diff --git a/base/sysimg.jl b/base/sysimg.jl
index 11a239a5f93bd..28def08e155e2 100644
--- a/base/sysimg.jl
+++ b/base/sysimg.jl
@@ -422,6 +422,24 @@ end
 INCLUDE_STATE = 3 # include = include_from_node1
 include("precompile.jl")
 
+@noinline function test_clone_f(a)
+    s = zero(eltype(a))
+    @inbounds @simd for i in 1:length(a)
+        s += a[i]
+    end
+    return s
+end
+
+@noinline function test_clone_g(a, n)
+    s = zero(eltype(a))
+    for i in 1:n
+        s += test_clone_f(a)
+    end
+    return s
+end
+
+test_clone_g(Float64[], 1)
+
 end # baremodule Base
 
 using Base
diff --git a/src/Makefile b/src/Makefile
index 8d18c458985bb..a2c26af230d8f 100644
--- a/src/Makefile
+++ b/src/Makefile
@@ -54,7 +54,7 @@ endif
 LLVMLINK :=
 
 ifeq ($(JULIACODEGEN),LLVM)
-SRCS += codegen jitlayers disasm debuginfo llvm-simdloop llvm-ptls llvm-gcroot llvm-lower-handlers cgmemmgr
+SRCS += codegen jitlayers disasm debuginfo llvm-simdloop llvm-ptls llvm-gcroot llvm-lower-handlers llvm-mv cgmemmgr
 FLAGS += -I$(shell $(LLVM_CONFIG_HOST) --includedir)
 LLVM_LIBS := all
 ifeq ($(USE_POLLY),1)
diff --git a/src/dump.c b/src/dump.c
index 244b3a12e5b3a..ca58c07e36789 100644
--- a/src/dump.c
+++ b/src/dump.c
@@ -223,6 +223,11 @@ JL_DLLEXPORT int jl_running_on_valgrind(void)
     return RUNNING_ON_VALGRIND;
 }
 
+STATIC_INLINE uint64_t i32_to_i64(uint64_t hi, uint64_t lo)
+{
+    return (hi << 32) | lo;
+}
+
 static void jl_load_sysimg_so(void)
 {
 #ifndef _OS_WINDOWS_
@@ -244,6 +249,24 @@ static void jl_load_sysimg_so(void)
         *sysimg_gvars[tls_offset_idx - 1] =
             (jl_value_t*)(uintptr_t)(jl_tls_offset == -1 ? 0 : jl_tls_offset);
 #endif
+#if defined(_CPU_X86_64_) || defined(_CPU_X86_)
+        // WIP
+        typedef void (*dispatch_t)(size_t, size_t*, void***, size_t**);
+        dispatch_t dispatchf = (dispatch_t)jl_dlsym_e(jl_sysimg_handle,
+                                                      "jl_dispatch_sysimg_fvars");
+        if (dispatchf) {
+            size_t nfunc = 0;
+            void **fptrs = NULL;
+            size_t *fidxs = NULL;
+            dispatchf(jl_test_cpu_feature(JL_X86_avx2), &nfunc, &fptrs, &fidxs);
+            if (nfunc && fptrs && fidxs) {
+                for (size_t i = 0; i < nfunc; i++) {
+                    size_t fi = fidxs[i];
+                    sysimg_fvars[fi] = fptrs[i];
+                }
+            }
+        }
+#endif
 
 #ifdef _OS_WINDOWS_
         sysimage_base = (intptr_t)jl_sysimg_handle;
diff --git a/src/jitlayers.cpp b/src/jitlayers.cpp
index 9256b935a2f22..d014c1f2a6f5b 100644
--- a/src/jitlayers.cpp
+++ b/src/jitlayers.cpp
@@ -177,6 +177,7 @@ void addOptimizationPasses(PassManager *PM)
     // Let the InstCombine pass remove the unnecessary load of
     // safepoint address first
     PM->add(createLowerPTLSPass(imaging_mode));
+    PM->add(createJuliaMVPass());
     PM->add(createSROAPass());                 // Break up aggregate allocas
 #ifndef INSTCOMBINE_BUG
     PM->add(createInstructionCombiningPass()); // Cleanup for scalarrepl.
