/Statistic.h>
#include <llvm/ADT/StringMap.h>
#include <llvm/Analysis/TargetLibraryInfo.h>
#include <llvm/Analysis/TargetTransformInfo.h>
#include <llvm/ExecutionEngine/Orc/CompileUtils.h>
#include <llvm/ExecutionEngine/Orc/ExecutionUtils.h>
#include <llvm/ExecutionEngine/Orc/DebugObjectManagerPlugin.h>
#if JL_LLVM_VERSION >= 210000
#  include <llvm/ExecutionEngine/Orc/SelfExecutorProcessControl.h>
#endif
#include <llvm/ExecutionEngine/Orc/TargetProcess/JITLoaderGDB.h>
#if JL_LLVM_VERSION >= 200000
#include <llvm/ExecutionEngine/Orc/AbsoluteSymbols.h>
#include <llvm/ExecutionEngine/Orc/EHFrameRegistrationPlugin.h>
#endif
#if JL_LLVM_VERSION >= 180000
#include <llvm/ExecutionEngine/Orc/Debugging/DebugInfoSupport.h>
#include <llvm/ExecutionEngine/Orc/Debugging/PerfSupportPlugin.h>
#include <llvm/ExecutionEngine/Orc/TargetProcess/JITLoaderPerf.h>
#endif
#if JL_LLVM_VERSION >= 190000
#include <llvm/ExecutionEngine/Orc/Debugging/VTuneSupportPlugin.h>
#include <llvm/ExecutionEngine/Orc/TargetProcess/JITLoaderVTune.h>
#endif
#include <llvm/ExecutionEngine/Orc/ExecutorProcessControl.h>
#include <llvm/IR/Verifier.h>
#include <llvm/Support/DynamicLibrary.h>
#include <llvm/Support/FormattedStream.h>
#include <llvm/Support/TimeProfiler.h>
#include <llvm/Support/SmallVectorMemoryBuffer.h>
#include <llvm/Support/raw_ostream.h>
#include <llvm/Transforms/Utils/Cloning.h>
#include <llvm/Transforms/Utils/ModuleUtils.h>
#include <llvm/Bitcode/BitcodeWriter.h>

#include <llvm/ExecutionEngine/JITLink/JITLink.h>
#if JL_LLVM_VERSION >= 210000
#include <llvm/ExecutionEngine/JITLink/EHFrameSupport.h>
#include <llvm/ExecutionEngine/Orc/Shared/WrapperFunctionUtils.h>
#endif
#include <llvm/ExecutionEngine/Orc/ObjectFileInterface.h>
#include <llvm/ExecutionEngine/Orc/DebugUtils.h>
#include <llvm/Object/MachO.h>
#include <llvm/Object/ObjectFile.h>

// target machine computation
#include <llvm/CodeGen/TargetSubtargetInfo.h>
#include <llvm/MC/TargetRegistry.h>
#include <llvm/Target/TargetOptions.h>
#include <llvm/TargetParser/Host.h>
#include <llvm/Support/TargetSelect.h>
#include <llvm/Object/SymbolSize.h>

using namespace llvm;

#include "jitlayers.h"
#include "julia_assert.h"
#include "processor.h"
#include "julia-task-dispatcher.h"

#if JL_LLVM_VERSION >= 180000
# include <llvm/ExecutionEngine/Orc/Debugging/DebuggerSupportPlugin.h>
#else
# include <llvm/ExecutionEngine/Orc/DebuggerSupportPlugin.h>
#endif
# include <llvm/ExecutionEngine/JITLink/EHFrameSupport.h>
# include <llvm/ExecutionEngine/JITLink/JITLinkMemoryManager.h>
# include <llvm/ExecutionEngine/Orc/MapperJITLinkMemoryManager.h>
# include <llvm/ExecutionEngine/SectionMemoryManager.h>

#define DEBUG_TYPE "julia_jitlayers"

STATISTIC(LinkedGlobals, "Number of globals linked");
STATISTIC(SpecFPtrCount, "Number of specialized function pointers compiled");
STATISTIC(UnspecFPtrCount, "Number of unspecialized function pointers compiled");
STATISTIC(ModulesAdded, "Number of modules added to the JIT");
STATISTIC(ModulesOptimized, "Number of modules optimized by the JIT");
STATISTIC(OptO0, "Number of modules optimized at level -O0");
STATISTIC(OptO1, "Number of modules optimized at level -O1");
STATISTIC(OptO2, "Number of modules optimized at level -O2");
STATISTIC(OptO3, "Number of modules optimized at level -O3");
STATISTIC(InternedGlobals, "Number of global constants interned in the string pool");

#ifdef _COMPILER_MSAN_ENABLED_
// TODO: This should not be necessary on ELF x86_64, but LLVM's implementation
// of the TLS relocations is currently broken, so enable this unconditionally.
#define MSAN_EMUTLS_WORKAROUND 1

// See https://github.com/google/sanitizers/wiki/MemorySanitizerJIT
namespace msan_workaround {

extern "C" {
    extern __thread unsigned long long __msan_param_tls[];
    extern __thread unsigned int __msan_param_origin_tls[];
    extern __thread unsigned long long __msan_retval_tls[];
    extern __thread unsigned int __msan_retval_origin_tls;
    extern __thread unsigned long long __msan_va_arg_tls[];
    extern __thread unsigned int __msan_va_arg_origin_tls[];
    extern __thread unsigned long long __msan_va_arg_overflow_size_tls;
    extern __thread unsigned int __msan_origin_tls;
}

enum class MSanTLS
{
    param = 1,             // __msan_param_tls
    param_origin,          //__msan_param_origin_tls
    retval,                // __msan_retval_tls
    retval_origin,         //__msan_retval_origin_tls
    va_arg,                // __msan_va_arg_tls
    va_arg_origin,         // __msan_va_arg_origin_tls
    va_arg_overflow_size,  // __msan_va_arg_overflow_size_tls
    origin,                //__msan_origin_tls
};

static void *getTLSAddress(void *control)
{
    auto tlsIndex = static_cast<MSanTLS>(reinterpret_cast<uintptr_t>(control));
    switch(tlsIndex)
    {
    case MSanTLS::param: return reinterpret_cast<void *>(&__msan_param_tls);
    case MSanTLS::param_origin: return reinterpret_cast<void *>(&__msan_param_origin_tls);
    case MSanTLS::retval: return reinterpret_cast<void *>(&__msan_retval_tls);
    case MSanTLS::retval_origin: return reinterpret_cast<void *>(&__msan_retval_origin_tls);
    case MSanTLS::va_arg: return reinterpret_cast<void *>(&__msan_va_arg_tls);
    case MSanTLS::va_arg_origin: return reinterpret_cast<void *>(&__msan_va_arg_origin_tls);
    case MSanTLS::va_arg_overflow_size: return reinterpret_cast<void *>(&__msan_va_arg_overflow_size_tls);
    case MSanTLS::origin: return reinterpret_cast<void *>(&__msan_origin_tls);
    default:
        assert(false && "BAD MSAN TLS INDEX");
        return nullptr;
    }
}
}
#endif

#ifdef _OS_OPENBSD_
extern "C" {
    __int128 __divti3(__int128, __int128);
    __int128 __modti3(__int128, __int128);
    unsigned __int128 __udivti3(unsigned __int128, unsigned __int128);
    unsigned __int128 __umodti3(unsigned __int128, unsigned __int128);
}
#endif

// Snooping on which functions are being compiled, and how long it takes
extern "C" JL_DLLEXPORT_CODEGEN
void jl_dump_compiles_impl(void *s)
{
    **jl_ExecutionEngine->get_dump_compiles_stream() = (ios_t*)s;
}
extern "C" JL_DLLEXPORT_CODEGEN
void jl_dump_llvm_opt_impl(void *s)
{
    **jl_ExecutionEngine->get_dump_llvm_opt_stream() = (ios_t*)s;
}

static void decorate_module(Module &M) JL_NOTSAFEPOINT;

// convert local roots into global roots, if they are needed
static void jl_promote_method_roots(jl_codegen_output_t &out, jl_method_instance_t *mi) JL_CANSAFEPOINT
{
    JL_GC_PROMISE_ROOTED(out.temporary_roots); // rooted by caller
    if (jl_array_dim0(out.temporary_roots) == 0)
        return;
    jl_method_t *m = mi->def.method;
    if (jl_is_method(m))
        // the method might have a root for this already; use it if so
        JL_LOCK(&m->writelock);
    for (size_t i = 0; i < jl_array_dim0(out.temporary_roots); i++) {
        jl_value_t *val = jl_array_ptr_ref(out.temporary_roots, i);
        auto ref = out.global_targets.find((void*)val);
        if (ref == out.global_targets.end())
            continue;
        auto get_global_root = [val, m]() JL_CANSAFEPOINT {
            if (jl_is_globally_rooted(val))
                return val;
            if (jl_is_method(m) && m->roots) {
                size_t j, len = jl_array_dim0(m->roots);
                for (j = 0; j < len; j++) {
                    jl_value_t *mval = jl_array_ptr_ref(m->roots, j);
                    if (jl_egal(mval, val)) {
                        return mval;
                    }
                }
            }
            return jl_as_global_root(val, 1);
        };
        jl_value_t *mval = get_global_root();
        if (mval != val) {
            GlobalVariable *GV = ref->second;
            out.global_targets.erase(ref);
            auto mref = out.global_targets.find((void*)mval);
            if (mref != out.global_targets.end()) {
                GV->replaceAllUsesWith(mref->second);
                GV->eraseFromParent();
            }
            else {
                out.global_targets[(void*)mval] = GV;
            }
        }
    }
    if (jl_is_method(m))
        JL_UNLOCK(&m->writelock);
}

StringRef jl_codegen_output_t::strip_linux(StringRef name)
{
    if (TargetTriple.isOSLinux()) {
        if (name[0] == '@')
            return name.drop_front();
    }
    return name;
}

std::string jl_codegen_output_t::make_name(jl_symbol_prefix_t type, jl_invoke_api_t api,
                                           StringRef orig_name)
{
    return make_name(jl_symbol_prefix(type, api), orig_name);
}

static std::atomic<size_t> global_name_counter;

template<class... Ts>
static std::string make_name_unique(Ts... args) JL_NOTSAFEPOINT
{
    std::string name;
    raw_string_ostream s{name};
    (s << ... << args);
    s << global_name_counter.fetch_add(1, memory_order_relaxed);
    return name;
}

std::string jl_codegen_output_t::make_name(StringRef prefix, StringRef orig_name)
{
    if (params->unique_names)
        return make_name_unique(prefix, strip_linux(orig_name), "_");
    return names(prefix, strip_linux(orig_name), "_");
}

std::string jl_codegen_output_t::make_name(StringRef orig_name)
{
    if (params->unique_names)
        return make_name_unique(strip_linux(orig_name));
    return names(strip_linux(orig_name));
}

// TODO: Don't repeat so much work in this and `emit_call_specfun_other`
// TODO: just take jl_invoke_api_t argument instead of specsig?
StringRef jl_codegen_output_t::get_call_target(jl_code_instance_t *ci, bool specsig,
                                               bool always_inline)
{
    jl_invoke_api_t api = specsig ? JL_INVOKE_SPECSIG : JL_INVOKE_ARGS;
    auto it = call_targets.find({ci, api});
    if (it != call_targets.end()) {
        it->second.external_linkage |= !always_inline;
        it->second.private_linkage |= always_inline;
        return it->second.decl->getName();
    }
    std::string protoname = make_name(JL_SYMBOL_SPECPTR_PROTO, api,
                                      name_from_method_instance(jl_get_ci_mi(ci)));
    jl_codegen_call_target_t &target = call_targets[{ci, api}];
    target.external_linkage = !always_inline;
    target.private_linkage = always_inline;
    if (specsig) {
        jl_method_instance_t *mi = jl_get_ci_mi(ci);
        bool is_opaque_closure =
            jl_is_method(mi->def.value) && mi->def.method->is_for_opaque_closure;
        jl_returninfo_t info =
            get_specsig_function(*this, &get_module(), nullptr, protoname, get_ci_abi(ci),
                                 ci->rettype, is_opaque_closure);
        target.decl = cast<Function>(info.decl.getCallee());
    }
    else {
        target.decl = get_or_emit_fptr1(protoname, &get_module());
    }
    return target.decl->getName();
}

jl_emitted_output_t jl_codegen_output_t::finish(std::unique_ptr<LLVMContext> ctx,
                                                std::unique_ptr<Module> mod,
                                                orc::SymbolStringPool &SSP)
{
    auto info = std::make_unique<jl_linker_info_t>();
    auto intern = [&](StringRef name) JL_NOTSAFEPOINT {
        SmallString<128> buf;
        Mangler::getNameWithPrefix(buf, name, DL);
        return SSP.intern(buf);
    };

    // Mangle and intern each part of the linking metadata, before all the
    // pointers to LLVM values are invalidated.
    for (auto &[ci, funcs] : ci_funcs) {
        info->ci_funcs[ci] = {funcs.invoke_api,
                              funcs.invoke ? intern(funcs.invoke->getName()) : nullptr,
                              funcs.specptr ? intern(funcs.specptr->getName()) : nullptr};
    }
    for (auto &[call, target] : call_targets)
        info->call_targets[call] = intern(target.decl->getName());
    for (auto [val, gv] : global_targets) {
        info->global_targets[val] = intern(gv->getName());
    }

    return {std::move(ctx), std::move(mod), std::move(info)};
}

// Return a specptr that is ABI-compatible with `from_abi` which invokes `codeinst`.
//
// If `codeinst` is NULL, the returned specptr instead performs a standard `apply_generic`
// call via a dynamic dispatch.
extern "C" JL_DLLEXPORT_CODEGEN
void *jl_jit_abi_converter_impl(jl_task_t *ct, jl_abi_t from_abi,
                                jl_code_instance_t *codeinst)
{
    void *target = nullptr;
    bool target_specsig = false;
    jl_callptr_t invoke = nullptr;
    if (codeinst != nullptr) {
        uint8_t specsigflags;
        jl_method_instance_t *mi = jl_get_ci_mi(codeinst);
        void *specptr = nullptr;
        jl_read_codeinst_invoke(codeinst, &specsigflags, &invoke, &specptr, /* waitcompile */ 1);
        if (invoke != nullptr) {
            if (invoke == jl_fptr_const_return_addr) {
                target = nullptr;
                target_specsig = false;
            }
            else if (invoke == jl_fptr_args_addr) {
                assert(specptr != nullptr);
                if (!from_abi.specsig && jl_subtype(codeinst->rettype, from_abi.rt))
                    return specptr; // no adapter required

                target = specptr;
                target_specsig = false;
            }
            else if (specsigflags & JL_CI_FLAGS_SPECPTR_SPECIALIZED) {
                assert(specptr != nullptr);
                if (from_abi.specsig && jl_egal(mi->specTypes, from_abi.sigt) && jl_egal(codeinst->rettype, from_abi.rt))
                    return specptr; // no adapter required

                target = specptr;
                target_specsig = true;
            }
        }
    }

    orc::ThreadSafeModule result_m;
    std::string gf_thunk_name;
    auto ctx = std::make_unique<LLVMContext>();
    auto mod = jl_create_llvm_module("gfthunk", *ctx, jl_ExecutionEngine->getDataLayout(),
                                     jl_ExecutionEngine->getTargetTriple());
    jl_codegen_output_t out{*mod};
    // root the wrapper types that `mark_julia_const` mints for egality-pinned
    // (`TypeEgal`) argument slots while the thunk is emitted
    out.temporary_roots = jl_alloc_array_1d(jl_array_any_type, 0);
    JL_GC_PUSH1(&out.temporary_roots);
    {
        ctx->setDiscardValueNames(true);
        out.imaging_mode = 0;
        if (target) {
            Value *llvmtarget = literal_static_pointer_val((void*)target, PointerType::get(*ctx, 0));
            gf_thunk_name = emit_abi_converter(out, from_abi, codeinst, llvmtarget, target_specsig);
        }
        else if (invoke == jl_fptr_const_return_addr) {
            assert(codeinst);   // Convince the static analyzer
            gf_thunk_name = emit_abi_constreturn(out, from_abi, codeinst->rettype_const);
        }
        else {
            Value *llvminvoke = invoke ? literal_static_pointer_val((void*)invoke, PointerType::get(*ctx, 0)) : nullptr;
            gf_thunk_name = emit_abi_dispatcher(out, from_abi, codeinst, llvminvoke);
        }
    }
    auto &ES = jl_ExecutionEngine->getExecutionSession();
    auto emitted = out.finish(std::move(ctx), std::move(mod), *ES.getSymbolStringPool());
    out.temporary_roots = nullptr;
    out.temporary_roots_set.clear();
    JL_GC_POP();
    jl_ExecutionEngine->addOutput(std::move(emitted));
    uintptr_t Addr = jl_ExecutionEngine->getFunctionAddress(gf_thunk_name);
    assert(Addr);
    return (void*)Addr;
}

