diff --git a/benchmark/mjit_send_cfunc.yml b/benchmark/mjit_send_cfunc.yml
new file mode 100644
index 0000000000..b5f9c897ec
--- /dev/null
+++ b/benchmark/mjit_send_cfunc.yml
@@ -0,0 +1,6 @@
+prelude: |
+  def mjit_send_cfunc
+    self.class
+  end
+benchmark: mjit_send_cfunc
+loop_count: 100000000
diff --git a/tool/ruby_vm/views/_mjit_compile_send.erb b/tool/ruby_vm/views/_mjit_compile_send.erb
index 0e033a4ebe..255648bc2d 100644
--- a/tool/ruby_vm/views/_mjit_compile_send.erb
+++ b/tool/ruby_vm/views/_mjit_compile_send.erb
@@ -15,15 +15,18 @@
 % # compiler: Use captured cc to avoid race condition
     const struct rb_callcache *captured_cc = captured_cc_entries(status)[call_data_index(cd, body)];
 %
-    const rb_iseq_t *iseq;
+% # compiler: Inline send insn where some supported fastpath is used.
+    const rb_iseq_t *iseq = NULL;
     const CALL_INFO ci = cd->ci;
-    if (!status->compile_info->disable_send_cache && has_valid_method_type(captured_cc)
-        // CC_SET_FASTPATH in vm_callee_setup_arg
-        && !(vm_ci_flag(ci) & VM_CALL_TAILCALL) // inlining non-tailcall path
-        && vm_cc_cme(captured_cc)->def->type == VM_METHOD_TYPE_ISEQ
-        && fastpath_applied_iseq_p(ci, captured_cc, iseq = def_iseq_ptr(vm_cc_cme(captured_cc)->def))) {
-
-        int param_size = iseq->body->param.size;
+    if (!status->compile_info->disable_send_cache && has_valid_method_type(captured_cc) && (
+%       # `CC_SET_FASTPATH(cc, vm_call_cfunc, TRUE)` in `vm_call_method_each_type`
+        vm_cc_cme(captured_cc)->def->type == VM_METHOD_TYPE_CFUNC
+%       # `CC_SET_FASTPATH(cc, vm_call_iseq_setup_func(...), vm_call_iseq_optimizable_p(...))` in `vm_callee_setup_arg`,
+%       # and support only non-VM_CALL_TAILCALL path inside it
+        || (vm_cc_cme(captured_cc)->def->type == VM_METHOD_TYPE_ISEQ
+            && fastpath_applied_iseq_p(ci, captured_cc, iseq = def_iseq_ptr(vm_cc_cme(captured_cc)->def))
+            && !(vm_ci_flag(ci) & VM_CALL_TAILCALL))
+    )) {
         int sp_inc = (int)sp_inc_of_sendish(ci);
         fprintf(f, "{\n");
 
@@ -40,7 +43,7 @@
 <%= render 'mjit_compile_pc_and_sp', locals: { insn: insn } -%>
 
 % # JIT: If ISeq is inlinable, call the inlined method without pushing a frame.
-        if (status->inlined_iseqs != NULL && status->inlined_iseqs[pos] == iseq->body) {
+        if (iseq && status->inlined_iseqs != NULL && iseq->body == status->inlined_iseqs[pos]) {
             fprintf(f, "    {\n");
             fprintf(f, "        VALUE orig_self = reg_cfp->self;\n");
             fprintf(f, "        reg_cfp->self = stack[%d];\n", b->stack_size + sp_inc - 1);
@@ -49,7 +52,7 @@
             fprintf(f, "    }\n");
         }
         else {
-% # JIT: Forked `vm_sendish` to inline various things
+% # JIT: Forked `vm_sendish` (except method_explorer = vm_search_method_wrap) to inline various things
             fprintf(f, "    {\n");
             fprintf(f, "        VALUE val;\n");
             fprintf(f, "        struct rb_calling_info calling;\n");
@@ -58,20 +61,28 @@
 % else
             fprintf(f, "        calling.block_handler = VM_BLOCK_HANDLER_NONE;\n");
 % end
-            fprintf(f, "        calling.argc = %d;\n", vm_ci_argc(ci));
+            fprintf(f, "        calling.kw_splat = %d;\n", IS_ARGS_KW_SPLAT(ci) > 0);
             fprintf(f, "        calling.recv = stack[%d];\n", b->stack_size + sp_inc - 1);
+            fprintf(f, "        calling.argc = %d;\n", vm_ci_argc(ci));
 
-%           # fastpath_applied_iseq_p checks rb_simple_iseq_p, which ensures has_opt == FALSE
-            fprintf(f, "        vm_call_iseq_setup_normal(ec, reg_cfp, &calling, cc_cme, 0, %d, %d);\n", param_size, iseq->body->local_table_size);
-            if (iseq->body->catch_except_p) {
-                fprintf(f, "        VM_ENV_FLAGS_SET(ec->cfp->ep, VM_FRAME_FLAG_FINISH);\n");
-                fprintf(f, "        val = vm_exec(ec, TRUE);\n");
+            if (vm_cc_cme(captured_cc)->def->type == VM_METHOD_TYPE_CFUNC) {
+%               # TODO: optimize this more
+                fprintf(f, "        CALL_DATA cd = (CALL_DATA)0x%"PRIxVALUE";\n", operands[0]);
+                fprintf(f, "        val = vm_call_cfunc(ec, reg_cfp, &calling, cd);\n");
             }
-            else {
-                fprintf(f, "        if ((val = mjit_exec(ec)) == Qundef) {\n");
-                fprintf(f, "            VM_ENV_FLAGS_SET(ec->cfp->ep, VM_FRAME_FLAG_FINISH);\n"); // This is vm_call0_body's code after vm_call_iseq_setup
-                fprintf(f, "            val = vm_exec(ec, FALSE);\n");
-                fprintf(f, "        }\n");
+            else { // VM_METHOD_TYPE_ISEQ
+%               # fastpath_applied_iseq_p checks rb_simple_iseq_p, which ensures has_opt == FALSE
+                fprintf(f, "        vm_call_iseq_setup_normal(ec, reg_cfp, &calling, cc_cme, 0, %d, %d);\n", iseq->body->param.size, iseq->body->local_table_size);
+                if (iseq->body->catch_except_p) {
+                    fprintf(f, "        VM_ENV_FLAGS_SET(ec->cfp->ep, VM_FRAME_FLAG_FINISH);\n");
+                    fprintf(f, "        val = vm_exec(ec, TRUE);\n");
+                }
+                else {
+                    fprintf(f, "        if ((val = mjit_exec(ec)) == Qundef) {\n");
+                    fprintf(f, "            VM_ENV_FLAGS_SET(ec->cfp->ep, VM_FRAME_FLAG_FINISH);\n"); // This is vm_call0_body's code after vm_call_iseq_setup
+                    fprintf(f, "            val = vm_exec(ec, FALSE);\n");
+                    fprintf(f, "        }\n");
+                }
             }
             fprintf(f, "        stack[%d] = val;\n", b->stack_size + sp_inc - 1);
             fprintf(f, "    }\n");