@@ -1094,7 +1095,7 @@ static void jl_gen_llvm_globaldata(llvm::Module *mod, ValueToValueMapTy &VMap,
     ArrayType *fvars_type = ArrayType::get(T_pvoidfunc, jl_sysimg_fvars.size());
     addComdat(new GlobalVariable(*mod,
                                  fvars_type,
-                                 true,
+                                 false,
                                  GlobalVariable::ExternalLinkage,
                                  MapValue(ConstantArray::get(fvars_type, ArrayRef<Constant*>(jl_sysimg_fvars)), VMap),
                                  "jl_sysimg_fvars"));
diff --git a/src/jitlayers.h b/src/jitlayers.h
index 92a318c2e3fbf..6a05192bc38b1 100644
--- a/src/jitlayers.h
+++ b/src/jitlayers.h
@@ -249,6 +249,7 @@ JL_DLLEXPORT extern LLVMContext &jl_LLVMContext;
 Pass *createLowerPTLSPass(bool imaging_mode);
 Pass *createLowerGCFramePass();
 Pass *createLowerExcHandlersPass();
+Pass *createJuliaMVPass();
 // Whether the Function is an llvm or julia intrinsic.
 static inline bool isIntrinsicFunction(Function *F)
 {
diff --git a/src/llvm-mv.cpp b/src/llvm-mv.cpp
new file mode 100644
index 0000000000000..fa6e764d60904
--- /dev/null
+++ b/src/llvm-mv.cpp
@@ -0,0 +1,272 @@
+// This file is a part of Julia. License is MIT: https://julialang.org/license
+
+// Function multi-versioning
+#define DEBUG_TYPE "julia_mv"
+#undef DEBUG
+
+// LLVM pass to clone function for different archs
+
+#include "llvm-version.h"
+#include "support/dtypes.h"
+
+#include <llvm/Pass.h>
+#include <llvm/IR/Module.h>
+#include <llvm/IR/Function.h>
+#include <llvm/IR/Instructions.h>
+#include <llvm/IR/Constants.h>
+#include <llvm/IR/LLVMContext.h>
+#include <llvm/Analysis/LoopInfo.h>
+#if JL_LLVM_VERSION >= 30700
+#include <llvm/IR/LegacyPassManager.h>
+#else
+#include <llvm/PassManager.h>
+#endif
+#include <llvm/IR/MDBuilder.h>
+#include <llvm/IR/IRBuilder.h>
+#include <llvm/Transforms/Utils/Cloning.h>
+#include "fix_llvm_assert.h"
+
+#include "julia.h"
+#include "julia_internal.h"
+
+#include <unordered_map>
+#include <vector>
+
+using namespace llvm;
+
+extern std::pair<MDNode*,MDNode*> tbaa_make_child(const char *name, MDNode *parent=nullptr, bool isConstant=false);
+extern "C" void jl_dump_llvm_value(void *v);
+
+namespace {
+
+struct JuliaMV: public ModulePass {
+    static char ID;
+    JuliaMV()
+        : ModulePass(ID)
+    {}
+
+private:
+    bool runOnModule(Module &M) override;
+    void getAnalysisUsage(AnalysisUsage &AU) const override
+    {
+        AU.addRequired<LoopInfoWrapperPass>();
+        AU.setPreservesAll();
+    }
+    bool shouldClone(Function &F);
+    bool checkUses(Function &F, Constant *fary);
+    bool checkUses(Function &F, Constant *V, Constant *fary, bool &inFVars);
+    bool checkConstantUse(Function &F, Constant *V, Constant *fary, bool &inFVars);
+};
+
+bool JuliaMV::shouldClone(Function &F)
+{
+    if (F.empty())
+        return false;
+    auto &LI = getAnalysis<LoopInfoWrapperPass>(F).getLoopInfo();
+    if (!LI.empty())
+        return true;
+    for (auto &bb: F) {
+        for (auto &I: bb) {
+            if (auto call = dyn_cast<CallInst>(&I)) {
+                if (auto callee = call->getCalledFunction()) {
+                    auto name = callee->getName();
+                    if (name.startswith("llvm.muladd.") || name.startswith("llvm.fma.")) {
+                        return true;
+                    }
+                }
+            }
+        }
+    }
+    return false;
+}
+
+bool JuliaMV::checkUses(Function &F, Constant *fary)