  // lock for places where only single threaded behavior is implemented, so we need GC support
static jl_mutex_t jitlock;

// Lock hierarchy here:
//   jitlock is outermost, can contain others and allows GC
//   ThreadSafeContext locks are next, they should not be nested
//   jl_ExecutionEngine internal locks are exclusive to this list, since OrcJIT promises to never hold a lock over a materialization unit:
//        construct a query object from a query set and query handler
//        lock the session
//        lodge query against requested symbols, collect required materializers (if any)
//        unlock the session
//        dispatch materializers (if any)
//     However, this guarantee relies on Julia releasing all TSC locks before causing any materialization units to be dispatched
//     as materialization may need to acquire TSC locks.

static void jl_publish_compiled_ci(jl_code_instance_t *ci,
                                   const jl_codeinst_funcs_t<void *> &addrs) JL_NOTSAFEPOINT
{
    void *spec = addrs.specptr;
    jl_callptr_t invoke = addrs.invoke_api == JL_INVOKE_SPECSIG ?
                              (jl_callptr_t)addrs.invoke :
                              jl_invoke_api_callptr(addrs.invoke_api);

    void *prev = nullptr;
    if (jl_atomic_cmpswap_acqrel(&ci->specptr.fptr, &prev, spec)) {
        // only set specsig and invoke if we were the first to set specptr
        // Clear compilation state bits, then set SPECPTR_SPECIALIZED if needed
        if (addrs.invoke_api == JL_INVOKE_SPECSIG)
            jl_atomic_fetch_or_relaxed(&ci->flags, JL_CI_FLAGS_SPECPTR_SPECIALIZED);
        // we might overwrite invokeptr here; that's ok, anybody who relied on the identity
        // of invokeptr either assumes that specptr was null, doesn't care about specptr, or
        // will wait until flags has 0b10 set before reloading invoke
        jl_atomic_store_release(&ci->invoke, invoke);
        // Set INVOKE_MATCHES_SPECPTR to signal completion
        jl_atomic_fetch_or_relaxed(&ci->flags, JL_CI_FLAGS_INVOKE_MATCHES_SPECPTR);
    }
    else {
        // someone else beat us, don't commit any results
        while (!(jl_atomic_load_acquire(&ci->flags) & JL_CI_FLAGS_INVOKE_MATCHES_SPECPTR))
            jl_cpu_pause();
    }
}

static void jl_do_dump_compile(jl_code_instance_t *codeinst, uint64_t time) JL_NOTSAFEPOINT
{
    jl_method_instance_t *mi = jl_get_ci_mi(codeinst);
    if (jl_is_method(mi->def.method)) {
        auto stream = *jl_ExecutionEngine->get_dump_compiles_stream();
        if (stream) {
            ios_printf(stream, "%" PRIu64 "\t\"", time);
            jl_static_show((JL_STREAM *)stream, mi->specTypes);
            ios_printf(stream, "\"\n");
        }
    }

    float orig_time = julia_half_to_float(jl_atomic_load_relaxed(&codeinst->time_compile));
    jl_atomic_store_relaxed(&codeinst->time_compile,
                            julia_double_to_half(orig_time + time * 1e-9));
}

extern "C" JL_DLLEXPORT_CODEGEN void
jl_emit_codeinsts_to_jit_impl(jl_code_instance_t **codeinsts, jl_code_info_t **srcs, int len)
{
    if (len == 0)
        return;

    JL_TIMING(CODEINST_COMPILE, CODEINST_COMPILE);
    const char *name = name_from_method_instance(jl_get_ci_mi(codeinsts[len - 1]));
    auto ctx = std::make_unique<LLVMContext>();
    auto &dl = jl_ExecutionEngine->getDataLayout();
    auto &tt = jl_ExecutionEngine->getTargetTriple();
    auto mod = jl_create_llvm_module(name, *ctx, dl, tt);
    jl_codegen_output_t out{*mod};
    out.get_context().setDiscardValueNames(true);
    out.imaging_mode = false;
    JL_GC_PUSH1(&out.temporary_roots);

    for (int i = 0; i < len; ++i) {
        jl_code_instance_t *codeinst = codeinsts[i];
        jl_code_info_t *src = srcs[i];
        jl_method_instance_t *mi = jl_get_ci_mi(codeinst);

        if (jl_atomic_load_relaxed(&codeinst->invoke))
            continue;

        out.temporary_roots = jl_alloc_array_1d(jl_array_any_type, 0);
        out.temporary_roots_set.clear();
        if (!jl_emit_codeinst(out, codeinst, src)) { // contains safepoints
            JL_GC_POP();
            return;
        }

        // contains safepoints
        jl_promote_method_roots(out, mi);
        emit_always_inline(out, jl_get_method_ir); // contains safepoints

        // Non-opaque-closure MethodInstances are considered globally rooted
        // through their methods, but for OC, we need to create a global root
        // here.
        if (jl_is_method(mi->def.value) && mi->def.method->is_for_opaque_closure)
            jl_as_global_root((jl_value_t*)mi, 1);
    }

    out.temporary_roots = nullptr;
    out.temporary_roots_set.clear();
    JL_GC_POP();

    if (out.ci_funcs.empty())
        return;

    emit_llvmcall_modules(out);

    auto &ES = jl_ExecutionEngine->getExecutionSession();
    jl_emitted_output_t emitted =
        out.finish(std::move(ctx), std::move(mod), *ES.getSymbolStringPool());
    jl_ExecutionEngine->addOutput(std::move(emitted));
}

extern "C" JL_DLLEXPORT_CODEGEN
int jl_compile_codeinst_impl(jl_code_instance_t *ci)
{
    int newly_compiled = 0;
    if (!jl_is_compiled_codeinst(ci)) {
        ++SpecFPtrCount;
        uint64_t start = jl_typeinf_timing_begin();
        jl_ExecutionEngine->publishCIs(ci, true);
        jl_typeinf_timing_end(start, 0);
        newly_compiled = 1;
    }
    return newly_compiled;
}

extern "C" JL_DLLEXPORT_CODEGEN
void jl_generate_fptr_for_unspecialized_impl(jl_code_instance_t *unspec)
{
    if (jl_atomic_load_relaxed(&unspec->invoke) != NULL) {
        return;
    }
    auto ct = jl_current_task;
    bool timed = (ct->reentrant_timing & 1) == 0;
    if (timed)
        ct->reentrant_timing |= 1;
    uint64_t compiler_start_time = 0;
    uint8_t measure_compile_time_enabled = jl_atomic_load_relaxed(&jl_measure_compile_time_enabled);
    if (measure_compile_time_enabled)
        compiler_start_time = jl_hrtime();
    jl_code_info_t *src = NULL;
    JL_GC_PUSH1(&src);
    jl_method_t *def = jl_get_ci_mi(unspec)->def.method;
    if (jl_is_method(def)) {
        src = (jl_code_info_t*)def->source;
        if (src && (jl_value_t*)src != jl_nothing)
            src = jl_uncompress_ir(def, NULL, (jl_value_t*)src);
    }
    else {
        jl_method_instance_t *mi = jl_get_ci_mi(unspec);
        jl_code_instance_t *uninferred = jl_cached_uninferred(jl_atomic_load_relaxed(&mi->cache), 1);
        assert(uninferred);
        src = (jl_code_info_t*)jl_atomic_load_relaxed(&uninferred->inferred);
        assert(src);
    }
    if (src) {
        // TODO: first prepare recursive_compile_graph(unspec, src) before taking this lock to avoid recursion?
        JL_LOCK(&jitlock); // TODO: use a better lock
        if (!jl_is_compiled_codeinst(unspec)) {
            assert(jl_is_code_info(src));
            ++UnspecFPtrCount;
            jl_svec_t *edges = (jl_svec_t*)src->edges;
            if (jl_is_svec(edges)) {
                jl_gc_write_atomic(unspec, unspec->edges, jl_svec_t, edges, release); // n.b. this assumes the field was always empty svec(), which is not entirely true
            }
            jl_debuginfo_t *debuginfo = src->debuginfo;
            jl_gc_write_atomic(unspec, unspec->debuginfo, jl_debuginfo_t, debuginfo, release); // n.b. this assumes the field was previously NULL, which is not entirely true
            jl_emit_codeinsts_to_jit(&unspec, &src, 1);
            jl_ExecutionEngine->publishCIs(unspec, true);
        }
        JL_UNLOCK(&jitlock); // Might GC
    }
    JL_GC_POP();
    jl_callptr_t null = nullptr;
    // if we hit a codegen bug (or ran into a broken generated function or llvmcall), fall back to the interpreter as a last resort
    jl_atomic_cmpswap(&unspec->invoke, &null, jl_fptr_interpret_call_addr);
    if (timed) {
        if (measure_compile_time_enabled) {
            auto end = jl_hrtime();
            jl_atomic_fetch_add_relaxed(&jl_cumulative_compile_time, end - compiler_start_time);
        }
        ct->reentrant_timing &= ~1ull;
    }
}


// get a native disassembly for a compiled method
extern "C" JL_DLLEXPORT_CODEGEN
jl_value_t *jl_dump_method_asm_impl(jl_method_instance_t *mi, size_t world,
        char emit_mc, char getwrapper, const char* asm_variant, const char *debuginfo, char binary)
{
    // printing via disassembly
    jl_code_instance_t *codeinst = jl_compile_method_internal(mi, world);
    if (codeinst) {
        uintptr_t fptr = (uintptr_t)jl_atomic_load_acquire(&codeinst->invoke);
        uintptr_t specfptr = (uintptr_t)jl_atomic_load_relaxed(&codeinst->specptr.fptr);
        if (getwrapper || specfptr == 0)
            specfptr = fptr;
        if (specfptr != 0)
            return jl_dump_fptr_asm(specfptr, emit_mc, asm_variant, debuginfo, binary);
    }
    return jl_an_empty_string;
}

#if JL_LLVM_VERSION >= 180000
CodeGenOptLevel CodeGenOptLevelFor(int optlevel)
{
#ifdef DISABLE_OPT
    return CodeGenOptLevel::None;
#else
    return optlevel == 0 ? CodeGenOptLevel::None :
        optlevel == 1 ? CodeGenOptLevel::Less :
        optlevel == 2 ? CodeGenOptLevel::Default :
        CodeGenOptLevel::Aggressive;
#endif
}
#else
CodeGenOpt::Level CodeGenOptLevelFor(int optlevel)
{
#ifdef DISABLE_OPT
    return CodeGenOpt::None;
#else
    return optlevel == 0 ? CodeGenOpt::None :
        optlevel == 1 ? CodeGenOpt::Less :
        optlevel == 2 ? CodeGenOpt::Default :
        CodeGenOpt::Aggressive;
#endif
}
#endif

static auto countBasicBlocks(const Function &F) JL_NOTSAFEPOINT
{
    return std::distance(F.begin(), F.end());
}

static constexpr size_t N_optlevels = 4;

static void selectOptLevel(Module &M) JL_NOTSAFEPOINT {
    size_t opt_level = std::max(static_cast<int>(jl_options.opt_level), 0);
    do {
        if (jl_generating_output()) {
            opt_level = 0;
            break;
        }
        size_t opt_level_min = std::max(static_cast<int>(jl_options.opt_level_min), 0);
        for (auto &F : M) {
            if (!F.isDeclaration()) {
                Attribute attr = F.getFnAttribute("julia-optimization-level");
                StringRef val = attr.getValueAsString();
                if (val != "") {
                    size_t ol = (size_t)val[0] - '0';
                    if (ol < opt_level)
                        opt_level = ol;
                }
            }
        }
        if (opt_level < opt_level_min)
            opt_level = opt_level_min;
    } while (0);
    // currently -O3 is max
    opt_level = std::min(opt_level, N_optlevels - 1);
    M.addModuleFlag(Module::Warning, "julia.optlevel", opt_level);
}

static bool isJITLinkEHFrameSection(StringRef Name) JL_NOTSAFEPOINT
{
    // EH-frame sections are handled by the EH-frame registration plugin. Its
    // post-allocation graph state is not suitable for generic section range
    // walks here.
    return Name == ".eh_frame" || Name == "__eh_frame" || Name.ends_with(",__eh_frame");
}

void JLDebuginfoPlugin::notifyMaterializingWithInfo(
    orc::MaterializationResponsibility &MR, jitlink::LinkGraph &G,
    MemoryBufferRef InputObject, std::unique_ptr<jl_linker_info_t> LinkerInfo)
{
    auto NewBuffer =
        MemoryBuffer::getMemBufferCopy(InputObject.getBuffer(), G.getName());
    // Re-parsing the InputObject is wasteful, but for now, this lets us
    // reuse the existing debuginfo.cpp code. Should look into just
    // directly pulling out all the information required in a JITLink pass
    // and just keeping the required tables/DWARF sections around (perhaps
    // using the LLVM DebuggerSupportPlugin as a reference).
    auto NewObj =
        cantFail(object::ObjectFile::createObjectFile(NewBuffer->getMemBufferRef()));

