diff --git a/src/Makefile b/src/Makefile
index eb4b5d1266df9d..7972840a9656ed 100644
--- a/src/Makefile
+++ b/src/Makefile
@@ -56,7 +56,7 @@ RUNTIME_SRCS += jitlayers aotcompile debuginfo disasm llvm-simdloop llvm-muladd
 	llvm-final-gc-lowering llvm-pass-helpers llvm-late-gc-lowering \
 	llvm-lower-handlers llvm-gc-invariant-verifier llvm-propagate-addrspaces \
 	llvm-multiversioning llvm-alloc-opt cgmemmgr llvm-api llvm-remove-addrspaces \
-	llvm-remove-ni
+	llvm-remove-ni llvm-julia-licm
 FLAGS += -I$(shell $(LLVM_CONFIG_HOST) --includedir)
 LLVM_LIBS := all
 ifeq ($(USE_POLLY),1)
diff --git a/src/aotcompile.cpp b/src/aotcompile.cpp
index 9d90a691557013..1a6f27bd318c19 100644
--- a/src/aotcompile.cpp
+++ b/src/aotcompile.cpp
@@ -694,8 +694,10 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level,
     // LoopRotate strips metadata from terminator, so run LowerSIMD afterwards
     PM->add(createLowerSimdLoopPass()); // Annotate loop marked with "loopinfo" as LLVM parallel loop
     PM->add(createLICMPass());
+    PM->add(createJuliaLICMPass());
     PM->add(createLoopUnswitchPass());
     PM->add(createLICMPass());
+    PM->add(createJuliaLICMPass());
     // Subsequent passes not stripping metadata from terminator
     PM->add(createInstSimplifyLegacyPass());
     PM->add(createIndVarSimplifyPass());
diff --git a/src/jitlayers.h b/src/jitlayers.h
index 7305736f6de649..8dd45c1f939f52 100644
--- a/src/jitlayers.h
+++ b/src/jitlayers.h
@@ -238,6 +238,7 @@ Pass *createGCInvariantVerifierPass(bool Strong);
 Pass *createPropagateJuliaAddrspaces();
 Pass *createRemoveJuliaAddrspacesPass();
 Pass *createRemoveNIPass();
+Pass *createJuliaLICMPass();
 Pass *createMultiVersioningPass();
 Pass *createAllocOptPass();
 // Whether the Function is an llvm or julia intrinsic.
diff --git a/src/llvm-julia-licm.cpp b/src/llvm-julia-licm.cpp
new file mode 100644
index 00000000000000..c7ef9d0468e09a
--- /dev/null
+++ b/src/llvm-julia-licm.cpp
@@ -0,0 +1,134 @@
+// This file is a part of Julia. License is MIT: https://julialang.org/license
+
+#include "llvm-version.h"
+
+#include <llvm/Analysis/LoopInfo.h>
+#include <llvm/Analysis/LoopPass.h>
+#include "llvm/Analysis/LoopIterator.h"
+#include <llvm/IR/Dominators.h>
+#include <llvm/IR/LegacyPassManager.h>
+#include <llvm/Transforms/Utils/LoopUtils.h>
+
+#include "llvm-pass-helpers.h"
+
+#define DEBUG_TYPE "julia-licm"
+
+using namespace llvm;
+
+/*
+ * Julia LICM pass.
+ * This takes care of some julia intrinsics that is safe to move around/out of loops but
+ * can't be handled by LLVM's LICM. These intrinsics can be moved outside of
+ * loop context as well but it is inside a loop where they matter the most.
+ */
+
+namespace {
+
+struct JuliaLICMPass : public LoopPass, public JuliaPassContext {
+    static char ID;
+    JuliaLICMPass() : LoopPass(ID) {};
+
+    bool runOnLoop(Loop *L, LPPassManager &LPM) override
+    {
+        // Get the preheader block to move instructions into,
+        // required to run this pass.
+        BasicBlock *preheader = L->getLoopPreheader();
+        if (!preheader)
+            return false;
+        BasicBlock *header = L->getHeader();
+        initFunctions(*header->getModule());
+        // Also require `gc_preserve_begin_func` whereas
+        // `gc_preserve_end_func` is optional since the input to
+        // `gc_preserve_end_func` must be from `gc_preserve_begin_func`.
+        if (!gc_preserve_begin_func)
+            return false;
+        auto LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
+        auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
+
+        // Lazy initialization of exit blocks insertion points.
+        bool exit_pts_init = false;
+        SmallVector<Instruction*, 8> _exit_pts;
+        auto get_exit_pts = [&] () -> ArrayRef<Instruction*> {
+            if (!exit_pts_init) {
+                exit_pts_init = true;
+                SmallVector<BasicBlock*, 8> exit_bbs;
+                L->getUniqueExitBlocks(exit_bbs);
+                for (BasicBlock *bb: exit_bbs) {
+                    _exit_pts.push_back(&*bb->getFirstInsertionPt());
+                }
+            }
+            return _exit_pts;
+        };
+
+        bool changed = false;
+        // Scan in the right order so that we'll hoist the `begin`
+        // before we consider sinking `end`.
+        LoopBlocksRPO worklist(L);
+        worklist.perform(LI);
+        for (auto *bb : worklist) {
+            for (BasicBlock::iterator II = bb->begin(), E = bb->end(); II != E;) {
+                auto call = dyn_cast<CallInst>(&*II++);
+                if (!call)
+                    continue;
+                auto callee = call->getCalledValue();
+                assert(callee);
+                // It is always legal to extend the preserve period
+                // so we only need to make sure it is legal to move/clone
+                // the calls.
+                // If all the input arguments dominates the whole loop we can
+                // hoist the `begin` and if a `begin` dominates the loop the
+                // corresponding `end` can be moved to the loop exit.
+                if (callee == gc_preserve_begin_func) {
+                    bool canhoist = true;
+                    for (Use &U : call->arg_operands()) {
+                        // Check if all arguments are generated outside the loop
+                        auto origin = dyn_cast<Instruction>(U.get());
+                        if (!origin)
+                            continue;
+                        if (!DT->properlyDominates(origin->getParent(), header)) {
+                            canhoist = false;
+                            break;
+                        }
+                    }
+                    if (!canhoist)
+                        continue;
+                    call->moveBefore(preheader->getTerminator());
+                    changed = true;
+                }
+                else if (callee == gc_preserve_end_func) {
+                    auto begin = cast<Instruction>(call->getArgOperand(0));
+                    if (!DT->properlyDominates(begin->getParent(), header))
+                        continue;
+                    changed = true;
+                    auto exit_pts = get_exit_pts();
+                    if (exit_pts.empty()) {
+                        call->eraseFromParent();
+                        continue;
+                    }
+                    call->moveBefore(exit_pts[0]);
+                    for (unsigned i = 1; i < exit_pts.size(); i++) {
+                        // Clone exit
+                        CallInst::Create(call, {}, exit_pts[i]);
+                    }
+                }
+            }
+        }
+        return changed;
+    }
+
+    void getAnalysisUsage(AnalysisUsage &AU) const override
+    {
+        getLoopAnalysisUsage(AU);
+    }
+};
+
+char JuliaLICMPass::ID = 0;
+static RegisterPass<JuliaLICMPass>
+        Y("JuliaLICM", "LICM for julia specific intrinsics.",
+          false, false);
+}
+
+Pass *createJuliaLICMPass()
+{
+    return new JuliaLICMPass();
+}