+{
+    bool inFVars = false;
+    bool res = checkUses(F, &F, fary, inFVars);
+    return res && inFVars;
+}
+
+bool JuliaMV::checkConstantUse(Function &F, Constant *V, Constant *fary, bool &inFVars)
+{
+    if (V == fary) {
+        inFVars = true;
+        return true;
+    }
+    if (auto cexpr = dyn_cast<ConstantExpr>(V)) {
+        if (cexpr->getOpcode() == Instruction::BitCast) {
+            return checkUses(F, V, fary, inFVars);
+        }
+    }
+    return false;
+}
+
+bool JuliaMV::checkUses(Function &F, Constant *V, Constant *fary, bool &inFVars)
+{
+    for (auto *user: V->users()) {
+        if (isa<Instruction>(user))
+            continue;
+        auto *C = dyn_cast<Constant>(user);
+        if (!C || !checkConstantUse(F, C, fary, inFVars)) {
+            return false;
+        }
+    }
+    return true;
+}
+
+static Function *getFunction(Value *v)
+{
+    if (auto f = dyn_cast<Function>(v))
+        return f;
+    if (auto c = dyn_cast<ConstantExpr>(v)) {
+        if (c->getOpcode() == Instruction::BitCast) {
+            return getFunction(c->getOperand(0));
+        }
+    }
+    return nullptr;
+}
+
+static void addFeatures(Function *F)
+{
+    auto attr = F->getFnAttribute("target-features");
+    std::string feature =
+        "+avx2,+avx,+fma,+popcnt,+sse,+sse2,+sse3,+sse4.1,+sse4.2,+ssse3";
+    if (attr.isStringAttribute()) {
+        feature += ",";
+        feature += attr.getValueAsString();
+    }
+    F->addFnAttr("target-features", feature);
+}
+
+bool JuliaMV::runOnModule(Module &M)
+{
+    MDNode *tbaa_const = tbaa_make_child("jtbaa_const", nullptr, true).first;
+    GlobalVariable *fvars = M.getGlobalVariable("jl_sysimg_fvars");
+    // This makes sure this only runs during sysimg generation
+    if (!fvars || !fvars->hasInitializer())
+        return true;
+    auto *fary = dyn_cast<ConstantArray>(fvars->getInitializer());
+    if (!fary)
+        return true;
+    LLVMContext &ctx = M.getContext();
+    ValueToValueMapTy VMap;
+    for (auto &F: M) {
+        if (shouldClone(F) && checkUses(F, fary)) {
+            Function *NF = Function::Create(cast<FunctionType>(F.getValueType()),
+                                            F.getLinkage(), F.getName() + ".avx2", &M);
+            NF->copyAttributesFrom(&F);
+            VMap[&F] = NF;
+        }
+    }
+    std::unordered_map<Function*,size_t> idx_map;
+    size_t nf = fary->getNumOperands();
+    for (size_t i = 0; i < nf; i++) {
+        if (Function *ele = getFunction(fary->getOperand(i))) {
+            auto it = VMap.find(ele);
+            if (it != VMap.end()) {
+                idx_map[ele] = i;
+            }
+        }
+    }
+    for (auto I: idx_map) {
+        auto oldF = I.first;
+        auto newF = cast<Function>(VMap[oldF]);
+        Function::arg_iterator DestI = newF->arg_begin();
+        for (Function::const_arg_iterator J = oldF->arg_begin(); J != oldF->arg_end(); ++J) {
+            DestI->setName(J->getName());
+            VMap[&*J] = &*DestI++;
+        }
+        SmallVector<ReturnInst*,8> Returns;
+        CloneFunctionInto(newF, oldF, VMap, false, Returns);
+        addFeatures(newF);
+    }
+    std::vector<Constant*> ptrs;
+    std::vector<Constant*> idxs;
+    auto T_void = Type::getVoidTy(ctx);
+    auto T_pvoidfunc = FunctionType::get(T_void, false)->getPointerTo();
+    auto T_size = (sizeof(size_t) == 8 ? Type::getInt64Ty(ctx) : Type::getInt32Ty(ctx));
+    for (auto I: idx_map) {
+        auto oldF = I.first;