    {
        std::lock_guard<std::mutex> lock{PluginMutex};
        assert(PendingObjs.count(&MR) == 0);
        PendingObjs[&MR] = std::unique_ptr<JITObjectInfo>(new JITObjectInfo{
            std::move(NewBuffer), std::move(NewObj), {}, std::move(LinkerInfo)});
    }
}

// TODO: analysis disabled since we aren't able to annotate that it was safe to lock
// std::mutex here because we asserted !jl_gcunsaferegion, so we don't need to assert jl_notsafepoint
Error JLDebuginfoPlugin::notifyEmitted(MaterializationResponsibility &MR) JL_NO_SAFEPOINT_ANALYSIS // NOLINT[julia-first-decl-annotations]
{
    {
        std::lock_guard<std::mutex> lock(PluginMutex);
        auto It = PendingObjs.find(&MR);
        if (It == PendingObjs.end())
            return Error::success();

        auto NewInfo = PendingObjs[&MR].get();
        auto getLoadAddress = [NewInfo](const StringRef &Name) -> uint64_t {
            auto result = NewInfo->SectionLoadAddresses.find(Name);
            if (result == NewInfo->SectionLoadAddresses.end()) {
                LLVM_DEBUG({
                    dbgs() << "JLDebuginfoPlugin: No load address found for section '"
                        << Name << "'\n";
                });
                return 0;
            }
            return result->second;
        };

        jl_register_jit_object(*NewInfo->Object, getLoadAddress, *NewInfo->LinkerInfo);
        PendingObjs.erase(&MR);
    }

    return Error::success();
}

Error JLDebuginfoPlugin::notifyFailed(MaterializationResponsibility &MR)
{
    std::lock_guard<std::mutex> lock(PluginMutex);
    PendingObjs.erase(&MR);
    return Error::success();
}

Error JLDebuginfoPlugin::notifyRemovingResources(JITDylib &JD, orc::ResourceKey K)
{
    return Error::success();
}

void JLDebuginfoPlugin::notifyTransferringResources(JITDylib &JD, orc::ResourceKey DstKey,
                                 orc::ResourceKey SrcKey) {}

void JLDebuginfoPlugin::modifyPassConfig(MaterializationResponsibility &MR, jitlink::LinkGraph &,
                      jitlink::PassConfiguration &PassConfig)
{
    std::lock_guard<std::mutex> lock(PluginMutex);
    auto It = PendingObjs.find(&MR);
    if (It == PendingObjs.end())
        return;

    JITObjectInfo &Info = *It->second;
    PassConfig.PostAllocationPasses.push_back([&Info, this](jitlink::LinkGraph &G) -> Error {
        std::lock_guard<std::mutex> lock(PluginMutex);
        for (const jitlink::Section &Sec : G.sections()) {
#if defined(_OS_DARWIN_)
            // Canonical JITLink section names have the segment name included, e.g.
            // "__TEXT,__text" or "__DWARF,__debug_str". There are some special internal
            // sections without a comma separator, which we can just ignore.
            size_t SepPos = Sec.getName().find(',');
            if (SepPos >= 16 || (Sec.getName().size() - (SepPos + 1) > 16)) {
                LLVM_DEBUG({
                    dbgs() << "JLDebuginfoPlugin: Ignoring section '" << Sec.getName()
                           << "'\n";
                });
                continue;
            }
            auto SecName = Sec.getName().substr(SepPos + 1);
#else
            auto SecName = Sec.getName();
#endif
            if (isJITLinkEHFrameSection(SecName))
                continue;
            if (Sec.blocks().empty())
                continue;
            // https://github.com/llvm/llvm-project/commit/118e953b18ff07d00b8f822dfbf2991e41d6d791
            Info.SectionLoadAddresses[SecName] = jitlink::SectionRange(Sec).getStart().getValue();
        }
        return Error::success();
    });
}

namespace {

using namespace llvm::orc;

class JLMemoryUsagePlugin : public ObjectLinkingLayer::Plugin {
private:
    _Atomic(size_t)* jit_bytes_size;

public:

    JLMemoryUsagePlugin(_Atomic(size_t)* jit_bytes_size)
        : jit_bytes_size(jit_bytes_size) {}

    Error notifyFailed(orc::MaterializationResponsibility &MR) override {
        return Error::success();
    }
    Error notifyRemovingResources(JITDylib &JD, orc::ResourceKey K) override
    {
        return Error::success();
    }
    void notifyTransferringResources(JITDylib &JD, orc::ResourceKey DstKey,
                                     orc::ResourceKey SrcKey) override {}

    void modifyPassConfig(orc::MaterializationResponsibility &,
                          jitlink::LinkGraph &,
                          jitlink::PassConfiguration &Config) override {
        Config.PostAllocationPasses.push_back([this](jitlink::LinkGraph &G) {
            // `G.blocks()` is exactly the union of the sections' blocks, so a
            // single pass over the (non-EH-frame) sections counts every block
            // once; `graph_size == code_size + data_size`
            size_t graph_size = 0;
            size_t code_size = 0;
            size_t data_size = 0;
            for (auto &section : G.sections()) {
                if (isJITLinkEHFrameSection(section.getName()))
                    continue;
                size_t secsize = 0;
                for (auto block : section.blocks()) {
                    secsize += block->getSize();
                }
                if ((section.getMemProt() & orc::MemProt::Exec) == orc::MemProt::None) {
                    data_size += secsize;
                } else {
                    code_size += secsize;
                }
                graph_size += secsize;
            }
            (void) code_size;
            (void) data_size;
            jl_atomic_fetch_add_relaxed(this->jit_bytes_size, graph_size);
            jl_timing_counter_inc(JL_TIMING_COUNTER_JITSize, graph_size);
            jl_timing_counter_inc(JL_TIMING_COUNTER_JITCodeSize, code_size);
            jl_timing_counter_inc(JL_TIMING_COUNTER_JITDataSize, data_size);
            return Error::success();
        });
    }
};
} // namespace anonymous

class JLMaterializationUnit : public orc::MaterializationUnit {
public:
    // Must hold LinkerMutex when calling Create and until the
    // MaterializationUnit has been added to the JITDylib.
    static JLMaterializationUnit Create(JuliaOJIT &JIT, ObjectLinkingLayer &OL,
                                        jl_emitted_output_t Out) JL_NOTSAFEPOINT
    {
        Interface I;
        auto &Syms = I.SymbolFlags;
        SmallSet<SymbolStringPtr, 2> CISyms;
        for (auto &[CI, Funcs] : Out.linker_info->ci_funcs) {
            if (Funcs.invoke)
                CISyms.insert(Funcs.invoke);
            if (Funcs.specptr)
                CISyms.insert(Funcs.specptr);

            // If we discover that another thread added this CI to the JIT
            // first, we'll still add the original symbols to CISyms (so they
            // will be filtered out of the MU Interface), but we won't register them in
            // CISymbols.
            jl_callptr_t Expected = NULL;
            CISymbolPtr Unique{};
            if (jl_atomic_cmpswap_relaxed(&CI->invoke, &Expected,
                                          jl_fptr_wait_for_compiled_addr))
                Unique = JIT.makeUniqueCIName(CI, Funcs);
            if (Unique.invoke)
                Syms[Unique.invoke] = JITSymbolFlags::Callable | JITSymbolFlags::Exported;
            if (Unique.specptr)
                Syms[Unique.specptr] = JITSymbolFlags::Callable | JITSymbolFlags::Exported;
        }

        // Tell ORC about all the other definition in this module.  When
        // linker_info contains enough information to produce the full
        // Interface, remove this.
        auto SSP = JIT.getExecutionSession().getSymbolStringPool();
        for (auto &G : Out.module->global_objects()) {
            if (G.isDeclaration() || !G.hasExternalLinkage())
                continue;
            auto Flags = JITSymbolFlags::Exported;
            if (isa<Function>(&G))
                Flags |= JITSymbolFlags::Callable;
            auto S = JIT.mangle(G.getName());
            if (CISyms.contains(S))
                continue;
            Syms[S] = Flags;
        }

        return JLMaterializationUnit{JIT, OL, std::move(Out), std::move(I)};
    }

    // During materialization: finalizers disabled, GC safe
    void materialize(std::unique_ptr<MaterializationResponsibility> R) JL_CANSAFEPOINT_ENTER_LEAVE override // NOLINT[julia-first-decl-annotations]
    {
        auto &ES = R->getExecutionSession();

        std::unique_ptr<MemoryBuffer> Obj;
        uint64_t start_time = jl_hrtime();
        {
            TimeTraceScope CompileScope("JIT Compile", Out.module->getModuleIdentifier());
            // Embeds the optlevel, CPU, and features into the module, so they form part of
            // the cache key.
            selectOptLevel(*Out.module);
            Out.module->addModuleFlag(Module::Warning, "julia.cpu",
                                      MDString::get(*Out.ctx, JIT.getTargetCPU()));
            Out.module->addModuleFlag(Module::Warning, "julia.cpu.features",
                                      MDString::get(*Out.ctx,
                                                    JIT.getTargetFeatureString()));
            auto Compile = [this]() JL_CANSAFEPOINT_ENTER_LEAVE {
                JIT.optimizeModule(*Out.module);
                return JIT.compileModule(*Out.module);
            };
            // The jl_dump_llvm_opt hook records timing and before/after IR
            // statistics from inside the optimizer, so a warm cache hit would
            // keep it from ever firing. Bypass the object cache while the
            // hook is installed.
            if (*JIT.get_dump_llvm_opt_stream())
                Obj = Compile();
            else
                Obj = JIT.OCache.get(*Out.module, Compile);
            if (!Obj) {
                R->failMaterialization();
                return;
            }
            // Save some memory
            auto Ctx = std::move(Out.ctx);
            auto M = std::move(Out.module);
        }
        uint64_t end_time = jl_hrtime();

        for (auto [CI, _] : Out.linker_info->ci_funcs) {
            JL_GC_PROMISE_ROOTED(CI);
            jl_do_dump_compile(CI, end_time - start_time);
        }

        auto G = jitlink::createLinkGraphFromObject(Obj->getMemBufferRef(),
                                                    ES.getSymbolStringPool());
        if (!G) {
#ifndef __clang_gcanalyzer__ // reportError runs an unknown callback, which cannot be annotated as safe here (but is)
            ES.reportError(G.takeError());
#endif
            R->failMaterialization();
            return;
        }

        // Causes the invoke/specptr to be published when the symbols are emitted
        SmallVector<jl_code_instance_t *> CIs;
        for (auto [CI, _] : Out.linker_info->ci_funcs)
            CIs.push_back(CI);

        jl_task_t *ct = jl_current_task;
        uint8_t gc_state = jl_gc_unsafe_enter(ct->ptls);
        JIT.publishCIs(CIs);
        jl_gc_unsafe_leave(ct->ptls, gc_state);

        if (!JIT.linkOutput(*R, Obj->getMemBufferRef(), **G, std::move(Out.linker_info)))
            return;
        OL.emit(std::move(R), std::move(*G), std::move(Obj));
    }

    StringRef getName() const override JL_NOTSAFEPOINT
    {
        return Out.module->getName();
    }

    void discard(const JITDylib &JD, const SymbolStringPtr &Name) override {}

protected:
    JLMaterializationUnit(JuliaOJIT &JIT, ObjectLinkingLayer &OL, jl_emitted_output_t Out,
                          Interface I) JL_NOTSAFEPOINT : orc::MaterializationUnit(I),
                                                         JIT(JIT),
                                                         OL(OL),
                                                         Out(std::move(Out))
    {
    }

private:
    JuliaOJIT &JIT;
    ObjectLinkingLayer &OL;
    jl_emitted_output_t Out;
};

class JLTrampolineMaterializationUnit : public orc::MaterializationUnit {
public:
    JLTrampolineMaterializationUnit(JuliaOJIT &JIT, ObjectLinkingLayer &OL,
                                    SymbolStringPtr Sym, jl_code_instance_t *CI,
                                    jl_invoke_api_t API) JL_NOTSAFEPOINT
      : orc::MaterializationUnit({{{JIT.mangle(*Sym),
                                    JITSymbolFlags::Exported | JITSymbolFlags::Callable}},
                                  {}}),
        JIT(JIT),
        OL(OL),
        Sym(Sym),
        CI(CI),
        API(API)
    {
        assert(API == JL_INVOKE_ARGS || API == JL_INVOKE_SPECSIG);
    };

    // During materialization: finalizers disabled, GC safe
    void materialize(std::unique_ptr<MaterializationResponsibility> R) JL_CANSAFEPOINT_ENTER_LEAVE override // NOLINT[julia-first-decl-annotations]
    {
        auto Ctx = std::make_unique<LLVMContext>();
        auto Mod =
            jl_create_llvm_module(*Sym, *Ctx, JIT.getDataLayout(), JIT.getTargetTriple());
        jl_codegen_output_t Out{*Mod};

        jl_task_t *ct = jl_current_task;
        uint8_t state = jl_gc_unsafe_enter(ct->ptls);
        Function *F = emit_tojlinvoke(CI, "", Out);
        if (API == JL_INVOKE_SPECSIG)
            F = emit_specsig_to_fptr1(Out, CI, F); // may safepoint
        jl_gc_unsafe_leave(ct->ptls, state);
        F->setLinkage(GlobalValue::ExternalLinkage);
        F->setName(*Sym);

        std::unique_lock Lock{JIT.LinkerMutex};
        if (auto Err = R->replace(
                std::make_unique<JLMaterializationUnit>(JLMaterializationUnit::Create(
                    JIT, OL,
                    Out.finish(std::move(Ctx), std::move(Mod),
                               *R->getExecutionSession().getSymbolStringPool()))))) {
            R->getExecutionSession().reportError(std::move(Err));
            R->failMaterialization();
        }
    }

    StringRef getName() const override JL_NOTSAFEPOINT { return *Sym; }

    void discard(const JITDylib &JD, const SymbolStringPtr &Name) override {}

private:
    JuliaOJIT &JIT;
    ObjectLinkingLayer &OL;
    SymbolStringPtr Sym;
    jl_code_instance_t *CI;
    jl_invoke_api_t API;
};

#if defined(LLVM_SHLIB)
namespace JLEHFrames {
static Error registerEHFrames(orc::ExecutorAddrRange EHFrameSection) {
    register_eh_frames(EHFrameSection.Start.toPtr<uint8_t *>(), static_cast<size_t>(EHFrameSection.size()));
    return Error::success();
}

static Error deregisterEHFrames(orc::ExecutorAddrRange EHFrameSection) {
    deregister_eh_frames(EHFrameSection.Start.toPtr<uint8_t *>(), static_cast<size_t>(EHFrameSection.size()));
    return Error::success();
}
}
#if JL_LLVM_VERSION < 210000
class JLEHFrameRegistrar final : public jitlink::EHFrameRegistrar {
public:
    Error registerEHFrames(orc::ExecutorAddrRange EHFrameSection) override {
        return JLEHFrames::registerEHFrames(EHFrameSection);
    }

