diff --git a/src/cgutils.cpp b/src/cgutils.cpp
index 8ce84acb30901..b96d054651785 100644
--- a/src/cgutils.cpp
+++ b/src/cgutils.cpp
@@ -999,8 +999,7 @@ static LoadInst *emit_nthptr_recast(jl_codectx_t &ctx, Value *v, Value *idx, MDN
     return load;
 }
 
-static Value *boxed(jl_codectx_t &ctx, const jl_cgval_t &v,  bool is_promotable=false);
-
+static Value *boxed(jl_codectx_t &ctx, const jl_cgval_t &v,  bool is_promotable=false, jl_count_box_type log_reason=JL_DONT_LOG_BOX);
 static Value *emit_typeof(jl_codectx_t &ctx, Value *v, bool maybenull);
 
 static jl_cgval_t emit_typeof(jl_codectx_t &ctx, const jl_cgval_t &p, bool maybenull)
@@ -2672,6 +2671,7 @@ static Value *emit_arraylen_prim(jl_codectx_t &ctx, const jl_cgval_t &tinfo)
         }
     }
     ++EmittedArraylen;
+    // TODO(PR): HERE?
     Value *t = boxed(ctx, tinfo);
     Value *addr = ctx.builder.CreateStructGEP(ctx.types().T_jlarray,
             emit_bitcast(ctx, decay_derived(ctx, t), ctx.types().T_pjlarray),
@@ -2723,6 +2723,7 @@ static Value *emit_arrayptr_internal(jl_codectx_t &ctx, const jl_cgval_t &tinfo,
 
 static Value *emit_arrayptr(jl_codectx_t &ctx, const jl_cgval_t &tinfo, bool isboxed = false)
 {
+    // TODO(PR): here?
     Value *t = boxed(ctx, tinfo);
     return emit_arrayptr_internal(ctx, tinfo, decay_derived(ctx, t), AddressSpace::Loaded, isboxed);
 }
@@ -2809,6 +2810,7 @@ static Value *emit_array_nd_index(
         const jl_cgval_t *argv, size_t nidxs, jl_value_t *inbounds)
 {
     ++EmittedArrayNdIndex;
+    // TODO(PR): here?
     Value *a = boxed(ctx, ainfo);
     Value *i = Constant::getNullValue(getSizeTy(ctx.builder.getContext()));
     Value *stride = ConstantInt::get(getSizeTy(ctx.builder.getContext()), 1);
@@ -3303,7 +3305,7 @@ static void recursively_adjust_ptr_type(llvm::Value *Val, unsigned FromAS, unsig
 // dynamically-typed value is required (e.g. argument to unknown function).
 // if it's already a pointer it's left alone.
 // Returns ctx.types().T_prjlvalue
-static Value *boxed(jl_codectx_t &ctx, const jl_cgval_t &vinfo, bool is_promotable)
+static Value *boxed(jl_codectx_t &ctx, const jl_cgval_t &vinfo, bool is_promotable, jl_count_box_type log_reason)
 {
     jl_value_t *jt = vinfo.typ;
     if (jt == jl_bottom_type || jt == NULL)
@@ -3320,6 +3322,17 @@ static Value *boxed(jl_codectx_t &ctx, const jl_cgval_t &vinfo, bool is_promotab
         return vinfo.V;
     }
 