+        auto idx = I.second;
+        auto newF = cast<Function>(VMap[oldF]);
+        ptrs.push_back(ConstantExpr::getBitCast(newF, T_pvoidfunc));
+        auto offset = ConstantInt::get(T_size, idx);
+        idxs.push_back(offset);
+        for (auto user: oldF->users()) {
+            auto inst = dyn_cast<Instruction>(user);
+            if (!inst)
+                continue;
+            auto encloseF = inst->getParent()->getParent();
+            if (VMap.find(encloseF) != VMap.end())
+                continue;
+            Value *slot = ConstantExpr::getBitCast(fvars, T_pvoidfunc->getPointerTo());
+            slot = GetElementPtrInst::Create(T_pvoidfunc, slot, {offset}, "", inst);
+            Instruction *ptr = new LoadInst(slot, "", inst);
+            ptr->setMetadata(llvm::LLVMContext::MD_tbaa, tbaa_const);
+            ptr = new BitCastInst(ptr, oldF->getType(), "", inst);
+            inst->replaceUsesOfWith(oldF, ptr);
+        }
+    }
+    ArrayType *fvars_type = ArrayType::get(T_pvoidfunc, ptrs.size());
+    auto ptr_gv = new GlobalVariable(M, fvars_type, true, GlobalVariable::InternalLinkage,
+                                     ConstantArray::get(fvars_type, ptrs));
+    ArrayType *idxs_type = ArrayType::get(T_size, idxs.size());
+    auto idx_gv = new GlobalVariable(M, idxs_type, true, GlobalVariable::InternalLinkage,
+                                     ConstantArray::get(idxs_type, idxs));
+
+    // TODO
+    std::vector<Type*> dispatch_args(0);
+    dispatch_args.push_back(T_size); // hasavx2
+    dispatch_args.push_back(T_size->getPointerTo());
+    dispatch_args.push_back(fvars_type->getPointerTo()->getPointerTo());
+    dispatch_args.push_back(idxs_type->getPointerTo()->getPointerTo());
+    Function *dispatchF = Function::Create(FunctionType::get(T_void, dispatch_args, false),
+                                           Function::ExternalLinkage,
+                                           "jl_dispatch_sysimg_fvars", &M);
+    IRBuilder<> builder(ctx);
+    BasicBlock *b0 = BasicBlock::Create(ctx, "top", dispatchF);
+    builder.SetInsertPoint(b0);
+    DebugLoc noDbg;
+    builder.SetCurrentDebugLocation(noDbg);
+
+    std::vector<Argument*> args;
+    for (auto &arg: dispatchF->args())
+        args.push_back(&arg);
+
+    auto sz_arg = args[1];
+    auto fvars_arg = args[2];
+    auto idxs_arg = args[3];
+
+    builder.CreateStore(ConstantInt::get(T_size, ptrs.size()), sz_arg);
+
+    BasicBlock *match_bb = BasicBlock::Create(ctx, "match");
+    BasicBlock *fail_bb = BasicBlock::Create(ctx, "fail");
+    builder.CreateCondBr(builder.CreateICmpEQ(args[0], ConstantInt::get(T_size, 1)),
+                         match_bb, fail_bb);
+
+    dispatchF->getBasicBlockList().push_back(match_bb);
+    builder.SetInsertPoint(match_bb);
+    builder.CreateStore(ptr_gv, fvars_arg);
+    builder.CreateStore(idx_gv, idxs_arg);
+    builder.CreateRetVoid();
+
+    dispatchF->getBasicBlockList().push_back(fail_bb);
+    builder.SetInsertPoint(fail_bb);
+    builder.CreateStore(ConstantPointerNull::get(fvars_type->getPointerTo()), fvars_arg);
+    builder.CreateStore(ConstantPointerNull::get(idxs_type->getPointerTo()), idxs_arg);
+    builder.CreateRetVoid();
+
+    return true;
+}
+
+char JuliaMV::ID = 0;
+static RegisterPass<JuliaMV> X("JuliaMV", "JuliaMV Pass",
+                               false /* Only looks at CFG */,
+                               false /* Analysis Pass */);
+
+}
+
+Pass *createJuliaMVPass()
+{
+    return new JuliaMV();
+}