    Error deregisterEHFrames(orc::ExecutorAddrRange EHFrameSection) override {
        return JLEHFrames::deregisterEHFrames(EHFrameSection);
    }
};
#else
namespace JLEHFrames {
    static auto
    registerEHFrameSectionAllocAction(const char *ArgData, size_t ArgSize) {
        using namespace llvm::orc::shared;
        return WrapperFunction<SPSError(SPSExecutorAddrRange)>::handle(
            ArgData, ArgSize, registerEHFrames)
            .release();
    }

    static auto
    deregisterEHFrameSectionAllocAction(const char *ArgData, size_t ArgSize) {
        using namespace llvm::orc::shared;
        return WrapperFunction<SPSError(SPSExecutorAddrRange)>::handle(
            ArgData, ArgSize, deregisterEHFrames)
            .release();
    }
}
#endif
#endif

// A simple forwarding class, since OrcJIT v2 needs a unique_ptr, while we have a shared_ptr
namespace {
class ForwardingMemoryManager : public RuntimeDyld::MemoryManager {
private:
    std::shared_ptr<RuntimeDyld::MemoryManager> MemMgr;

public:
    ForwardingMemoryManager(std::shared_ptr<RuntimeDyld::MemoryManager> MemMgr) : MemMgr(MemMgr) {}
    ForwardingMemoryManager(ForwardingMemoryManager &) = delete;
    virtual ~ForwardingMemoryManager() {
        assert(!MemMgr);
    }
    virtual uint8_t *allocateCodeSection(uintptr_t Size, unsigned Alignment,
                                     unsigned SectionID,
                                     StringRef SectionName) override {
        return MemMgr->allocateCodeSection(Size, Alignment, SectionID, SectionName);
    }
    virtual uint8_t *allocateDataSection(uintptr_t Size, unsigned Alignment,
                                     unsigned SectionID,
                                     StringRef SectionName,
                                     bool IsReadOnly) override {
        return MemMgr->allocateDataSection(Size, Alignment, SectionID, SectionName, IsReadOnly);
    }
    virtual void reserveAllocationSpace(uintptr_t CodeSize, Align CodeAlign,
                                        uintptr_t RODataSize, Align RODataAlign,
                                        uintptr_t RWDataSize, Align RWDataAlign) override {
        return MemMgr->reserveAllocationSpace(CodeSize, CodeAlign, RODataSize, RODataAlign, RWDataSize, RWDataAlign);
    }
    virtual bool needsToReserveAllocationSpace() override {
        return MemMgr->needsToReserveAllocationSpace();
    }
    virtual void registerEHFrames(uint8_t *Addr, uint64_t LoadAddr,
                                  size_t Size) override {
        return MemMgr->registerEHFrames(Addr, LoadAddr, Size);
    }
    virtual void deregisterEHFrames() override { /* not actually supported or allowed with this */ }
    virtual bool finalizeMemory(std::string *ErrMsg = nullptr) override {
        bool b = false;
        if (MemMgr.use_count() == 2)
            b = MemMgr->finalizeMemory(ErrMsg);
        MemMgr.reset();
        return b;
    }
    virtual void notifyObjectLoaded(RuntimeDyld &RTDyld,
                                    const object::ObjectFile &Obj) override {
        return MemMgr->notifyObjectLoaded(RTDyld, Obj);
    }
};
}  // anonymous namespace

namespace {
    static std::unique_ptr<TargetMachine> createTargetMachine() JL_NOTSAFEPOINT {
        TargetOptions options = TargetOptions();

        Triple TheTriple(sys::getProcessTriple());
        // use ELF because RuntimeDyld COFF i686 support didn't exist
        // use ELF because RuntimeDyld COFF X86_64 doesn't seem to work (fails to generate function pointers)?
        bool force_elf = TheTriple.isOSWindows();
#ifdef FORCE_ELF
        force_elf = true;
#endif
        if (force_elf) {
            TheTriple.setObjectFormat(Triple::ELF);
        }
        //options.PrintMachineCode = true; //Print machine code produced during JIT compiling
#if defined(MSAN_EMUTLS_WORKAROUND)
        options.EmulatedTLS = true;
        options.ExplicitEmulatedTLS = true;
#endif
#if defined(_CPU_RISCV64_)
        // we set these manually to avoid LLVM defaulting to soft-float
#if defined(__riscv_float_abi_double)
        options.MCOptions.ABIName = "lp64d";
#elif defined(__riscv_float_abi_single)
        options.MCOptions.ABIName = "lp64f";
#else
        options.MCOptions.ABIName = "lp64";
#endif
#endif
        auto [TheCPU, FeaturesStr] = jl_get_llvm_target(jl_options.cpu_target, jl_generating_output());
        std::string errorstr;
        const Target *TheTarget = TargetRegistry::lookupTarget("", TheTriple, errorstr);
        if (!TheTarget) {
            jl_errorf("Internal problem with process triple %s lookup: %s", TheTriple.str().c_str(), errorstr.c_str());
            return nullptr;
        }
        // Allocate a target...
        std::optional<CodeModel::Model> codemodel =
#ifdef _P64
            // Make sure we are using the large code model on 64bit
            // Let LLVM pick a default suitable for jitting on 32bit
            CodeModel::Large;
#else
            None;
#endif
        if (TheTriple.isAArch64())
            codemodel = CodeModel::Small;
#if JL_LLVM_VERSION < 200000
        else if (TheTriple.isRISCV()) {
            // RISC-V only supports large code model from LLVM 20
            // https://github.com/llvm/llvm-project/pull/70308
            codemodel = CodeModel::Medium;
        }
#endif
        // Generate simpler code for JIT
        Reloc::Model relocmodel = Reloc::Static;
        if (TheTriple.isRISCV()) {
            // until large code model is supported, use PIC for RISC-V
            // https://github.com/llvm/llvm-project/issues/106203
            relocmodel = Reloc::PIC_;
        }
        auto optlevel = CodeGenOptLevelFor(jl_options.opt_level);
        auto TM = TheTarget->createTargetMachine(
#if JL_LLVM_VERSION < 210000
                TheTriple.getTriple(),
#else
                TheTriple,
#endif
                TheCPU, FeaturesStr,
                options,
                relocmodel,
                codemodel,
                optlevel,
                true // JIT
                );
        assert(TM && "Failed to select target machine -"
                     " Is the LLVM backend for this CPU enabled?");
        fixupTM(*TM);
        return std::unique_ptr<TargetMachine>(TM);
    }

    typedef NewPM PassManager;

    orc::JITTargetMachineBuilder createJTMBFromTM(TargetMachine &TM, int optlevel) JL_NOTSAFEPOINT {
        return orc::JITTargetMachineBuilder(TM.getTargetTriple())
            .setCPU(TM.getTargetCPU().str())
            .setFeatures(TM.getTargetFeatureString())
            .setOptions(TM.Options)
            .setRelocationModel(TM.getRelocationModel())
            .setCodeModel(TM.getCodeModel())
            .setCodeGenOptLevel(CodeGenOptLevelFor(optlevel));
    }

    struct TMCreator {
        orc::JITTargetMachineBuilder JTMB;

        TMCreator(TargetMachine &TM, int optlevel) JL_NOTSAFEPOINT
            : JTMB(createJTMBFromTM(TM, optlevel)) {}
        ~TMCreator() JL_NOTSAFEPOINT = default;

        std::unique_ptr<TargetMachine> operator()() JL_NOTSAFEPOINT {
            auto TM = cantFail(JTMB.createTargetMachine());
            fixupTM(*TM);
            return TM;
        }
    };

    struct PMCreator {
        orc::JITTargetMachineBuilder JTMB;
        OptimizationLevel O;
        SmallVector<std::function<void()>, 0> &printers;
        std::mutex &llvm_printing_mutex;
        bool cache_enabled;
        PMCreator(TargetMachine &TM, int optlevel, SmallVector<std::function<void()>, 0> &printers, std::mutex &llvm_printing_mutex, bool cache_enabled) JL_NOTSAFEPOINT
            : JTMB(createJTMBFromTM(TM, optlevel)), O(getOptLevel(optlevel)), printers(printers), llvm_printing_mutex(llvm_printing_mutex), cache_enabled(cache_enabled) {}
        ~PMCreator() JL_NOTSAFEPOINT = default;

        auto operator()() JL_NOTSAFEPOINT {
            auto TM = cantFail(JTMB.createTargetMachine());
            fixupTM(*TM);
            auto options = OptimizationOptions::defaults();
            // It is unsafe to embed the specific TLS offset into the output
            // when the cache is enabled.
            options.tls_getters = cache_enabled;
            auto NPM = std::make_unique<NewPM>(std::move(TM), O, options);
            // TODO this needs to be locked, as different resource pools may add to the printer vector at the same time
            {
                std::lock_guard<std::mutex> lock(llvm_printing_mutex);
                printers.push_back([NPM = NPM.get()]() JL_NOTSAFEPOINT {
                    NPM->printTimers();
                });
            }
            return NPM;
        }
    };

    template<size_t N>
    struct sizedOptimizerT {
        sizedOptimizerT(TargetMachine &TM, SmallVector<std::function<void()>, 0> &printers, std::mutex &llvm_printing_mutex, bool cache_enabled) JL_NOTSAFEPOINT {
            for (size_t i = 0; i < N; i++) {
                PMs[i] = std::make_unique<JuliaOJIT::ResourcePool<std::unique_ptr<PassManager>>>(PMCreator(TM, i, printers, llvm_printing_mutex, cache_enabled));
            }
        }

        void operator()(Module &M) JL_NOTSAFEPOINT {
            auto PoolIdx = cast<ConstantInt>(cast<ConstantAsMetadata>(M.getModuleFlag("julia.optlevel"))->getValue())->getZExtValue();
            assert(PoolIdx < N && "Invalid optimization pool index");

            uint64_t start_time = 0;

            struct Stat {
                std::string name;
                uint64_t insts;
                uint64_t bbs;

                void dump(ios_t *stream) JL_NOTSAFEPOINT {
                    ios_printf(stream, "    \"%s\":\n", name.c_str());
                    ios_printf(stream, "        instructions: %u\n", insts);
                    ios_printf(stream, "        basicblocks: %zd\n", bbs);
                }

                Stat(Function &F) JL_NOTSAFEPOINT : name(F.getName().str()), insts(F.getInstructionCount()), bbs(countBasicBlocks(F)) {}

                ~Stat() JL_NOTSAFEPOINT = default;
            };
            SmallVector<Stat, 8> before_stats;
            {
                if (*jl_ExecutionEngine->get_dump_llvm_opt_stream()) {
                    for (auto &F : M.functions()) {
                        if (F.isDeclaration() || F.getName().starts_with(JL_SYM_INVOKE_SPECSIG)) {
                            continue;
                        }
                        // Each function is printed as a YAML object with several attributes
                        before_stats.emplace_back(F);
                    }

                    start_time = jl_hrtime();
                }
            }

            {
                JL_TIMING(LLVM_JIT, JIT_Opt);
                TimeTraceScope OptimizeScope("JIT Optimize", M.getModuleIdentifier());
                //Run the optimization
                (****PMs[PoolIdx]).run(M);
                assert(!verifyLLVMIR(M));
            }

            {
                // Print optimization statistics as a YAML object
                // Looks like:
                // -
                //   before:
                //     "foo":
                //       instructions: uint64
                //       basicblocks: uint64
                //    "bar":
                //       instructions: uint64
                //       basicblocks: uint64
                //   time_ns: uint64
                //   optlevel: int
                //   after:
                //     "foo":
                //       instructions: uint64
                //       basicblocks: uint64
                //    "bar":
                //       instructions: uint64
                //       basicblocks: uint64
                if (auto stream = *jl_ExecutionEngine->get_dump_llvm_opt_stream()) {
                    uint64_t end_time = jl_hrtime();
                    ios_printf(stream, "- \n");

                    // Print LLVM function statistic _before_ optimization
                    ios_printf(stream, "  before: \n");
                    for (auto &s : before_stats) {
                        s.dump(stream);
                    }
                    ios_printf(stream, "  time_ns: %" PRIu64 "\n", end_time - start_time);
                    ios_printf(stream, "  optlevel: %d\n", PoolIdx);

                    // Print LLVM function statistics _after_ optimization
                    ios_printf(stream, "  after: \n");
                    for (auto &F : M.functions()) {
                        if (F.isDeclaration() || F.getName().starts_with(JL_SYM_INVOKE_SPECSIG)) {
                            continue;
                        }
                        Stat(F).dump(stream);
                    }
                }
            }
            ++ModulesOptimized;
            switch (PoolIdx) {
                case 0:
                    ++OptO0;
                    break;
                case 1:
                    ++OptO1;
                    break;
                case 2:
                    ++OptO2;
                    break;
                case 3:
                    ++OptO3;
                    break;
                default:
                    // Change this if we ever gain other optlevels
                    llvm_unreachable("optlevel is between 0 and 3!");
            }
        }
    private:
        std::array<std::unique_ptr<JuliaOJIT::ResourcePool<std::unique_ptr<PassManager>>>, N> PMs;
    };

    // shim for converting a unique_ptr to a TransformFunction to a TransformFunction
    template <typename T>
    struct IRTransformRef {
        IRTransformRef(T &transform) JL_NOTSAFEPOINT : transform(transform) {}
        ~IRTransformRef() JL_NOTSAFEPOINT = default;
        OptimizerResultT operator()(orc::ThreadSafeModule TSM, orc::MaterializationResponsibility &R) JL_NOTSAFEPOINT {
            TSM.withModuleDo([&](Module &M) JL_NOTSAFEPOINT {
                transform(M, R);
            });
            return std::move(TSM);
        }
    private:
        T &transform;
    };

    template<size_t N>
    struct CompilerT : orc::IRCompileLayer::IRCompiler {

        CompilerT(orc::IRSymbolMapper::ManglingOptions MO, TargetMachine &TM) JL_NOTSAFEPOINT
            : orc::IRCompileLayer::IRCompiler(MO) {
            for (size_t i = 0; i < N; ++i) {
                TMs[i] = std::make_unique<JuliaOJIT::ResourcePool<std::unique_ptr<TargetMachine>>>(TMCreator(TM, i));
            }
        }

        Expected<std::unique_ptr<MemoryBuffer>> operator()(Module &M) override {
            JL_TIMING(LLVM_JIT, JIT_Compile);
            size_t PoolIdx;
            if (auto opt_level = M.getModuleFlag("julia.optlevel")) {
                PoolIdx = cast<ConstantInt>(cast<ConstantAsMetadata>(opt_level)->getValue())->getZExtValue();
            }
            else {
                PoolIdx = jl_options.opt_level;
            }
            assert(PoolIdx < N && "Invalid optimization level for compiler!");

            auto TM = **TMs[PoolIdx];
            if (M.getDataLayout().isDefault())
                M.setDataLayout((*TM)->createDataLayout());