+    if (log_reason != JL_DONT_LOG_BOX) {
+        Function *F;
+        if (log_reason == JL_COUNT_BOX_INPUTS) {
+            F = prepare_call(jl_log_box_func_INPUTS);
+        } else {
+            F = prepare_call(jl_log_box_func_RETURNS);
+        }
+        ctx.builder.CreateCall(F,
+            literal_pointer_val(ctx, (jl_value_t*)jt));
+    }
+
     Value *box;
     if (vinfo.TIndex) {
         SmallBitVector skip_none;
diff --git a/src/codegen.cpp b/src/codegen.cpp
index a4773acb3fbea..d2f6d1cce9a47 100644
--- a/src/codegen.cpp
+++ b/src/codegen.cpp
@@ -889,6 +889,20 @@ static const auto jlegalx_func = new JuliaFunction{
             AttributeSet(),
             None); },
 };
+static const auto jl_log_box_func_INPUTS = new JuliaFunction{
+   XSTR(jl_nhd_log_box_input),
+   [](LLVMContext &C) {
+       return FunctionType::get(getVoidTy(C), {getInt8PtrTy(C)}, false);
+   },
+   nullptr,
+};
+static const auto jl_log_box_func_RETURNS = new JuliaFunction{
+   XSTR(jl_nhd_log_box_return),
+   [](LLVMContext &C) {
+       return FunctionType::get(getVoidTy(C), {getInt8PtrTy(C)}, false);
+   },
+   nullptr,
+};
 static const auto jl_alloc_obj_func = new JuliaFunction{
     "julia.gc_alloc_obj",
     [](LLVMContext &C) {
@@ -4072,7 +4086,7 @@ static CallInst *emit_jlcall(jl_codectx_t &ctx, FunctionCallee theFptr, Value *t
     if (theF)
         theArgs.push_back(theF);
     for (size_t i = 0; i < nargs; i++) {
-        Value *arg = boxed(ctx, argv[i]);
+        Value *arg = boxed(ctx, argv[i], false, JL_COUNT_BOX_INPUTS); // log the boxes
         theArgs.push_back(arg);
     }
     CallInst *result = ctx.builder.CreateCall(TheTrampoline, theArgs);
@@ -4144,6 +4158,9 @@ static jl_cgval_t emit_call_specfun_other(jl_codectx_t &ctx, jl_method_instance_
         jl_cgval_t arg = argv[i];
         if (isboxed) {
             assert(at == ctx.types().T_prjlvalue && et == ctx.types().T_prjlvalue);
+            // Note(NHD): I don't think this needs a log: this is only boxing it if it's actually
+            // needed by the callee, for already resolved static dispatch, which cannot be
+            // avoided.
             argvals[idx] = boxed(ctx, arg);
         }
         else if (et->isAggregateType()) {
@@ -6723,7 +6740,8 @@ static Function *gen_invoke_wrapper(jl_method_instance_t *lam, jl_value_t *jlret
             break;
         }
     }
-    ctx.builder.CreateRet(boxed(ctx, retval));
+    // NOTE(PR): log this box for the return values
+    ctx.builder.CreateRet(boxed(ctx, retval, false, JL_COUNT_BOX_RETURNS));
     return w;
 }
 
@@ -7618,7 +7636,11 @@ static jl_llvm_functions_t
                 }
             }
             else {
-                Value *argp = boxed(ctx, theArg);
+                // This boxes the args?
+                // TODO(PR): What's this? It didn't seem to have any affect on hash
+                // I *think* this function is also for static, already resolved dispatch,
+                // so there's nothing that could be avoided here, and nothing to log.
+                Value *argp = boxed(ctx, theArg); //, false, true);
                 ctx.builder.CreateStore(argp, vi.boxroot);
             }
         }
@@ -7996,7 +8018,8 @@ static jl_llvm_functions_t
             Type *retty = f->getReturnType();
             switch (returninfo.cc) {
             case jl_returninfo_t::Boxed:
-                retval = boxed(ctx, retvalinfo); // skip the gcroot on the return path
+                // NOTE(PR): here? return values?
+                retval = boxed(ctx, retvalinfo, false, JL_COUNT_BOX_RETURNS); // skip the gcroot on the return path
                 break;
             case jl_returninfo_t::Register:
                 if (type_is_ghost(retty))
@@ -8224,7 +8247,8 @@ static jl_llvm_functions_t
                     else if (VN->getType() == ctx.types().T_prjlvalue) {
                         // Includes the jl_is_uniontype(phiType) && !TindexN case
                         // TODO: if convert_julia_type says it is wasted effort and to skip it, is it worth using Constant::getNullValue(ctx.types().T_prjlvalue) (dynamically)?
-                        V = boxed(ctx, val);
+                        // NOTE(PR): This one is the boxing for hash.
+                        V = boxed(ctx, val, false, JL_COUNT_BOX_INPUTS);
                     }
                     else {
                         // must be careful to emit undef here (rather than a bitcast or
diff --git a/src/gc-alloc-profiler.cpp b/src/gc-alloc-profiler.cpp
index 1bcbeb2189f5f..ed5c0ddf578d2 100644
--- a/src/gc-alloc-profiler.cpp
+++ b/src/gc-alloc-profiler.cpp
@@ -7,6 +7,7 @@
 
 #include <string>
 #include <vector>
+#include <map>
 
 using std::string;
 using std::vector;
@@ -74,10 +75,59 @@ jl_raw_backtrace_t get_raw_backtrace() JL_NOTSAFEPOINT {
     };
 }
 
+#include <utility>
+#include <numeric>
+#include <iostream>
+
 // == exported interface ==
 
 extern "C" {  // Needed since these functions doesn't take any arguments.
 
