Optimize Int8 and UInt8 boxing. (JuliaLang#36991)

src/cgutils.cpp

@@ -2186,6 +2186,17 @@ static Value *as_value(jl_codectx_t &ctx, Type *to, const jl_cgval_t &v)
     return emit_unbox(ctx, to, v, v.typ);
 }
 
+static Value *load_i8box(jl_codectx_t &ctx, Value *v, jl_datatype_t *ty)
+{
+    auto jvar = ty == jl_int8_type ? jlboxed_int8_cache : jlboxed_uint8_cache;
+    Constant *gv = prepare_global_in(jl_Module, jvar);
+    Value *idx[] = {ConstantInt::get(T_int32, 0), ctx.builder.CreateZExt(v, T_int32)};
+    auto slot = ctx.builder.CreateInBoundsGEP(gv, idx);
+    return tbaa_decorate(tbaa_const, maybe_mark_load_dereferenceable(
+            ctx.builder.CreateAlignedLoad(T_pjlvalue, slot, sizeof(void*)), false,
+            (jl_value_t*)ty));
+}
+
 // some types have special boxing functions with small-value caches
 // Returns T_prjlvalue
 static Value *_boxed_special(jl_codectx_t &ctx, const jl_cgval_t &vinfo, Type *t)
@@ -2210,7 +2221,7 @@ static Value *_boxed_special(jl_codectx_t &ctx, const jl_cgval_t &vinfo, Type *t
     assert(jl_is_datatype(jb));
     Value *box = NULL;
     if (jb == jl_int8_type)
-        box = track_pjlvalue(ctx, call_with_attrs(ctx, box_int8_func, as_value(ctx, t, vinfo)));
+        box = track_pjlvalue(ctx, load_i8box(ctx, as_value(ctx, t, vinfo), jb));
     else if (jb == jl_int16_type)
         box = call_with_attrs(ctx, box_int16_func, as_value(ctx, t, vinfo));
     else if (jb == jl_int32_type)
@@ -2223,7 +2234,7 @@ static Value *_boxed_special(jl_codectx_t &ctx, const jl_cgval_t &vinfo, Type *t
     //  box = ctx.builder.CreateCall(box_float64_func, as_value(ctx, t, vinfo);
     // for Float64, fall through to generic case below, to inline alloc & init of Float64 box. cheap, I know.
     else if (jb == jl_uint8_type)
-        box = track_pjlvalue(ctx, call_with_attrs(ctx, box_uint8_func, as_value(ctx, t, vinfo)));
+        box = track_pjlvalue(ctx, load_i8box(ctx, as_value(ctx, t, vinfo), jb));
     else if (jb == jl_uint16_type)
         box = call_with_attrs(ctx, box_uint16_func, as_value(ctx, t, vinfo));
     else if (jb == jl_uint32_type)

src/codegen.cpp

@@ -404,6 +404,18 @@ static const auto jlgetworld_global = new JuliaVariable{
     [](LLVMContext &C) { return (Type*)T_size; },
 };
 
+static const auto jlboxed_int8_cache = new JuliaVariable{
+    "jl_boxed_int8_cache",
+    true,
+    [](LLVMContext &C) { return (Type*)ArrayType::get(T_pjlvalue, 256); },
+};
+
+static const auto jlboxed_uint8_cache = new JuliaVariable{
+    "jl_boxed_uint8_cache",
+    true,
+    [](LLVMContext &C) { return (Type*)ArrayType::get(T_pjlvalue, 256); },
+};
+
 static const auto jltls_states_func = new JuliaFunction{
     "julia.ptls_states",
     [](LLVMContext &C) { return FunctionType::get(PointerType::get(T_ppjlvalue, 0), false); },
@@ -746,8 +758,6 @@ static const auto box_##ct##_func = new JuliaFunction{                        \
             {at}, false); },                                                  \
     attrs,                                                                    \
 }
-BOX_FUNC(int8, T_pjlvalue, T_int8, get_attrs_sext);
-BOX_FUNC(uint8, T_pjlvalue, T_int8, get_attrs_zext);
 BOX_FUNC(int16, T_prjlvalue, T_int16, get_attrs_sext);
 BOX_FUNC(uint16, T_prjlvalue, T_int16, get_attrs_zext);
 BOX_FUNC(int32, T_prjlvalue, T_int32, get_attrs_sext);

src/datatype.c

@@ -819,15 +819,15 @@ JL_DLLEXPORT jl_value_t *jl_box_char(uint32_t x)
     return v;
 }
 
-static jl_value_t *boxed_int8_cache[256];
+JL_DLLEXPORT jl_value_t *jl_boxed_int8_cache[256];
 JL_DLLEXPORT jl_value_t *jl_box_int8(int8_t x)
 {
-    return boxed_int8_cache[(uint8_t)x];
+    return jl_boxed_int8_cache[(uint8_t)x];
 }
-static jl_value_t *boxed_uint8_cache[256];
+JL_DLLEXPORT jl_value_t *jl_boxed_uint8_cache[256];
 JL_DLLEXPORT jl_value_t *jl_box_uint8(uint8_t x)
 {
-    return boxed_uint8_cache[x];
+    return jl_boxed_uint8_cache[x];
 }
 
 void jl_init_int32_int64_cache(void)
@@ -845,7 +845,7 @@ void jl_init_int32_int64_cache(void)
 #endif
     }
     for(i=0; i < 256; i++) {
-        boxed_uint8_cache[i] = jl_permbox8(jl_uint8_type, i);
+        jl_boxed_uint8_cache[i] = jl_permbox8(jl_uint8_type, i);
     }
 }
 
@@ -856,7 +856,7 @@ void jl_init_box_caches(void)
         boxed_char_cache[i] = jl_permbox32(jl_char_type, i << 24);
     }
     for(i=0; i < 256; i++) {
-        boxed_int8_cache[i] = jl_permbox8(jl_int8_type, i);
+        jl_boxed_int8_cache[i] = jl_permbox8(jl_int8_type, i);
     }
     for(i=0; i < NBOX_C; i++) {
         boxed_int16_cache[i]  = jl_permbox16(jl_int16_type, i-NBOX_C/2);