            SmallVector<char, 0> ObjBufferSV;
            {
                raw_svector_ostream ObjStream(ObjBufferSV);
                legacy::PassManager PM;
                MCContext *Ctx;
                if ((*TM)->addPassesToEmitMC(PM, Ctx, ObjStream))
                    return make_error<StringError>("Target does not support MC emission",
                                                   inconvertibleErrorCode());
                PM.run(M);
            }

            // OrcJIT requires that all modules / files have unique names:
            // https://llvm.org/doxygen/namespacellvm_1_1orc.html#a1f5a1bc60c220cdccbab0f26b2a425e1
            auto name = (M.getModuleIdentifier() + "-jitted-" +
                         Twine(jl_atomic_fetch_add_relaxed(&bufcounter, 1)))
                            .str();
            return std::make_unique<SmallVectorMemoryBuffer>(std::move(ObjBufferSV), name,
                                                             false);
        }

        std::array<std::unique_ptr<JuliaOJIT::ResourcePool<std::unique_ptr<TargetMachine>>>, N> TMs;
        _Atomic(size_t) bufcounter{0};
    };
}

struct JuliaOJIT::OptimizerT {
    OptimizerT(TargetMachine &TM, SmallVector<std::function<void()>, 0> &printers, std::mutex &llvm_printing_mutex, bool cache_enabled)
        : opt(TM, printers, llvm_printing_mutex, cache_enabled) {}
    void operator()(Module &M) JL_NOTSAFEPOINT {
        opt(M);
    }
    void operator()(Module &M, orc::MaterializationResponsibility &R) JL_NOTSAFEPOINT {
        return opt(M);
    }
private:
    struct sizedOptimizerT<N_optlevels> opt;
};

struct JuliaOJIT::JITPointersT {
    JITPointersT(SharedBytesT &SharedBytes, std::mutex &Lock) JL_NOTSAFEPOINT
        : SharedBytes(SharedBytes), Lock(Lock) {}

    void operator()(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;
                GV.replaceAllUsesWith(Shared);
                GV.eraseFromParent();
            }
        }
    }
    void operator()(Module &M, orc::MaterializationResponsibility &R) JL_NOTSAFEPOINT {
        return operator()(M);
    }

private:
    // optimize memory by turning long strings into memoized copies, instead of
    // making a copy per object file of output.
    // 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,
        // as long as they have global unnamedaddr,
        // but currently we shouldn't be emitting those
        // except in imaging mode, and we don't want to
        // do this optimization there.
        if (GV.hasExternalLinkage() || !GV.hasGlobalUnnamedAddr()) {
            return nullptr;
        }
        if (!GV.hasInitializer()) {
            return nullptr;
        }
        if (!GV.isConstant()) {
            return nullptr;
        }
        auto CDS = dyn_cast<ConstantDataSequential>(GV.getInitializer());
        if (!CDS) {
            return nullptr;
        }
        StringRef Data = CDS->getRawDataValues();
        if (Data.size() < 16) {
            // Cutoff, since we don't want to intern small strings
            return nullptr;
        }
        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());
    }

    SharedBytesT &SharedBytes;
    std::mutex &Lock;
};


struct JuliaOJIT::DLSymOptimizer {
    DLSymOptimizer(bool named) JL_NOTSAFEPOINT {
        this->named = named;
#define INIT_RUNTIME_LIBRARY(libname, handle) \
        do { \
            auto libidx = (uintptr_t) libname; \
            if (libidx >= runtime_symbols.size()) { \
                runtime_symbols.resize(libidx + 1); \
            } \
            runtime_symbols[libidx].first = handle; \
        } while (0)

        INIT_RUNTIME_LIBRARY(NULL, jl_RTLD_DEFAULT_handle);
        INIT_RUNTIME_LIBRARY(JL_EXE_LIBNAME, jl_exe_handle);
        INIT_RUNTIME_LIBRARY(JL_LIBJULIA_INTERNAL_DL_LIBNAME, jl_libjulia_internal_handle);
        INIT_RUNTIME_LIBRARY(JL_LIBJULIA_DL_LIBNAME, jl_libjulia_handle);

#undef INIT_RUNTIME_LIBRARY
    }
    ~DLSymOptimizer() JL_NOTSAFEPOINT = default;

    void *lookup_symbol(void *libhandle, const char *fname) JL_NOTSAFEPOINT {
        void *addr;
        jl_dlsym(libhandle, fname, &addr, 0, 1);
        return addr;
    }

    // TODO: analysis disabled since we aren't able to annotate that it was safe to lock
    // std::mutex here because we asserted !jl_gcunsaferegion, so we don't need to assert jl_notsafepoint
    void *lookup(const char *libname, const char *fname) JL_CANSAFEPOINT_ENTER_LEAVE JL_NO_SAFEPOINT_ANALYSIS {
        StringRef lib(libname);
        StringRef f(fname);
        std::lock_guard<std::mutex> lock(symbols_mutex);
        auto uit = user_symbols.find(lib);
        if (uit == user_symbols.end()) {
            jl_task_t *ct = jl_current_task;
            int8_t gc_state = jl_gc_unsafe_enter(ct->ptls);
            void *handle = jl_get_library_(libname, 0);
            jl_gc_unsafe_leave(ct->ptls, gc_state);
            if (!handle)
                return nullptr;
            uit = user_symbols.insert(std::make_pair(lib, std::make_pair(handle, StringMap<void*>()))).first;
        }
        auto &symmap = uit->second.second;
        auto it = symmap.find(f);
        if (it != symmap.end()) {
            return it->second;
        }
        void *handle = lookup_symbol(uit->second.first, fname);
        symmap[f] = handle;
        return handle;
    }

    void *lookup(uintptr_t libidx, const char *fname) JL_NOTSAFEPOINT {
        std::lock_guard<std::mutex> lock(symbols_mutex);
        runtime_symbols.resize(std::max(runtime_symbols.size(), libidx + 1));
        auto it = runtime_symbols[libidx].second.find(fname);
        if (it != runtime_symbols[libidx].second.end()) {
            return it->second;
        }
        auto handle = lookup_symbol(runtime_symbols[libidx].first, fname);
        runtime_symbols[libidx].second[fname] = handle;
        return handle;
    }

    void operator()(Module &M) JL_CANSAFEPOINT_ENTER_LEAVE {
        for (auto &GV : M.globals()) {
            auto Name = GV.getName();
            if (Name.starts_with("jlplt") && Name.ends_with("got")) {
                auto fname = GV.getAttribute("julia.fname").getValueAsString().str();
                void *addr;
                if (GV.hasAttribute("julia.libname")) {
                    auto libname = GV.getAttribute("julia.libname").getValueAsString().str();
                    addr = lookup(libname.data(), fname.data());
                } else {
                    assert(GV.hasAttribute("julia.libidx") && "PLT entry should have either libname or libidx attribute!");
                    auto libidx = (uintptr_t)std::stoull(GV.getAttribute("julia.libidx").getValueAsString().str());
                    addr = lookup(libidx, fname.data());
                }
                if (addr) {
                    Function *Thunk = nullptr;
                    if (!GV.isDeclaration()) {
                        Thunk = cast<Function>(GV.getInitializer()->stripPointerCasts());
                        assert(++Thunk->uses().begin() == Thunk->uses().end() && "Thunk should only have one use in PLT initializer!");
                        assert(Thunk->hasLocalLinkage() && "Thunk should not have non-local linkage!");
                    }
                    else {
                        GV.setLinkage(GlobalValue::PrivateLinkage);
                    }
                    auto init = ConstantExpr::getIntToPtr(ConstantInt::get(M.getDataLayout().getIntPtrType(M.getContext()), (uintptr_t)addr), GV.getValueType());
                    if (named) {
                        auto T = GV.getValueType();
                        assert(T->isPointerTy());
                        init = GlobalAlias::create(T, 0, GlobalValue::PrivateLinkage, GV.getName() + ".jit", init, &M);
                    }
                    GV.setInitializer(init);
                    GV.setConstant(true);
                    GV.setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
                    if (Thunk) {
                        Thunk->eraseFromParent();
                    }
                }
            }
        }

        for (auto &F : M) {
            for (auto &BB : F) {
                SmallVector<Instruction *, 0> to_delete;
                for (auto &I : make_early_inc_range(BB)) {
                    auto CI = dyn_cast<CallInst>(&I);
                    if (!CI)
                        continue;
                    auto Callee = CI->getCalledFunction();
                    if (!Callee || Callee->getName() != XSTR(jl_load_and_lookup))
                        continue;
                    // Long-winded way of extracting fname without needing a second copy in an attribute
                    auto fname = cast<ConstantDataArray>(cast<GlobalVariable>(CI->getArgOperand(1)->stripPointerCasts())->getInitializer())->getAsCString();
                    auto libarg = CI->getArgOperand(0)->stripPointerCasts();
                    // Should only use in store and phi node
                    // Note that this uses the raw output of codegen,
                    // which is why we can assume this
                    assert(++++CI->use_begin() == CI->use_end());
                    void *addr;
                    if (auto GV = dyn_cast<GlobalVariable>(libarg)) {
                        // Can happen if the library is the empty string, just give up when that happens
                        if (isa<ConstantAggregateZero>(GV->getInitializer()))
                            continue;
                        auto libname = cast<ConstantDataArray>(GV->getInitializer())->getAsCString();
                        addr = lookup(libname.data(), fname.data());
                    } else {
                        // Can happen if we fail the compile time dlfind i.e when we try a symbol that doesn't exist in libc
                        if (dyn_cast<ConstantPointerNull>(libarg))
                            continue;
                        assert(cast<ConstantExpr>(libarg)->getOpcode() == Instruction::IntToPtr && "libarg should be either a global variable or a integer index!");
                        libarg = cast<ConstantExpr>(libarg)->getOperand(0);
                        auto libidx = cast<ConstantInt>(libarg)->getZExtValue();
                        addr = lookup(libidx, fname.data());
                    }
                    if (addr) {
                        auto init = ConstantExpr::getIntToPtr(ConstantInt::get(M.getDataLayout().getIntPtrType(M.getContext()), (uintptr_t)addr), CI->getType());
                        if (named) {
                            auto T = CI->getType();
                            assert(T->isPointerTy());
                            init = GlobalAlias::create(T, 0, GlobalValue::PrivateLinkage, CI->getName() + ".jit", init, &M);
                        }
                        // DCE and SimplifyCFG will kill the branching structure around
                        // the call, so we don't need to worry about removing everything
                        for (auto user : make_early_inc_range(CI->users())) {
                            if (auto SI = dyn_cast<StoreInst>(user)) {
                                to_delete.push_back(SI);
                            } else {
                                auto PHI = cast<PHINode>(user);
                                PHI->replaceAllUsesWith(init);
                                to_delete.push_back(PHI);
                            }
                        }
                        to_delete.push_back(CI);
                    }
                }
                for (auto I : to_delete) {
                    I->eraseFromParent();
                }
            }
        }
    }

    std::mutex symbols_mutex;
    StringMap<std::pair<void *, StringMap<void *>>> user_symbols;
    SmallVector<std::pair<void *, StringMap<void *>>, 0> runtime_symbols;
    bool named;
};

void optimizeDLSyms(Module &M) {
    JuliaOJIT::DLSymOptimizer(true)(M);
}

void fixupTM(TargetMachine &TM) {
    auto TheTriple = TM.getTargetTriple();
    if (jl_options.opt_level < 2) {
        // Try GlobalISel on AArch64 - it's the default in LLVM at -O0 and
        // is apparently generally faster than SelectionDAG while producing good code.
        // Use fallback mode so unsupported patterns fall back to SelectionDAG.
        // Note: Requires RemoveJuliaAddrspacesPass to run before codegen
        // because GlobalISel doesn't handle Julia's custom address spaces.
        if (TheTriple.isAArch64()) {
            TM.setGlobalISel(true);
            TM.setGlobalISelAbort(GlobalISelAbortMode::Disable);
            TM.setFastISel(false);
        }
        else if (!TheTriple.isARM() && !TheTriple.isPPC64())
            TM.setFastISel(true);
        else    // FastISel seems to be buggy Ref #13321
            TM.setFastISel(false);
    }
}

llvm::DataLayout jl_create_datalayout(TargetMachine &TM) {
    // Mark our address spaces as non-integral
    auto jl_data_layout = TM.createDataLayout();
    jl_data_layout = DataLayout(jl_data_layout.getStringRepresentation() + "-ni:10:11:12:13");
    return jl_data_layout;
}

JuliaOJIT::JuliaOJIT()
  : TM(createTargetMachine()),
    DL(jl_create_datalayout(*TM)),
    ES(cantFail(orc::SelfExecutorProcessControl::Create(nullptr, std::make_unique<::JuliaTaskDispatcher>()))),
    SessionJD(ES.createBareJITDylib("JuliaSession")),
    GlobalJD(ES.createBareJITDylib("JuliaGlobals")),
    JD(ES.createBareJITDylib("JuliaOJIT")),
    DLSymOpt(std::make_unique<DLSymOptimizer>(false)),
    OCache(),
    MemMgr(createJITLinkMemoryManager()),
    ObjectLayer(ES, *MemMgr),
    CompileLayer(ES, ObjectLayer, std::make_unique<CompilerT<N_optlevels>>(orc::irManglingOptionsFromTargetOptions(TM->Options), *TM)),
    JITPointers(std::make_unique<JITPointersT>(SharedBytes, SharedBytesMutex)),
    JITPointersLayer(ES, CompileLayer, IRTransformRef(*JITPointers)),
    Optimizers(std::make_unique<OptimizerT>(*TM, PrintLLVMTimers, llvm_printing_mutex, OCache.isEnabled())),
    OptimizeLayer(ES, JITPointersLayer, IRTransformRef(*Optimizers)),
    DebuginfoPlugin(std::make_shared<JLDebuginfoPlugin>())
{
#if JL_LLVM_VERSION < 210000
# if defined(LLVM_SHLIB)
    // When dynamically linking against LLVM, use our custom EH frame registration code
    // also used with RTDyld to inform both our and the libc copy of libunwind.
    auto ehRegistrar = std::make_unique<JLEHFrameRegistrar>();
# else
    auto ehRegistrar = std::make_unique<jitlink::InProcessEHFrameRegistrar>();
# endif
    ObjectLayer.addPlugin(std::make_unique<EHFrameRegistrationPlugin>(
        ES, std::move(ehRegistrar)));
#else
    // LLVM 21+ removed EHFrameRegistrar. Use our own plugin for custom registration
    // when dynamically linking, plus the built-in plugin for standard registration.
# if defined(LLVM_SHLIB)
    ObjectLayer.addPlugin(std::make_unique<EHFrameRegistrationPlugin>(
        ExecutorAddr::fromPtr(JLEHFrames::registerEHFrameSectionAllocAction),
        ExecutorAddr::fromPtr(JLEHFrames::deregisterEHFrameSectionAllocAction)));
# else
    ObjectLayer.addPlugin(cantFail(EHFrameRegistrationPlugin::Create(ES)));
# endif
#endif

