Add HIDDEN ijl_small_typeof as an optimization

This allows us to avoid a linker indirection on, e.g., Linux and macOS
where symbol interposition means that an extra load is incurred to
support possible relocation of the symbol to a pre-empting library.

src/gc.c

@@ -2427,7 +2427,7 @@ FORCE_INLINE void gc_mark_outrefs(jl_ptls_t ptls, jl_gc_markqueue_t *mq, void *_
             vtag == (jl_vararg_tag << 4)) {
             // these objects have pointers in them, but no other special handling
             // so we want these to fall through to the end
-            vtag = (uintptr_t)jl_small_typeof[vtag / sizeof(*jl_small_typeof)];
+            vtag = (uintptr_t)ijl_small_typeof[vtag / sizeof(*ijl_small_typeof)];
         }
         else if (vtag < jl_max_tags << 4) {
             // these objects either have specialing handling
@@ -2532,7 +2532,7 @@ FORCE_INLINE void gc_mark_outrefs(jl_ptls_t ptls, jl_gc_markqueue_t *mq, void *_
                     objprofile_count(jl_string_type, bits == GC_OLD_MARKED, dtsz);
             }
             else {
-                jl_datatype_t *vt = jl_small_typeof[vtag / sizeof(*jl_small_typeof)];
+                jl_datatype_t *vt = ijl_small_typeof[vtag / sizeof(*ijl_small_typeof)];
                 size_t dtsz = jl_datatype_size(vt);
                 if (update_meta)
                     gc_setmark(ptls, o, bits, dtsz);

src/jltypes.c

@@ -20,6 +20,7 @@ extern "C" {
 #endif
 
 _Atomic(jl_value_t*) cmpswap_names JL_GLOBALLY_ROOTED;
+jl_datatype_t *ijl_small_typeof[(jl_max_tags << 4) / sizeof(*ijl_small_typeof)]; // 16-bit aligned, like the GC
 
 // compute empirical max-probe for a given size
 #define max_probe(size) ((size) <= 1024 ? 16 : (size) >> 6)
@@ -2528,8 +2529,13 @@ static jl_tvar_t *tvar(const char *name)
                           (jl_value_t*)jl_any_type);
 }
 
+void export_jl_small_typeof(void)
+{
+    memcpy(&jl_small_typeof, &ijl_small_typeof, sizeof(jl_small_typeof));
+}
+
 #define XX(name) \
-    jl_small_typeof[(jl_##name##_tag << 4) / sizeof(*jl_small_typeof)] = jl_##name##_type; \
+    ijl_small_typeof[(jl_##name##_tag << 4) / sizeof(*ijl_small_typeof)] = jl_##name##_type; \
     jl_##name##_type->smalltag = jl_##name##_tag;
 void jl_init_types(void) JL_GC_DISABLED
 {
@@ -3351,6 +3357,8 @@ void jl_init_types(void) JL_GC_DISABLED
 
     // override the preferred layout for a couple types
     jl_lineinfonode_type->name->mayinlinealloc = 0; // FIXME: assumed to be a pointer by codegen
+
+    export_jl_small_typeof();
 }
 
 static jl_value_t *core(const char *name)
@@ -3431,6 +3439,8 @@ void post_boot_hooks(void)
             }
         }
     }
+
+    export_jl_small_typeof();
 }
 
 void post_image_load_hooks(void) {

src/julia.h

@@ -750,16 +750,23 @@ enum jl_small_typeof_tags {
     jl_bitstags_first = jl_char_tag, // n.b. bool is not considered a bitstype, since it can be compared by pointer
     jl_max_tags = 64
 };
-#ifndef JL_LIBRARY_EXPORTS
-JL_DLLIMPORT
-#endif
-extern jl_datatype_t *jl_small_typeof[(jl_max_tags << 4) / sizeof(jl_datatype_t*)];
+extern JL_DLLIMPORT jl_datatype_t *jl_small_typeof[(jl_max_tags << 4) / sizeof(jl_datatype_t*)];
+#ifndef JL_LIBRARY_EXPORTS_INTERNAL
 static inline jl_value_t *jl_to_typeof(uintptr_t t)
 {
     if (t < (jl_max_tags << 4))
         return (jl_value_t*)jl_small_typeof[t / sizeof(*jl_small_typeof)];
     return (jl_value_t*)t;
 }
+#else
+extern JL_HIDDEN jl_datatype_t *ijl_small_typeof[(jl_max_tags << 4) / sizeof(jl_datatype_t*)];
+static inline jl_value_t *jl_to_typeof(uintptr_t t)
+{
+    if (t < (jl_max_tags << 4))
+        return (jl_value_t*)ijl_small_typeof[t / sizeof(*ijl_small_typeof)];
+    return (jl_value_t*)t;
+}
+#endif
 
 
 // kinds

src/staticdata.c

@@ -2839,6 +2839,7 @@ JL_DLLEXPORT void jl_set_sysimg_so(void *handle)
 #endif
 
 extern void rebuild_image_blob_tree(void);
+extern void export_jl_small_typeof(void);
 
 static void jl_restore_system_image_from_stream_(ios_t *f, jl_image_t *image, jl_array_t *depmods, uint64_t checksum,
                                 /* outputs */    jl_array_t **restored,         jl_array_t **init_order,
@@ -2914,9 +2915,10 @@ static void jl_restore_system_image_from_stream_(ios_t *f, jl_image_t *image, jl
             *tag = jl_read_value(&s);
         }
 #define XX(name) \
-        jl_small_typeof[(jl_##name##_tag << 4) / sizeof(*jl_small_typeof)] = jl_##name##_type;
+        ijl_small_typeof[(jl_##name##_tag << 4) / sizeof(*ijl_small_typeof)] = jl_##name##_type;
         JL_SMALL_TYPEOF(XX)
 #undef XX
+        export_jl_small_typeof();
         jl_global_roots_table = (jl_array_t*)jl_read_value(&s);
         // set typeof extra-special values now that we have the type set by tags above
         jl_astaggedvalue(jl_nothing)->header = (uintptr_t)jl_nothing_type | jl_astaggedvalue(jl_nothing)->header;