+extern uint64_t num_boxes_inputs = 0;
+extern uint64_t num_boxes_returns = 0;
+
+JL_DLLEXPORT uint64_t jl_nhd_total_boxes() {
+    return num_boxes_inputs + num_boxes_returns;
+}
+
+JL_DLLEXPORT uint64_t jl_nhd_boxes_inputs() {
+    return num_boxes_inputs;
+}
+JL_DLLEXPORT uint64_t jl_nhd_boxes_returns() {
+    return num_boxes_returns;
+}
+
+static float extra_allocs_rate = 0.0f;
+JL_DLLEXPORT void jl_nhd_set_extra_allocs_rate(float rate) {
+    extra_allocs_rate = rate;
+}
+
+JL_DLLEXPORT void jl_nhd_log_box_input(jl_datatype_t* type) {
+    // Randomly, with a probability of `extra_allocs_rate`, we will allocate some number of
+    // extra objects. This is to measure the impact of reducing the number of allocations.
+    // If the rate is >1, we may allocate more than once.
+    // We pick a random number between 0 and extra_allocs_rate, then round it, and allocate
+    // that many extra objects.
+    // TODO(PR): ... Dunno why sometimes we get an invalid type in here....
+    if (jl_is_datatype(type)) {
+        float num_extra_allocs = extra_allocs_rate;
+        jl_task_t *ct = jl_current_task;                                \
+        while (num_extra_allocs > 1) {
+            num_extra_allocs--;
+            jl_gc_alloc(ct->ptls, jl_datatype_size(type), type);
+        }
+        // For the last one, we use a random float to decide whether to allocate or not.
+        float sample = float(rand()) / float(RAND_MAX);
+        if (sample < num_extra_allocs) {
+            jl_gc_alloc(ct->ptls, jl_datatype_size(type), type);
+        }
+    }
+    num_boxes_inputs++;
+}
+JL_DLLEXPORT void jl_nhd_log_box_return(jl_value_t* _type) {
+    num_boxes_returns++;
+}
+
 JL_DLLEXPORT void jl_start_alloc_profile(double sample_rate) {
     // We only need to do this once, the first time this is called.
     size_t nthreads = jl_atomic_load_acquire(&jl_n_threads);
diff --git a/src/gc-alloc-profiler.h b/src/gc-alloc-profiler.h
index fcd8e45caa2d8..a20c304d43021 100644
--- a/src/gc-alloc-profiler.h
+++ b/src/gc-alloc-profiler.h
@@ -17,6 +17,12 @@ extern "C" {
 // Forward-declaration to avoid dependency in header file.
 struct jl_raw_alloc_t;  // Defined in gc-alloc-profiler.cpp
 
+enum jl_count_box_type {
+    JL_DONT_LOG_BOX,
+    JL_COUNT_BOX_INPUTS,
+    JL_COUNT_BOX_RETURNS,
+};
+
 typedef struct {
     struct jl_raw_alloc_t *allocs;
     size_t num_allocs;
diff --git a/src/julia.h b/src/julia.h
index a34c1f06d0cc1..3ea503af568d2 100644
--- a/src/julia.h
+++ b/src/julia.h
@@ -1439,6 +1439,8 @@ JL_DLLEXPORT const char *jl_typename_str(jl_value_t *v) JL_NOTSAFEPOINT;
 JL_DLLEXPORT const char *jl_typeof_str(jl_value_t *v) JL_NOTSAFEPOINT;
 JL_DLLEXPORT int jl_type_morespecific(jl_value_t *a, jl_value_t *b);
 
+JL_DLLEXPORT void jl_nhd_log_box(const char* key);
+
 STATIC_INLINE int jl_is_dispatch_tupletype(jl_value_t *v) JL_NOTSAFEPOINT
 {
     return jl_is_datatype(v) && ((jl_datatype_t*)v)->isdispatchtuple;