    ObjectLayer.addPlugin(DebuginfoPlugin);
    ObjectLayer.addPlugin(std::make_unique<JLMemoryUsagePlugin>(&jit_bytes_size));

    SetVector<void*> libhandles;
    // Make sure that libjulia-internal is loaded and placed first in the
    // DynamicLibrary order so that calls to runtime intrinsics are resolved
    // to the correct library when multiple libjulia-*'s have been loaded
    // (e.g. when we `ccall` into a PackageCompiler.jl-created shared library)
    libhandles.insert(jl_libjulia_internal_handle);
    libhandles.insert(jl_libjulia_handle);
    // Make sure SectionMemoryManager::getSymbolAddressInProcess can resolve
    // symbols in the program as well. The nullptr argument to the function
    // tells DynamicLibrary to load the program); not a library.
    libhandles.insert(jl_exe_handle);
#ifdef _OS_WINDOWS_
    // Find where compiler symbols (assumed by LLVM) are linked from
    // by looking for an exported data symbol, or by typical name.
    // libgcc_s_seh-1 doesn't export any data, so we have to hard-code a name.
    // libwinpthreads-1 exports a single symbol: the pthread_key_dest table.
    libhandles.insert(jl_dlopen("libgcc_s_seh-1.dll", JL_RTLD_NOLOAD));
    libhandles.insert(jl_find_dynamic_library_by_addr((void*)&_pthread_key_dest, /* throw_err */ 1, 0));
    // Add system C libraries explicitly too.
    // Unlike posix, these aren't automatically handled by recursive search from libjulia-internal.
    libhandles.insert(jl_ntdll_handle);
    libhandles.insert(jl_kernel32_handle);
    libhandles.insert(jl_crtdll_handle);
    libhandles.insert(jl_winsock_handle);
#endif
    for (void *h : libhandles) {
        if (h == nullptr)
            continue;
        std::string ErrorStr;
        sys::DynamicLibrary dylib = sys::DynamicLibrary::addPermanentLibrary(h, &ErrorStr);
        if (!ErrorStr.empty())
            report_fatal_error(llvm::Twine("FATAL: unable to dlopen libjulia dependency\n") + ErrorStr);
        GlobalJD.addGenerator(
            std::make_unique<orc::DynamicLibrarySearchGenerator>(
            dylib,
            DL.getGlobalPrefix()));
    }

    // Resolve non-lock free atomic functions in the libatomic1 library.
    // This is the library that provides support for c11/c++11 atomic operations.
    auto TT = getTargetTriple();
    const char *const libatomic = TT.isOSLinux() || TT.isOSFreeBSD() ?
        "libatomic.so.1" : TT.isOSWindows() ?
        "libatomic-1.dll" : nullptr;
    if (libatomic) {
        static void *atomic_hdl = jl_load_dynamic_library(libatomic, JL_RTLD_LOCAL, 0);
        if (atomic_hdl != NULL) {
            std::string ErrorStr;
            sys::DynamicLibrary dylib = sys::DynamicLibrary::addPermanentLibrary(atomic_hdl, &ErrorStr);
            if (!ErrorStr.empty())
                report_fatal_error(llvm::Twine("FATAL: unable to dlopen libjulia dependency\n") + ErrorStr);
            GlobalJD.addGenerator(
              std::make_unique<orc::DynamicLibrarySearchGenerator>(
                  dylib,
                  DL.getGlobalPrefix(),
                  [&](const orc::SymbolStringPtr &S) {
                        const char *const atomic_prefix = "__atomic_";
                        return (*S).starts_with(atomic_prefix);
                  }));
        }
    }

    JD.addToLinkOrder(GlobalJD, orc::JITDylibLookupFlags::MatchExportedSymbolsOnly);
#if defined(_OS_WINDOWS_)
    // TODO: why does Windows CI hang without this?
    JD.addToLinkOrder(SessionJD, orc::JITDylibLookupFlags::MatchExportedSymbolsOnly);
#endif

    orc::SymbolAliasMap jl_crt = {
        // Float16 conversion routines
#if defined(_CPU_X86_64_) && defined(_OS_DARWIN_)
        // LLVM 16 reverted to soft-float ABI for passing half on x86_64 Darwin
        // https://github.com/llvm/llvm-project/commit/2bcf51c7f82ca7752d1bba390a2e0cb5fdd05ca9
        { mangle("__gnu_h2f_ieee"), { mangle("julia_half_to_float"),  JITSymbolFlags::Exported } },
        { mangle("__extendhfsf2"),  { mangle("julia_half_to_float"),  JITSymbolFlags::Exported } },
        { mangle("__gnu_f2h_ieee"), { mangle("julia_float_to_half"),  JITSymbolFlags::Exported } },
        { mangle("__truncsfhf2"),   { mangle("julia_float_to_half"),  JITSymbolFlags::Exported } },
        { mangle("__truncdfhf2"),   { mangle("julia_double_to_half"), JITSymbolFlags::Exported } },
#else
        { mangle("__gnu_h2f_ieee"), { mangle("julia__gnu_h2f_ieee"),  JITSymbolFlags::Exported } },
        { mangle("__extendhfsf2"),  { mangle("julia__gnu_h2f_ieee"),  JITSymbolFlags::Exported } },
        { mangle("__gnu_f2h_ieee"), { mangle("julia__gnu_f2h_ieee"),  JITSymbolFlags::Exported } },
        { mangle("__truncsfhf2"),   { mangle("julia__gnu_f2h_ieee"),  JITSymbolFlags::Exported } },
        { mangle("__truncdfhf2"),   { mangle("julia__truncdfhf2"),    JITSymbolFlags::Exported } },
#endif
        // BFloat16 conversion routines
        { mangle("__truncsfbf2"),   { mangle("julia__truncsfbf2"),    JITSymbolFlags::Exported } },
        { mangle("__truncdfbf2"),   { mangle("julia__truncdfbf2"),    JITSymbolFlags::Exported } },
    };
    cantFail(GlobalJD.define(orc::symbolAliases(jl_crt)));

#ifdef _OS_OPENBSD_
    orc::SymbolMap i128_crt;

    i128_crt[mangle("__divti3")] = JITEvaluatedSymbol::fromPointer(&__divti3, JITSymbolFlags::Exported);
    i128_crt[mangle("__modti3")] = JITEvaluatedSymbol::fromPointer(&__modti3, JITSymbolFlags::Exported);
    i128_crt[mangle("__udivti3")] = JITEvaluatedSymbol::fromPointer(&__udivti3, JITSymbolFlags::Exported);
    i128_crt[mangle("__umodti3")] = JITEvaluatedSymbol::fromPointer(&__umodti3, JITSymbolFlags::Exported);

    cantFail(GlobalJD.define(orc::absoluteSymbols(i128_crt)));
#endif

#ifdef MSAN_EMUTLS_WORKAROUND
    orc::SymbolMap msan_crt;
    msan_crt[mangle("__emutls_get_address")] = {ExecutorAddr::fromPtr(msan_workaround::getTLSAddress), JITSymbolFlags::Exported};
    msan_crt[mangle("__emutls_v.__msan_param_tls")] = {ExecutorAddr::fromPtr(
        reinterpret_cast<void *>(static_cast<uintptr_t>(msan_workaround::MSanTLS::param))), JITSymbolFlags::Exported};
    msan_crt[mangle("__emutls_v.__msan_param_origin_tls")] = {ExecutorAddr::fromPtr(
        reinterpret_cast<void *>(static_cast<uintptr_t>(msan_workaround::MSanTLS::param_origin))), JITSymbolFlags::Exported};
    msan_crt[mangle("__emutls_v.__msan_retval_tls")] = {ExecutorAddr::fromPtr(
        reinterpret_cast<void *>(static_cast<uintptr_t>(msan_workaround::MSanTLS::retval))), JITSymbolFlags::Exported};
    msan_crt[mangle("__emutls_v.__msan_retval_origin_tls")] = {ExecutorAddr::fromPtr(
        reinterpret_cast<void *>(static_cast<uintptr_t>(msan_workaround::MSanTLS::retval_origin))), JITSymbolFlags::Exported};
    msan_crt[mangle("__emutls_v.__msan_va_arg_tls")] = {ExecutorAddr::fromPtr(
        reinterpret_cast<void *>(static_cast<uintptr_t>(msan_workaround::MSanTLS::va_arg))), JITSymbolFlags::Exported};
    msan_crt[mangle("__emutls_v.__msan_va_arg_origin_tls")] = {ExecutorAddr::fromPtr(
        reinterpret_cast<void *>(static_cast<uintptr_t>(msan_workaround::MSanTLS::va_arg_origin))), JITSymbolFlags::Exported};
    msan_crt[mangle("__emutls_v.__msan_va_arg_overflow_size_tls")] = {ExecutorAddr::fromPtr(
        reinterpret_cast<void *>(static_cast<uintptr_t>(msan_workaround::MSanTLS::va_arg_overflow_size))), JITSymbolFlags::Exported};
    msan_crt[mangle("__emutls_v.__msan_origin_tls")] = {ExecutorAddr::fromPtr(
        reinterpret_cast<void *>(static_cast<uintptr_t>(msan_workaround::MSanTLS::origin))), JITSymbolFlags::Exported};
    cantFail(GlobalJD.define(orc::absoluteSymbols(msan_crt)));
#endif
#ifdef _COMPILER_ASAN_ENABLED_
    // this is a hack to work around a bad assertion:
    //   /workspace/srcdir/llvm-project/llvm/lib/ExecutionEngine/Orc/Core.cpp:3028: llvm::Error llvm::orc::ExecutionSession::OL_notifyResolved(llvm::orc::MaterializationResponsibility&, const SymbolMap&): Assertion `(KV.second.getFlags() & ~JITSymbolFlags::Common) == (I->second & ~JITSymbolFlags::Common) && "Resolving symbol with incorrect flags"' failed.
    static int64_t jl___asan_globals_registered;
    orc::SymbolMap asan_crt;
    asan_crt[mangle("___asan_globals_registered")] = {ExecutorAddr::fromPtr(&jl___asan_globals_registered), JITSymbolFlags::Common | JITSymbolFlags::Exported};
    cantFail(JD.define(orc::absoluteSymbols(asan_crt)));
#endif

#if defined(_COMPILER_GCC_) && __GNUC__ < 14
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
#endif
    SessionJD.addGenerator(
      cantFail(orc::DynamicLibrarySearchGenerator::GetForCurrentProcess(
        DL.getGlobalPrefix())));
#if defined(_COMPILER_GCC_) && __GNUC__ < 14
#pragma GCC diagnostic pop
#endif

    if (jl_is_timing_trace) {
        PrintLLVMTimers.push_back([]() JL_NOTSAFEPOINT {
            if (timeTraceProfilerEnabled()) {
                StringRef FileName = jl_timing_trace_file.empty() ?
                    StringRef("julia_time_trace.json") : StringRef(jl_timing_trace_file);
                if (auto E = timeTraceProfilerWrite(FileName, "")) {
                    handleAllErrors(std::move(E), [](const StringError &SE) JL_NOTSAFEPOINT {
                        errs() << SE.getMessage() << "\n";
                    });
                }
                timeTraceProfilerCleanup();
            }
        });
    }
}

JuliaOJIT::~JuliaOJIT() = default;

ThreadSafeContext JuliaOJIT::makeContext()
{
    auto ctx = std::make_unique<LLVMContext>();
    return orc::ThreadSafeContext(std::move(ctx));
}

orc::SymbolStringPtr JuliaOJIT::mangle(StringRef Name)
{
    std::string MangleName = getMangledName(Name);
    return ES.intern(MangleName);
}

void JuliaOJIT::addGlobalMapping(StringRef Name, uint64_t Addr)
{
    cantFail(JD.define(orc::absoluteSymbols({{mangle(Name), {ExecutorAddr::fromPtr((void*)Addr), JITSymbolFlags::Exported}}})));
}


#ifdef ENABLE_TIMINGS
static void timing_print_module_names(jl_timing_block_t *block,
                                      Module &M) JL_NOTSAFEPOINT
{
    for (auto &f : M) {
        if (!f.isDeclaration()) {
            jl_timing_puts(block, f.getName().str().c_str());
        }
    }
}
#endif

void JuliaOJIT::addOutput(jl_emitted_output_t O)
{
    JL_TIMING(LLVM_JIT, JIT_Total);
    ++ModulesAdded;
#ifdef ENABLE_TIMINGS
    timing_print_module_names(JL_TIMING_DEFAULT_BLOCK, *O.module);
#endif
    std::unique_lock Lock{LinkerMutex};
    auto MU = std::make_unique<JLMaterializationUnit>(
        JLMaterializationUnit::Create(*this, ObjectLayer, std::move(O)));
    ExitOnError check{"Failed to add objectfile to JIT!"};
    check(JD.define(MU, JD.getDefaultResourceTracker()));
}

orc::JITDylib& JuliaOJIT::createJITDylib(StringRef NamePrefix)
{
    // Create a new JITDylib with unique name
    std::string dylib_name = (NamePrefix + "_" + Twine(jl_atomic_fetch_add_relaxed(&jitcounter, 1))).str();
    JITDylib &NewJD = ES.createBareJITDylib(dylib_name);
    NewJD.addToLinkOrder(GlobalJD, orc::JITDylibLookupFlags::MatchExportedSymbolsOnly);
    NewJD.addToLinkOrder(SessionJD, orc::JITDylibLookupFlags::MatchExportedSymbolsOnly);
    return NewJD;
}

Error JuliaOJIT::addExternalModule(orc::JITDylib &JD, orc::ThreadSafeModule TSM, bool ShouldOptimize)
{
    if (auto Err = TSM.withModuleDo([&](Module &M) JL_NOTSAFEPOINT -> Error {
            auto PostOptDL = TM->createDataLayout(); // excludes ni tags stripped by optzns
            if (M.getDataLayout().isDefault())
                M.setDataLayout(PostOptDL);
            if (M.getDataLayout() != PostOptDL)
                return make_error<StringError>(
                    "Added modules have incompatible data layouts: " +
                    M.getDataLayout().getStringRepresentation() + " (module) vs " +
                    PostOptDL.getStringRepresentation() + " (jit)",
                inconvertibleErrorCode());
            // OrcJIT requires that all modules / files have unique names:
            M.setModuleIdentifier((M.getModuleIdentifier() + Twine("-") + Twine(jl_atomic_fetch_add_relaxed(&jitcounter, 1))).str());
            return Error::success();
        }))
        return Err;

    //if (ShouldOptimize)
    //    return OptimizeLayer.add(JD, std::move(TSM));
    return CompileLayer.add(JD.getDefaultResourceTracker(), std::move(TSM));
}

Error JuliaOJIT::addObjectFile(orc::JITDylib &JD, std::unique_ptr<MemoryBuffer> Obj) {
    assert(Obj && "Can not add null object");
    return ObjectLayer.add(JD.getDefaultResourceTracker(), std::move(Obj));
}

SmallVector<uint64_t> JuliaOJIT::findSymbols(ArrayRef<StringRef> Names)
{
    // assert(MemMgr.use_count() == 1); (true single-threaded, but slightly race-y to assert it with concurrent threads)
    DenseMap<orc::NonOwningSymbolStringPtr, size_t> Unmangled;
    orc::SymbolLookupSet Exports;
    for (StringRef Name : Names) {
        auto Mangled = ES.intern(getMangledName(Name));
        Unmangled[NonOwningSymbolStringPtr(Mangled)] = Unmangled.size();
        Exports.add(std::move(Mangled));
    }
    SymbolMap Syms = cantFail(::safelookup(ES, orc::makeJITDylibSearchOrder(ArrayRef(&JD)), std::move(Exports)));
    SmallVector<uint64_t> Addrs(Names.size());
    for (auto it : Syms) {
        Addrs[Unmangled.at(orc::NonOwningSymbolStringPtr(it.first))] = it.second.getAddress().getValue();
    }
    return Addrs;
}

Expected<ExecutorSymbolDef> JuliaOJIT::findJDSymbol(JITDylib &JD, StringRef Name, bool ExternalJDOnly)
{
    orc::JITDylib *SearchOrders[2] = {&JD, &GlobalJD};
    ArrayRef<orc::JITDylib*> SearchOrder = ArrayRef<orc::JITDylib*>(&SearchOrders[0], ExternalJDOnly ? 1 : 2);
    auto Sym = ::safelookup(ES, SearchOrder, getMangledName(Name));
    return Sym;
}

uint64_t JuliaOJIT::getGlobalValueAddress(StringRef Name)
{
    auto addr = findJDSymbol(JD, Name, true);
    if (!addr) {
        consumeError(addr.takeError());
        return 0;
    }
    return addr->getAddress().getValue();
}

uint64_t JuliaOJIT::getFunctionAddress(StringRef Name)
{
    auto addr = findJDSymbol(JD, Name, true);
    if (!addr) {
        consumeError(addr.takeError());
        return 0;
    }
    return addr->getAddress().getValue();
}

void JuliaOJIT::publishCIs(ArrayRef<jl_code_instance_t *> CIs, bool Wait)
{
    orc::SymbolLookupSet Exports;
    {
        std::unique_lock Lock{LinkerMutex};
        for (auto CI : CIs) {
            auto It = CISymbols.find(CI);
            if (It == CISymbols.end())
                return;
            auto CISym = It->second;
            if (CISym.invoke)
                Exports.add(CISym.invoke);
            if (CISym.specptr)
                Exports.add(CISym.specptr);
        }
    }

    JuliaTaskDispatcher::future<void> F;
    auto Callback = [this, CIs = SmallVector<jl_code_instance_t *, 1>(CIs),
                     P = Wait ? std::optional(F.get_promise()) :
                                std::nullopt](Expected<SymbolMap> SymsE) JL_NOTSAFEPOINT {
        std::unique_lock Lock{LinkerMutex};
        if (!SymsE) {
            errs() << "Internal error: Lookup failed: " << SymsE.takeError() << "\n";
            if (P)
                P->set_value();
            abort();
        }
        auto Syms = std::move(*SymsE);
        for (auto [i, CI] : llvm::enumerate(CIs)) {
            jl_codeinst_funcs_t<void *> Addrs{};
            const auto &S = CISymbols.at(CIs[i]);
            Addrs.invoke_api = S.invoke_api;
            if (S.invoke)
                Addrs.invoke = (void *)Syms.at(S.invoke).getAddress().getValue();
            if (S.specptr)
                Addrs.specptr = (void *)Syms.at(S.specptr).getAddress().getValue();
            jl_publish_compiled_ci(CI, Addrs);
        }
        if (P)
            P->set_value();
    };
    ES.lookup(LookupKind::Static, orc::makeJITDylibSearchOrder(ArrayRef(&JD)),
              std::move(Exports), SymbolState::Ready, std::move(Callback),
              NoDependenciesToRegister);

    if (Wait)
        F.get(static_cast<JuliaTaskDispatcher &>(
            ES.getExecutorProcessControl().getDispatcher()));
}

void JuliaOJIT::registerCI(jl_code_instance_t *CI)
{
#ifndef JL_NDEBUG
    std::unique_lock Lock{LinkerMutex};
    assert(!CISymbols.contains(CI));
#endif
}

void JuliaOJIT::unregisterCI(jl_code_instance_t *CI)
{
    std::unique_lock Lock{LinkerMutex};
    CISymbols.erase(CI);
}

#define addAbsoluteToMap(map,name) \
    (map[mangle(#name)] = {ExecutorAddr::fromPtr(&name), JITSymbolFlags::Exported | JITSymbolFlags::Callable}, orc::ExecutorAddr::fromPtr(&name))

void JuliaOJIT::enableJITDebuggingSupport()
{
    orc::SymbolMap GDBFunctions;
    addAbsoluteToMap(GDBFunctions,llvm_orc_registerJITLoaderGDBAllocAction);
    auto registerJITLoaderGDBWrapper = addAbsoluteToMap(GDBFunctions,llvm_orc_registerJITLoaderGDBWrapper);
    cantFail(JD.define(orc::absoluteSymbols(GDBFunctions)));
    (void)registerJITLoaderGDBWrapper;
    if (TM->getTargetTriple().isOSBinFormatMachO()) {
        auto RegisterSym = cantFail(
            safelookup(ES, {&JD}, ES.intern("_llvm_orc_registerJITLoaderGDBAllocAction")));
        ObjectLayer.addPlugin(
            std::make_unique<GDBJITDebugInfoRegistrationPlugin>(RegisterSym.getAddress()));
    }
#ifndef _COMPILER_ASAN_ENABLED_ // TODO: Fix duplicated sections spam #51794
    else if (TM->getTargetTriple().isOSBinFormatELF()) {
        //EPCDebugObjectRegistrar doesn't take a JITDylib, so we have to directly provide the call address
        ObjectLayer.addPlugin(std::make_unique<orc::DebugObjectManagerPlugin>(ES, std::make_unique<orc::EPCDebugObjectRegistrar>(ES, registerJITLoaderGDBWrapper)));
    }
#endif
}

void JuliaOJIT::enableIntelJITEventListener()
{
#if JL_LLVM_VERSION >= 190000
    if (TM->getTargetTriple().isOSBinFormatELF()) {
        orc::SymbolMap VTuneFunctions;
        auto RegisterImplAddr = addAbsoluteToMap(VTuneFunctions,llvm_orc_registerVTuneImpl);
        auto UnregisterImplAddr = addAbsoluteToMap(VTuneFunctions,llvm_orc_unregisterVTuneImpl);
        ObjectLayer.addPlugin(cantFail(DebugInfoPreservationPlugin::Create()));
        //ObjectLayer.addPlugin(cantFail(VTuneSupportPlugin::Create(ES.getExecutorProcessControl(),
        //                           JD, /*EmitDebugInfo=*/true,
        //                           /*TestMode=*/false)));
        bool EmitDebugInfo = true;
        ObjectLayer.addPlugin(std::make_unique<VTuneSupportPlugin>(
            ES.getExecutorProcessControl(), RegisterImplAddr, UnregisterImplAddr, EmitDebugInfo));
    }
#endif
}

void JuliaOJIT::enableOProfileJITEventListener()
{
    // implement when available in LLVM
}

void JuliaOJIT::enablePerfJITEventListener()
{
#if JL_LLVM_VERSION >= 180000
    if (TM->getTargetTriple().isOSBinFormatELF()) {
        orc::SymbolMap PerfFunctions;
        auto StartAddr = addAbsoluteToMap(PerfFunctions,llvm_orc_registerJITLoaderPerfStart);
        auto EndAddr = addAbsoluteToMap(PerfFunctions,llvm_orc_registerJITLoaderPerfEnd);
        auto ImplAddr = addAbsoluteToMap(PerfFunctions,llvm_orc_registerJITLoaderPerfImpl);
        cantFail(JD.define(orc::absoluteSymbols(PerfFunctions)));
        ObjectLayer.addPlugin(cantFail(DebugInfoPreservationPlugin::Create()));
        //ObjectLayer.addPlugin(cantFail(PerfSupportPlugin::Create(
        //    ES.getExecutorProcessControl(), *JD, true, true)));
        bool EmitDebugInfo = true, EmitUnwindInfo = true;
        ObjectLayer.addPlugin(std::make_unique<PerfSupportPlugin>(
            ES.getExecutorProcessControl(), StartAddr, EndAddr, ImplAddr, EmitDebugInfo, EmitUnwindInfo));
    }
#endif
}

const DataLayout& JuliaOJIT::getDataLayout() const
{
    return DL;
}

std::string JuliaOJIT::getMangledName(StringRef Name)
{
    SmallString<128> FullName;
    Mangler::getNameWithPrefix(FullName, Name, DL);
    return FullName.str().str();
}

std::string JuliaOJIT::getMangledName(const GlobalValue *GV)
{
    return getMangledName(GV->getName());
}

CISymbolPtr JuliaOJIT::makeUniqueCIName(jl_code_instance_t *CI, const CISymbolPtr &Funcs)
{
    orc::SymbolStringPtr wrapper, specialized;
    if (Funcs.invoke)
        wrapper = ES.intern(Names(*Funcs.invoke, "#"));
    if (Funcs.specptr)
        specialized = ES.intern(Names(*Funcs.specptr, "#"));
    CISymbolPtr Ret{Funcs.invoke_api, wrapper, specialized};
    if (CISymbols.contains(CI)) {
        errs() << "Attempting to register CodeInstance that was already added to JIT!\n";
        abort();
    }
    CISymbols[CI] = Ret;
    return Ret;
}

// Convenience function to get a map from string pool symbols to symbols in this
// LinkGraph that participate in linking (defined and external).
static DenseMap<orc::SymbolStringPtr, jitlink::Symbol *>
linkGraphSymbols(jitlink::LinkGraph &G) JL_NOTSAFEPOINT
{
    DenseMap<orc::SymbolStringPtr, jitlink::Symbol *> Syms;
    auto AddSyms = [&](auto Symbols) JL_NOTSAFEPOINT {
        for (auto S : Symbols)
            if (S->getName())
                Syms[S->getName()] = S;
    };
    AddSyms(G.defined_symbols());
    AddSyms(G.external_symbols());
    return Syms;
}

static jitlink::Symbol *
findLinkGraphSymbolByName(jitlink::LinkGraph &G,
                          const orc::SymbolStringPtr &Name) JL_NOTSAFEPOINT
{
    if (auto *Sym = G.findDefinedSymbolByName(Name))
        return Sym;
    if (auto *Sym = G.findExternalSymbolByName(Name))
        return Sym;
    if (auto *Sym = G.findAbsoluteSymbolByName(Name))
        return Sym;
    return nullptr;
}

static void retargetLinkGraphEdges(jitlink::LinkGraph &G, jitlink::Symbol &From,
                                   jitlink::Symbol &To) JL_NOTSAFEPOINT
{
    struct RetargetEdgeVisitor {
        jitlink::Symbol &From;
        jitlink::Symbol &To;

        bool visitEdge(jitlink::LinkGraph &, jitlink::Block *,
                       jitlink::Edge &Edge) JL_NOTSAFEPOINT
        {
            if (&Edge.getTarget() != &From)
                return false;
            Edge.setTarget(To);
            return true;
        }
    };
    jitlink::visitExistingEdges(G, RetargetEdgeVisitor{From, To});
}

static jitlink::Symbol *
makeAnonymousLinkGraphSymbol(jitlink::LinkGraph &G,
                             jitlink::Symbol &Sym) JL_NOTSAFEPOINT
{
    assert(Sym.isDefined());
    auto &Anon = G.addAnonymousSymbol(Sym.getBlock(), Sym.getOffset(),
                                      Sym.getSize(), Sym.isCallable(),
                                      Sym.isLive());
    Anon.setTargetFlags(Sym.getTargetFlags());
    retargetLinkGraphEdges(G, Sym, Anon);
    G.removeDefinedSymbol(Sym);
    return &Anon;
}

static jitlink::Symbol *
renameLinkGraphSymbol(jitlink::LinkGraph &G, jitlink::Symbol &Sym,
                      const orc::SymbolStringPtr &Name) JL_NOTSAFEPOINT
{
    if (Sym.getName() == Name)
        return &Sym;
    if (!Sym.isExternal()) {
        Sym.setName(Name);
        return &Sym;
    }
    // External symbols are keyed by name inside LinkGraph, so retarget rather
    // than renaming in place and leaving the external symbol map stale.
    auto *Dest = findLinkGraphSymbolByName(G, Name);
    if (!Dest) {
        Dest = &G.addExternalSymbol(Name, Sym.getSize(), Sym.isWeaklyReferenced());
        Dest->setCallable(Sym.isCallable());
        Dest->setTargetFlags(Sym.getTargetFlags());
    }
    else if (Dest->isExternal()) {
        Dest->setWeaklyReferenced(Dest->isWeaklyReferenced() && Sym.isWeaklyReferenced());
        Dest->setCallable(Dest->isCallable() || Sym.isCallable());
        Dest->setTargetFlags(Dest->getTargetFlags() | Sym.getTargetFlags());
    }
    retargetLinkGraphEdges(G, Sym, *Dest);
    return Dest;
}

bool JuliaOJIT::linkOutput(orc::MaterializationResponsibility &MR, MemoryBufferRef ObjBuf,
                           jitlink::LinkGraph &G, std::unique_ptr<jl_linker_info_t> Info)
{
    std::unique_lock Lock{LinkerMutex};

    auto Syms = linkGraphSymbols(G);

    // Rename the defined CI functions.
    auto RenameDef = [&](const SymbolStringPtr &Orig, const SymbolStringPtr &Dest)
                             JL_NOTSAFEPOINT {
        auto It = Syms.find(Orig);
        assert(It != Syms.end());
        It->second = renameLinkGraphSymbol(G, *It->second, Dest);
    };
    SmallSet<SymbolStringPtr, 2> OwnedSyms;
    for (auto &KV : MR.getSymbols())
        OwnedSyms.insert(KV.first);
    for (auto &[CI, Funcs] : Info->ci_funcs) {
        auto &S = CISymbols.at(CI);
        if (Funcs.invoke)
            RenameDef(Funcs.invoke, S.invoke);
        if (Funcs.specptr)
            RenameDef(Funcs.specptr, S.specptr);
    }

    // Pre-pass: find CI equivalents, and build EquivMap for use in the main
    // pass to memoize findCompatibleCI.
    DenseMap<jl_code_instance_t *, jl_code_instance_t *> EquivMap;
    for (auto &[Call, T] : Info->call_targets) {
        auto [CI, API] = Call;
        JL_GC_PROMISE_ROOTED(CI);
        if (!Syms.contains(T))
            continue;
        if (EquivMap.contains(CI))
            continue;
        if (!jl_mi_cache_has_ci(jl_get_ci_mi(CI), CI)) {
            jl_code_instance_t *Equiv = findCompatibleCI(CI);
            if (Equiv != CI)
                EquivMap[CI] = Equiv;
        }
    }

    // Rename referenced CIs in the workqueue.
    for (auto &[Call, T] : Info->call_targets) {
        auto [CI, API] = Call;
        auto It = Syms.find(T);
        if (It == Syms.end())
            continue;
        if (!It->second->isExternal()) {
            // Non-primary call target bodies are local copies. Keep them out
            // of ORC's public symbols instead of publishing duplicate CI
            // definitions under the target's global name.
            if (!OwnedSyms.contains(It->second->getName()))
                It->second = makeAnonymousLinkGraphSymbol(G, *It->second);
            continue;
        }
        JL_GC_PROMISE_ROOTED(CI);
        auto Dest = linkCallTarget(MR, CI, API, EquivMap);
        if (!Dest)
            return false;
        if (auto *DestSym = findLinkGraphSymbolByName(G, Dest);
            DestSym && !DestSym->isExternal() && !OwnedSyms.contains(Dest))
            makeAnonymousLinkGraphSymbol(G, *DestSym);
        It->second = renameLinkGraphSymbol(G, *It->second, Dest);
    }
    SmallSet<SymbolStringPtr, 0> KnownCISyms;
    for (auto &KV : CISymbols) {
        auto &S = KV.second;
        if (S.invoke)
            KnownCISyms.insert(S.invoke);
        if (S.specptr)
            KnownCISyms.insert(S.specptr);
    }
    SmallVector<jitlink::Symbol *, 0> DefinedSyms;
    for (auto *Sym : G.defined_symbols())
        DefinedSyms.push_back(Sym);
    // Another thread may have claimed a CI after this module was emitted but
    // before this materialization unit registered its interface. Any body that
    // remains in this graph for that CI must not be exported here.
    for (auto *Sym : DefinedSyms) {
        if (Sym->hasName() && KnownCISyms.contains(Sym->getName()) &&
            !OwnedSyms.contains(Sym->getName()))
            makeAnonymousLinkGraphSymbol(G, *Sym);
    }

    // Rename globals and add mappings
    // TODO: don't leak when we have a way to GC code
    void **Ptrs;
    #ifdef __clang_analyzer__
    // hide this "leak" from clang-sa analysis
    extern void** make_new_pointers(size_t) JL_NOTSAFEPOINT;
    Ptrs = make_new_pointers(Info->global_targets.size());
    #else
    Ptrs = new void *[Info->global_targets.size()];
    #endif
    size_t i = 0;
    orc::SymbolMap GlobalSyms;
    for (auto &[Addr, Orig] : Info->global_targets) {
        auto Sym = ES.intern(Names(*Orig, "#"));
        auto It = Syms.find(Orig);
        if (It == Syms.end())
            continue;
        It->second = renameLinkGraphSymbol(G, *It->second, Sym);
        Ptrs[i] = Addr;
        GlobalSyms[Sym] = {ExecutorAddr::fromPtr(Ptrs + i), JITSymbolFlags::Exported};
        ++i;
        ++LinkedGlobals;
    }
    cantFail(JD.define(orc::absoluteSymbols(std::move(GlobalSyms))));

    DebuginfoPlugin->notifyMaterializingWithInfo(MR, G, ObjBuf, std::move(Info));
    return true;
}

// Must hold LinkerMutex.
orc::SymbolStringPtr JuliaOJIT::linkCallTarget(orc::MaterializationResponsibility &MR,
                                               jl_code_instance_t *CI, jl_invoke_api_t API,
                                               const DenseMap<jl_code_instance_t *, jl_code_instance_t *> &EquivMap)
{
    {
        auto It = EquivMap.find(CI);
        if (It != EquivMap.end())
            CI = It->second;
    }
    {
        auto It = CISymbols.find(CI);
        if (It != CISymbols.end() && It->second.invoke_api == API)
            return It->second.specptr;
    }

    CISymbolPtr *Sym = linkCISymbol(CI);

    // If !Sym: The target CI was not compiled, so generate a tojlinvoke
    // trampoline that will cause it to be compiled.
    // TODO: replace this with a GOT/PLT mechanism that avoids the jl_invoke
    // after it has been compiled.
    //
    // We also generate a tojlinvoke to handle args1 -> specsig.
    CISymbolPtr Trampoline;
    if (!Sym || Sym->invoke_api != API) {
        auto TSym = ES.intern(Names("tojlinvoke#", name_from_method_instance(jl_get_ci_mi(CI)), "#"));
        Trampoline.specptr = mangle(*TSym);
        Trampoline.invoke_api = API;
        Sym = &Trampoline;
        auto Err = JD.define(std::make_unique<JLTrampolineMaterializationUnit>(
            *this, ObjectLayer, TSym, CI, API));
        if (Err) {
#ifndef __clang_analyzer__ // reportError calls an arbitrary function, which the static analyzer thinks might be a safepoint
            MR.getExecutionSession().reportError(std::move(Err));
#endif
            MR.failMaterialization();
            return {};
        }
    }

    assert(Sym->invoke_api == API);
    return Sym->specptr;
}

jl_code_instance_t *JuliaOJIT::findCompatibleCI(jl_code_instance_t *ci)
{
    // add_codeinsts_to_jit! may have added an equivalent ci to the JIT, but
    // the invoke itself won't be updated.
    auto mi = jl_get_ci_mi(ci);
    for (auto ci2 = jl_atomic_load_relaxed(&mi->cache); ci2;
         ci2 = jl_atomic_load_relaxed(&ci2->next)) {
        if (ci2 != ci && jl_is_ci_equiv(ci, ci2, 0) &&
            (CISymbols.contains(ci2) || jl_atomic_load_relaxed(&ci2->invoke))) {
            return ci2;
        }
    }
    return ci;
}

CISymbolPtr *JuliaOJIT::linkCISymbol(jl_code_instance_t *CI)
{
    uint8_t Flags;
    jl_callptr_t Invoke;
    void *SpecPtr;

    // Tell the analyzer no safepoint is possible with waitcompile = 0
#ifdef __clang_safetyanalysis__
    #define jl_read_codeinst_invoke jl_read_codeinst_invoke_nosafepoint
#endif
    void jl_read_codeinst_invoke(jl_code_instance_t *, uint8_t *, jl_callptr_t *, void **, int) JL_NOTSAFEPOINT;
    jl_read_codeinst_invoke(CI, &Flags, &Invoke, &SpecPtr, 0);
#ifdef __clang_safetyanalysis__
    #undef jl_read_codeinst_invoke
#endif

    if (!(Flags & JL_CI_FLAGS_INVOKE_MATCHES_SPECPTR))
        return nullptr;

    // TODO: Remove specialized CI flag entirely?
    jl_invoke_api_t API = jl_callptr_invoke_api(Invoke);
    assert((API == JL_INVOKE_SPECSIG) == bool(Flags & JL_CI_FLAGS_SPECPTR_SPECIALIZED));

    orc::SymbolStringPtr InvokeSym;
    SymbolMap Symbols;
    const char *Name = jl_symbol_name(jl_get_ci_mi(CI)->def.method->name);

    auto SpecSym = mangle(Names(jl_symbol_prefix(JL_SYMBOL_SPECPTR_IMG, API), "#", Name));
    Symbols[SpecSym] = {ExecutorAddr::fromPtr(SpecPtr), JITSymbolFlags::Exported};
    if (API == JL_INVOKE_SPECSIG) {
        InvokeSym = mangle(Names(jl_symbol_prefix(JL_SYMBOL_INVOKE_IMG, API), "#", Name));
        Symbols[InvokeSym] = {ExecutorAddr::fromPtr(Invoke), JITSymbolFlags::Exported};
    }
    cantFail(JD.define(orc::absoluteSymbols(Symbols)));

    auto &CISym = CISymbols[CI] = {API, InvokeSym, SpecSym};
    return &CISym;
}

void JuliaOJIT::optimizeModule(Module &M)
{
    if (!OCache.isEnabled())
        optimizeDLSyms(M);
    (*Optimizers)(M);
    if (!OCache.isEnabled())
        (*JITPointers)(M);
    // Windows needs some inline asm to help
    // build unwind tables, if they have any functions to decorate
    if (!M.functions().empty())
        decorate_module(M);
}

std::unique_ptr<MemoryBuffer> JuliaOJIT::compileModule(Module &M)
{
    // Treat this as if one of the passes might contain a safepoint
    // even though that shouldn't be the case and might be unwise
    Expected<std::unique_ptr<MemoryBuffer>> Obj = CompileLayer.getCompiler()(M);
    if (!Obj) {
#ifndef __clang_analyzer__ // reportError calls an arbitrary function, which the static analyzer thinks might be a safepoint
        ES.reportError(Obj.takeError());
#endif
        errs() << "Failed to add module to JIT!\n";
        errs() << "Dumping failing module\n" << M << "\n";
        return {};
    }
    return std::move(*Obj);
}

size_t JuliaOJIT::getTotalBytes() const
{
    auto bytes = jl_atomic_load_relaxed(&jit_bytes_size);
    return bytes;
}

void JuliaOJIT::addBytes(size_t bytes)
{
    jl_atomic_fetch_add_relaxed(&jit_bytes_size, bytes);
}

void JuliaOJIT::printTimers()
{
    for (auto &printer : PrintLLVMTimers) {
        printer();
    }
    reportAndResetTimings();
}

void JuliaOJIT::optimizeDLSyms(Module &M) {
    (*DLSymOpt)(M);
}

void JuliaOJIT::shutdown()
{
    OCache.shutdown();
}

JuliaOJIT *jl_ExecutionEngine;

//TargetMachine pass-through methods

std::unique_ptr<TargetMachine> JuliaOJIT::cloneTargetMachine() const
{
    auto NewTM = std::unique_ptr<TargetMachine>(getTarget()
        .createTargetMachine(
#if JL_LLVM_VERSION < 210000
            getTargetTriple().str(),
#else
            getTargetTriple(),
#endif
            getTargetCPU(),
            getTargetFeatureString(),
            getTargetOptions(),
            TM->getRelocationModel(),
            TM->getCodeModel(),
            TM->getOptLevel()));
    fixupTM(*NewTM);
    return NewTM;
}

const Triple& JuliaOJIT::getTargetTriple() const {
    return TM->getTargetTriple();
}
StringRef JuliaOJIT::getTargetFeatureString() const {
    return TM->getTargetFeatureString();
}
StringRef JuliaOJIT::getTargetCPU() const {
    return TM->getTargetCPU();
}
const TargetOptions &JuliaOJIT::getTargetOptions() const {
    return TM->Options;
}
const Target &JuliaOJIT::getTarget() const {
    return TM->getTarget();
}
TargetIRAnalysis JuliaOJIT::getTargetIRAnalysis() const {
    return TM->getTargetIRAnalysis();
}

static void decorate_module(Module &M) {
    auto TT = Triple(M.getTargetTriple());
    if (TT.isOSWindows() && TT.getArch() == Triple::x86_64) {
        // Add special values used by debuginfo to build the UnwindData table registration for Win64
        // This used to be GV, but with https://reviews.llvm.org/D100944 we no longer can emit GV into `.text`
        // and with JITLink it became difficult to change the content afterwards, but we
        // would prefer that this simple content wasn't recompiled in every single module,
        // so we emit the necessary PLT trampoline as inline assembly.
        // This is somewhat duplicated with the .pdata section, but we haven't been able to
        // use that yet due to relocation issues.
#define ASM_USES_ELF // use ELF or COFF syntax based on FORCE_ELF
        StringRef inline_asm(
    ".section"
#if JL_LLVM_VERSION >= 180000
        " .ltext,\"ax\",@progbits\n"
#else
        " .text\n"
#endif
    ".globl __julia_personality\n"
    "\n"
#ifdef ASM_USES_ELF
    ".type __UnwindData,@object\n"
#else
    ".def __UnwindData\n"
    ".scl 2\n"
    ".type 0\n"
    ".endef\n"
#endif
    ".p2align        2, 0x90\n"
    "__UnwindData:\n"
    "  .byte 0x09;\n" // version info, UNW_FLAG_EHANDLER
    "  .byte 4;\n"    // size of prolog (bytes)
    "  .byte 2;\n"    // count of unwind codes (slots)
    "  .byte 0x05;\n" // frame register (rbp) = rsp
    "  .byte 4;\n"    // second instruction
    "  .byte 0x03;\n" // mov RBP, RSP
    "  .byte 1;\n"    // first instruction
    "  .byte 0x50;\n" // push RBP
    "  .int __catchjmp - "
#if JL_LLVM_VERSION >= 180000
    ".ltext;\n" // Section-relative offset (if using COFF and JITLink, this can be relative to __ImageBase instead, though then we could possibly use pdata/xdata directly then)
#else
    ".text;\n"
#endif
    ".size __UnwindData, 12\n"
    "\n"
#ifdef ASM_USES_ELF
    ".type __catchjmp,@function\n"
#else
    ".def __catchjmp\n"
    ".scl 2\n"
    ".type 32\n"
    ".endef\n"
#endif
    ".p2align        2, 0x90\n"
    "__catchjmp:\n"
    "  movabsq $__julia_personality, %rax\n"
    "  jmpq *%rax\n"
    ".size __catchjmp, . - __catchjmp\n"
    "\n");
        M.appendModuleInlineAsm(inline_asm);
    }
#undef ASM_USES_ELF
}

extern "C" JL_DLLEXPORT_CODEGEN
void jl_decorate_llvm_module_impl(LLVMModuleRef m) JL_NOTSAFEPOINT
{
    decorate_module(*unwrap(m));
}

// helper function for adding a DLLImport (dlsym) address to the execution engine
void add_named_global(StringRef name, void *addr)
{
    jl_ExecutionEngine->addGlobalMapping(name, (uint64_t)(uintptr_t)addr);
}

extern "C" JL_DLLEXPORT_CODEGEN
size_t jl_jit_total_bytes_impl(void)
{
    return jl_ExecutionEngine->getTotalBytes();
}

extern "C" JL_DLLEXPORT_CODEGEN
const char *jl_objcache_disabled_notice_impl(void) JL_CANSAFEPOINT_ENTER_LEAVE
{
    return jl_ExecutionEngine->objCacheDisabledNotice();
}

// API for adding bytes to record being owned by the JIT
void jl_jit_add_bytes(size_t bytes)
{
    jl_ExecutionEngine->addBytes(bytes);
}


extern "C" JL_DLLEXPORT_CODEGEN
void jl_jit_register_ci_impl(jl_code_instance_t *ci)
{
    jl_ExecutionEngine->registerCI(ci);
}

extern "C" JL_DLLEXPORT_CODEGEN
void jl_jit_unregister_ci_impl(jl_code_instance_t *ci)
{
    jl_ExecutionEngine->unregisterCI(ci);
}